Dependent Randomized Rounding via Exchange Properties of Combinatorial Structures [1] 学习笔记
Author: Sijin Yu
文章目录
- Dependent Randomized Rounding via Exchange Properties of Combinatorial Structures [1] 学习笔记
- 1. 原文的重要结论
- Rounding 1: Randomized Swap Rounding
- Theorem 1
- Definition 1: Matroid Polytopes
- Lemma 1
- Definition 2: Matroid Intersection
- Definition 3: Equivalent Elements in Matroids
- Definition 4: Submodular Function
- Definition 5: Multilinear Extension
- Theorem 2 (Chernoff-type Concentration Bounds for any Linear Function)
- Theorem 3 (Exponential Bounds for Monotone Submodular Functions)
- Rounding 2: Randomized Swap Rounding for Matroids
- Definition 6: (For Rounding 3) Exchange Digraph
- Definition 7: (For Rounding 3) Irreducible Directed Cycle
- Lemma 2: (For Rounding 3)
- Lemma 3: (For Rounding 3, Variant of Lemma 2)
- Rounding 3: Rounding for Matroid Intersection
- Theorem 4 (Chernoff-type Concentration Bounds for Matroid Intersection Case)
- Theorem 5 (Constant Approximation Algorithm for Matroid Intersection)
- Theorem 6
- Theorem 7 (Approximation Algorithm for Matroid Bases with Packing Constraints)
- Max-Min Submodular Allocation
- Theorem 8 (Approximation Algorithm for Max-Min Submodular Allocation)
- References
1. 原文的重要结论
Rounding 1: Randomized Swap Rounding
- x x x 为 P P P 的顶点的凸组合, 即 x = ∑ i = 1 n α i v i x=\sum_{i=1}^n\alpha_i\mathbf{v}_i x=∑i=1nαivi, 其中 v i \mathbf v_i vi 是 P P P 的顶点, 且 ∑ i α i = 1 \sum_i \alpha_i=1 ∑iαi=1.
- 令 w 1 = v 1 , β 1 = α 1 \mathbf w_1=\mathbf v_1, \beta_1=\alpha_1 w1=v1,β1=α1, 在每一步, 使用参数 β i + 1 = β i + α i + 1 \beta_{i+1}=\beta_i+\alpha_{i+1} βi+1=βi+αi+1 将两个顶点 w i , v i + 1 \mathbf w_i, \mathbf v_{i+1} wi,vi+1 合并为新的顶点 w i + 1 \mathbf w_{i+1} wi+1, 使 E [ β i + 1 w i + 1 ] = β i w i + α i + 1 v i + 1 \mathbf E[\beta_{i+1}\mathbf w{i+1}]=\beta_i\mathbf w_i+\alpha_{i+1}\mathbf v_{i+1} E[βi+1wi+1]=βiwi+αi+1vi+1 (the merge operation).
- 在 n − 1 n-1 n−1 步后, 得到一个顶点 X = w n X=\mathbf w_n X=wn 使得 E [ X ] = ∑ i = 1 n α i v i = x \mathbf E[X]=\sum^n_{i=1}\alpha_i\mathbf v_i =x E[X]=∑i=1nαivi=x
Theorem 1
-
令 P P P 为一多面体, 其顶点为 { 0 , 1 } N \{0,1\}^N {0,1}N. 则以下两个性质是等价的 [2]:
(1) 对于任意的 x ∈ P x\in P x∈P, 存在一个 P P P 的顶点的概率分布, 使得一个从这一分布采样的随机顶点 X X X 满足 E [ X ] = x \mathbf E[X]=x E[X]=x 并且它的各分量 { X i } i ∈ N \{X_i\}_{i\in N} {Xi}i∈N 是负相关的.
(2) P P P 是一个拟阵基多面体的投影, 即存在一个拟阵 M = ( N ′ , I ) \mathcal M=(N',\mathcal I) M=(N′,I) 且 N ⊆ N ′ N\subseteq N' N⊆N′, 则 X X X 是 P P P 的顶点, 当且仅当 X = 1 B ∩ N X=\mathbf 1_{B\cap N} X=1B∩N 对某个 M \mathcal M M 的基 B B B 成立.
-
Let P P P be a polytope with vertices in { 0 , 1 } N \{0,1\}^N {0,1}N. Then the following two properties are equivalent [2]:
(1) For any x ∈ P x\in P x∈P, there exists a probability distribution over vertices of P P P such that a random vertex X X X drawn from this distribution satisfies E [ X ] = x \mathbf E[X]=x E[X]=x and the coordinates { X i } i ∈ N \{X_i\}_{i\in N} {Xi}i∈N are negatively correlated.
(2) P P P is a projection of a matroid base polytope, in the sense that there is a matroid M = ( N ′ , I ) \mathcal M=(N',\mathcal I) M=(N′,I) such that N ⊆ N ′ N\subseteq N' N⊆N′ and X X X is a vertex of P P P iff X = 1 B ∩ N X=\mathbf 1_{B\cap N} X=1B∩N for some base B B B of M \mathcal M M.
Definition 1: Matroid Polytopes
-
给定一个拟阵 M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) 和秩函数 r : 2 N → Z + r: 2^N\to\Z_+ r:2N→Z+, 拟阵多面体 P ( M ) P(\mathcal M) P(M) 定义为 M \mathcal M M 的独立集的特征向量的凸包, 或等价地定义为:
P ( M ) = conv { 1 I : I ∈ I } = { x ≥ 0 : ∀ S ; ∑ i ∈ S x i ≤ r ( S ) } P(\mathcal M)=\text{conv}\{\mathbf1_{I}:I\in\mathcal I\}=\{x\geq0:\forall S;\sum_{i\in S}x_i\leq r(S)\} P(M)=conv{1I:I∈I}={x≥0:∀S;i∈S∑xi≤r(S)} -
Given a matroid M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) with rank function r : 2 N → Z + r: 2^N\to\Z_+ r:2N→Z+, the matroid polytope P ( M ) P(\mathcal M) P(M) is defined as the convex hull of characteristic vectors of the independent sets of M \mathcal M M, or equivalently:
P ( M ) = conv { 1 I : I ∈ I } = { x ≥ 0 : ∀ S ; ∑ i ∈ S x i ≤ r ( S ) } P(\mathcal M)=\text{conv}\{\mathbf1_{I}:I\in\mathcal I\}=\{x\geq0:\forall S;\sum_{i\in S}x_i\leq r(S)\} P(M)=conv{1I:I∈I}={x≥0:∀S;i∈S∑xi≤r(S)}
Lemma 1
- 令 M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) 为一拟阵并令 B 1 , B 2 B_1,B_2 B1,B2 为基. 对于任意的 i ∈ B 1 − B 2 i\in B_1-B_2 i∈B1−B2 总存在 j ∈ B 2 − B 1 j\in B_2-B_1 j∈B2−B1, 使得 B 1 − i + j B_1-i+j B1−i+j 和 B 2 − j + i B_2-j+i B2−j+i 也为基.
