6 密度估计的方法
# 本节内容中约束条件需要由某些一致性定理保证,这里没有给出。
SVM用于密度估计问题
密度估计问题可以表述为:
求解概率密度
p ( t , a ) : ∫ − ∞ x p ( t , α ) d t = F ( x ) p(t,a): \int_{-\infty}^x p(t, \alpha) dt = F(x) p(t,a):∫−∞xp(t,α)dt=F(x)
已知 ( x i ) (x_i) (xi) 为 满足 F ( x ) F(x) F(x) 分布的随机独立样本集,并据此可给出经验分布函数
F l ( x ) = 1 l ∑ i = 1 l θ ( x − x i ) F_l(x) = \dfrac{1}{l} \sum\limits_{i=1}^l \theta(x-x_i) Fl(x)=l1i=1∑lθ(x−xi)
θ \theta θ 为阶跃函数。
考虑使用一致度量来定义概率分布与经验分布之间的距离
ρ ( F ( x ) , F l ( x ) ) = sup x ∣ F ( x ) − F l ( x ) ∣ \rho(F(x),F_l(x)) = \sup\limits_x |F(x) - F_l(x)| ρ(F(x),Fl(x))=xsup∣F(x)−Fl(x)∣
由某再生核希尔伯特空间(Reproducing Kernel Hilbert Space, RKHS)的模定义正则化泛函
Ω ( f ) = ( f , f ) H \Omega(f) = (f,f)_H Ω(f)=(f,f)H
首先先构造RKHS。设有对称正定核
K ( x , y ) = ∑ i = 1 ∞ λ i ϕ i ( x ) ϕ i ( y ) K(x,y) = \sum\limits_{i=1}^\infty \lambda_i \phi_i(x)\phi_i(y) K(x,y)=i=1∑∞λiϕi(x)ϕi(y)
定义的希尔伯特空间H的内积 ( f , g ) H (f,g)_H (f,g)H 需要满足再生特性
( f ( x ) , K ( x , y ) ) H = f ( y ) , ∀ f ∈ H (f(x),K(x,y))_H = f(y), \forall f \in H (f(x),K(x,y))H=f(y),∀f∈H
考虑函数集
f ( x , c ) = ∑ i = 1 ∞ c i ϕ i ( x ) f(x,c) = \sum\limits_{i=1}^\infty c_i \phi_i(x) f(x,c)=i=1∑∞ciϕi(x)
及定义在其上的内积
( f ( x , c ) , f ( x , d ) ) H = ∑ i = 1 ∞ c i d i λ i (f(x,c), f(x,d))_H = \sum\limits_{i=1}^\infty \dfrac{c_i d_i}{\lambda_i} (f(x,c),f(x,d))H=i=1∑∞λicidi
有
( f ( x , c ) , K ( x , y ) ) H = ( f ( x , c ) , f ( x , λ i ϕ i ( y ) ) ) H = ∑ i = 1 ∞ c i λ i ϕ i ( y ) λ i = f ( y ) (f(x,c), K(x,y))_H = (f(x,c), f(x, \lambda_i \phi_i(y)))_H = \sum\limits_{i=1}^\infty \dfrac{c_i \lambda_i \phi_i(y)}{\lambda_i} = f(y) (f(x,c),K(x,y))H=(f(x,c),f(x,λiϕi(y)))H=i=1∑∞λiciλiϕi(y)=f(y)
因此上述定义满足再生特性, f , K , ( f , g ) H f,K,(f,g)_H f,K,(f,g)H 构成了RKHS。
寻找如下形式解
f ( x ) = ∑ i = 1 l β i K ( x i , x ) = ∑ k = 1 ∞ ∑ i = 1 l β i λ k ϕ k ( x i ) ϕ k ( x ) , β i ≥ 0 f(x) = \sum\limits_{i=1}^l \beta_i K(x_i, x) = \sum\limits_{k=1}^\infty \sum\limits_{i=1}^l \beta_i \lambda_k \phi_k(x_i) \phi_k(x),\ \ \ \beta_i \ge 0 f(x)=i=1∑lβiK(xi,x)=k=1∑∞i=1∑lβiλkϕk(xi)ϕk(x), βi≥0
使用方法P(Phillips 残差方法, 1962)解决密度估计问题,即最小化泛函
Ω ( f ) = ( f , f ) H , sup x ∣ F ( x ) − F l ( x ) ∣ = σ l \Omega(f) = (f,f)_H,\ \ \ \sup\limits_x |F(x) - F_l(x)| = \sigma_l Ω(f)=