1 矩阵\(Y=f(x)\)对标量x求导

     矩阵Y是一个\(m\times n\)的矩阵，对标量x求导，相当于矩阵中每个元素对x求导

\[\frac{dY}{dx}=\begin{bmatrix}\dfrac{df_{11}(x)}{dx} & \ldots & \dfrac{df_{1n}(x)}{dx} \\ \vdots & \ddots &\vdots \\ \dfrac{df_{m1}(x)}{dx} & \ldots & \dfrac{df_{mn}(x)}{dx} \end{bmatrix}\]

2 标量y=f(x)对矩阵X求导

     注意与上面不同，这次括号内是求偏导，\(X\)是是一个\(m\times n\)的矩阵，函数\(y=f(x)\)对矩阵\(X\)中的每个元素求偏导，对\(m\times n\)矩阵求导后还是\(m\times n\)矩阵

\[\frac{dy}{dX} = \begin{bmatrix}\dfrac{\partial f}{\partial x_{11}} & \ldots & \dfrac{\partial f}{\partial x_{1n}}\\ \vdots & \ddots & \vdots \\\dfrac{\partial f}{\partial x_{m1}} & \ldots & \dfrac{\partial f}{\partial x_{mn}}\end{bmatrix}\]

3 函数矩阵Y对矩阵X求导

矩阵\(Y=F(x)\)对每一个\(X\)的元素求导，构成一个超级矩阵

\[F(x)=\begin{bmatrix}f_{11}(x) & \ldots &  f_{1n}(x)\\ \vdots & \ddots &\vdots \\ f_{m1}(x) & \ldots & f_{mn}(x) \end{bmatrix}\]

\[X=\begin{bmatrix}x_{11} & \ldots &  x_{1s}\\ \vdots & \ddots &\vdots \\ x_{r1} & \ldots & x_{rs}\end{bmatrix}\]

\[\frac{dF}{dX} = \begin{bmatrix}\dfrac{\partial F}{\partial x_{11}} & \ldots & \dfrac{\partial F}{\partial x_{1s}}\\ \vdots & \ddots & \vdots \\\dfrac{\partial F}{\partial x_{r1}} & \ldots & \dfrac{\partial F}{\partial x_{rs}}\end{bmatrix}\]

其中

\[\frac{\partial F}{\partial x_{ij}} = \begin{bmatrix}\dfrac{\partial f_{11}}{\partial x_{ij}} & \ldots & \dfrac{\partial f_{1n}}{\partial x_{ij}}\\ \vdots & \ddots & \vdots \\\dfrac{\partial f_{m1}}{\partial x_{ij}} & \ldots & \dfrac{\partial f_{mn}}{\partial x_{ij}}\end{bmatrix}\]

4 向量导数

若\(m\times 1\)向量函数\(y=[y_1,y_2,…,y_m]^T\),其中，\(y_1,y_2,…,y_m\)是向量的标量函数。\(x\)是\(n\times 1\)向量。则有

\[\frac{\partial Y}{\partial X^T} = \begin{bmatrix}\dfrac{\partial y_1}{\partial x_1} & \ldots & \dfrac{\partial y_1}{\partial x_n}\\ \vdots & \ddots & \vdots \\\dfrac{\partial y_m}{\partial x_1} & \ldots & \dfrac{\partial y_m}{\partial x_n}\end{bmatrix}\]

这是一个\(m\times n\)矩阵，称作向量函数\(y\)的Jacobi矩阵。

若\(y=[x_1,x_2,…,x_n]\),则有

\[\frac{\partial{x^T}}{\partial{x}}=I \tag{$1$}\]

其中，\(I\)是单位矩阵。

若\(A\)和\(y\)均与向量\(x\)无关，则

\[\frac{\partial{x^TAy}}{\partial{x}}=\frac{\partial{x^T}}{\partial{x}}Ay=Ay \tag{$1$}\]

注意到：\(y^TAx=<A^Ty,x>=<x,A^Ty>=x^TA^Ty\), 向量内积的公式，故

\[\frac{\partial{y^TAx}}{\partial{x}}=\frac{\partial{x^TA^Ty}}{\partial{x}}=A^Ty \tag{$2$}\]

由于\(x^TAx=\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}A_{ij}x_ix_j\)

可求出梯度\(\frac{\partial{x^TAx}}{\partial{x}}\)的第k个分量为

\[\bigg[\frac{\partial{x^TAx}}{\partial{x}}\bigg]_k=\frac{\partial}{\partial{x_k}}\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}A_{ij}x_ix_j=\sum\limits_{i=1}^nA_{ik}x_i+\sum\limits_{j=1}^{n}A_{kj}x_j\]