- Let M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) be a matroid and let B 1 , B 2 B_1,B_2 B1,B2 be bases. For any i ∈ B 1 − B 2 i\in B_1-B_2 i∈B1−B2 there exists j ∈ B 2 − B 1 j\in B_2-B_1 j∈B2−B1 such that B 1 − i + j B_1-i+j B1−i+j and B 2 − j + i B_2-j+i B2−j+i are also bases.
Definition 2: Matroid Intersection
- 令 M 1 = ( N , I 1 ) \mathcal M_1=(N,\mathcal I_1) M1=(N,I1) 和 M 2 = ( N , I 2 ) \mathcal M_2=(N,\mathcal I_2) M2=(N,I2) 为两个拟阵. 由 Edmonds 等人的工作 [2], 相交多面体 P ( M 1 ) ∩ P ( M 2 ) P(\mathcal M_1)\cap P(\mathcal M_2) P(M1)∩P(M2) 是整数的, 且它的顶点和集合 I ∈ I 1 ∩ I 2 I\in\mathcal I_1\cap\mathcal I_2 I∈I1∩I2 相关.
- Let M 1 = ( N , I 1 ) \mathcal M_1=(N,\mathcal I_1) M1=(N,I1) and M 2 = ( N , I 2 ) \mathcal M_2=(N,\mathcal I_2) M2=(N,I2) be two matroids. It follows from the work of Edmonds [2] that the intersection polytope P ( M 1 ) ∩ P ( M 2 ) P(\mathcal M_1)\cap P(\mathcal M_2) P(M1)∩P(M2) is integral and its vertices correspond exactly to sets I ∈ I 1 ∩ I 2 I\in\mathcal I_1\cap\mathcal I_2 I∈I1∩I2.
Definition 3: Equivalent Elements in Matroids
- 在一个拟阵 M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) 中, 两个元素 i , j ∈ N i,j\in N i,j∈N 是等价的, 若对于任意的不包含 i , j i,j i,j 的集合 A ∈ I A\in \mathcal I A∈I, A + i ∈ I A+i\in\mathcal I A+i∈I 当且仅当 A + j ∈ I A+j\in \mathcal I A+j∈I.
- Two elements i , j ∈ N i,j\in N i,j∈N are equivalent in a matroid M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) if for any set A ∈ I A\in \mathcal I A∈I not containing i , j i, j i,j, A + i ∈ I A+i\in\mathcal I A+i∈I if and only if A + j ∈ I A+j\in \mathcal I A+j∈I.
Remark
对于任意一个不包含 i , j i,j i,j 的集合 A ∈ I A\in\mathcal I A∈I, 增加 i i i 后依然独立, 和增加 j j j 后依然独立是等价的, 因此 i i i 和 j j j 对 A A A 的"扩张"的影响是等效的, 所以 i i i 和 j j j 等价.
Definition 4: Submodular Function
- 一个函数 f : 2 N → R f:2^N\to\R f:2N→R 是次模的, 若给定任意的 A , B ⊆ N A,B\subseteq N A,B⊆N, 都有 f ( A ) + f ( B ) ≥ f ( A ∪ B ) + f ( A ∩ B ) f(A)+f(B)\geq f(A\cup B)+f(A\cap B) f(A)+f(B)≥f(A∪B)+f(A∩B).
- A function f : 2 N → R f:2^N\to\R f:2N→R is submodular if for any A , B ⊆ N A,B\subseteq N A,B⊆N, f ( A ) + f ( B ) ≥ f ( A ∪ B ) + f ( A ∩ B ) f(A)+f(B)\geq f(A\cup B)+f(A\cap B) f(A)+f(B)≥f(A∪B)+f(A∩B).
Definition 5: Multilinear Extension
-
对函数 f : 2 N → R f:2^N\to \R f:2N→R, 其 multilinear extension 为函数 F : [ 0 , 1 ] n → R F:[0,1]^n\to\R F:[0,1]n→R, 定义为:
F ( x ) = ∑ S ⊆ N f ( S ) ∏ i ∈ S x i ∏ i ∈ N − S ( 1 − x i ) F(x)=\sum_{S\subseteq N}f(S)\prod_{i\in S}x_i\prod_{i\in N-S}(1-x_i) F(x)=S⊆N∑f(S)i∈S∏xii∈N−S∏(1−xi) -
The multilinear extension of f : 2 N → R f:2^N\to \R f:2N→R is the function F : [ 0 , 1 ] n → R F:[0,1]^n\to\R F:[0,1]n→R defined by:
F ( x ) = ∑ S ⊆ N f ( S ) ∏ i ∈ S x i ∏ i ∈ N − S ( 1 − x i ) F(x)=\sum_{S\subseteq N}f(S)\prod_{i\in S}x_i\prod_{i\in N-S}(1-x_i) F(x)=S⊆N∑f(S)i∈S∏xii∈N−S∏(1−xi)
Remark
F ( ⋅ ) F(\cdot) F(⋅) 可以看作 f ( ⋅ ) f(\cdot) f(⋅) 的期望
Theorem 2 (Chernoff-type Concentration Bounds for any Linear Function)
-
令 ( x 1 , . . . , x n ) ∈ P ( M ) (x_1,...,x_n)\in P(\mathcal M) (x1,...,xn)∈P(M) 为拟阵多面体内的一分数解, ( X 1 , . . . , X n ) ∈ P ( M ) ∩ { 0 , 1 } n (X_1,...,X_n)\in P(\mathcal M)\cap\{0,1\}^n (X1,...,Xn)∈P(M)∩{0,1}n 为使用 randomized swap rounding 得到的一整数解. 则对于任意的线性函数 Z = ∑ a i X i Z=\sum a_iX_i Z=∑aiXi, 其中 a i ∈ [ 0 , 1 ] a_i\in[0,1] ai∈[0,1], 有 E [ Z ] = ∑ a i x i \mathbf E[Z]=\sum a_ix_i E[Z]=∑aixi 且:
-
若 δ > 0 \delta>0 δ>0 和 μ ≥ E [ Z ] \mu\geq\mathbf E[Z] μ≥E[Z], 则
Pr [ Z ≥ ( 1 + δ ) μ ] ≤ ( e δ ( 1 + δ ) 1 + δ ) μ . \text{Pr}[Z\geq(1+\delta)\mu]\leq\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^\mu. Pr[Z≥(1+δ)μ]≤((1+δ)1+δeδ)μ. -
若 δ ∈ [ 0 , 1 ] \delta\in[0,1] δ∈[0,1] 和 μ ≤ E [ Z ] \mu\leq\mathbf E[Z] μ≤E[Z], 则
Pr [ Z ≤ ( 1 − δ ) μ ] ≤ e − μ δ 2 / 2 . \text{Pr}[Z\leq(1-\delta)\mu]\leq e^{-\mu\delta^2/2}. Pr[Z≤(1−δ)μ]≤e−μδ2/2.
