—

0:00
0:00

NO.002 · 技术笔记 · Math

关于矩阵求导(其一)

DATE
READ
25 MIN
VIEWS
–

概念准备

在进行矩阵求导之前,我们首先需要做一些关于概念上的铺垫。主要分为以下两个部分:

  • 函数与变元的种类
  • 分子布局与分母布局

函数与变元的种类

这里的种类指的就是标量、向量和矩阵这三种。在本文后续内容中对于这三种形式的符号表达都遵循如下定义:

  • 标量,用正文字体小写字母表示,例如:
x,yx, y
  • 向量,用粗体小写字母表示,例如:
xn×1=(x1,x2,⋯ ,xn)T\mathbf{x}_{n \times 1} = \begin{pmatrix} x_1, x_2, \cdots, x_n \end{pmatrix}^T
  • 矩阵,用粗体大写字母表示,例如:
Xm×n=(x11x12⋯x1nx21x22⋯x2n⋮⋮⋱⋮xm1xm2⋯xmn)\mathbf{X}_{m \times n} = \begin{pmatrix} x_{11} & x_{12} & \cdots & x_{1n} \\ x_{21} & x_{22} & \cdots & x_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ x_{m1} & x_{m2} & \cdots & x_{mn} \end{pmatrix}

接下来是具体分类:

  1. 函数为标量,称之为实值标量函数,根据变元种类分为以下三种:

    • 变元为标量,例:
    f(x)=x2f(x) = x^2
    • 变元为向量, 例:
    x3×1=(x1,x2,x3)Tf(x)=x12+x22+x32\begin{aligned} & \mathbf{x}_{3 \times 1} = \begin{pmatrix} x_1, x_2, x_3 \end{pmatrix}^T \\[5mm] & f(\mathbf{x}) = x_1^2 + x_2^2 + x_3^2 \end{aligned}
    • 变元为矩阵, 例:
    X2×3=(x11x12x13x21x22x23)f(X)=x112+x12+x13+x21+x22+x23\begin{aligned} \mathbf{X}_{2 \times 3} &= \begin{pmatrix} x_{11} & x_{12} & x_{13} \\ x_{21} & x_{22} & x_{23} \end{pmatrix} \\[5mm] f(\mathbf{X}) &= x_{11}^2 + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} \end{aligned}
  2. 函数为向量,称之为实向量函数,根据变元种类分为以下三种:

    • 变元为标量, 例:
    f3×1(x)=(f1(x)f2(x)f3(x))=(xx2x3)\mathbf{f}_{3 \times 1}(x) = \begin{pmatrix} f_1(x) \\ f_2(x) \\ f_3(x) \end{pmatrix} = \begin{pmatrix} x \\ x^2 \\ x^3 \end{pmatrix}
    • 变元为向量, 例:
    x3×1=(x1x2x3)Tf3×1(x)=(f1(x)f2(x)f3(x))=(x1+x2x2+x3x3+x1)\begin{aligned} \mathbf{x}_{3 \times 1} &= \begin{pmatrix} x_1 & x_2 & x_3 \end{pmatrix}^T \\[5mm] \mathbf{f}_{3 \times 1}(\mathbf{x}) &= \begin{pmatrix} f_1(\mathbf{x}) \\ f_2(\mathbf{x}) \\ f_3(\mathbf{x}) \end{pmatrix} = \begin{pmatrix} x_1 + x_2 \\ x_2 + x_3 \\ x_3 + x_1 \end{pmatrix} \end{aligned}
    • 变元为矩阵, 例:
    X2×3=(x11x12x13x21x22x23)f3×1(X)=(f1(X)f2(X)f3(X))=(x11+x12+x212x13+x21+x222x22+x23+x232)\begin{aligned} \mathbf{X}_{2 \times 3} &= \begin{pmatrix} x_{11} & x_{12} & x_{13} \\ x_{21} & x_{22} & x_{23} \end{pmatrix} \\[7mm] \mathbf{f}_{3 \times 1}(\mathbf{X}) &= \begin{pmatrix} f_1(\mathbf{\mathbf{X}}) \\ f_2(\mathbf{\mathbf{X}}) \\ f_3(\mathbf{\mathbf{X}}) \end{pmatrix} \\[7mm] &= \begin{pmatrix} x_{11} + x_{12} + x_{21}^2 \\ x_{13} + x_{21} + x_{22}^2 \\ x_{22} + x_{23} + x_{23}^2 \end{pmatrix} \end{aligned}
  3. 函数为矩阵, 称之为实矩阵函数,根据变元种类分为以下三种:

    • 变元为标量, 例:
    F2×3(x)=(f11(x)f12(x)f13(x)f21(x)f22(x)f23(x))=(x2x3xx22x23x3)\begin{aligned} \mathbf{F}_{2 \times 3}(x) &= \begin{pmatrix} f_{11}(x) & f_{12}(x) & f_{13}(x) \\ f_{21}(x) & f_{22}(x) & f_{23}(x) \end{pmatrix} \\[5mm] &= \begin{pmatrix} x & 2x & 3x \\ x^2 & 2x^2 & 3x^3 \end{pmatrix} \end{aligned}
    • 变元为向量, 例:
    x3×1=(x11x12x13)TF2×3(x)=(f11(x)f12(x)f13(x)f21(x)f22(x)f23(x))=(x11+x12+x132x11+x12+x133x11+x12+x134x11+x12+x135x11+x12+x136x11+x12+x13)\begin{aligned} \mathbf{x}_{3 \times 1} &= \begin{pmatrix} x_{11} & x_{12} & x_{13} \end{pmatrix}^T \\[5mm] \mathbf{F}_{2 \times 3}(\mathbf{x}) &= \begin{pmatrix} f_{11}(\mathbf{x}) & f_{12}(\mathbf{x}) & f_{13}(\mathbf{x}) \\ f_{21}(\mathbf{x}) & f_{22}(\mathbf{x}) & f_{23}(\mathbf{x}) \end{pmatrix} \\[5mm] &= \begin{pmatrix} x_{11} + x_{12} + x_{13} & 2x_{11} + x_{12} + x_{13} & 3x_{11} + x_{12} + x_{13} \\ 4x_{11} + x_{12} + x_{13} & 5x_{11} + x_{12} + x_{13} & 6x_{11} + x_{12} + x_{13} \end{pmatrix} \end{aligned}
    • 变元为矩阵, 例:
    X2×3=(x11x12x13x21x22x23)F3×2(X)=(f11(X)f12(X)f21(X)f22(X)f31(X)f32(X))=(x11+x12+x13+x21+x22+x232x11+x12+x13+x21+x22+x233x11+x12+x13+x21+x22+x234x11+x12+x13+x21+x22+x235x11+x12+x13+x21+x22+x236x11+x12+x13+x21+x22+x23)\begin{aligned} \mathbf{X}_{2 \times 3} &= \begin{pmatrix} x_{11} & x_{12} & x_{13} \\ x_{21} & x_{22} & x_{23} \end{pmatrix} \\[7mm] \mathbf{F}_{3 \times 2}(\mathbf{X}) &= \begin{pmatrix} f_{11}(\mathbf{X}) & f_{12}(\mathbf{X}) \\ f_{21}(\mathbf{X}) & f_{22}(\mathbf{X}) \\ f_{31}(\mathbf{X}) & f_{32}(\mathbf{X}) \end{pmatrix} \\[7mm] &= \begin{pmatrix} x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} & 2x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} \\ 3x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} & 4x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} \\ 5x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} & 6x_{11} + x_{12} + x_{13} + x_{21} + x_{22} + x_{23} \end{pmatrix} \end{aligned}

