1. pre
ref: https://zhuanlan.zhihu.com/p/263777564
起初是课上看到 \(\frac{\partial w^TRw}{\partial w}\)，但没认真听，后面不知道怎么求导...
2. 简单情况
首先，针对函数\(f\)：
\[f(x_1,x_2,x_3)=x_1^2+x_1x_2+x_2x_3 \tag{e.g.10} \]可以分别对 \(x_1,x_2,x_3\) 求偏导
\[\left\{ \begin{align*} \frac{\partial f}{\partial x_1} & = 2x_1+x_2 \\\\ \frac{\partial f}{\partial x_2} & = x_1+x_3 \\\\ \frac{\partial f}{\partial x_3} & = x_2 \end{align*} \right. \]可以把得到的结果写成列向量形式：
\[{\frac{\partial f(x)}{\partial x_{3\times1}}}= {\left[\begin{matrix}{{\frac{\partial f}{\partial x_{1}}}}\\ {{\frac{\partial f}{\partial x_{2}}}}\\ {{\frac{\partial f}{\partial x_{3}}}}\end{matrix}\right]}= {\left[\begin{matrix}{2x_{1}+x_{2}}\\ {x_{1}+x_{3}}\\ {x_{2}}\end{matrix}\right]} \tag{1} \]或者也可以以行向量形式展开：
\[\frac{\partial f(x)}{\partial x_{3\times1}^T}= \left[ \frac{\partial f}{\partial x_1}, \frac{\partial f}{\partial x_2}, \frac{\partial f}{\partial x_3} \right] = \left[ 2x_1+x_2, x_1+x_3, x_2 \right] \\\\ \tag{2} \]3. 两种布局
至此函数\(f\)都是类似标量那样只有一个，实际上函数也可以多个组合成向量或者矩阵，引申出分子布局跟分母布局
分子布局：分子是列向量形式，分母是行向量形式，如式2
分母布局：分母是列向量形式，分子是行向量形式，如式1
\[\frac{\partial f^T_{2\times1}(x)}{\partial x_{3\times1}}= \left[ \begin{matrix} {{\frac{\partial f_{1}}{\partial x_{1}}}}& {{\frac{\partial f_{2}}{\partial x_{1}}}}\\ {{\frac{\partial f_{1}}{\partial x_{2}}}}& {{\frac{\partial f_{2}}{\partial x_{2}}}}\\ {{\frac{\partial f_{1}}{\partial x_{3}}}}& {{\frac{\partial f_{2}}{\partial x_{3}}}}\\ \end{matrix}\right]_{3\times 2} \\\\ \tag{4} \]4. 实值标量函数-向量变元
实值标量函数\(f(\pmb{x})\) 对 向量变元\(\pmb{x}=[x_1,x_2,\cdots,x_n]^T\) 求导，有两种形式：
4.1. 行向量偏导形式（又称行偏导向量形式）
\[\text{D}_{x}f(x)= \frac{\partial f(x)}{\partial x^T}= \left[ \frac{\partial f}{\partial x_1}, \frac{\partial f}{\partial x_2}, \cdots, \frac{\partial f}{\partial x_n} \right] \\\\ \tag{5} \]4.2. 梯度向量形式（又称列向量偏导形式、列偏导向量形式）
\[\nabla_{x}f(x)= \frac{\partial f(x)}{\partial x}= \left[ \frac{\partial f}{\partial x_1}, \frac{\partial f}{\partial x_2}, \cdots, \frac{\partial f}{\partial x_n} \right]^T \\\\ \tag{6} \]这俩互为转置，也就是说对行向量求偏导，相当于对相应行向量求偏导后，结果再转置
5. 实值标量函数-矩阵变元
实值标量函数 \(f(X)\) 对 矩阵变元 \(X_{m\times n}=(x_{ij})_{i=1,j=1}^{m,n}\)
首先，有个符号 \(\text{vec}(X)\)，将矩阵X按列堆栈来向量化，就是将矩阵X的第 1 列，第 2 列，直到第 n 列取出来，然后按顺序组成一个列向量，即：
\[\text{vec}({X)}= \left[ x_{11},x_{21},\cdots,x_{m1},x_{12},x_{22},\cdots,x_{m2},\cdots,x_{1n},x_{2n},\cdots,x_{mn} \right]^T \\\\ \tag{7} \]5.1. 行向量偏导形式（又称行偏导向量形式）
即先把矩阵变元 X 按 vec 向量化，转换成向量变元，再对该向量变元使用式5：
\[\begin{align} \text{D}_{\text{vec}X}f(X)&= \frac{\partial f(X)}{\partial \text{vec}^T(X)} \\\\ &= \left[ \frac{\partial f}{\partial x_{11}},\frac{\partial f}{\partial x_{21}},\cdots,\frac{\partial f}{\partial x_{m1}},\frac{\partial f}{\partial x_{12}},\frac{\partial f}{\partial x_{22}},\cdots,\frac{\partial f}{\partial x_{m2}},\cdots,\frac{\partial f} {\partial x_{1n}},\frac{\partial f}{\partial x_{2n}},\cdots,\frac{\partial f}{\partial x_{mn}} \right] \end{align} \\\\ \tag{8} \]5.2. Jacobian矩阵形式