进一步说, 对于任意的次模函数 f : { 0 , 1 } n → R f:\{0,1\}^n\to\R f:{0,1}n→R 和它的 multilinear extension F : [ 0 , 1 ] n → R F:[0,1]^n\to\R F:[0,1]n→R,
E [ f ( X 1 , . . . , X n ) ] ≥ F ( x 1 , . . . , x n ) . \mathbf E[f(X_1,...,X_n)]\geq F(x_1,...,x_n). E[f(X1,...,Xn)]≥F(x1,...,xn). -
-
Let ( x 1 , . . . , x n ) ∈ P ( M ) (x_1,...,x_n)\in P(\mathcal M) (x1,...,xn)∈P(M) be a fractional solution in the matroid polytope and ( X 1 , . . . , X n ) ∈ P ( M ) ∩ { 0 , 1 } n (X_1,...,X_n)\in P(\mathcal M)\cap\{0,1\}^n (X1,...,Xn)∈P(M)∩{0,1}n an integral solution obtained using randomized swap rounding. Then for any linear function Z = ∑ a i X i Z=\sum a_iX_i Z=∑aiXi with a i ∈ [ 0 , 1 ] a_i\in[0,1] ai∈[0,1], we have E [ Z ] = ∑ a i x i \mathbf E[Z]=\sum a_ix_i E[Z]=∑aixi and:
-
If δ > 0 \delta>0 δ>0 and μ ≥ E [ Z ] \mu\geq\mathbf E[Z] μ≥E[Z], then
Pr [ Z ≥ ( 1 + δ ) μ ] ≤ ( e δ ( 1 + δ ) 1 + δ ) μ . \text{Pr}[Z\geq(1+\delta)\mu]\leq\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^\mu. Pr[Z≥(1+δ)μ]≤((1+δ)1+δeδ)μ. -
If δ ∈ [ 0 , 1 ] \delta\in[0,1] δ∈[0,1] and μ ≤ E [ Z ] \mu\leq\mathbf E[Z] μ≤E[Z], then
Pr [ Z ≤ ( 1 − δ ) μ ] ≤ e − μ δ 2 / 2 . \text{Pr}[Z\leq(1-\delta)\mu]\leq e^{-\mu\delta^2/2}. Pr[Z≤(1−δ)μ]≤e−μδ2/2.
Moreover, for any submodular function f : { 0 , 1 } n → R f:\{0,1\}^n\to\R f:{0,1}n→R and its multilinear extension F : [ 0 , 1 ] n → R F:[0,1]^n\to\R F:[0,1]n→R,
E [ f ( X 1 , . . . , X n ) ] ≥ F ( x 1 , . . . , x n ) . \mathbf E[f(X_1,...,X_n)]\geq F(x_1,...,x_n). E[f(X1,...,Xn)]≥F(x1,...,xn). -
Theorem 3 (Exponential Bounds for Monotone Submodular Functions)
-
令 f : { 0 , 1 } n → R + f:\{0,1\}^n\to\R_+ f:{0,1}n→R+ 为单调次模函数, 其边际值在 [ 0 , 1 ] [0,1] [0,1] 内, F : [ 0 , 1 ] n → R + F:[0,1]^n\to\R_+ F:[0,1]n→R+ 为它的 multilinear extension. 令 ( x 1 , . . . , x n ) ∈ P ( M ) (x_1,...,x_n)\in P(\mathcal M) (x1,...,xn)∈P(M) 为拟阵多面体内的一点且 ( X 1 , . . . , X n ) ∈ { 0 , 1 } n (X_1,...,X_n)\in\{0,1\}^n (X1,...,Xn)∈{0,1}n 为从它进行 randomized swap rounding 获得的一随机解. 令 μ 0 = F ( x 1 , . . . , x n ) \mu_0=F(x_1,...,x_n) μ0=F(x1,...,xn) 和 δ > 0 \delta>0 δ>0. 则
Pr [ f ( X 1 , . . . , X n ) ≤ ( 1 − δ ) μ 0 ] ≤ e − μ 0 δ 2 / 8 . \text{Pr}[f(X_1,...,X_n)\leq(1-\delta)\mu_0]\leq e^{-\mu_0\delta^2/8}. Pr[f(X1,...,Xn)≤(1−δ)μ0]≤e−μ0δ2/8. -
Let f : { 0 , 1 } n → R + f:\{0,1\}^n\to\R_+ f:{0,1}n→R+ be a monotone submodular function with marginal values in [ 0 , 1 ] [0,1] [0,1], and F : [ 0 , 1 ] n → R + F:[0,1]^n\to\R_+ F:[0,1]n→R+ its multilinear extension. Let ( x 1 , . . . , x n ) ∈ P ( M ) (x_1,...,x_n)\in P(\mathcal M) (x1,...,xn)∈P(M) be a point in a matroid polytope and ( X 1 , . . . , X n ) ∈ { 0 , 1 } n (X_1,...,X_n)\in\{0,1\}^n (X1,...,Xn)∈{0,1}n a random solution obtained from it by randomized swap rounding. Let μ 0 = F ( x 1 , . . . , x n ) \mu_0=F(x_1,...,x_n) μ0=F(x1,...,xn) and δ > 0 \delta>0 δ>0. Then
Pr [ f ( X 1 , . . . , X n ) ≤ ( 1 − δ ) μ 0 ] ≤ e − μ 0 δ 2 / 8 . \text{Pr}[f(X_1,...,X_n)\leq(1-\delta)\mu_0]\leq e^{-\mu_0\delta^2/8}. Pr[f(X1,...,Xn)≤(1−δ)μ0]≤e−μ0δ2/8.