最后关于函数和变元的种类,可以总结如下:

函数 \ 变元标量变元向量变元矩阵变元
实值标量函数f(x)f(x)f(x)f(\mathbf{x})f(X)f\mathbf{(X)}
实向量函数f(x)\mathbf{f}(x)f(x)\mathbf{f}(\mathbf{x})f(X)\mathbf{f}(\mathbf{X})
实矩阵函数F(x)\mathbf{F}(x)F(x)\mathbf{F}(\mathbf{x}) F(X)\mathbf{F}(\mathbf{X})

分子布局与分母布局

进行矩阵求导的时候实际上是将函数中的每一个分量依次对变元中的每一个分量求偏导,并将所有求偏导的结果排列起来,根据不同的排列规则分为分子布局 (Numerator Layout) 和 分母布局 (Denominator Layout)。举一个例子:

x=(x1x2⋮xn)n×1,∂f∂x=(∂f∂x1∂f∂x2⋮∂f∂xn)n×1\mathbf{x} = \begin{pmatrix} x_1 \\ x_2 \\ \vdots \\ x_n \end{pmatrix}_{n \times 1} , \qquad \frac{\partial{f}}{\partial{\mathbf{x}}} = \begin{pmatrix} \frac{\partial{f}}{\partial{x_1}} \\ \frac{\partial{f}}{\partial{x_2}} \\ \vdots \\ \frac{\partial{f}}{\partial{x_n}} \end{pmatrix}_{n \times 1}

这里的 f(x)f(\mathbf{x}) 是一个实值标量函数,而变元 x\mathbf{x} 是一个 n×1n \times 1 的列向量,求完导后的结果也是一个 n×1n \times 1 的列向量。如果把这里求导的式子 ∂f∂x\frac{\partial{f}}{\partial{\mathbf{x}}} 视作分式的话可以发现,求导结果的维度与分母 x\mathbf{x} 一致,所以这里采用的是分母布局。再举一个例子:

x=(x1x2⋮xn)n×1,f=(f1f2⋮fm)m×1,∂f∂x=(∂f1∂x1∂f2∂x1⋯∂fm∂x1∂f1∂x2∂f2∂x2⋯∂fm∂x2⋮⋮⋱⋮∂f1∂xn∂f2∂xn⋯∂fm∂xn)n×m\mathbf{x} = \begin{pmatrix} x_1 \\ x_2 \\ \vdots \\ x_n \end{pmatrix}_{n \times 1} , \qquad \mathbf{f} = \begin{pmatrix} f_1 \\ f_2 \\ \vdots \\ f_m \end{pmatrix}_{m \times 1} , \qquad \frac{\partial{\mathbf{f}}}{\partial{\mathbf{x}}} = \large \begin{pmatrix} \frac{\partial{\color{royalblue}f_1}}{\partial{\color{orange}x_1}} & \frac{\partial{\color{royalblue}{f_2}}}{\partial{x_1}} & \color{royalblue}{\cdots} & \frac{\partial{\color{royalblue}{f_m}}}{\partial{x_1}} \\ \frac{\partial{f_1}}{\partial{\color{orange}{x_2}}} & \frac{\partial{f_2}}{\partial{x_2}} & \cdots & \frac{\partial{f_m}}{\partial{x_2}} \\ \color{orange}{\vdots} & \vdots & \ddots & \vdots \\ \frac{\partial{f_1}}{\partial{\color{orange}{x_n}}} & \frac{\partial{f_2}}{\partial{x_n}} & \cdots & \frac{\partial{f_m}}{\partial{x_n}} \end{pmatrix}_{\color{orange}{n} \times \color{royalblue}{m}}

这里的 f(x)\mathbf{f}(\mathbf{x}) 是一个 m×1m \times 1 的实向量函数,而变元 x\mathbf{x} 是一个 n×1n \times 1 的列向量,求导的结果是一个 n×m\color{orange}{n} \times \color{royalblue}{m} 的矩阵。我们可以看到这个结果符合变元 x\mathbf{x} 的排列方式(橙色部分),而函数 f\mathbf{f} 则转变为行向量的排列方式(蓝色部分),也就是转置了。这里采用的还是分母布局。如果采用分子布局的话:

∂f∂x=(∂f1∂x1∂f1∂x2⋯∂f1∂xn∂f2∂x1∂f2∂x2⋯∂f2∂xn⋮⋮⋱⋮∂fm∂x1∂fm∂x2⋯∂fm∂xn)m×n\frac{\partial{\mathbf{f}}}{\partial{\mathbf{x}}} = \large \begin{pmatrix} \frac{\partial{\color{royalblue}{f_1}}}{\partial{\color{orange}{x_1}}} & \frac{\partial{f_1}}{\partial{\color{orange}{x_2}}} & \color{orange}{\cdots} & \frac{\partial{f_1}}{\partial{\color{orange}{x_n}}} \\ \frac{\partial{\color{royalblue}{f_2}}}{\partial{x_1}} & \frac{\partial{f_2}}{\partial{x_2}} & \cdots & \frac{\partial{f_2}}{\partial{x_n}} \\ \color{royalblue}{\vdots} & \vdots & \ddots & \vdots \\ \frac{\partial{\color{royalblue}{f_m}}}{\partial{x_1}} & \frac{\partial{f_m}}{\partial{x_2}} & \cdots & \frac{\partial{f_m}}{\partial{x_n}} \end{pmatrix}_{\color{royalblue}{m} \times \color{orange}{n}}

结果就是一个 m×n\color{royalblue}{m} \times \color{orange}{n} 的矩阵,变元 x\mathbf{x} 发生了转置。

函数 \ 变元标量变元向量变元矩阵变元
实值标量函数f(x)f(x)f(x)\color{red}{f(\mathbf{x})}f(X)\color{red}{f\mathbf{(X)}}
实向量函数f(x)\color{red}{\mathbf{f}(x)}f(x)\color{red}{\mathbf{f}(\mathbf{x})}f(X)\mathbf{f}(\mathbf{X})
实矩阵函数F(x)\color{red}{\mathbf{F}(x)}F(x)\mathbf{F}(\mathbf{x}) F(X)\mathbf{F}(\mathbf{X})

本文后续关于求导的部分只讨论上表中标注为红色的情况,那么关于求导的布局可以通俗地总结为:矩阵求导结果是什么布局由分子、分母中没有发生转置的那一个决定。


矩阵求导公式

以下表中所有的向量初始状态都默认为列向量。例如:

x=(x1⋮xn)n×1\mathbf{x} = \begin{pmatrix} x_1 \\ \vdots \\ x_n \end{pmatrix}_{n \times 1}

向量对向量 (Vector-by-vector)

