Dependent Randomized Rounding via Exchange Properties of Combinatorial Structures 学习笔记

Dependent Randomized Rounding via Exchange Properties of Combinatorial Structures [1] 学习笔记

Author: Sijin Yu

1. 原文的重要结论

Rounding 1: Randomized Swap Rounding

• $x$ 为 $P$ 的顶点的凸组合, 即 $x={\sum }_{i=1}^n{\alpha }_i{\mathbf{v}}_i$, 其中 ${\mathbf{v}}_i$ 是 $P$ 的顶点, 且 ${\sum }_i{\alpha }_i=1$.
• 令 ${\mathbf{w}}_1={\mathbf{v}}_1,{\beta }_1={\alpha }_1$, 在每一步, 使用参数 ${\beta }_{i+1}={\beta }_i+{\alpha }_{i+1}$ 将两个顶点 ${\mathbf{w}}_i,{\mathbf{v}}_{i+1}$ 合并为新的顶点 ${\mathbf{w}}_{i+1}$, 使 $\mathbf{E}[{\beta }_{i+1}\mathbf{w}i+1]={\beta }_i{\mathbf{w}}_i+{\alpha }_{i+1}{\mathbf{v}}_{i+1}$ (the merge operation).
• 在 $n-1$ 步后, 得到一个顶点 $X={\mathbf{w}}_n$ 使得 $\mathbf{E}[X]={\sum }_{i=1}^n{\alpha }_i{\mathbf{v}}_i=x$

Theorem 1

• 令 $P$ 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 $x\in P$, 存在一个 $P$ 的顶点的概率分布, 使得一个从这一分布采样的随机顶点 $X$ 满足 $\mathbf{E}[X]=x$ 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) $P$ 是一个拟阵基多面体的投影, 即存在一个拟阵 $\mathcal{M}=(N^{\prime },\mathcal{I})$ 且 $N\subseteq N^{\prime }$, 则 $X$ 是 $P$ 的顶点, 当且仅当 $X=1_{B\cap N}$ 对某个 $\mathcal{M}$ 的基 $B$ 成立.

• Let $P$ 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 $x\in P$, there exists a probability distribution over vertices of $P$ such that a random vertex $X$ drawn from this distribution satisfies $\mathbf{E}[X]=x$ and the coordinates { X i } i ∈ N \{X_i\}_{i\in N} {Xi​}i∈N​ are negatively correlated.

  (2) $P$ is a projection of a matroid base polytope, in the sense that there is a matroid $\mathcal{M}=(N^{\prime },\mathcal{I})$ such that $N\subseteq N^{\prime }$ and $X$ is a vertex of $P$ iff $X=1_{B\cap N}$ for some base $B$ of $\mathcal{M}$.

Definition 1: Matroid Polytopes

• 给定一个拟阵 $\mathcal{M}=(N,\mathcal{I})$ 和秩函数 $r:2^N\to {\mathbb{Z}}_{+}$, 拟阵多面体 $P(\mathcal{M})$ 定义为 $\mathcal{M}$ 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 $\mathcal{M}=(N,\mathcal{I})$ with rank function $r:2^N\to {\mathbb{Z}}_{+}$, the matroid polytope $P(\mathcal{M})$ is defined as the convex hull of characteristic vectors of the independent sets of $\mathcal{M}$, 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 $\mathcal{M}=(N,\mathcal{I})$ 为一拟阵并令 $B_1,B_2$ 为基. 对于任意的 $i\in B_1-B_2$ 总存在 $j\in B_2-B_1$, 使得 $B_1-i+j$ 和 $B_2-j+i$ 也为基.
• Let $\mathcal{M}=(N,\mathcal{I})$ be a matroid and let $B_1,B_2$ be bases. For any $i\in B_1-B_2$ there exists $j\in B_2-B_1$ such that $B_1-i+j$ and $B_2-j+i$ are also bases.

Definition 2: Matroid Intersection

• 令 ${\mathcal{M}}_1=(N,{\mathcal{I}}_1)$ 和 ${\mathcal{M}}_2=(N,{\mathcal{I}}_2)$ 为两个拟阵. 由 Edmonds 等人的工作 [2], 相交多面体 $P({\mathcal{M}}_1)\cap P({\mathcal{M}}_2)$ 是整数的, 且它的顶点和集合 $I\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ 相关.
• Let ${\mathcal{M}}_1=(N,{\mathcal{I}}_1)$ and ${\mathcal{M}}_2=(N,{\mathcal{I}}_2)$ be two matroids. It follows from the work of Edmonds [2] that the intersection polytope $P({\mathcal{M}}_1)\cap P({\mathcal{M}}_2)$ is integral and its vertices correspond exactly to sets $I\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$.