Remark
该不等式是结束点 f ( X 1 , . . . , X n ) f(X_1,...,X_n) f(X1,...,Xn) 与起始点 μ 0 = F ( x 1 , . . . , x n ) \mu_0=F(x_1,...,x_n) μ0=F(x1,...,xn) 的比较, 这往往是应用中我们希望去比较的, 因此这样的一个 bound 是我们所需要的.
此外, 这一不等式被证出的难点在于, bound 与 n n n 无关.
Rounding 2: Randomized Swap Rounding for Matroids
令 M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) 为一个秩为 d = r ( N ) d=r(N) d=r(N) 的拟阵, 并令 n = ∣ N ∣ n=|N| n=∣N∣. Randomized Swap Rounding 将一个点 x ∈ P ( M ) x\in P(\mathcal M) x∈P(M) round 到一个基 x ∈ B ( M ) x\in B(\mathcal M) x∈B(M) 上. 将 x x x 预先表示为基的凸组合, 即 x = ∑ l = 1 m β l 1 B l x=\sum^m_{l=1}\beta_l\mathbf 1_{B_l} x=∑l=1mβl1Bl, 且 ∑ l = 1 m β l = 1 , β l > 0 \sum^m_{l=1}\beta_l=1,\beta_l>0 ∑l=1mβl=1,βl>0. 凸组合可以根据 [3] 找到. rounding 将进行 n − 1 n-1 n−1 次迭代, 在第 1 1 1 步我们将两个基 B 1 , B 2 B_1,B_2 B1,B2 (随机的) 组合为 C 2 C_2 C2, 并用 ( β 1 + β 2 ) 1 C 2 (\beta_1+\beta_2)\mathbf 1_{C_2} (β1+β2)1C2 替代线性组合中的 β 1 1 B 1 + β 2 1 B 2 \beta_1\mathbf 1_{B_1}+\beta_2\mathbf 1_{B_2} β11B1+β21B2. 在第 k k k 步, C k C_k Ck 和 B k + 1 B_{k+1} Bk+1 组合为一个新的基 C k + 1 C_{k+1} Ck+1, 并用 ( ∑ l = 1 k + 1 β l ) 1 C k + 1 \left( \sum^{k+1}_{l=1}\beta_l \right)\mathbf 1_{C_{k+1}} (∑l=1k+1βl)1Ck+1 替代线性组合中的 ( ∑ l = 1 k β l ) 1 C k + β k + 1 1 B k + 1 \left( \sum^k_{l=1}\beta_l \right)\mathbf 1_{C_k} +\beta_{k+1}\mathbf1_{B_{k+1}} (∑l=1kβl)1Ck+βk+11Bk+1. 最后 C n C_n Cn 被返回. ( x x x 则被 round 为整数点 1 C n \mathbf 1_{C_n} 1Cn)
算法: 输入 β 1 , B 1 , β 2 , B 2 \beta_1,B_1,\beta_2,B_2 β1,B1,β2,B2, 将基 B 1 B_1 B1 和 B 2 B_2 B2 组合, 返回新的基.
Algorithm MergeBases( β 1 , B 1 , β 2 , B 2 \beta_1,B_1,\beta_2,B_2 β1,B1,β2,B2):
While( B 1 ≠ B 2 B_1\neq B_2 B1=B2) do:
Pick i ∈ B 1 − B 2 i\in B_1-B_2 i∈B1−B2 and find j ∈ B 2 − B 1 j\in B_2-B_1 j∈B2−B1 such that B 1 − i + j ∈ I B_1-i+j\in \mathcal I B1−i+j∈I and B 2 − j + i ∈ I B_2-j+i\in\mathcal I B2−j+i∈I;
With probability β 1 β 1 + β 2 \frac{\beta_1}{\beta_1+\beta_2} β1+β2β1 do:
B 2 ← B 2 − j + i B_2\gets B_2-j+i B2←B2−j+i;
Else:
B 1 ← B 1 − i + j B_1\gets B_1-i+j B1←B1−i+j;
EndWhile;
Output B 1 B_1 B1.
Remark
该算法的意思是让 B 1 B_1 B1 和 B 2 B_2 B2 不断靠拢, 并以线性系数 β 1 \beta_1 β1 和 β 2 \beta_2 β2 作为其靠拢的概率权重.
该算法由 Lemma 1 保证.
算法: 输入分数点 x x x, 输出将其 round 到的基 C n C_n Cn.
Algorithm SwapRound( x = ∑ l = 1 n β l 1 B l x=\sum_{l=1}^n\beta_l\mathbf 1_{B_l} x=∑l=1nβl1Bl):
C 1 = B 1 C_1=B_1 C1=B1;
For( k = 1 k=1 k=1 to n − 1 n-1 n−1) do:
C k + 1 = C_{k+1}= Ck+1= MergeBases( ∑ l = 1 k β l , C k , β k + 1 , B k + 1 \sum_{l=1}^k\beta_l,C_k,\beta_{k+1},B_{k+1} ∑l=1kβl,Ck,βk+1,Bk+1);
EndFor;
Output C n C_n Cn.
Definition 6: (For Rounding 3) Exchange Digraph
-
考虑两个拟阵 M 1 = ( N , I 1 ) \mathcal M_1=(N,\mathcal I_1) M1=(N,I1) 和 M 2 = ( N , I 2 ) \mathcal M_2=(N,\mathcal I_2) M2=(N,I2). 对于 I ∈ I 1 ∩ I 2 I\in\mathcal I_1\cap\mathcal I_2 I∈I1∩I2, 我们定义两个有向图 D M 1 ( I ) D_{\mathcal M_1}(I) DM1(I) 和 D M 2 ( I ) D_{\mathcal M_2}(I) DM2(I) 如下.