CoditionExpressionNumerator layoutDenominator layoutProof
a\large \mathbf{a} is not a function of x\large \mathbf{x}∂a∂x\Large \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}}0\large \mathbf{0}0\large \mathbf{0}Proof 1-1
∂x∂x\Large \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}}I\large \mathbf{I}I\large \mathbf{I}
A\large \mathbf{A} is not a funtion of x\large \mathbf{x}∂Ax∂x\Large \frac{\partial{\mathbf{Ax}}}{\partial{\mathbf{x}}}A\large \mathbf{A}AT\large \mathbf{A}^TProof 1-2
A\large \mathbf{A} is not a funtion of x\large \mathbf{x}∂xTA∂x\Large \frac{\partial{\mathbf{x}^T\mathbf{A}}}{\partial{\mathbf{x}}}AT\large \mathbf{A}^TA\large \mathbf{A}Proof 1-3
a\large a is not a function of x\large \mathbf{x},
u=u(x)\mathbf{u} = \mathbf{u}(\mathbf{x})
∂au∂x\Large \frac{\partial{a\mathbf{u}}}{\partial{\mathbf{x}}}aa ∂u∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}aa ∂u∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}Proof 1-4
v=v(x),  u=u(x)v = v(\mathbf{x}), \; \mathbf{u} = \mathbf{u}(\mathbf{x})∂vu∂x\Large \frac{\partial{v \mathbf{u}}}{\partial{\mathbf{x}}}vv ∂u∂x+\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + u\mathbf{u} ∂v∂x\large \frac{\partial{v}}{\partial{\mathbf{x}}}vv ∂u∂x+∂v∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \frac{\partial{v}}{\partial{\mathbf{x}}} uT\mathbf{u}^TProof 1-5
A\large \mathbf{A} is not a function of x\large \mathbf{x},
u=u(x)\mathbf{u} = \mathbf{u}(\mathbf{x})
∂Au∂x\Large \frac{\partial{\mathbf{Au}}}{\partial{\mathbf{x}}}A\mathbf{A} ∂u∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}∂u∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} AT\mathbf{A}^TProof 1-6
u=u(x),  v=v(x)\mathbf{u} = \mathbf{u}(\mathbf{x}), \; \mathbf{v} = \mathbf{v}(\mathbf{x})∂(u+v)∂x\Large \frac{\partial{(\mathbf{u} + \mathbf{v})}}{\partial{\mathbf{x}}}∂u∂x+∂v∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}}∂u∂x+∂v∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}}
u=u(x)\mathbf{u} = \mathbf{u}(\mathbf{x})∂g(u)∂x\Large \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}}∂g(u)∂u∂u∂x\large \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}∂u∂x∂g(u)∂x\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}}Proof 1-7
u=u(x)\mathbf{u} = \mathbf{u}(\mathbf{x})∂f(g(u))∂x\Large \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{x}}}∂f(g)∂g∂g(u)∂u∂u∂x\large \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}∂u∂x∂g(u)∂u∂f(g)∂g\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}}Proof 1-8

标量对向量 (Scalar-by-vector)

CoditionExpressionNumerator layout
(row vector)
Denominator layout
(column vector)
Proof
a\large a is not a function of x\large \mathbf{x}∂a∂x\Large \frac{\partial{a}}{\partial{\mathbf{x}}}0T\large \mathbf{0}^T0\large \mathbf{0}
a\large a is not a function of x\large \mathbf{x},
u=u(x)u = u(\mathbf{x})
∂au∂x\Large \frac{\partial{au}}{\partial{\mathbf{x}}}a∂u∂x\large a \frac{\partial{u}}{\partial{\mathbf{x}}}a∂u∂x\large a \frac{\partial{u}}{\partial{\mathbf{x}}}
u=u(x),  v=v(x)u = u(\mathbf{x}), \; v = v(\mathbf{x})∂(u+v)∂x\Large \frac{\partial{(u + v)}}{\partial{\mathbf{x}}}∂u∂x+∂v∂x\large \frac{\partial{u}}{\partial{\mathbf{x}}} + \frac{\partial{v}}{\partial{\mathbf{x}}}∂u∂x+∂v∂x\large \frac{\partial{u}}{\partial{\mathbf{x}}} + \frac{\partial{v}}{\partial{\mathbf{x}}}
u=u(x),  v=v(x)u = u(\mathbf{x}), \; v = v(\mathbf{x})∂uv∂x\Large \frac{\partial{uv}}{\partial{\mathbf{x}}}u∂v∂x+v∂u∂x\large u \frac{\partial{v}}{\partial{\mathbf{x}}} + v \frac{\partial{u}}{\partial{\mathbf{x}}}u∂v∂x+v∂u∂x\large u \frac{\partial{v}}{\partial{\mathbf{x}}} + v \frac{\partial{u}}{\partial{\mathbf{x}}}
u=u(x)u = u(\mathbf{x})∂g(u)∂x\Large \frac{\partial{g}(u)}{\partial{\mathbf{x}}}∂g(u)∂u∂u∂x\large \frac{\partial{g}(u)}{\partial{u}} \frac{\partial{u}}{\partial{\mathbf{x}}}∂g(u)∂u∂u∂x\large \frac{\partial{g}(u)}{\partial{u}} \frac{\partial{u}}{\partial{\mathbf{x}}}
u=u(x)u = u(\mathbf{x})∂f(g(u))∂x\Large \frac{\partial{f(g(u))}}{\partial{\mathbf{x}}}∂f(g)∂g∂g(u)∂u∂u∂x\large \frac{\partial{f}(g)}{\partial{g}} \frac{\partial{g}(u)}{\partial{u}} \frac{\partial{u}}{\partial{\mathbf{x}}}∂f(g)∂g∂g(u)∂u∂u∂x\large \frac{\partial{f}(g)}{\partial{g}} \frac{\partial{g}(u)}{\partial{u}} \frac{\partial{u}}{\partial{\mathbf{x}}}
u=u(x),  v=v(x)\mathbf{u} = \mathbf{u}(\mathbf{x}), \; \mathbf{v} = \mathbf{v}(\mathbf{x})∂(u⋅v)∂x\Large \frac{\partial{(\mathbf{u \cdot v})}}{\partial{\mathbf{x}}}uT∂v∂x+vT∂u∂x\mathbf{u}^T \large \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} + \normalsize \mathbf{v}^T \large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}∂u∂xv+∂v∂xu\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \normalsize \mathbf{v} + \large \frac{\partial{\mathbf{v}}} {\partial{\mathbf{x}}} \normalsize \mathbf{u}Proof 2-1
u=u(x),  v=v(x)\mathbf{u} = \mathbf{u}(\mathbf{x}), \; \mathbf{v} = \mathbf{v}(\mathbf{x}),
A\mathbf{A} is not a function of x\mathbf{x}
∂(u⋅Av)∂x\Large \frac{\partial{(\mathbf{u} \cdot \mathbf{Av})}}{\partial{\mathbf{x}}}uTA∂v∂x+vTAT∂u∂x\mathbf{u}^T \mathbf{A} \large \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} + \normalsize \mathbf{v}^T \mathbf{A}^T \large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}∂u∂xAv+∂v∂xATu\large \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \normalsize \mathbf{A} \mathbf{v} + \large \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} \normalsize \mathbf{A}^T \mathbf{u}Proof 2-2
\large ∂2f∂x∂xT\Large \frac{\partial^2{f}}{\partial{\mathbf{x}}\partial{\mathbf{x}}^T}HT\mathbf{H}^TH\mathbf{H}Proof 2-3
a\large \mathbf{a} is not a function of x\large \mathbf{x}∂(a⋅x)∂x\Large \frac{\partial{(\mathbf{a \cdot x})}}{\partial{\mathbf{x}}} aT\mathbf{a}^T a\mathbf{a}Proof 2-4
A,  a\mathbf{A, \; a} are not functions of x\mathbf{x}∂(aTAx)∂x\Large \frac{\partial{(\mathbf{a}^T \mathbf{Ax})}}{\partial{\mathbf{x}}}aTA\mathbf{a}^T \mathbf{A}ATa\mathbf{A}^T \mathbf{a}Proof 2-5
A\mathbf{A} is not a function of x\mathbf{x}∂(xTAx)∂x\Large \frac{\partial({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}}}xT(A+AT)\mathbf{x}^T (\mathbf{A} + \mathbf{A}^T)(A+AT)x(\mathbf{A} + \mathbf{A}^T) \mathbf{x}Proof 2-6
A\mathbf{A} is not a function of x\mathbf{x}∂2(xTAx)∂x∂xT\Large \frac{\partial^2({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}\partial{\mathbf{x}^T}}}A+AT\mathbf{A} + \mathbf{A}^TA+AT\mathbf{A} + \mathbf{A}^TProof 2-7
\large ∂∥x∥2∂x\Large \frac{\partial{\Vert \mathbf{x} \Vert^2}}{\partial{\mathbf{x}}}2xT2 \mathbf{x}^T2x2 \mathbf{x}Proof 2-8
a,b\large \mathbf{a}, \mathbf{b} are not functions of x\large \mathbf{x}∂(aTxxTb)∂x\Large \frac{\partial({\mathbf{a}^T \mathbf{xx}^T \mathbf{b}})}{\partial{\mathbf{x}}}xT(abT+baT)\mathbf{x}^T (\mathbf{ab}^T + \mathbf{ba}^T)(abT+baT)x(\mathbf{ab}^T + \mathbf{ba}^T) \mathbf{x}Proof 2-9
a\large \mathbf{a} is not a function of x\large \mathbf{x}∂∥x−a∥∂x\Large \frac{\partial{\Vert \mathbf{x - a} \Vert}}{\partial{\mathbf{x}}}(x−a)T∥x−a∥\large \frac{(\mathbf{x - a})^T}{\Vert \mathbf{x - a} \Vert}x−a∥x−a∥\large \frac{\mathbf{x - a}}{\Vert \mathbf{x - a} \Vert}Proof 2-10

