核函式匯出的核矩陣性質的證明

ykjs_發表於2020-11-25

定理中說核矩陣是半正定且對稱的。我們先證其對稱性:
( κ ( x 1 , x 1 ) κ ( x 1 , x 2 ) κ ( x 1 , x 3 ) ⋯ κ ( x 1 , x n ) κ ( x 2 , x 1 ) κ ( x 2 , x 2 ) κ ( x 2 , x 3 ) ⋯ κ ( x 2 , x n ) ⋮ ⋮ ⋮ ⋱ ⋮ κ ( x n , x 1 ) κ ( x n , x 2 ) κ ( x n , x 3 ) ⋯ κ ( x n , x n ) ) \begin{pmatrix} \kappa(x_1,x_1) & \kappa(x_1,x_2) & \kappa(x_1,x_3) & \cdots & \kappa(x_1,x_n) \\ \kappa(x_2,x_1) & \kappa(x_2,x_2) & \kappa(x_2,x_3) & \cdots & \kappa(x_2,x_n) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \kappa(x_n,x_1) & \kappa(x_n,x_2) & \kappa(x_n,x_3) & \cdots & \kappa(x_n,x_n) \\ \end{pmatrix} κ(x1,x1)κ(x2,x1)κ(xn,x1)κ(x1,x2)κ(x2,x2)κ(xn,x2)κ(x1,x3)κ(x2,x3)κ(xn,x3)κ(x1,xn)κ(x2,xn)κ(xn,xn) κ ( x i , x j ) \kappa(x_i,x_j) κ(xi,xj)按定義展開得到 κ ( x i , x j ) = ϕ ( x i ) T ϕ ( x j ) \kappa(x_i,x_j)=\phi(x_i)^T\phi(x_j) κ(xi,xj)=ϕ(xi)Tϕ(xj)由於 κ ( x i , x j ) \kappa(x_i,x_j) κ(xi,xj)是一個 1 × 1 1\times1 1×1的矩陣(或者理解成一個數),取轉置得到 κ ( x i , x j ) = κ ( x i , x j ) T = ϕ ( x j ) T ϕ ( x i ) = κ ( x j , x i ) \kappa(x_i,x_j)=\kappa(x_i,x_j)^T=\phi(x_j)^T\phi(x_i)=\kappa(x_j,x_i) κ(xi,xj)=κ(xi,xj)T=ϕ(xj)Tϕ(xi)=κ(xj,xi)即證明了核矩陣的對稱性,下證其半正定性:
將按照定義展開的 κ ( x i , x j ) \kappa(x_i,x_j) κ(xi,xj)帶入核矩陣中
( ϕ ( x 1 ) T ϕ ( x 1 ) ϕ ( x 1 ) T ϕ ( x 2 ) ϕ ( x 1 ) T ϕ ( x 3 ) ⋯ ϕ ( x 1 ) T ϕ ( x n ) ϕ ( x 2 ) T ϕ ( x 1 ) ϕ ( x 2 ) T ϕ ( x 2 ) ϕ ( x 3 ) T ϕ ( x 2 ) ⋯ ϕ ( x 2 ) T ϕ ( x n ) ⋮ ⋮ ⋮ ⋱ ⋮ ϕ ( x n ) T ϕ ( x 1 ) ϕ ( x n ) T ϕ ( x 2 ) ϕ ( x n ) T ϕ ( x 3 ) ⋯ ϕ ( x n ) T ϕ ( x n ) ) \begin{pmatrix} \phi(x_1)^T\phi(x_1) & \phi(x_1)^T\phi(x_2) & \phi(x_1)^T\phi(x_3) & \cdots & \phi(x_1)^T\phi(x_n) \\ \phi(x_2)^T\phi(x_1) & \phi(x_2)^T\phi(x_2) & \phi(x_3)^T\phi(x_2) & \cdots & \phi(x_2)^T\phi(x_n) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \phi(x_n)^T\phi(x_1) & \phi(x_n)^T\phi(x_2) & \phi(x_n)^T\phi(x_3) & \cdots & \phi(x_n)^T\phi(x_n) \\ \end{pmatrix} ϕ(x1)Tϕ(x1)ϕ(x2)Tϕ(x1)ϕ(xn)Tϕ(x1)ϕ(x1)Tϕ(x2)ϕ(x2)Tϕ(x2)ϕ(xn)Tϕ(x2)ϕ(x1)Tϕ(x3)ϕ(x3)Tϕ(x2)ϕ(xn)Tϕ(x3)ϕ(x1)Tϕ(xn)ϕ(x2)Tϕ(xn)ϕ(xn)Tϕ(xn)取其 k k k階順序主子式 ( 1 ≤ k ≤ n ) (1\leq k\leq n) (1kn)