- 对每个 i ∈ I , j ∈ N − I i\in I, j\in N-I i∈I,j∈N−I, 且 I + j − i ∈ I 1 I+j-i\in\mathcal I_1 I+j−i∈I1, 令边 ( i , j ) ∈ D M 1 ( I ) (i,j)\in D_{\mathcal M_1}(I) (i,j)∈DM1(I);
- 对每个 i ∈ I , j ∈ N − I i\in I,j\in N-I i∈I,j∈N−I, 且 I + j − i ∈ I 2 I+j-i\in \mathcal I_2 I+j−i∈I2, 令边 ( i , j ) ∈ D M 2 ( I ) (i,j)\in D_{\mathcal M_2}(I) (i,j)∈DM2(I).
我们定义一个有向图 D M 1 , M 2 ( I ) = D M 1 ( I ) ∪ D M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I)=D_{\mathcal M_1}(I)\cup D_{\mathcal M_2}(I) DM1,M2(I)=DM1(I)∪DM2(I).
-
Consider two matroids M 1 = ( N , I 1 ) \mathcal M_1=(N,\mathcal I_1) M1=(N,I1) and M 2 = ( N , I 2 ) \mathcal M_2=(N,\mathcal I_2) M2=(N,I2). For I ∈ I 1 ∩ I 2 I\in\mathcal I_1\cap\mathcal I_2 I∈I1∩I2, we define two digraphs D M 1 ( I ) D_{\mathcal M_1}(I) DM1(I) and D M 2 ( I ) D_{\mathcal M_2}(I) DM2(I) as follows.
- For each i ∈ I , j ∈ N − I i\in I, j\in N-I i∈I,j∈N−I with I + j − i ∈ I 1 I+j-i\in\mathcal I_1 I+j−i∈I1, we have an arc ( i , j ) ∈ D M 1 ( I ) (i,j)\in D_{\mathcal M_1}(I) (i,j)∈DM1(I);
- For each i ∈ I , j ∈ N − I i\in I, j\in N-I i∈I,j∈N−I with I + j − i ∈ I 2 I+j-i\in\mathcal I_2 I+j−i∈I2, we have an arc ( i , j ) ∈ D M 2 ( I ) (i,j)\in D_{\mathcal M_2}(I) (i,j)∈DM2(I).
We define a digraph D M 1 , M 2 ( I ) = D M 1 ( I ) ∪ D M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I)=D_{\mathcal M_1}(I)\cup D_{\mathcal M_2}(I) DM1,M2(I)=DM1(I)∪DM2(I).
Definition 7: (For Rounding 3) Irreducible Directed Cycle
- 一个 D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) 里的有向圈 C C C 是不可约的, 如果在点集 V ( C ) V(C) V(C) 上 C ∩ D M 1 ( I ) C\cap D_{\mathcal M_1}(I) C∩DM1(I) 是 D M 1 ( I ) D_{\mathcal M_1}(I) DM1(I) 里的唯一完美匹配, 且 C ∩ D M 2 ( I ) C\cap D_{\mathcal M_2}(I) C∩DM2(I) 是 D M 2 ( I ) D_{\mathcal M_2}(I) DM2(I) 里的唯一完美匹配. 否则, C C C 是可约的.
- A directed cycle C C C in D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) is irreducible if D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) is the unique perfect matching in D M 1 ( I ) D_{\mathcal M_1}(I) DM1(I) and C ∩ D M 2 ( I ) C\cap D_{\mathcal M_2}(I) C∩DM2(I) is the unique perfect matching in D M 2 ( I ) D_{\mathcal M_2}(I) DM2(I) on the vertex set V ( C ) V (C) V(C). Otherwise, C C C is reducible.
Lemma 2: (For Rounding 3)
- 令 M l = ( N , I l ) , l = 1 , 2 \mathcal M_l=(N,\mathcal I_l),l=1,2 Ml=(N,Il),l=1,2 为在全集 N N N 上的拟阵. 假设 I , J ∈ I 1 ∩ I 2 I,J\in \mathcal I_1\cap\mathcal I_2 I,J∈I1∩I2 且 ∣ I ∣ = ∣ J ∣ |I|=|J| ∣I∣=∣J∣. 则存在一个整数 s ≥ 0 s\geq0 s≥0 和一个 D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) 里的不可约有向圈的 collection { C 1 , . . . , C m } \{C_1,...,C_m\} {C1,...,Cm} (允许重复), 且其只使用 I Δ J I\Delta J IΔJ 中的元素, 使得每个 I Δ J I\Delta J IΔJ 里的元素都在这些有向圈里出现 2 s 2^s 2s 次.
- Let M l = ( N , I l ) , l = 1 , 2 \mathcal M_l=(N,\mathcal I_l),l=1,2 Ml=(N,Il),l=1,2, be matroids on ground set N N N. Suppose that I , J ∈ I 1 ∩ I 2 I,J\in \mathcal I_1\cap\mathcal I_2 I,J∈I1∩I2 and ∣ I ∣ = ∣ J ∣ |I|=|J| ∣I∣=∣J∣. Then there is an integer s ≥ 0 s\geq0 s≥0 and a collection of irreducible directed cycles { C 1 , . . . , C m } \{C_1,...,C_m\} {C1,...,Cm} (allowing repetition) in D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I), using only elements of I Δ J I\Delta J IΔJ, so that each element of I Δ J I\Delta J IΔJ appears in exactly 2 s 2^s 2s of the directed cycles.
Remark
collection 是允许出现重复元素的集合, 可翻译为多集.
I Δ J I\Delta J IΔJ 定义为 ( I − J ) ∪ ( J − I ) (I-J)\cup(J-I) (I−J)∪(J−I).