向量对标量 (Vector-by-scalar)

CoditionExpressionNumerator layoutDenominator layoutProof
a\large \mathbf{a} is not a function of x\large x∂a∂x\Large \frac{\partial{\mathbf{a}}}{\partial{x}} 0\large \mathbf{0}0\large \mathbf{0}
a\large a is not a function of x\large x,
u=u(x)\mathbf{u} = \mathbf{u}(x)
∂au∂x\Large \frac{\partial{a\mathbf{u}}}{\partial{x}}a∂u∂x\normalsize a \large \frac{\partial{\mathbf{u}}}{\partial{x}}a∂u∂x\normalsize a \large \frac{\partial{\mathbf{u}}}{\partial{x}}
A\large \mathbf{A} is not a function of x\large x.
u=u(x)\mathbf{u} = \mathbf{u}(x)
∂Au∂x\Large \frac{\partial{\mathbf{Au}}}{\partial{x}}A∂u∂x\normalsize \mathbf{A} \large \frac{\partial{\mathbf{u}}}{\partial{x}}∂u∂xAT\large \frac{\partial{\mathbf{u}}}{\partial{x}} \normalsize \mathbf{A}^TProof 3-1
u=u(x)\mathbf{u} = \mathbf{u}(x)∂uT∂x\Large \frac{\partial{\mathbf{u}^T}}{\partial{x}}(∂u∂x)T\large (\frac{\partial{\mathbf{u}}}{\partial{x}})^T(∂u∂x)T\large (\frac{\partial{\mathbf{u}}}{\partial{x}})^T
u=u(x),v=v(x)\mathbf{u} = \mathbf{u}(x), \mathbf{v} = \mathbf{v}(x)∂(u+v)∂x\Large \frac{\partial(\mathbf{u} + \mathbf{v})}{\partial{x}}∂u∂x+∂v∂x\large \frac{\partial{\mathbf{u}}}{\partial{x}} + \frac{\partial{\mathbf{v}}}{\partial{x}}∂u∂x+∂v∂x\large \frac{\partial{\mathbf{u}}}{\partial{x}} + \frac{\partial{\mathbf{v}}}{\partial{x}}
u=u(x),v=v(x)\mathbf{u} = \mathbf{u}(x), \mathbf{v} = \mathbf{v}(x)∂(uTv)∂x\Large \frac{\partial(\mathbf{u}^T \mathbf{v})}{\partial{x}}(∂u∂x)Tv+uT∂v∂x\large (\frac{\partial{\mathbf{u}}}{\partial{x}})^T \normalsize \mathbf{v} + \mathbf{u}^T \large \frac{\partial{\mathbf{v}}}{\partial{x}}∂u∂xv+uT(∂v∂x)T\large \frac{\partial{\mathbf{u}}}{\partial{x}} \normalsize \mathbf{v} + \mathbf{u}^T \large (\frac{\partial{\mathbf{v}}}{\partial{x}})^TProof 3-2
u=u(x)\mathbf{u} = \mathbf{u}(x)∂g(u)∂x\Large \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{x}}∂g(u)∂u∂u∂x\large \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}}∂u∂x∂g(u)∂u\large \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}}Proof 3-3
u=u(x)\mathbf{u} = \mathbf{u}(x)∂f(g(u))∂x\Large \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{x}}∂f(g)∂g∂g(u)∂u∂u∂x\large \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}}∂u∂x∂g(u)∂u∂f(g)∂g\large \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}}Proof 3-4
U=U(x),v=v(x)\mathbf{U} = \mathbf{U}(x), \mathbf{v} = \mathbf{v}(x)∂Uv∂x\Large \frac{\partial{\mathbf{U}\mathbf{v}}}{\partial{x}}∂U∂xv+U∂v∂x\large \frac{\partial{\mathbf{U}}}{\partial{x}} \normalsize \mathbf{v} + \mathbf{U} \large \frac{\partial{\mathbf{v}}}{\partial{x}}vT∂U∂x+∂v∂xUT\large \normalsize \mathbf{v}^T \large \frac{\partial{\mathbf{U}}}{\partial{x}} + \frac{\partial{\mathbf{v}}}{\partial{x}} \normalsize \mathbf{U}^TProof 3-5

推导过程

向量对向量 (Vector-by-vector)

  • Proof 1-1
    • ∂a∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}}},   a\; \large \mathbf{a} is not a function of x\large \mathbf{x}

      a=(a1⋯am)m×1T,x=(x1⋯xn)n×1T\mathbf{a} = \begin{pmatrix} a_1 & \cdots & a_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Numerator layout:

      ∂a∂x=(∂a1∂x1⋯∂a1∂xn⋮⋱⋮∂am∂x1⋯∂am∂xn)m×n=(0⋯0⋮⋱⋮0⋯0)m×n=0m×n\frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} = \large \begin{pmatrix} \frac{\partial{a_1}}{\partial{x_1}} & \cdots & \frac{\partial{a_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{a_m}}{\partial{x_1}} & \cdots & \frac{\partial{a_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} = \begin{pmatrix} 0 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 0 \end{pmatrix}_{m \times n} = \mathbf{0}_{m \times n}

      Denominator layout:

      ∂a∂x=(∂a1∂x1⋯∂am∂x1⋮⋱⋮∂a1∂xn⋯∂am∂xn)n×m=(0⋯0⋮⋱⋮0⋯0)n×m=0n×m\frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} = \large \begin{pmatrix} \frac{\partial{a_1}}{\partial{x_1}} & \cdots & \frac{\partial{a_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{a_1}}{\partial{x_n}} & \cdots & \frac{\partial{a_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} = \begin{pmatrix} 0 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 0 \end{pmatrix}_{n \times m} = \mathbf{0}_{n \times m}

Back to table



  • Proof 1-2
    • ∂Ax∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{Ax}}}{\partial{\mathbf{x}}}},   A\; \large \mathbf{A} is not a funtion of x\large \mathbf{x}

