矩阵求导之 ∂ A ( x ) v ( x ) ∂ x \frac{\partial \bold{A}(\bold{x})\bold{v}(\bold{x})}{\partial\bold{x}} ∂x∂A(x)v(x)
数学-矩阵计算中指出,vector对vector的偏导为矩阵,且当 A \bold{A} A与 x \bold{x} x无关时, ∂ A v ( x ) ∂ x = A ∂ v ∂ x \frac{\partial \bold{A}\bold{v}(\bold{x})}{\partial\bold{x}}=A\frac{\partial \bold{v}}{\partial \bold{x}} ∂x∂Av(x)=A∂x∂v(分子布局下, [ ∂ v ∂ x ] p , q = ∂ v p ∂ x q [\frac{\partial \bold{v}}{\partial \bold{x}}]_{p,q}=\frac{\partial \bold{v}_p}{\partial \bold{x}_q} [∂x∂v]p,q=∂xq∂vp)。当 A = A ( x ) A=A(\bold{x}) A=A(x)时,上式不成立,在分子布局下,有如下结论。
1. 结论:
∂ A ( x ) v ( x ) ∂ x = ∂ A ∂ x v + A ∂ v ∂ x ⊤ \frac{\partial \bold{A}(\bold{x})\bold{v}(\bold{x})}{\partial\bold{x}} = \frac{\partial\bold{A}}{\partial\bold{x}}\bold{v}+\bold{A}\frac{\partial\bold{v}}{\partial{\bold{x}^\top}} ∂x∂A(x)v(x)=∂x∂Av+A∂x⊤∂v
其中 ∂ A ∂ x \frac{\partial\bold{A}}{\partial\bold{x}} ∂x∂A、 ∂ v ∂ x ⊤ \frac{\partial\bold{v}}{\partial{\bold{x}^\top}} ∂x⊤∂v均为分子布局, ∂ A ∂ x v \frac{\partial\bold{A}}{\partial\bold{x}}\bold{v} ∂x∂Av为分块矩阵相乘,后使用普通的矩阵乘法。
2. ∂ A ∂ x v \frac{\partial\bold{A}}{\partial\bold{x}}\bold{v} ∂x∂Av项
∂ A ∂ x v = [ ∂ A ∂ x 1 , ∂ A ∂ x 2 , … , ∂ A ∂ x n ] v = [ ∂ A ∂ x 1 v , ∂ A ∂ x 2 v , … , ∂ A ∂ x n v ] \frac{\partial\bold{A}}{\partial\bold{x}}\bold{v}=[\frac{\partial\bold{A}}{\partial{x}_1},\frac{\partial\bold{A}}{\partial{x}_2},\ldots,\frac{\partial\bold{A}}{\partial{x}_n}]\bold{v}=[\frac{\partial\bold{A}}{\partial{x}_1}\bold{v},\frac{\partial\bold{A}}{\partial{x}_2}\bold{v},\ldots,\frac{\partial\bold{A}}{\partial{x}_n}\bold{v}] ∂x∂Av=[∂x1∂A,∂x2∂A,…,∂xn∂A]v=[∂x1∂Av,∂x2∂Av,…,∂xn∂Av]
其中, ∂ A ∂ x i \frac{\partial\bold{A}}{\partial{x}_i} ∂xi∂A为 R n , n \mathbb{R}^{n,n} Rn,n矩阵, [ ∂ A ∂ x i ] p , q = ∂ a p , q ∂ x i [\frac{\partial\bold{A}}{\partial{x_i}}]_{p,q}=\frac{\partial a_{p,q}}{\partial x_{i}} [∂xi∂A]p,q=∂xi∂ap,q。 ∂ A ∂ x i v \frac{\partial\bold{A}}{\partial{x}_i}\bold{v} ∂xi∂Av为 R n \mathbb{R}^{n} Rn列向量。