Lemma 3: (For Rounding 3, Variant of Lemma 2)
- 令 M l = ( N , I l ) , l = 1 , 2 \mathcal M_l=(N,\mathcal I_l),l=1,2 Ml=(N,Il),l=1,2 为在全集 N N N 上的拟阵. 假设 I , J ∈ I 1 ∩ I 2 I,J\in \mathcal I_1\cap\mathcal I_2 I,J∈I1∩I2 且 ∣ I ∣ = ∣ J ∣ |I|=|J| ∣I∣=∣J∣. 则我们可以在多项式时间内寻在 D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) 找到一个不可约有向圈的 collection { C 1 , . . . , C m } , m ≤ ∣ I Δ J ∣ \{C_1,...,C_m\},m\leq|I\Delta J| {C1,...,Cm},m≤∣IΔJ∣, 与参数 γ i ≥ 0 , ∑ i = 1 m γ i = 1 \gamma_i\geq0,\sum^m_{i=1}\gamma_i=1 γi≥0,∑i=1mγi=1, 使得对某些 γ > 0 \gamma>0 γ>0, ∑ i = 1 m γ i 1 V ( C i ) = γ 1 I Δ J \sum^m_{i=1}\gamma_i\mathbf1_{V(C_i)}=\gamma\mathbf1_{I\Delta J} ∑i=1mγi1V(Ci)=γ1IΔJ.
- Let M l = ( N , I l ) , l = 1 , 2 \mathcal M_l=(N,\mathcal I_l),l=1,2 Ml=(N,Il),l=1,2, be matroids on ground set N N N. Suppose that I , J ∈ I 1 ∩ I 2 I,J\in \mathcal I_1\cap\mathcal I_2 I,J∈I1∩I2 and ∣ I ∣ = ∣ J ∣ |I|=|J| ∣I∣=∣J∣. Then we can find in polynomial time a collection of irreducible directed cycles { C 1 , . . . , C m } , m ≤ ∣ I Δ J ∣ \{C_1,...,C_m\},m\leq|I\Delta J| {C1,...,Cm},m≤∣IΔJ∣, in D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I), with coefficients γ i ≥ 0 , ∑ i = 1 m γ i = 1 \gamma_i\geq0,\sum^m_{i=1}\gamma_i=1 γi≥0,∑i=1mγi=1, such that for some γ > 0 \gamma>0 γ>0, ∑ i = 1 m γ i 1 V ( C i ) = γ 1 I Δ J \sum^m_{i=1}\gamma_i\mathbf1_{V(C_i)}=\gamma\mathbf1_{I\Delta J} ∑i=1mγi1V(Ci)=γ1IΔJ.
Rounding 3: Rounding for Matroid Intersection
给定两个拟阵 M 1 \mathcal M_1 M1, M 2 \mathcal M_2 M2 和一个分数解 x ∈ P ( M 1 ) ∩ P ( M 2 ) x\in P(\mathcal M_1)\cap P(\mathcal M_2) x∈P(M1)∩P(M2), 将 x x x round 为一整数解 X X X, 它表示一个在两个拟阵中都独立的集合, 且 E [ X ] = x \mathbf E[X]=x E[X]=x. 假定 I , J ∈ I 1 ∩ I 2 I,J\in\mathcal I_1\cap \mathcal I_2 I,J∈I1∩I2 为两个基, 令 x x x 为 I I I 和 J J J 的线性组合 x = α I + β J x=\alpha I+\beta J x=αI+βJ.
算法: 输入 x x x 的线性组合参数 α , I , β , J \alpha,I,\beta,J α,I,β,J, 输出 x x x 被 round 到的基.
Algorithm mergeIntersectionBases( α , I , β , J \alpha,I,\beta,J α,I,β,J):
While( I ≠ J I\neq J I=J) do:
Generate a collection of irreducible cycles in D M 1 , M 2 ( I ) D_{\mathcal M_1,\mathcal M_2}(I) DM1,M2(I) such that
∑ γ i 1 V ( C i ) = γ 1 I Δ J , ∑ γ i = 1 , γ i ≥ 0 \sum\gamma_i\mathbf1_{V(C_i)}=\gamma\mathbf1_{I\Delta J},\sum\gamma_i=1,\gamma_i\geq0 ∑γi1V(Ci)=γ1IΔJ,∑γi=1,γi≥0;
and a collection of irreducible cycles in D M 1 , M 2 ( J ) D_{\mathcal M_1,\mathcal M_2}(J) DM1,M2(J) such that
∑ δ i 1 V ( C i ′ ) = δ 1 I Δ J , ∑ δ i = 1 , δ i ≥ 0 \sum\delta_i\mathbf1_{V(C_i')}=\delta\mathbf1_{I\Delta J},\sum\delta_i=1,\delta_i\geq0 ∑δi1V(Ci′)=δ1IΔJ,∑δi=1,δi≥0;
With probability β δ γ i α γ + β δ \frac{\beta\delta\gamma_i}{\alpha\gamma+\beta\delta} αγ+βδβδγi for each i i i:
Let I : = I Δ V ( C i ) I:=I\Delta V(C_i) I:=IΔV(Ci);
Else with probability α γ δ i α γ + β δ \frac{\alpha\gamma\delta_i}{\alpha\gamma+\beta\delta} αγ+βδαγδi for each i i i:
Let J : = J Δ V ( C i ′ ) J:=J\Delta V(C_i') J:=JΔV(Ci′).
EndWhile;
Output I I I.
Theorem 4 (Chernoff-type Concentration Bounds for Matroid Intersection Case)
-
令 ( X 1 , . . . , X n ) (X_1,...,X_n) (X1,...,Xn) 为从 ( x 1 , . . . , x n ) ∈ P ( M 1 ) ∩ P ( M 2 ) (x_1,...,x_n)\in P(\mathcal M_1)\cap P(\mathcal M_2) (x1,...,xn)∈P(M1)∩P(M2) 开始进行 randomized swap rounding for matroid intersection 得到的整点. 令 Q Q Q 为 M 1 \mathcal M_1 M1 和 M 2 \mathcal M_2 M2 里等价的元素的集合. 则对于任意的线性函数 Z = ∑ i ∈ Q a i X i Z=\sum_{i\in Q}a_iX_i Z=∑i∈QaiXi, 其中 a i ∈ [ 0 , 1 ] , E [ Z ] = ∑ i ∈ Q a i x i a_i\in[0,1],\mathbf E[Z]=\sum_{i\in Q}a_i x_i ai∈[0,1],E[Z]=∑i∈Qaixi,
-
若 δ ≥ 0 \delta\geq0 δ≥0 且 μ ≥ E [ Z ] \mu\geq\mathbf E[Z] μ≥E[Z], 则
Pr [ Z ≥ ( 1 + δ ) μ ] ≤ ( e δ ( 1 + δ ) 1 + δ ) μ . \text{Pr}[Z\geq(1+\delta)\mu]\leq\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^\mu. Pr[Z≥(1+δ)μ]≤((1+δ)1+δeδ)μ. -
若 δ ∈ [ 0 , 1 ] \delta\in[0,1] δ∈[0,1] 且 μ ≤ E [ Z ] \mu\leq\mathbf E[Z] μ≤E[Z], 则
Pr [ Z ≤ ( 1 − δ ) μ ] ≤ e − μ δ 2 / 2 . \text{Pr}[Z\leq(1-\delta)\mu]\leq e^{-\mu\delta^2/2}. Pr[Z≤(1−δ)μ]≤e−μδ2/2.