      A=(a11⋯a1n⋮⋱⋮am1⋯amn)m×n,x=(x1⋯xn)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix}_{m \times n} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Denominator layout:

      ∂Ax∂x=∂∂x{(a11⋯a1n⋮⋱⋮am1⋯amn)m×n(x1⋮xn)n×1}=∂∂x(∑i=1na1ixi⋮∑i=1namixi)m×1=(∂∑i=1na1ixi∂x1⋯∂∑i=1namixi∂x1⋮⋱⋮∂∑i=1na1ixi∂xn⋯∂∑i=1namixi∂xn)n×m=(a11⋯am1⋮⋱⋮a1n⋯amn)n×m=AT\begin{aligned} \frac{\partial{\mathbf{Ax}}}{\partial{\mathbf{x}}} &= \frac{\partial}{\partial{\mathbf{x}}} \left\{ \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix}_{m \times n} \begin{pmatrix} x_1 \\ \vdots \\ x_n \end{pmatrix}_{n \times 1} \right\} \\[10mm] &= \frac{\partial}{\partial{\mathbf{x}}} \large \begin{pmatrix} \sum_{i=1}^n a_{1i}x_i \\ \vdots \\ \sum_{i=1}^n a_{mi}x_i \end{pmatrix}_{m \times 1} \\[10mm] &= \Large \begin{pmatrix} \frac{\partial{\sum_{i=1}^n a_{1i}x_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^n a_{mi}x_i}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{\sum_{i=1}^n a_{1i}x_i}}{\partial{x_n}} & \cdots & \frac{\partial{\sum_{i=1}^n a_{mi}x_i}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \begin{pmatrix} a_{11} & \cdots & a_{m1} \\ \vdots & \ddots & \vdots \\ a_{1n} & \cdots & a_{mn} \end{pmatrix}_{n \times m} \\[10mm] &= \mathbf{A}^T \end{aligned}

Back to table



  • Proof 1-3
    • ∂xTA∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{x}^T\mathbf{A}}}{\partial{\mathbf{x}}}},   A\; \large \mathbf{A} is not a funtion of x\large \mathbf{x}

      A=(a11⋯a1m⋮⋱⋮an1⋯anm)n×m,x=(x1⋯xn)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{n1} & \cdots & a_{nm} \end{pmatrix}_{n \times m} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Denominator layout:

      ∂xTA∂x=∂∂x{(x1⋯xn)1×n(a11⋯a1m⋮⋱⋮an1⋯anm)n×m}=∂∂x(∑i=1nai1xi⋯∑i=1naimxi)1×m=(∂∑i=1nai1xi∂x1⋯∂∑i=1naimxi∂x1⋮⋱⋮∂∑i=1nai1xi∂xn⋯∂∑i=1naimxi∂xn)n×m=(a11⋯a1m⋮⋱⋮an1⋯anm)n×m=A\begin{aligned} \frac{\partial{\mathbf{x}^T\mathbf{A}}}{\partial{\mathbf{x}}} &= \frac{\partial}{\partial{\mathbf{x}}} \left\{ \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{1 \times n} \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{n1} & \cdots & a_{nm} \end{pmatrix}_{n \times m} \right\} \\[10mm] &= \frac{\partial}{\partial{\mathbf{x}}} \begin{pmatrix} \sum_{i=1}^n a_{i1}x_i & \cdots & \sum_{i=1}^n a_{im}x_i \end{pmatrix}_{1 \times m} \\[5mm] &= \Large \begin{pmatrix} \frac{\partial{\sum_{i=1}^n a_{i1}x_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^n a_{im}x_i}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{\sum_{i=1}^n a_{i1}x_i}}{\partial{x_n}} & \cdots & \frac{\partial{\sum_{i=1}^n a_{im}x_i}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{n1} & \cdots & a_{nm} \end{pmatrix}_{n \times m} \\[10mm] &= \mathbf{A} \end{aligned}

Back to table



  • Proof 1-4
    • ∂au∂x\color{royalblue}{\LARGE \frac{\partial{a\mathbf{u}}}{\partial{\mathbf{x}}}},   a\; \large a is not a function of x,  u=u(x)x, \; \mathbf{u} = \mathbf{u}(\mathbf{x})

      u=(u1⋯um)m×1T,x=(x1⋯xn)n×1T\mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Denominator layout:

      ∂au∂x=(∂au1∂x1⋯∂aum∂x1⋮⋱⋮∂au1∂xn⋯∂aum∂xn)n×m=a(∂u1∂x1⋯∂um∂x1⋮⋱⋮∂u1∂xn⋯∂um∂xn)n×m=a∂u∂x\begin{aligned} \frac{\partial{a\mathbf{u}}}{\partial{\mathbf{x}}} &= \large \begin{pmatrix} \frac{\partial{au_1}}{\partial{x_1}} & \cdots & \frac{\partial{au_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{au_1}}{\partial{x_n}} & \cdots & \frac{\partial{au_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= a \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= a \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}

Back to table



  • Proof 1-5
    • ∂vu∂x\color{royalblue}{\LARGE \frac{\partial{v \mathbf{u}}}{\partial{\mathbf{x}}}},   v=v(x),  u=u(x)\; v = v(\mathbf{x}), \; \mathbf{u} = \mathbf{u}(\mathbf{x})

      u=(u1⋯um)m×1T,x=(x1⋯xn)n×1T\mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Numerator layout:

      ∂vu∂x=(∂vu1∂x1⋯∂vu1∂xn⋮⋱⋮∂vum∂x1⋯∂vu1∂xn)m×n=(∂v∂x1u1+v∂u1∂x1⋯∂v∂xnu1+v∂u1∂xn⋮⋱⋮∂v∂x1um+v∂um∂x1⋯∂v∂xnum+v∂um∂xn)m×n=(∂v∂x1u1⋯∂v∂xnu1⋮⋱⋮∂v∂x1um⋯∂v∂xnum)m×n+(v∂u1∂x1⋯v∂u1∂xn⋮⋱⋮v∂um∂x1⋯v∂um∂xn)m×n=(u1⋮um)m×1(∂v∂x1⋯∂v∂xn)1×n+v(∂u1∂x1⋯∂u1∂xn⋮⋱⋮∂um∂x1⋯∂um∂xn)m×n=u∂v∂x+v∂u∂x\begin{aligned} \frac{\partial{v \mathbf{u}}}{\partial{\mathbf{x}}} &= \large \begin{pmatrix} \frac{\partial{v u_1}}{\partial{x_1}} & \cdots & \frac{\partial{v u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{v u_m}}{\partial{x_1}} & \cdots & \frac{\partial{v u_1}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} u_1 + v \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{v}}{\partial{x_n}} u_1 + v \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{v}}{\partial{x_1}} u_m + v \frac{\partial{u_m}}{\partial{x_1}} & \cdots & \frac{\partial{v}}{\partial{x_n}} u_m + v \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} u_1 & \cdots & \frac{\partial{v}}{\partial{x_n}} u_1 \\ \vdots & \ddots & \vdots \\ \frac{\partial{v}}{\partial{x_1}} u_m & \cdots & \frac{\partial{v}}{\partial{x_n}} u_m \end{pmatrix}_{m \times n} \quad + \quad \large \begin{pmatrix} v \frac{\partial{u_1}}{\partial{x_1}} & \cdots & v \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ v \frac{\partial{u_m}}{\partial{x_1}} & \cdots & v \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \large \begin{pmatrix} u_1 \\ \vdots \\ u_m \end{pmatrix}_{m \times 1} \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} & \cdots & \frac{\partial{v}}{\partial{x_n}} \end{pmatrix}_{1 \times n} \quad + \quad v \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_m}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \mathbf{u} \frac{\partial{v}}{\partial{\mathbf{x}}} + v \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}