∣ ϕ ( x 1 ) T ϕ ( x 1 ) ϕ ( x 1 ) T ϕ ( x 2 ) ϕ ( x 1 ) T ϕ ( x 3 ) ⋯ ϕ ( x 1 ) T ϕ ( x k ) ϕ ( x 2 ) T ϕ ( x 1 ) ϕ ( x 2 ) T ϕ ( x 2 ) ϕ ( x 3 ) T ϕ ( x 2 ) ⋯ ϕ ( x 2 ) T ϕ ( x k ) ⋮ ⋮ ⋮ ⋱ ⋮ ϕ ( x n ) T ϕ ( x 1 ) ϕ ( x n ) T ϕ ( x 2 ) ϕ ( x n ) T ϕ ( x 3 ) ⋯ ϕ ( x n ) T ϕ ( x n ) ∣ \begin{vmatrix} \phi(x_1)^T\phi(x_1) & \phi(x_1)^T\phi(x_2) & \phi(x_1)^T\phi(x_3) & \cdots & \phi(x_1)^T\phi(x_k) \\ \phi(x_2)^T\phi(x_1) & \phi(x_2)^T\phi(x_2) & \phi(x_3)^T\phi(x_2) & \cdots & \phi(x_2)^T\phi(x_k) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \phi(x_n)^T\phi(x_1) & \phi(x_n)^T\phi(x_2) & \phi(x_n)^T\phi(x_3) & \cdots & \phi(x_n)^T\phi(x_n) \\ \end{vmatrix} ϕ(x1)Tϕ(x1)ϕ(x2)Tϕ(x1)ϕ(xn)Tϕ(x1)ϕ(x1)Tϕ(x2)ϕ(x2)Tϕ(x2)ϕ(xn)Tϕ(x2)ϕ(x1)Tϕ(x3)ϕ(x3)Tϕ(x2)ϕ(xn)Tϕ(x3)ϕ(x1)Tϕ(xk)ϕ(x2)Tϕ(xk)ϕ(xn)Tϕ(xn) k = 1 k=1 k=1時,顯然此時的順序主子式就是一個無窮維向量的轉置與自身的內積,必然大於等於0。
k = 2 k=2 k=2時,順序主子式為
∣ ϕ ( x 1 ) T ϕ ( x 1 ) ϕ ( x 1 ) T ϕ ( x 2 ) ϕ ( x 2 ) T ϕ ( x 1 ) ϕ ( x 2 ) T ϕ ( x 2 ) ∣ = ( ϕ ( x 1 ) − ϕ ( x 2 ) ) T ( ϕ ( x 1 ) − ϕ ( x 2 ) ) \begin{vmatrix} \phi(x_1)^T\phi(x_1) & \phi(x_1)^T\phi(x_2) \\ \phi(x_2)^T\phi(x_1) & \phi(x_2)^T\phi(x_2) \end{vmatrix} =(\phi(x_1)-\phi(x_2))^T(\phi(x_1)-\phi(x_2)) ϕ(x1)Tϕ(x1)ϕ(x2)Tϕ(x1)ϕ(x1)Tϕ(x2)ϕ(x2)Tϕ(x2)=(ϕ(x1)ϕ(x2))T(ϕ(x1)ϕ(x2))也同樣可以看作是一個無窮維向量的轉置與自身的內積,是非負的。
k ≥ 2 k\geq2 k2時,我們將順序主子式中的第一列和第二列同時減去第三列得到
( ϕ ( x 1 ) − ϕ ( x 3 ) ) ( ϕ ( x 2 ) − ϕ ( x 3 ) ) ∣ ϕ ( x 1 ) T ϕ ( x 1 ) T ϕ ( x 1 ) T ϕ ( x 3 ) ⋯ ϕ ( x 1 ) T ϕ ( x k ) ϕ ( x 2 ) T ϕ ( x 2 ) T ϕ ( x 3 ) T ϕ ( x 2 ) ⋯ ϕ ( x 2 ) T ϕ ( x k ) ⋮ ⋮ ⋮ ⋱ ⋮ ϕ ( x n ) T ϕ ( x n ) T ϕ ( x n ) T ϕ ( x 3 ) ⋯ ϕ ( x n ) T ϕ ( x n ) ∣ (\phi(x_1)-\phi(x_3))(\phi(x_2)-\phi(x_3))\begin{vmatrix} \phi(x_1)^T & \phi(x_1)^T & \phi(x_1)^T\phi(x_3) & \cdots & \phi(x_1)^T\phi(x_k) \\ \phi(x_2)^T & \phi(x_2)^T & \phi(x_3)^T\phi(x_2) & \cdots & \phi(x_2)^T\phi(x_k) \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \phi(x_n)^T & \phi(x_n)^T & \phi(x_n)^T\phi(x_3) & \cdots & \phi(x_n)^T\phi(x_n) \\ \end{vmatrix} (ϕ(x1)ϕ(x3))(ϕ(x2)ϕ(x3))ϕ(x1)Tϕ(x2)Tϕ(xn)Tϕ(x1)Tϕ(x2)Tϕ(xn)Tϕ(x1)Tϕ(x3)ϕ(x3)Tϕ(x2)ϕ(xn)Tϕ(x3)ϕ(x1)Tϕ(xk)ϕ(x2)Tϕ(xk)ϕ(xn)Tϕ(xn)此時順序主子式的行列式有兩列線性相關,結果為0,所以根據定義,該矩陣為一半正定的矩陣,故得證。

相關文章