进一步说, 对于任意的次模函数 f : { 0 , 1 } Q → R f:\{0,1\}^Q\to\R f:{0,1}Q→R 和它的 multilinear extension F : [ 0 , 1 ] Q → R F:[0,1]^Q\to\R F:[0,1]Q→R,
E [ f ( X i : i ∈ Q ) ] ≥ F ( x i : i ∈ Q ) . \mathbf E[f(X_i:i\in Q)]\geq F(x_i:i\in Q). E[f(Xi:i∈Q)]≥F(xi:i∈Q). -
-
Let ( X 1 , . . . , X n ) (X_1,...,X_n) (X1,...,Xn) be obtained by randomized swap rounding for matroid intersection from a starting point ( x 1 , . . . , x n ) ∈ P ( M 1 ) ∩ P ( M 2 ) (x_1,...,x_n)\in P(\mathcal M_1)\cap P(\mathcal M_2) (x1,...,xn)∈P(M1)∩P(M2). Let Q Q Q be a set of elements equivalent either in M 1 \mathcal M_1 M1 or M 2 \mathcal M_2 M2. Then for any linear function Z = ∑ i ∈ Q a i X i Z=\sum_{i\in Q}a_iX_i Z=∑i∈QaiXi with a i ∈ [ 0 , 1 ] , E [ Z ] = ∑ i ∈ Q a i x i a_i\in[0,1],\mathbf E[Z]=\sum_{i\in Q}a_i x_i ai∈[0,1],E[Z]=∑i∈Qaixi and
-
If δ ≥ 0 \delta\geq0 δ≥0 and μ ≥ E [ Z ] \mu\geq\mathbf E[Z] μ≥E[Z], then
Pr [ Z ≥ ( 1 + δ ) μ ] ≤ ( e δ ( 1 + δ ) 1 + δ ) μ . \text{Pr}[Z\geq(1+\delta)\mu]\leq\left(\frac{e^\delta}{(1+\delta)^{1+\delta}}\right)^\mu. Pr[Z≥(1+δ)μ]≤((1+δ)1+δeδ)μ. -
If δ ∈ [ 0 , 1 ] \delta\in[0,1] δ∈[0,1], and μ ≤ E [ Z ] \mu\leq\mathbf E[Z] μ≤E[Z], then
Pr [ Z ≤ ( 1 − δ ) μ ] ≤ e − μ δ 2 / 2 . \text{Pr}[Z\leq(1-\delta)\mu]\leq e^{-\mu\delta^2/2}. Pr[Z≤(1−δ)μ]≤e−μδ2/2.
Moreover, for any submodular function f : { 0 , 1 } Q → R f:\{0,1\}^Q\to\R f:{0,1}Q→R and its multilinear extension F : [ 0 , 1 ] Q → R F:[0,1]^Q\to\R F:[0,1]Q→R,
E [ f ( X i : i ∈ Q ) ] ≥ F ( x i : i ∈ Q ) . \mathbf E[f(X_i:i\in Q)]\geq F(x_i:i\in Q). E[f(Xi:i∈Q)]≥F(xi:i∈Q). -
Remark
Theorem 4 给出, 在两个拟阵相交的约束下, 依然能基本达到一个拟阵情况下的 bound (Theorem 2).
Theorem 5 (Constant Approximation Algorithm for Matroid Intersection)
-
令 M 1 = ( N , I 1 ) , M 2 = ( N , I 2 ) \mathcal M_1=(N,\mathcal I_1),\mathcal M_2=(N,\mathcal I_2) M1=(N,I1),M2=(N,I2) 为两个拟阵且 C 1 , C 2 ⊂ 2 N \mathcal C_1,\mathcal C_2\subset2^N C1,C2⊂2N 为它们各自的等价元素的幂集. 令 f : { 0 , 1 } n → R + f:\{0,1\}^n\to\R_+ f:{0,1}n→R+, 使得
f ( S ) = ∑ C ∈ C 1 f 1 , C ( S ∩ C ) + ∑ C ∈ C 2 f 2 , C ( S ∩ C ) f(S)=\sum_{C\in\mathcal C_1}f_{1,C}(S\cap C)+\sum_{C\in\mathcal C_2}f_{2,C}(S\cap C) f(S)=C∈C1∑f1,C(S∩C)+C∈C2∑f2,C(S∩C)
其中 f 1 , C , f 2 , C f_{1,C},f_{2,C} f1,C,f2,C 是单调次模的. 则对于 max { f ( S ) : S ∈ I 1 ∩ I 2 } \max\{f(S):S\in\mathcal I_1\cap\mathcal I_2\} max{f(S):S∈I1∩I2} 存在一个 ( 1 − 1 / e ) (1-1/e) (1−1/e)-近似的算法. -
Let M 1 = ( N , I 1 ) , M 2 = ( N , I 2 ) \mathcal M_1=(N,\mathcal I_1),\mathcal M_2=(N,\mathcal I_2) M1=(N,I1),M2=(N,I2) be two matroids and C 1 , C 2 ⊂ 2 N \mathcal C_1,\mathcal C_2\subset2^N C1,C2⊂2N their respective families of equivalence classes. Let f : { 0 , 1 } n → R + f:\{0,1\}^n\to\R_+ f:{0,1}n→R+ be such that
f ( S ) = ∑ C ∈ C 1 f 1 , C ( S ∩ C ) + ∑ C ∈ C 2 f 2 , C ( S ∩ C ) f(S)=\sum_{C\in\mathcal C_1}f_{1,C}(S\cap C)+\sum_{C\in\mathcal C_2}f_{2,C}(S\cap C) f(S)=C∈C1∑f1,C(S∩C)+C∈C2∑f2,C(S∩C)
where f 1 , C , f 2 , C f_{1,C},f_{2,C} f1,C,f2,C are monotone submodular. Then there is a ( 1 − 1 / e ) (1-1/e) (1−1/e)-approximation for f : { 0 , 1 } n → R + f:\{0,1\}^n\to\R_+ f:{0,1}n→R+.