      Denominator layout:

      ∂vu∂x=(∂vu1∂x1⋯∂vum∂x1⋮⋱⋮∂vu1∂xn⋯∂vum∂xn)n×m=(∂v∂x1u1+v∂u1∂x1⋯∂v∂x1um+v∂um∂x1⋮⋱⋮∂v∂xnu1+v∂u1∂xn⋯∂v∂xnum+v∂um∂xn)n×m=(∂v∂x1u1⋯∂v∂x1um⋮⋱⋮∂v∂xnu1⋯∂v∂xnum)n×m+(v∂u1∂x1⋯v∂um∂x1⋮⋱⋮v∂u1∂xn⋯v∂um∂xn)n×m=(∂v∂x1⋮∂v∂xn)n×1(u1⋯um)1×m+v(∂u1∂x1⋯∂um∂x1⋮⋱⋮∂u1∂xn⋯∂um∂xn)n×m=∂v∂xuT+v∂u∂x\begin{aligned} \frac{\partial{v \mathbf{u}}}{\partial{\mathbf{x}}} &= \large \begin{pmatrix} \frac{\partial{v u_1}}{\partial{x_1}} & \cdots & \frac{\partial{v u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{v u_1}}{\partial{x_n}} & \cdots & \frac{\partial{v u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} u_1 + v \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{v}}{\partial{x_1}} u_m + v \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{v}}{\partial{x_n}} u_1 + v \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{v}}{\partial{x_n}} u_m + v \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} u_1 & \cdots & \frac{\partial{v}}{\partial{x_1}} u_m \\ \vdots & \ddots & \vdots \\ \frac{\partial{v}}{\partial{x_n}} u_1 & \cdots & \frac{\partial{v}}{\partial{x_n}} u_m \end{pmatrix}_{n \times m} \quad + \quad \large \begin{pmatrix} v \frac{\partial{u_1}}{\partial{x_1}} & \cdots & v \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ v \frac{\partial{u_1}}{\partial{x_n}} & \cdots & v \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{v}}{\partial{x_1}} \\ \vdots \\ \frac{\partial{v}}{\partial{x_n}} \end{pmatrix}_{n \times 1} \large \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{1 \times m} \quad + \quad v \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \frac{\partial{v}}{\partial{\mathbf{x}}} \mathbf{u}^T + v \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}

Back to table



  • Proof 1-6
    • ∂Au∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{Au}}}{\partial{\mathbf{x}}}},   A\; \mathbf{A} is not a function of x,  u=u(x)\mathbf{x}, \; \mathbf{u} = \mathbf{u}(\mathbf{x})

      A=(a11⋯a1m⋮⋱⋮am1⋯amm)m×m,u=(u1⋯um)m×1T,x=(x1⋯xn)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mm} \end{pmatrix}_{m \times m} , \qquad \mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Numerator layout:

      ∂Au∂x=∂∂x(a11⋯a1m⋮⋱⋮am1⋯amm)m×m(u1⋮um)m×1=∂∂x(∑i=1ma1iui⋮∑i=1mamiui)m×1=(∂∑i=1ma1iui∂x1⋯∂∑i=1ma1iui∂xn⋮⋱⋮∂∑i=1mamiui∂x1⋯∂∑i=1mamiui∂xn)m×n=(∑i=1ma1i∂ui∂x1⋯∑i=1ma1i∂ui∂xn⋮⋱⋮∑i=1mami∂ui∂xn⋯∑i=1mami∂ui∂xn)m×n=(a11⋯a1m⋮⋱⋮am1⋯amm)m×m(∂u1∂x1⋯∂u1∂xn⋮⋱⋮∂um∂xn⋯∂um∂xn)m×n=A∂u∂x\begin{aligned} \frac{\partial{\mathbf{Au}}}{\partial{\mathbf{x}}} &= \frac{\partial}{\partial{\mathbf{x}}} \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mm} \end{pmatrix}_{m \times m} \begin{pmatrix} u_1 \\ \vdots \\ u_m \end{pmatrix}_{m \times 1} \\[10mm] &= \frac{\partial}{\partial{\mathbf{x}}} \large \begin{pmatrix} \sum_{i=1}^m a_{1i} u_i \\ \vdots \\ \sum_{i=1}^m a_{mi} u_i \end{pmatrix}_{m \times 1} \\[10mm] &= \Large \begin{pmatrix} \frac{\partial{\sum_{i=1}^m a_{1i} u_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^m a_{1i} u_i}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{\sum_{i=1}^m a_{mi} u_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^m a_{mi} u_i}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \large \begin{pmatrix} \sum_{i=1}^m a_{1i} \frac{\partial{u_i}}{\partial{x_1}} & \cdots & \sum_{i=1}^m a_{1i} \frac{\partial{u_i}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \sum_{i=1}^m a_{mi} \frac{\partial{u_i}}{\partial{x_n}} & \cdots & \sum_{i=1}^m a_{mi} \frac{\partial{u_i}}{\partial{x_n}} \\ \end{pmatrix}_{m \times n} \\[10mm] &= \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mm} \end{pmatrix}_{m \times m} \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_m}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \mathbf{A} \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}



      Denominator layout:

      ∂Au∂x=∂∂x(a11⋯a1m⋮⋱⋮am1⋯amm)m×m(u1⋮um)m×1=∂∂x(∑i=1ma1iui⋮∑i=1mamiui)m×1=(∂∑i=1ma1iui∂x1⋯∂∑i=1mamiui∂x1⋮⋱⋮∂∑i=1ma1iui∂xn⋯∂∑i=1mamiui∂xn)n×m=(∑i=1ma1i∂ui∂x1⋯∑i=1mami∂ui∂x1⋮⋱⋮∑i=1ma1i∂ui∂xn⋯∑i=1mami∂ui∂xn)n×m=(∂u1∂x1⋯∂um∂x1⋮⋱⋮∂u1∂xn⋯∂um∂xn)n×m(a11⋯am1⋮⋱⋮a1m⋯amm)m×m=∂u∂xAT\begin{aligned} \frac{\partial{\mathbf{Au}}}{\partial{\mathbf{x}}} &= \frac{\partial}{\partial{\mathbf{x}}} \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mm} \end{pmatrix}_{m \times m} \begin{pmatrix} u_1 \\ \vdots \\ u_m \end{pmatrix}_{m \times 1} \\[10mm] &= \frac{\partial}{\partial{\mathbf{x}}} \large \begin{pmatrix} \sum_{i=1}^m a_{1i} u_i \\ \vdots \\ \sum_{i=1}^m a_{mi} u_i \end{pmatrix}_{m \times 1} \\[10mm] &= \Large \begin{pmatrix} \frac{\partial{\sum_{i=1}^m a_{1i} u_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^m a_{mi} u_i}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{\sum_{i=1}^m a_{1i} u_i}}{\partial{x_n}} & \cdots & \frac{\partial{\sum_{i=1}^m a_{mi} u_i}}{\partial{x_n}} \end{pmatrix}_{n \times m} \\[10mm] &= \large \begin{pmatrix} \sum_{i=1}^m a_{1i} \frac{\partial{u_i}}{\partial{x_1}} & \cdots & \sum_{i=1}^m a_{mi} \frac{\partial{u_i}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \sum_{i=1}^m a_{1i} \frac{\partial{u_i}}{\partial{x_n}} & \cdots & \sum_{i=1}^m a_{mi} \frac{\partial{u_i}}{\partial{x_n}} \\ \end{pmatrix}_{n \times m} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \begin{pmatrix} a_{11} & \cdots & a_{m1} \\ \vdots & \ddots & \vdots \\ a_{1m} & \cdots & a_{mm} \end{pmatrix}_{m \times m} \\[10mm] &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \mathbf{A}^T \end{aligned}