即先把矩阵变元 X 进行转置，再对转置后的每个位置的元素逐个求偏导，结果布局和转置布局一样。
\[\begin{align} \text{D}_{X}f(X)&= \frac{\partial f(X)}{\partial X^T_{m\times n}} \\\\ &= \left[ \begin{array}{cccc} \frac{\partial f}{\partial x_{11}} & \frac{\partial f}{\partial x_{21}} & \cdots & \frac{\partial f}{\partial x_{m 1}} \\ \frac{\partial f}{\partial x_{12}} & \frac{\partial f}{\partial x_{22}} & \cdots & \frac{\partial f}{\partial x_{m 2}} \\ \vdots & \vdots & \vdots & \vdots \\ \frac{\partial f}{\partial x_{1 n}} & \frac{\partial f}{\partial x_{2 n}} & \cdots & \frac{\partial f}{\partial x_{m n}} \end{array} \right]_{n\times m} \end{align} \\\\ \tag{9} \]5.3. 梯度向量形式（又称列向量偏导形式、列偏导向量形式）
即先把矩阵变元 X 按 vec 向量化，转换成向量变元，再对该变元使用式6：
\[\begin{align} \nabla_{vec\pmb{X}}f(\pmb{X})&= \frac{\partial f(\pmb{X})}{\partial vec\pmb{X}} \\\\ &= \left[ \frac{\partial f}{\partial x_{11}},\frac{\partial f}{\partial x_{21}},\cdots,\frac{\partial f}{\partial x_{m1}},\frac{\partial f}{\partial x_{12}},\frac{\partial f}{\partial x_{22}},\cdots,\frac{\partial f}{\partial x_{m2}},\cdots,\frac{\partial f} {\partial x_{1n}},\frac{\partial f}{\partial x_{2n}},\cdots,\frac{\partial f}{\partial x_{mn}} \right]^T \end{align} \\\\ \tag{10} \]5.4. 梯度矩阵形式
直接对原矩阵变元 X 的每个位置的元素逐个求偏导，结果布局和原矩阵布局一样。
\[\begin{align} \nabla_{\pmb{X}}f(\pmb{X})&= \frac{\partial f(\pmb{X})}{\partial \pmb{X}_{m\times n}} \\\\ &= \left[ \begin{array}{cccc} \frac{\partial f}{\partial x_{11}} & \frac{\partial f}{\partial x_{12}} & \cdots & \frac{\partial f}{\partial x_{1n}} \\ \frac{\partial f}{\partial x_{21}} & \frac{\partial f}{\partial x_{22}} & \cdots & \frac{\partial f}{\partial x_{2n}} \\ \vdots & \vdots & \vdots & \vdots \\ \frac{\partial f}{\partial x_{m1}} & \frac{\partial f}{\partial x_{m2}} & \cdots & \frac{\partial f}{\partial x_{mn}} \end{array} \right]_{m\times n} \end{align} \\\\ \tag{11} \]6. 小总结
- 式8与式10互为转置；式9与式11互为转置
- 函数为标量形式时，按照x布局对其中每个元素求偏导
- 若函数也为向量，则偏导的x应该是形式不同的向量，如行-列，列-行。实际上若都是同一种可以用公式6进行转化
7. post
事实上，当A代表nxn矩阵，X为nx1向量，有下列便捷的公式可用：
\[(AX)' = A^T \\ (XA)' = A^T \\ (X^TA)' = A \\ (AX^T)' = A \\ \]回到开头的问题，依公式可得：\(\frac{\partial w^TRw}{\partial w} = Rw + R^Tw = (R+R^T)w\)
实际上这公式也很好证明，比如第一个公式，可设
\[A = \begin{bmatrix} a_{11}& a_{12} \\ a_{21}& a_{22} \\ \end{bmatrix} \\ \\ X = \begin{bmatrix} x_{1} \\ x_{2} \\ \end{bmatrix} \]按照上面的公式3 (3. 两种布局)，可得
\[\frac{\partial AX }{\partial X^T} = \begin{bmatrix} \frac{\partial a_{11}x_1 + a_{22}x_2 }{\partial x_1}& \ldots \\ \ldots& \ldots \\ \end{bmatrix} = \begin{bmatrix} a_{11}& a_{12} \\ a_{21}& a_{22} \\ \end{bmatrix} = A \]其他同理可得
ps. 不知道多行公式为啥会有多个tag，知道的同学麻烦教一教QvQ