Definition 3: Equivalent Elements in Matroids

• 在一个拟阵 $\mathcal{M}=(N,\mathcal{I})$ 中, 两个元素 $i,j\in N$ 是等价的, 若对于任意的不包含 $i,j$ 的集合 $A\in \mathcal{I}$, $A+i\in \mathcal{I}$ 当且仅当 $A+j\in \mathcal{I}$.
• Two elements $i,j\in N$ are equivalent in a matroid $\mathcal{M}=(N,\mathcal{I})$ if for any set $A\in \mathcal{I}$ not containing $i,j$, $A+i\in \mathcal{I}$ if and only if $A+j\in \mathcal{I}$.

Remark

对于任意一个不包含 $i,j$ 的集合 $A\in \mathcal{I}$, 增加 $i$ 后依然独立, 和增加 $j$ 后依然独立是等价的, 因此 $i$ 和 $j$ 对 $A$ 的"扩张"的影响是等效的, 所以 $i$ 和 $j$ 等价.

Definition 4: Submodular Function

• 一个函数 $f:2^N\to \mathbb{R}$ 是次模的, 若给定任意的 $A,B\subseteq N$, 都有 $f(A)+f(B)\ge f(A\cup B)+f(A\cap B)$.
• A function $f:2^N\to \mathbb{R}$ is submodular if for any $A,B\subseteq N$, $f(A)+f(B)\ge f(A\cup B)+f(A\cap B)$.

Definition 5: Multilinear Extension

• 对函数 $f:2^N\to \mathbb{R}$, 其 multilinear extension 为函数 $F:[0,1{]}^n\to \mathbb{R}$, 定义为:
  $$F(x)=\sum _{S\subseteq N}f(S)\prod _{i\in S}{x}_i\prod _{i\in N-S}(1-x_i)$$

• The multilinear extension of $f:2^N\to \mathbb{R}$ is the function $F:[0,1{]}^n\to \mathbb{R}$ defined by:

$$F(x)=\sum _{S\subseteq N}f(S)\prod _{i\in S}{x}_i\prod _{i\in N-S}(1-x_i)$$

Remark

$F(\cdot )$ 可以看作 $f(\cdot )$ 的期望

Theorem 2 (Chernoff-type Concentration Bounds for any Linear Function)

• 令 $(x_1,...,x_n)\in P(\mathcal{M})$ 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 $Z=\sum a_i{X}_i$, 其中 $a_i\in [0,1]$, 有 $\mathbf{E}[Z]=\sum a_i{x}_i$ 且:

  • 若 $\delta >0$ 和 $\mu \ge \mathbf{E}[Z]$, 则
    $$\text{Pr}[Z\ge (1+\delta )\mu ]\le {\left(\frac{e^{\delta }}{(1+\delta {)}^{1+\delta }}\right)}^{\mu }.$$

  • 若 $\delta \in [0,1]$ 和 $\mu \le \mathbf{E}[Z]$, 则
    $$\text{Pr}[Z\le (1-\delta )\mu ]\le e^{-\mu {\delta }^2/2}.$$

  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 $F:[0,1{]}^n\to \mathbb{R}$,
  $$\mathbf{E}[f(X_1,...,X_n)]\ge F(x_1,...,x_n).$$

• Let $(x_1,...,x_n)\in P(\mathcal{M})$ 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 $Z=\sum a_i{X}_i$ with $a_i\in [0,1]$, we have $\mathbf{E}[Z]=\sum a_i{x}_i$ and:

  • If $\delta >0$ and $\mu \ge \mathbf{E}[Z]$, then
    $$\text{Pr}[Z\ge (1+\delta )\mu ]\le {\left(\frac{e^{\delta }}{(1+\delta {)}^{1+\delta }}\right)}^{\mu }.$$

  • If $\delta \in [0,1]$ and $\mu \le \mathbf{E}[Z]$, then
    $$\text{Pr}[Z\le (1-\delta )\mu ]\le e^{-\mu {\delta }^2/2}.$$