Back to table



  • Proof 1-7
    • ∂g(u)∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}}},   u=u(x)\; \mathbf{u} = \mathbf{u}(\mathbf{x})

      g=(g1⋯gp)p×1T,u=(u1⋯um)m×1T,x=(x1⋯xn)n×1T\mathbf{g} = \begin{pmatrix} g_1 & \cdots & g_p \end{pmatrix}_{p \times 1}^T , \qquad \mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T

      Numerator layout:

      ∂g(u)∂x=(∂g1∂x1⋯∂g1∂xn⋮⋱⋮∂gp∂x1⋯∂gp∂xn)p×n=(∂g1∂u∂u∂x1⋯∂g1∂u∂u∂xn⋮⋱⋮∂gp∂u∂u∂x1⋯∂gp∂u∂u∂xn)p×n=((∂g1∂u1⋯∂g1∂um)(∂u1∂x1⋮∂um∂x1)⋯(∂g1∂u1⋯∂g1∂um)(∂u1∂xn⋮∂um∂xn)⋮⋱⋮(∂gp∂u1⋯∂gp∂um)(∂u1∂x1⋮∂um∂x1)⋯(∂gp∂u1⋯∂gp∂um)(∂u1∂xn⋮∂um∂xn))p×n=(∑i=1m∂g1∂ui∂ui∂x1⋯∑i=1m∂g1∂ui∂ui∂xn⋮⋱⋮∑i=1m∂gp∂ui∂ui∂x1⋯∑i=1m∂gp∂ui∂ui∂xn)p×n=(∂g1∂u1⋯∂g1∂um⋮⋱⋮∂gp∂u1⋯∂gp∂um)p×m(∂u1∂x1⋯∂u1∂xn⋮⋱⋮∂um∂x1⋯∂um∂xn)m×n=∂g(u)∂u∂u∂x\begin{aligned} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}} &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{x_1}} & \cdots & \frac{\partial{g_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_p}}{\partial{x_1}} & \cdots & \frac{\partial{g_p}}{\partial{x_n}} \end{pmatrix}_{p \times n} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x_1}} & \cdots & \frac{\partial{g_1}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_p}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x_1}} & \cdots & \frac{\partial{g_p}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x_n}} \end{pmatrix}_{p \times n} \\[10mm] &= \large \begin{pmatrix} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_1}}{\partial{u_m}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} \\ \vdots \\ \frac{\partial{u_m}}{\partial{x_1}} \end{pmatrix} & \cdots & \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_1}}{\partial{u_m}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots \\ \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix} \\ \vdots & \ddots & \vdots \\ \begin{pmatrix} \frac{\partial{g_p}}{\partial{u_1}} & \cdots & \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} \\ \vdots \\ \frac{\partial{u_m}}{\partial{x_1}} \end{pmatrix} & \cdots & \begin{pmatrix} \frac{\partial{g_p}}{\partial{u_1}} & \cdots & \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots \\ \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix} \end{pmatrix}_{p \times n} \\[10mm] &= \large \begin{pmatrix} \sum_{i=1}^m \frac{\partial{g_1}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x_1}} & \cdots & \sum_{i=1}^m \frac{\partial{g_1}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \sum_{i=1}^m \frac{\partial{g_p}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x_1}} & \cdots & \sum_{i=1}^m \frac{\partial{g_p}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x_n}} \end{pmatrix}_{p \times n} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_1}}{\partial{u_m}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_p}}{\partial{u_1}} & \cdots & \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix}_{p \times m} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_m}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}



      Denominator layout:

      ∂g(u)∂x=(∂g1∂x1⋯∂gp∂x1⋮⋱⋮∂g1∂xn⋯∂gp∂xn)n×p=(∂u∂x1∂g1∂u⋯∂u∂x1∂gp∂u⋮⋱⋮∂u∂xn∂g1∂u⋯∂u∂xn∂gp∂u)n×p=((∂u1∂x1⋯∂um∂x1)(∂g1∂u1⋮∂g1∂um)⋯(∂u1∂x1⋯∂um∂x1)(∂gp∂u1⋮∂gp∂um)⋮⋱⋮(∂u1∂xn⋯∂um∂xn)(∂g1∂u1⋮∂g1∂um)⋯(∂u1∂xn⋯∂um∂xn)(∂gp∂u1⋮∂gp∂um))n×p=(∑i=1m∂ui∂x1∂g1∂ui⋯∑i=1m∂ui∂x1∂gp∂ui⋮⋱⋮∑i=1m∂ui∂xn∂g1∂ui⋯∑i=1m∂ui∂xn∂gp∂ui)n×p=(∂u1∂x1⋯∂um∂x1⋮⋱⋮∂u1∂xn⋯∂um∂xn)n×m(∂g1∂u1⋯∂gp∂u1⋮⋱⋮∂g1∂um⋯∂gp∂um)m×p=∂u∂x∂g(u)∂u\begin{aligned} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}} &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{x_1}} & \cdots & \frac{\partial{g_p}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_1}}{\partial{x_n}} & \cdots & \frac{\partial{g_p}}{\partial{x_n}} \end{pmatrix}_{n \times p} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{\mathbf{u}}}{\partial{x_1}} \frac{\partial{g_1}}{\partial{\mathbf{u}}} & \cdots & \frac{\partial{\mathbf{u}}}{\partial{x_1}} \frac{\partial{g_p}}{\partial{\mathbf{u}}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{\mathbf{u}}}{\partial{x_n}} \frac{\partial{g_1}}{\partial{\mathbf{u}}}& \cdots & \frac{\partial{\mathbf{u}}}{\partial{x_n}} \frac{\partial{g_p}}{\partial{\mathbf{u}}} \end{pmatrix}_{n \times p} \\[10mm] &= \large \begin{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_1}}{\partial{u_m}} \end{pmatrix} & \cdots & \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_p}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix} \\ \vdots & \ddots & \vdots \\ \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_1}}{\partial{u_m}} \end{pmatrix} & \cdots & \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_p}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix} \end{pmatrix}_{n \times p} \\[10mm] &= \large \begin{pmatrix} \sum_{i=1}^m \frac{\partial{u_i}}{\partial{x_1}} \frac{\partial{g_1}}{\partial{u_i}} & \cdots & \sum_{i=1}^m \frac{\partial{u_i}}{\partial{x_1}} \frac{\partial{g_p}}{\partial{u_i}} \\ \vdots & \ddots & \vdots \\ \sum_{i=1}^m \frac{\partial{u_i}}{\partial{x_n}} \frac{\partial{g_1}}{\partial{u_i}} & \cdots & \sum_{i=1}^m \frac{\partial{u_i}}{\partial{x_n}} \frac{\partial{g_p}}{\partial{u_i}} \end{pmatrix}_{n \times p} \\[10mm] &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_1}}{\partial{x_n}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{n \times m} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_p}}{\partial{u_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_1}}{\partial{u_m}} & \cdots & \frac{\partial{g_p}}{\partial{u_m}} \end{pmatrix}_{m \times p} \\[10mm] &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \end{aligned}

