次/超模 (supermodular/submodular)

次/超模 (supermodular/submodular)

定义

submodular function

Definition1: In mathemaics, a function $f:R^k\to R$ is submodular if

$$f(x\uparrow y)+f(x\downarrow y)\le f(x)+f(y)$$

for all $x,y\in R^k$, where $x\uparrow y$ denotes the componentwise maximum and $x\downarrow y$ the componentwise minimun of $x$ and $y$.

Definition2: If $f$ is twice continuously differentiable, then submodularity is equibalent to the condition

$$\frac{{\partial }^{2}f}{\partial z_i\partial z_j}\le 0\text{for all}i\ne j$$

submodular set function

Definition 1: A function $f:2^N\to R$ is submodular if for any $S,T\subseteq N$,

$$f(S\cup T)+f(S\cap T)\le f(S)+f(T)$$

Definition 2: Submodularity can be alternatively defined by

f ( S ∪ { j } ) − f ( S ) ≥ f ( T ∪ { j } ) − f ( T ) f(S\cup \{j\})-f(S)\geq f(T\cup \{j\})-f(T) f(S∪{j})−f(S)≥f(T∪{j})−f(T)

for all $S\subseteq T,j\notin T$

性质

• If $f(x)$ is submodular, then $-f(x)$ is supermodular

• a sum of submodular functions is a submodular function

• 复合函数，Topik (1978)

  

  e.g., 第四行表示，$f$ is concave increasing function, $g$ is submodular, then $f\circ g=f(g(\cdot ))$ is submodular.

Lovász Extension and Minimization of Submodular set Function

可以用于找到 submodular set function 的最小值

Definition: Givena set function $f:2^N\to R$, the Lovász Extension $f^L:[0,1{]}^N\to R$ is defined as $f^L(x)={\sum }_{j=1}^m{\lambda }_{j}f(S_j)$, where { S j } \{S_j\} {Sj​} is the unique decreasing series of sets $N=S_1\supset S_2...\supset S_m=\varnothing$ such that $x={\sum }_j{\lambda }_j{1}_{S_j}$ for ${\sum }_j{\lambda }_j=1,{\lambda }_j\ge 0$

submodular set function 的 Lovász Extension 总是凸的。同样的， 如果某个 set function 的 Lovász Extension 是凸的，那么这个 set function 一定是 submodular。

[minimization of submodular set function]

If $f:2^N\to R$ is a submodular set functin, then the minimizer of its Lovász Extension is dimain $[0,1{]}^N$ can be ontained at vertex points: ${\min }_{x\in [0,1{]}^N}{f}^L(x)={\min }_{S\subseteq N}f(S)$

Multilinear relaxation and Maximization of Submodular set Function

在很多情况下我们也想要最大化一个 submodular set function。例如，一个公司在有限预算下投放广告，人与人之间存在行为影响。公司需要决策将广告投放给哪些人能够最大化效用。将会互相影响的 consumers 团体作为一个 set，k 可以看作是公司在有限预算下能够投放的广告数。那么投放给哪些 consumers 能最大化效用，实际上就是找到 $k$ 个 sets 能覆盖最多的 consumers。这个简化的模型称为 Max-k-Cover problem:

• Given a set of sets { S j ⊆ N ∣ j ∈ A } \{S_j\subseteq N|j\in A\} {Sj​⊆N∣j∈A}, find $k$ sets which covers the most number of elements.

也可以给每个 consumer 的价值赋值。The Maximum Coverage Problem:

• Given a set of $S_1,S_2,...,S_m\subseteq N$. For each element $i\in N$, it has a value ${\nu }_i\ge 0$, and for each set $S\subseteq N$ the value function is defined as $V(S)={\sum }_{i\in S}{\nu }_i$. We need to select $k$ sets { S j ∣ j ∈ A } \{S_j|j\in A\} {Sj​∣j∈A}, and to maximize the value $V({\cup }_{j\in A}{S}_j)$.

这个问题的目标函数有次模性。（证明不是很难）由此衍生出来的还有 Assortment Optimization 问题

• 假设有 $N$ 个互相替代商品，要选择其中的一些商品做广告。广告数量（or 广告位）不多于 $K$ 个。一些论文里会假设投放广告的利润 $V(S)$ 是 sumodular，那么问题可以表示为 max ⁡   { V ( S ) : ∣ S ∣ ≤ K , S ⊆ N } \max\ \{V(S):|S|\leq K,S\subseteq N\} max {V(S):∣S∣≤K,S⊆N}

以上有数量约束 (cardinality constrained) 的最大化submodular set function 的问题都是 NP hard。

对于最大化问题，可以引入 Multilinear relaxation

Definition: Given set function $f:2^N\to R$, we define its multilinear relaxiation by rounding a countinous point $x\in [0,1{]}^N$ to { 0 , 1 } N \{0,1\}^N {0,1}N : $F(x)=E[f(\xi (x))]$, where $\xi (x)\in R^N$ takes value $\xi (x{)}_i=1$ with probability $x_i$, and $\xi (x{)}_i=0$ with probability $1-x_i$ independently

可以由 $\max F(x)$ 得出近似结果。（有一些有用的算法，有空的话可以补）

Supermodular stochastic order (Muller and Scarsini 2000)

Definition: A random vector $D_1$ is said to be smaller than the random vector $D_2$ in the supermodular order, written $D_1{\le }_{sm}{D}_2$, if $E[f(D_1)]\le E[f(D_2)]$ for all supermodular functon $f$ such that the expectation exists.

根据定义，Muller and Scarsini 得到结论：

• Let $D_1$ and $D_2$ be multivariate normal random vectors with parameters $D_1\sim N(\mu ,{\Sigma }_1)$ and $D_2\sim N(\mu ,{\Sigma }_2)$, where ${\Sigma }_1,{\Sigma }_2$ are covariance matrices such that ${\sigma }_{ii}^1={\sigma }_{ii}^2,{\sigma }_{ij}^1\le {\sigma }_{ij}^2$. Then $D_1{\le }_{sm}{D}_2$.