  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 $F:[0,1{]}^n\to \mathbb{R}$,
  $$\mathbf{E}[f(X_1,...,X_n)]\ge F(x_1,...,x_n).$$

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 $[0,1]$ 内, $F:[0,1{]}^n\to {\mathbb{R}}_{+}$ 为它的 multilinear extension. 令 $(x_1,...,x_n)\in P(\mathcal{M})$ 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 ${\mu }_0=F(x_1,...,x_n)$ 和 $\delta >0$. 则
  $$\text{Pr}[f(X_1,...,X_n)\le (1-\delta ){\mu }_0]\le e^{-{\mu }_0{\delta }^2/8}.$$

• 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 $[0,1]$, and $F:[0,1{]}^n\to {\mathbb{R}}_{+}$ its multilinear extension. Let $(x_1,...,x_n)\in P(\mathcal{M})$ 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 ${\mu }_0=F(x_1,...,x_n)$ and $\delta >0$. Then
  $$\text{Pr}[f(X_1,...,X_n)\le (1-\delta ){\mu }_0]\le e^{-{\mu }_0{\delta }^2/8}.$$

Remark

该不等式是结束点 $f(X_1,...,X_n)$ 与起始点 ${\mu }_0=F(x_1,...,x_n)$ 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 $n$ 无关.

Rounding 2: Randomized Swap Rounding for Matroids

令 $\mathcal{M}=(N,\mathcal{I})$ 为一个秩为 $d=r(N)$ 的拟阵, 并令 $n=|N|$. Randomized Swap Rounding 将一个点 $x\in P(\mathcal{M})$ round 到一个基 $x\in B(\mathcal{M})$ 上. 将 $x$ 预先表示为基的凸组合, 即 $x={\sum }_{l=1}^m{\beta }_l{1}_{B_l}$, 且 ${\sum }_{l=1}^m{\beta }_l=1,{\beta }_l>0$. 凸组合可以根据 [3] 找到. rounding 将进行 $n-1$ 次迭代, 在第 $1$ 步我们将两个基 $B_1,B_2$ (随机的) 组合为 $C_2$, 并用 $({\beta }_1+{\beta }_2)1_{C_2}$ 替代线性组合中的 ${\beta }_1{1}_{B_1}+{\beta }_2{1}_{B_2}$. 在第 $k$ 步, $C_k$ 和 $B_{k+1}$ 组合为一个新的基 $C_{k+1}$, 并用 $\left({\sum }_{l=1}^{k+1}{\beta }_l\right)1_{C_{k+1}}$ 替代线性组合中的 $\left({\sum }_{l=1}^k{\beta }_l\right)1_{C_k}+{\beta }_{k+1}{1}_{B_{k+1}}$. 最后 $C_n$ 被返回. ($x$ 则被 round 为整数点 $1_{C_n}$)

算法: 输入 ${\beta }_1,B_1,{\beta }_2,B_2$, 将基 $B_1$ 和 $B_2$ 组合, 返回新的基.

Algorithm MergeBases(${\beta }_1,B_1,{\beta }_2,B_2$):
While($B_1\ne B_2$) do:
Pick $i\in B_1-B_2$ and find $j\in B_2-B_1$ such that $B_1-i+j\in \mathcal{I}$ and $B_2-j+i\in \mathcal{I}$;
With probability $\frac{{\beta }_1}{{\beta }_1+{\beta }_2}$ do:
$B_2\leftarrow B_2-j+i$;
Else:
$B_1\leftarrow B_1-i+j$;
EndWhile;
Output $B_1$.

Remark

该算法的意思是让 $B_1$ 和 $B_2$ 不断靠拢, 并以线性系数 ${\beta }_1$ 和 ${\beta }_2$ 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 $x$, 输出将其 round 到的基 $C_n$.

Algorithm SwapRound($x={\sum }_{l=1}^n{\beta }_l{1}_{B_l}$):
$C_1=B_1$;
For($k=1$ to $n-1$) do:
$C_{k+1}=$ MergeBases(${\sum }_{l=1}^k{\beta }_l,C_k,{\beta }_{k+1},B_{k+1}$);
EndFor;
Output $C_n$.

Definition 6: (For Rounding 3) Exchange Digraph

• 考虑两个拟阵 ${\mathcal{M}}_1=(N,{\mathcal{I}}_1)$ 和 ${\mathcal{M}}_2=(N,{\mathcal{I}}_2)$. 对于 $I\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, 我们定义两个有向图 $D_{{\mathcal{M}}_1}(I)$ 和 $D_{{\mathcal{M}}_2}(I)$ 如下.