即有公式

\[\frac{\partial{x^TAx}}{\partial{x}}=Ax+A^Tx \tag{$3$}\]

特别地，若\(A\)为对称矩阵，则有\(\frac{\partial{x^TAx}}{\partial{x}}=2Ax\)

用上面三个公式，我们能够得到更多的实值函数\(f(x)\)相对于列向量\(x\)的几个常用梯度公式：

若\(f(x)=c\)为常数，则有梯度 \(\frac{\partial{c}}{\partial{x}}=0\)

线性法则：若\(f(x)\)和\(g(x)\)分别是向量\(x\)的实值函数，\(c_1\)和\(c_2\)为实常数，则有

\[\frac{\partial{[c_1f(x)+c_2g(x)]}}{\partial{x}}=c_1\frac{\partial{f(x)}}{\partial{x}}+c_2\frac{\partial{g(x)}}{\partial{x}}\]

乘积法则：若\(f(x)\)和\(g(x)\)都是向量\(x\)的实值函数，则

\[\frac{\partial{f(x)g(x)}}{\partial{x}}=g(x)\frac{\partial{f(x)}}{\partial{x}}+f(x)\frac{\partial{g(x)}}{\partial{x}}\]

若\(f(x)\),\(g(x)\)和\(h(x)\)都是向量\(x\)的实值函数，则

\[\frac{\partial{f(x)g(x)h(x)}}{\partial{x}}=g(x)h(x)\frac{\partial{f(x)}}{\partial{x}}+f(x)h(x)\frac{\partial{g(x)}}{\partial{x}}+f(x)g(x)\frac{\partial{h(x)}}{\partial{x}}\]

商法则：若\(g(x)\neq0\),则

\[\frac{\partial{f(x)/g(x)}}{\partial{x}}=\frac{1}{g^2(x)}\big[g(x)\frac{\partial{f(x)}}{\partial{x}}-f(x)\frac{\partial{g(x)}}{\partial{x}}\big]\]

链式法则：若\(y(x)\)是\(x\)的向量值函数，则

\[\frac{\partial{f(y(x))}}{\partial{x}}=\frac{\partial{y^T(x)}}{\partial{x}}\frac{\partial{f(y)}}{\partial{y}}\]

其中，\(\frac{\partial{y^T(x)}}{\partial{x}}\)为\(n\times n\)矩阵。

若\(n\times 1\)向量\(\alpha\) 与\(x\)是无关的常数向量，则

\[\frac{\partial{\alpha^Ty(x)}}{\partial{x}}=\frac{\partial{y^T(x)}}{\partial{x}}\alpha\]

\[\frac{\partial{y^T(x)\alpha}}{\partial{x}}=\frac{\partial{y^T(x)}}{\partial{x}}\alpha\]

令\(x\)为\(n\times 1\)向量，\(\alpha\)为\(m\times 1\)常数向量，\(A\)和\(B\)分别为\(m\times n\)和\(m\times m\)常数矩阵，且\(B])为对称矩阵，则

\[\frac{\partial{(\alpha-Ax)^TB(\alpha-Ax)}}{\partial{x}}=-2A^TB(\alpha-Ax)\]

5 迹函数的梯度矩阵

二次项目标函数可以利用矩阵的迹重写，因为一标量可以视为\(1\times 1\)矩阵。所以二次项目标函数的迹直接等于函数本身，即

\[f(x)=x^TAx=tr(x^TAx)=tr(Axx^T)\]

\[\frac{\partial{tr(A)}}{\partial{A}}=I \tag{$1$}\]

\[\frac{\partial{tr(AB)}}{\partial{A}}=B^T \tag{$2$}\]

由于\(tr(xy^T)=tr(yx^T)=X^Ty\)，所以

\[\frac{\partial{tr(xy^T)}}{\partial{x}}=\frac{\partial{tr(yx^T)}}{\partial{x}}=y \tag{$3$} \]

\(m\times m\)矩阵\(W\)可逆时，有

\[\frac{\partial{tr(W^{-1})}}{\partial{W}}=-(W^{-1})^T \tag{$4$}\]

另外几个公式：

\[\frac{\partial{f(A)}}{\partial{A^T}}=(\frac{\partial{f(A)}}{\partial{A}})^T\]

\[\frac{\partial{tr(ABA^TC)}}{\partial{A}}= CAB + C^TAB^T \]

\[\frac{\partial{|A|}}{\partial{A}}=|A|(A^{-1})^T\]