文章目录
次/超模 (supermodular/submodular)
定义
submodular function
Definition1: In mathemaics, a function f : R k → R f:R^k \rightarrow R f:Rk→R is submodular if
f ( x ↑ y ) + f ( x ↓ y ) ≤ f ( x ) + f ( y ) f(x\uparrow y)+f(x\downarrow y)\leq f(x)+f(y) f(x↑y)+f(x↓y)≤f(x)+f(y)
for all x , y ∈ R k x,y\in R^k x,y∈Rk, where x ↑ y x\uparrow y x↑y denotes the componentwise maximum and x ↓ y x\downarrow y x↓y the componentwise minimun of x x x and y y y.
Definition2: If f f f is twice continuously differentiable, then submodularity is equibalent to the condition
∂ 2 f ∂ z i ∂ z j ≤ 0 for all i ≠ j \frac{\partial^2 f}{\partial z_i \partial z_j}\leq 0 \ \ \text{for all} \ \ i\not=j ∂zi∂zj∂2f≤0 for all i=j
submodular set function
Definition 1: A function f : 2 N → R f: 2^N\rightarrow R f:2N→R is submodular if for any S , T ⊆ N S,T\subseteq N S,T⊆N,
f ( S ∪ T ) + f ( S ∩ T ) ≤ f ( S ) + f ( T ) f(S\cup T)+f(S\cap T)\leq f(S)+f(T) f(S∪T)+f(S∩T)≤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 ⊆ T , j ∉ T S\subseteq T, j\notin T S⊆T,j∈/T
性质
-
If f ( x ) f(x) f(x) is submodular, then − f ( x ) -f(x) −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 f f is concave increasing function, g g g is submodular, then f ∘ g = f ( g ( ⋅ ) ) f\circ g =f(g(\cdot)) f∘g=f(g(⋅)) 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 → R f : 2^N\rightarrow R f:2N→R, the Lovász Extension f L : [ 0 , 1 ] N → R f^L:[0,1]^N\rightarrow R fL:[0,1]N→R is defined as f L ( x ) = ∑ j = 1 m λ j f ( S j ) f^L(x)=\sum_{j=1}^m\lambda_jf(S_j) fL(x)=∑j=1mλjf(Sj), where { S j } \{S_j\} {Sj} is the unique decreasing series of sets N = S 1 ⊃ S 2 . . . ⊃ S m = ∅ N=S_1\supset S_2 ... \supset S_m= \emptyset N=S1⊃S2...⊃Sm=∅ such that x = ∑ j λ j 1 S j x=\sum_j \lambda_j \mathcal{1}_{S_j} x=∑jλj1Sj for ∑ j λ j = 1 , λ j ≥ 0 \sum_j \lambda_j=1, \lambda_j\geq 0 ∑jλj=1,λj≥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 → R f:2^N\rightarrow R f:2N→R is a submodular set functin, then the minimizer of its Lovász Extension is dimain [ 0 , 1 ] N [0,1]^N [0,1]N can be ontained at vertex points: min x ∈ [ 0 , 1 ] N f L ( x ) = min S ⊆ N f ( S ) \min_{x\in[0,1]^N} f^L(x)=\min_{S\subseteq N}f(S) minx∈[0,1]NfL(x)=minS⊆Nf(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 k 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 k 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 ⊆ N S_1,S_2,...,S_m\subseteq N S1,S2,...,Sm⊆N. For each element i ∈ N i\in N i∈N, it has a value ν i ≥ 0 \nu_i\geq 0 νi≥0, and for each set S ⊆ N S\subseteq N S⊆N the value function is defined as V ( S ) = ∑ i ∈ S ν i V(S)=\sum_{i\in S}\nu_i V(S)=∑i∈Sνi. We need to select k k k sets { S j ∣ j ∈ A } \{S_j|j\in A\} {Sj∣j∈A}, and to maximize the value V ( ∪ j ∈ A S j ) V(\cup_{j\in A}S_j) V(∪j∈ASj).
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 N 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 K K 个。一些论文里会假设投放广告的利润 V ( S ) V(S) 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 → R f: 2^N \rightarrow R f:2N→R, we define its multilinear relaxiation by rounding a countinous point x ∈ [ 0 , 1 ] N x\in[0,1]^N x∈[0,1]N to { 0 , 1 } N \{0,1\}^N {0,1}N : F ( x ) = E [ f ( ξ ( x ) ) ] F(x)=E[f(\xi(x))] F(x)=E[f(ξ(x))], where ξ ( x ) ∈ R N \xi(x)\in R^N ξ(x)∈RN takes value ξ ( x ) i = 1 \xi(x)_i = 1 ξ(x)i=1 with probability x i x_i xi, and ξ ( x ) i = 0 \xi(x)_i=0 ξ(x)i=0 with probability 1 − x i 1-x_i 1−xi independently
可以由 max F ( x ) \max F(x) maxF(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 D_1 D1 is said to be smaller than the random vector D 2 D_2 D2 in the supermodular order, written D 1 ≤ s m D 2 D_1\leq_{sm}D_2 D1≤smD2, if E [ f ( D 1 ) ] ≤ E [ f ( D 2 ) ] E[f(D_1)]\leq E[f(D_2)] E[f(D1)]≤E[f(D2)] for all supermodular functon f f f such that the expectation exists.
根据定义,Muller and Scarsini 得到结论:
- Let D 1 D_1 D1 and D 2 D_2 D2 be multivariate normal random vectors with parameters D 1 ∼ N ( μ , Σ 1 ) D_1\sim N(\mu,\Sigma_1) D1∼N(μ,Σ1) and D 2 ∼ N ( μ , Σ 2 ) D_2 \sim N(\mu,\Sigma_2) D2∼N(μ,Σ2), where Σ 1 , Σ 2 \Sigma_1,\Sigma_2 Σ1,Σ2 are covariance matrices such that σ i i 1 = σ i i 2 , σ i j 1 ≤ σ i j 2 \sigma_{ii}^1=\sigma_{ii}^2, \sigma_{ij}^1\leq\sigma_{ij}^2 σii1=σii2,σij1≤σij2. Then D 1 ≤ s m D 2 D_1\leq_{sm}D_2 D1≤smD2.