  • 对每个 $i\in I,j\in N-I$, 且 $I+j-i\in {\mathcal{I}}_1$, 令边 $(i,j)\in D_{{\mathcal{M}}_1}(I)$;
  • 对每个 $i\in I,j\in N-I$, 且 $I+j-i\in {\mathcal{I}}_2$, 令边 $(i,j)\in D_{{\mathcal{M}}_2}(I)$.

  我们定义一个有向图 $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)=D_{{\mathcal{M}}_1}(I)\cup D_{{\mathcal{M}}_2}(I)$.

• Consider two matroids ${\mathcal{M}}_1=(N,{\mathcal{I}}_1)$ and ${\mathcal{M}}_2=(N,{\mathcal{I}}_2)$. For $I\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, we define two digraphs $D_{{\mathcal{M}}_1}(I)$ and $D_{{\mathcal{M}}_2}(I)$ as follows.

  • For each $i\in I,j\in N-I$ with $I+j-i\in {\mathcal{I}}_1$, we have an arc $(i,j)\in D_{{\mathcal{M}}_1}(I)$;
  • For each $i\in I,j\in N-I$ with $I+j-i\in {\mathcal{I}}_2$, we have an arc $(i,j)\in D_{{\mathcal{M}}_2}(I)$.

  We define a digraph $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)=D_{{\mathcal{M}}_1}(I)\cup D_{{\mathcal{M}}_2}(I)$.

Definition 7: (For Rounding 3) Irreducible Directed Cycle

• 一个 $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ 里的有向圈 $C$ 是不可约的, 如果在点集 $V(C)$ 上 $C\cap D_{{\mathcal{M}}_1}(I)$ 是 $D_{{\mathcal{M}}_1}(I)$ 里的唯一完美匹配, 且 $C\cap D_{{\mathcal{M}}_2}(I)$ 是 $D_{{\mathcal{M}}_2}(I)$ 里的唯一完美匹配. 否则, $C$ 是可约的.
• A directed cycle $C$ in $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ is irreducible if $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ is the unique perfect matching in $D_{{\mathcal{M}}_1}(I)$ and $C\cap D_{{\mathcal{M}}_2}(I)$ is the unique perfect matching in $D_{{\mathcal{M}}_2}(I)$ on the vertex set $V(C)$. Otherwise, $C$ is reducible.

Lemma 2: (For Rounding 3)

• 令 ${\mathcal{M}}_l=(N,{\mathcal{I}}_l),l=1,2$ 为在全集 $N$ 上的拟阵. 假设 $I,J\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ 且 $|I|=|J|$. 则存在一个整数 $s\ge 0$ 和一个 $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ 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 $I\Delta J$ 中的元素, 使得每个 $I\Delta J$ 里的元素都在这些有向圈里出现 $2^s$ 次.
• Let ${\mathcal{M}}_l=(N,{\mathcal{I}}_l),l=1,2$, be matroids on ground set $N$. Suppose that $I,J\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ and $|I|=|J|$. Then there is an integer $s\ge 0$ and a collection of irreducible directed cycles { C 1 , . . . , C m } \{C_1,...,C_m\} {C1​,...,Cm​} (allowing repetition) in $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$, using only elements of $I\Delta J$, so that each element of $I\Delta J$ appears in exactly $2^s$ of the directed cycles.

Remark

collection 是允许出现重复元素的集合, 可翻译为多集.

$I\Delta J$ 定义为 $(I-J)\cup (J-I)$.

Lemma 3: (For Rounding 3, Variant of Lemma 2)

• 令 ${\mathcal{M}}_l=(N,{\mathcal{I}}_l),l=1,2$ 为在全集 $N$ 上的拟阵. 假设 $I,J\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ 且 $|I|=|J|$. 则我们可以在多项式时间内寻在 $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ 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 ${\gamma }_i\ge 0,{\sum }_{i=1}^m{\gamma }_i=1$, 使得对某些 $\gamma >0$, ${\sum }_{i=1}^m{\gamma }_i{1}_{V(C_i)}=\gamma 1_{I\Delta J}$.
• Let ${\mathcal{M}}_l=(N,{\mathcal{I}}_l),l=1,2$, be matroids on ground set $N$. Suppose that $I,J\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ and $|I|=|J|$. 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 $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$, with coefficients ${\gamma }_i\ge 0,{\sum }_{i=1}^m{\gamma }_i=1$, such that for some $\gamma >0$, ${\sum }_{i=1}^m{\gamma }_i{1}_{V(C_i)}=\gamma 1_{I\Delta J}$.