Theorem 6
- 考虑单调的次模函数
f
1
,
.
.
.
,
f
n
:
2
N
→
R
+
f_1,...,f_n:2^N\to\R_+
f1,...,fn:2N→R+, 和它们的 multilinear extensions
F
i
:
[
0
,
1
]
N
→
R
+
F_i:[0,1]^N\to\R_+
Fi:[0,1]N→R+ 和一个可解的多面体
P
⊆
[
0
,
1
]
N
P\subseteq[0,1]^N
P⊆[0,1]N. 存在一个算法, 给定
V
1
,
.
.
.
,
V
n
∈
R
+
V_1,...,V_n\in\R_+
V1,...,Vn∈R+,
- 要么找到一个点 x ∈ P x\in P x∈P 使得 F i ( x ) ≥ ( 1 − 1 / e ) V i F_i(x)\geq(1-1/e)V_i Fi(x)≥(1−1/e)Vi 对每个 i i i,
- 要么返回一个说明, 不存在点 x ∈ P x\in P x∈P 使得 F i ( x ) ≥ V i F_i(x)\geq V_i Fi(x)≥Vi 对所有 i i i.
- Consider monotone submodular functions
f
1
,
.
.
.
,
f
n
:
2
N
→
R
+
f_1,...,f_n:2^N\to\R_+
f1,...,fn:2N→R+, their multilinear extensions
F
i
:
[
0
,
1
]
N
→
R
+
F_i:[0,1]^N\to\R_+
Fi:[0,1]N→R+ and a solvable polytope
P
⊆
[
0
,
1
]
N
P\subseteq[0,1]^N
P⊆[0,1]N. There is an algorithm which, given
V
1
,
.
.
.
,
V
n
∈
R
+
V_1,...,V_n\in\R_+
V1,...,Vn∈R+,
- either finds a point x ∈ P x\in P x∈P such that F i ( x ) ≥ ( 1 − 1 / e ) V i F_i(x)\geq(1-1/e)V_i Fi(x)≥(1−1/e)Vi for each i i i,
- or returns a certificate that there is no point x ∈ P x\in P x∈P such that F i ( x ) ≥ V i F_i(x)\geq V_i Fi(x)≥Vi for all i i i.
Theorem 7 (Approximation Algorithm for Matroid Bases with Packing Constraints)
-
给定一个拟阵 M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) 和一个矩阵 A ∈ R m × N A\in\R^{m\times N} A∈Rm×N, 对以下问题存在一个 O ( log m / log log m ) O(\log m/\log\log m) O(logm/loglogm)-近似的算法.
min { λ : ∃ base B in M ; A ⋅ 1 B ≤ λ 1 } . \min\{\lambda:\exist\text{~base~$B$~in~$\mathcal M$;~}A\cdot\mathbf1_B\leq\lambda\mathbf 1\}. min{λ:∃ base B in M; A⋅1B≤λ1}. -
Given a matroid M = ( N , I ) \mathcal M=(N,\mathcal I) M=(N,I) and a matrix A ∈ R m × N A\in\R^{m\times N} A∈Rm×N , there is an O ( log m / log log m ) O(\log m/\log\log m) O(logm/loglogm)-approximation for the problem
min { λ : ∃ base B in M ; A ⋅ 1 B ≤ λ 1 } . \min\{\lambda:\exist\text{~base~$B$~in~$\mathcal M$;~}A\cdot\mathbf1_B\leq\lambda\mathbf 1\}. min{λ:∃ base B in M; A⋅1B≤λ1}.
Max-Min Submodular Allocation
- 给定 m m m 个项目和 n n n 个用户, 以及一个次模的函数 w i : { 0 , 1 } m → R + w_i:\{0,1\}^m\to\R_+ wi:{0,1}m→R+, 对不相交的集合 S 1 , . . . S n S_1,...S_n S1,...Sn 找到一个分配以最小化 min 1 ≤ i ≤ n w i ( S i ) \min_{1\leq i\leq n}w_i(S_i) min1≤i≤nwi(Si).
- Given m m m items and n n n agents with submodular valuation functions w i : { 0 , 1 } m → R + w_i:\{0,1\}^m\to\R_+ wi:{0,1}m→R+, find an allocation of disjoint sets S 1 , . . . S n S_1,...S_n S1,...Sn maximizing min 1 ≤ i ≤ n w i ( S i ) \min_{1\leq i\leq n}w_i(S_i) min1≤i≤nwi(Si).
Theorem 8 (Approximation Algorithm for Max-Min Submodular Allocation)
- 对于任意 ε > 0 \varepsilon>0 ε>0 和任意常数数量的用户 n > 2 n>2 n>2, 对 Max-Min Submodular Allocation (in the value oracle model) 问题, 存在一个 ( 1 − 1 / e − ε ) (1-1/e-\varepsilon) (1−1/e−ε)-近似算法.
- For any ε > 0 \varepsilon>0 ε>0 and any constant number of agents n > 2 n>2 n>2, there is a 1 − 1 / e − ε 1-1/e-\varepsilon 1−1/e−ε-approximation for Max-Min Submodular Allocation (in the value oracle model). This is the best possible factor independent of n n n, since a ( 1 − ( 1 − 1 / n ) n + ε ) (1 − (1 − 1/n) ^n + \varepsilon) (1−(1−1/n)n+ε)-approximation for any fixed n ≥ 2 n\geq2 n≥2 and ε > 0 \varepsilon>0 ε>0 would require exponentially many queries.
References
[1] Chekuri C, Vondrák J, Zenklusen R. Dependent randomized rounding via exchange properties of combinatorial structures[C]//2010 IEEE 51st Annual Symposium on Foundations of Computer Science. IEEE, 2010: 575-584.
[2] J. Edmonds. Matroids, submodular functions and certain polyhedra. Combinatorial Structures and Their Applications, Gordon and Breach, New York, 1970, 69–87.
[3] A. Schrijver. Combinatorial optimization - polyhedra and efficiency. Springer, 2003.