标量对向量求导

最新推荐文章于 2025-05-03 04:04:36 发布

原创最新推荐文章于 2025-05-03 04:04:36 发布 · 481 阅读

1 ·

CC 4.0 BY-SA版权

文章标签：

#机器学习 #算法 #人工智能

1.分子布局和分母布局

标量函数对列向量求导，如果导数的行数和分子的行数相同，称为分子布局，导数的行数和分子相同，则称为分子布局。

$\overrightarrow{y}=[y1,y2,y3,...,yn]^{T},f=f(\overrightarrow{y}), \frac{\partial f}{\partial \overrightarrow{y}}=[\frac{\partial f}{\partial y1},...,\frac{\partial f}{\partial yn}],\frac{\partial f}{\partial \overrightarrow{y}}=[\frac{\partial f}{\partial y1},...,\frac{\partial f}{\partial yn}]^{T},$

2.常用计算公式

2.1 f=Ay

$f=Ay,A=\begin{bmatrix} a11 & a12 \\ a21 & a22 \end{bmatrix},y=\begin{bmatrix} y1\\y2 \end{bmatrix}, f=\begin{bmatrix} f1\\ f2 \end{bmatrix}, f1=a11*y1+a12*y2,f2=a21*y1+a22*y2,$

$\frac{\partial f}{\partial y}=\begin{bmatrix} \frac{\partial f1}{\partial y1} & \frac{\partial f2}{\partial y1}\\ \frac{\partial f1}{\partial y2} & \frac{\partial f2}{\partial y2} \end{bmatrix}=\begin{bmatrix} a11 & a21\\ a12 & a22 \end{bmatrix}=A^{T}$

2.2 f=yTAy

$f=y^{^{T}}Ay,\frac{\partial f}{\partial y}=Ay+A^{T}y$

3.对向量求导的链式法则

$\overrightarrow{u}=\begin{bmatrix} u1\\ u2 \end{bmatrix}, \overrightarrow{y}=\begin{bmatrix} y1\\ y2\\ y3 \end{bmatrix}, f=f(u1(y1,y2,y3),u2(y1,y2,y3)),\frac{\partial f}{\partial y1}=\frac{\partial f}{\partial u1}*\frac{\partial u1}{\partial y1}+\frac{\partial f}{\partial u2}*\frac{\partial u2}{\partial y1}, \frac{\partial f}{\partial y2}=\frac{\partial f}{\partial u1}*\frac{\partial u1}{\partial y2}+\frac{\partial f}{\partial u2}*\frac{\partial u2}{\partial y2}, \frac{\partial f}{\partial y3}=\frac{\partial f}{\partial u1}*\frac{\partial u1}{\partial y3}+\frac{\partial f}{\partial u2}*\frac{\partial u2}{\partial y3},$

$\begin{bmatrix} \frac{\partial f}{\partial y1}\\ \frac{\partial f}{\partial y2}\\ \frac{\partial f}{\partial y3} \end{bmatrix}=\begin{bmatrix} \frac{\partial u1}{\partial y1} & \frac{\partial u2}{\partial y1}\\ \frac{\partial u1}{\partial y2} & \frac{\partial u2}{\partial y2} \\ \frac{\partial u1}{\partial y3} & \frac{\partial u2}{\partial y3} \end{bmatrix}\begin{bmatrix} \frac{\partial f}{\partial u1}\\ \frac{\partial f}{\partial u2}\\ \end{bmatrix}=\frac{\partial \overrightarrow{u}}{\partial \overrightarrow{y}}\frac{\partial \overrightarrow{f}}{\partial \overrightarrow{u}}$