Rounding 3: Rounding for Matroid Intersection

给定两个拟阵 ${\mathcal{M}}_1$, ${\mathcal{M}}_2$ 和一个分数解 $x\in P({\mathcal{M}}_1)\cap P({\mathcal{M}}_2)$, 将 $x$ round 为一整数解 $X$, 它表示一个在两个拟阵中都独立的集合, 且 $\mathbf{E}[X]=x$. 假定 $I,J\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ 为两个基, 令 $x$ 为 $I$ 和 $J$ 的线性组合 $x=\alpha I+\beta J$.

算法: 输入 $x$ 的线性组合参数 $\alpha ,I,\beta ,J$, 输出 $x$ 被 round 到的基.

Algorithm mergeIntersectionBases($\alpha ,I,\beta ,J$):
While($I\ne J$) do:
Generate a collection of irreducible cycles in $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(I)$ such that
$\sum {\gamma }_i{1}_{V(C_i)}=\gamma 1_{I\Delta J},\sum {\gamma }_i=1,{\gamma }_i\ge 0$;
and a collection of irreducible cycles in $D_{{\mathcal{M}}_1,{\mathcal{M}}_2}(J)$ such that
$\sum {\delta }_i{1}_{V(C_i^{\prime })}=\delta 1_{I\Delta J},\sum {\delta }_i=1,{\delta }_i\ge 0$;
With probability $\frac{\beta \delta {\gamma }_i}{\alpha \gamma +\beta \delta }$ for each $i$:
Let $I:=I\Delta V(C_i)$;
Else with probability $\frac{\alpha \gamma {\delta }_i}{\alpha \gamma +\beta \delta }$ for each $i$:
Let $J:=J\Delta V(C_i^{\prime })$.
EndWhile;
Output $I$.

Theorem 4 (Chernoff-type Concentration Bounds for Matroid Intersection Case)

• 令 $(X_1,...,X_n)$ 为从 $(x_1,...,x_n)\in P({\mathcal{M}}_1)\cap P({\mathcal{M}}_2)$ 开始进行 randomized swap rounding for matroid intersection 得到的整点. 令 $Q$ 为 ${\mathcal{M}}_1$ 和 ${\mathcal{M}}_2$ 里等价的元素的集合. 则对于任意的线性函数 $Z={\sum }_{i\in Q}{a}_i{X}_i$, 其中 $a_i\in [0,1],\mathbf{E}[Z]={\sum }_{i\in Q}{a}_i{x}_i$,

  • 若 $\delta \ge 0$ 且 $\mu \ge \mathbf{E}[Z]$, 则
    $$\text{Pr}[Z\ge (1+\delta )\mu ]\le {\left(\frac{e^{\delta }}{(1+\delta {)}^{1+\delta }}\right)}^{\mu }.$$

  • 若 $\delta \in [0,1]$ 且 $\mu \le \mathbf{E}[Z]$, 则
    $$\text{Pr}[Z\le (1-\delta )\mu ]\le e^{-\mu {\delta }^2/2}.$$

  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 $F:[0,1{]}^Q\to \mathbb{R}$,
  $$\mathbf{E}[f(X_i:i\in Q)]\ge F(x_i:i\in Q).$$

• Let $(X_1,...,X_n)$ be obtained by randomized swap rounding for matroid intersection from a starting point $(x_1,...,x_n)\in P({\mathcal{M}}_1)\cap P({\mathcal{M}}_2)$. Let $Q$ be a set of elements equivalent either in ${\mathcal{M}}_1$ or ${\mathcal{M}}_2$. Then for any linear function $Z={\sum }_{i\in Q}{a}_i{X}_i$ with $a_i\in [0,1],\mathbf{E}[Z]={\sum }_{i\in Q}{a}_i{x}_i$ and

  • If $\delta \ge 0$ and $\mu \ge \mathbf{E}[Z]$, then
    $$\text{Pr}[Z\ge (1+\delta )\mu ]\le {\left(\frac{e^{\delta }}{(1+\delta {)}^{1+\delta }}\right)}^{\mu }.$$

  • If $\delta \in [0,1]$, and $\mu \le \mathbf{E}[Z]$, then
    $$\text{Pr}[Z\le (1-\delta )\mu ]\le e^{-\mu {\delta }^2/2}.$$

  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 $F:[0,1{]}^Q\to \mathbb{R}$,
  $$\mathbf{E}[f(X_i:i\in Q)]\ge F(x_i:i\in Q).$$

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 ${\mathcal{M}}_1=(N,{\mathcal{I}}_1),{\mathcal{M}}_2=(N,{\mathcal{I}}_2)$ 为两个拟阵且 ${\mathcal{C}}_1,{\mathcal{C}}_2\subset 2^N$ 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
  $$f(S)=\sum _{C\in {\mathcal{C}}_1}{f}_{1,C}(S\cap C)+\sum _{C\in {\mathcal{C}}_2}{f}_{2,C}(S\cap C)$$
  其中 $f_{1,C},f_{2,C}$ 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 $(1-1/e)$-近似的算法.

• Let ${\mathcal{M}}_1=(N,{\mathcal{I}}_1),{\mathcal{M}}_2=(N,{\mathcal{I}}_2)$ be two matroids and ${\mathcal{C}}_1,{\mathcal{C}}_2\subset 2^N$ 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
  $$f(S)=\sum _{C\in {\mathcal{C}}_1}{f}_{1,C}(S\cap C)+\sum _{C\in {\mathcal{C}}_2}{f}_{2,C}(S\cap C)$$
  where $f_{1,C},f_{2,C}$ are monotone submodular. Then there is a $(1-1/e)$-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 $f_1,...,f_n:2^N\to {\mathbb{R}}_{+}$, 和它们的 multilinear extensions $F_i:[0,1{]}^N\to {\mathbb{R}}_{+}$ 和一个可解的多面体 $P\subseteq [0,1{]}^N$. 存在一个算法, 给定 $V_1,...,V_n\in {\mathbb{R}}_{+}$,
  • 要么找到一个点 $x\in P$ 使得 $F_i(x)\ge (1-1/e)V_i$ 对每个 $i$,
  • 要么返回一个说明, 不存在点 $x\in P$ 使得 $F_i(x)\ge V_i$ 对所有 $i$.
• Consider monotone submodular functions $f_1,...,f_n:2^N\to {\mathbb{R}}_{+}$, their multilinear extensions $F_i:[0,1{]}^N\to {\mathbb{R}}_{+}$ and a solvable polytope $P\subseteq [0,1{]}^N$. There is an algorithm which, given $V_1,...,V_n\in {\mathbb{R}}_{+}$,
  • either finds a point $x\in P$ such that $F_i(x)\ge (1-1/e)V_i$ for each $i$,
  • or returns a certificate that there is no point $x\in P$ such that $F_i(x)\ge V_i$ for all $i$.

Theorem 7 (Approximation Algorithm for Matroid Bases with Packing Constraints)

• 给定一个拟阵 $\mathcal{M}=(N,\mathcal{I})$ 和一个矩阵 $A\in {\mathbb{R}}^{m\times N}$, 对以下问题存在一个 $O(\log m/\log \log m)$-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 $\mathcal{M}=(N,\mathcal{I})$ and a matrix $A\in {\mathbb{R}}^{m\times N}$ , there is an $O(\log m/\log \log m)$-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 $m$ 个项目和 $n$ 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 $S_1,...S_n$ 找到一个分配以最小化 ${\min }_{1\le i\le n}{w}_i(S_i)$.
• Given $m$ items and $n$ 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 $S_1,...S_n$ maximizing ${\min }_{1\le i\le n}{w}_i(S_i)$.

Theorem 8 (Approximation Algorithm for Max-Min Submodular Allocation)

• 对于任意 $\epsilon >0$ 和任意常数数量的用户 $n>2$, 对 Max-Min Submodular Allocation (in the value oracle model) 问题, 存在一个 $(1-1/e-\epsilon )$-近似算法.
• For any $\epsilon >0$ and any constant number of agents $n>2$, there is a $1-1/e-\epsilon$-approximation for Max-Min Submodular Allocation (in the value oracle model). This is the best possible factor independent of $n$, since a $(1-(1-1/n{)}^n+\epsilon )$-approximation for any fixed $n\ge 2$ and $\epsilon >0$ 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.