Back to table



  • Proof 1-8
    • ∂f(g(u))∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{x}}}},   u=u(x)\; \mathbf{u} = \mathbf{u}(\mathbf{x})



      Numerator layout:

      ∂f(g(u))∂x=∂f(g(u))∂g(u)∂g(u)∂x=∂f(g)∂g∂g(u)∂u∂u∂x\begin{aligned} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{x}}} &= \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{g}(\mathbf{u})}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}} \\[5mm] &= \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \end{aligned}



      Denominator layout:

      ∂f(g(u))∂x=∂g(u)∂x∂f(g(u))∂g(u)=∂u∂x∂g(u)∂u∂f(g)∂g\begin{aligned} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{x}}} &= \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{x}}} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{g}(\mathbf{u})}} \\[5mm] &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \end{aligned}

Back to table

标量对向量 (Scalar-by-vector)



  • Proof 2-1
    • ∂(u⋅v)∂x\color{royalblue}{\LARGE \frac{\partial{(\mathbf{u \cdot v})}}{\partial{\mathbf{x}}}},   u=u(x),  v=v(x)\; \mathbf{u} = \mathbf{u}(\mathbf{x}), \; \mathbf{v} = \mathbf{v}(\mathbf{x})

      u=(u1⋯um)m×1T,v=(v1⋯vm)m×1T,x=(x1⋯xn)n×1T\mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{v} = \begin{pmatrix} v_1 & \cdots & v_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂(u⋅v)∂x=∂∂x∑i=1muivi=(∂∑i=1muivi∂x1⋯∂∑i=1muivi∂xn)1×n=(∑i=1m(∂ui∂x1vi+ui∂vi∂x1)⋯∑i=1m(∂ui∂xnvi+ui∂vi∂xn))1×n=(∑i=1m(∂ui∂x1vi)⋯∑i=1m(∂ui∂xnvi))1×n+(∑i=1mui(∂vi∂x1)⋯∑i=1mui(∂vi∂xn))1×n=(v1⋯vm)1×m(∂u1∂x1⋯∂u1∂xn⋮⋱⋮∂um∂x1⋯∂um∂xn)m×n+(u1⋯um)1×m(∂v1∂x1⋯∂v1∂xn⋮⋱⋮∂vm∂x1⋯∂vm∂xn)m×n=vT∂u∂x+uT∂v∂x\begin{aligned} \begin{split} \frac{\partial{(\mathbf{u \cdot v})}}{\partial{\mathbf{x}}} &= \frac{\partial}{\partial{\mathbf{x}}} \sum_{i=1}^m u_i v_i \\[5mm] &= \Large \begin{pmatrix} \frac{\partial{\sum_{i=1}^m u_i v_i}}{\partial{x_1}} & \cdots & \frac{\partial{\sum_{i=1}^m u_i v_i}}{\partial{x_n}} \end{pmatrix}_{1 \times n} \\[5mm] &= \large \begin{pmatrix} \sum_{i=1}^m \left(\frac{\partial{u_i}}{\partial{x_1}} v_i + u_i \frac{\partial{v_i}}{\partial{x_1}} \right) & \cdots & \sum_{i=1}^m \left(\frac{\partial{u_i}}{\partial{x_n}} v_i + u_i \frac{\partial{v_i}}{\partial{x_n}} \right) \end{pmatrix}_{1 \times n} \\[5mm] &= \large \begin{pmatrix} \sum_{i=1}^m \left(\frac{\partial{u_i}}{\partial{x_1}} v_i \right) & \cdots & \sum_{i=1}^m \left(\frac{\partial{u_i}}{\partial{x_n}} v_i \right) \end{pmatrix}_{1 \times n} \\ &+ \large \begin{pmatrix} \sum_{i=1}^m u_i \left(\frac{\partial{v_i}}{\partial{x_1}} \right) & \cdots & \sum_{i=1}^m u_i \left(\frac{\partial{v_i}}{\partial{x_n}} \right) \end{pmatrix}_{1 \times n} \\[5mm] &= \large \begin{pmatrix} v_1 & \cdots & v_m \end{pmatrix}_{1 \times m} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x_1}} & \cdots & \frac{\partial{u_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_m}}{\partial{x_1}} & \cdots & \frac{\partial{u_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\ &+ \large \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{1 \times m} \begin{pmatrix} \frac{\partial{v_1}}{\partial{x_1}} & \cdots & \frac{\partial{v_1}}{\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{v_m}}{\partial{x_1}} & \cdots & \frac{\partial{v_m}}{\partial{x_n}} \end{pmatrix}_{m \times n} \\[10mm] &= \mathbf{v}^T \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \mathbf{u}^T \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} \end{split} \end{aligned}



      Denominator layout:

      ∂(u⋅v)∂x={vT(∂u∂x)Numerator+uT(∂v∂x)Numerator}T=(∂u∂x)NumeratorT(vT)T+(∂v∂x)NumeratorT(uT)T=(∂u∂x)Denominatorv+(∂v∂x)Denominatoru=∂u∂xv+∂v∂xu\begin{aligned} \frac{\partial{(\mathbf{u \cdot v})}}{\partial{\mathbf{x}}} &= \left\{ \mathbf{v}^T \underset{\bf{Numerator}}{\left(\frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}\right)} + \mathbf{u}^T \underset{\bf{Numerator}}{\left(\frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}}\right)} \right\}^T \\[7mm] &= \underset{\bf{Numerator}}{\left(\frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}\right)}^T {(\mathbf{v}^T)}^T + \underset{\bf{Numerator}}{\left(\frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}}\right)}^T {(\mathbf{u}^T)}^T \\[7mm] &= \underset{\bf{Denominator}}{\left(\frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}}\right)} \mathbf{v} + \underset{\bf{Denominator}}{\left(\frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}}\right)} \mathbf{u} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \mathbf{v} + \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} \mathbf{u} \end{aligned}

Back to table



  • Proof 2-2
    • ∂(u⋅Av)∂x\color{royalblue}{\LARGE \frac{\partial{(\mathbf{u} \cdot \mathbf{Av})}}{\partial{\mathbf{x}}}},   u=u(x),  v=v(x),  A\; \mathbf{u} = \mathbf{u}(\mathbf{x}), \; \mathbf{v} = \mathbf{v}(\mathbf{x}), \; \mathbf{A} is not a function of x\mathbf{x}

      u=(u1⋯um)m×1T,v=(v1⋯vm)m×1TA=(a11⋯a1m⋮⋱⋮am1⋯amm)m×m,x=(x1⋯xn)n×1T\begin{aligned} & \mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{v} = \begin{pmatrix} v_1 & \cdots & v_m \end{pmatrix}_{m \times 1}^T \\[10mm] & \mathbf{A}= \begin{pmatrix} a_{11} & \cdots & a_{1m} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mm} \end{pmatrix}_{m \times m} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T \end{aligned}



      Numerator layout:

      ∂(u⋅Av)∂x=(Av)T∂u∂x+uT∂(Av)∂x=vTAT∂u∂x+uTA∂v∂x\begin{aligned} \frac{\partial{(\mathbf{u} \cdot \mathbf{Av})}}{\partial{\mathbf{x}}} &= (\mathbf{Av})^T \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \mathbf{u}^T \frac{\partial{(\mathbf{Av})}}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{v}^T \mathbf{A}^T \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} + \mathbf{u}^T \mathbf{A} \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} \end{aligned}



      Denominator layout:

      ∂(u⋅Av)∂x=∂u∂xAv+∂(Av)∂xu=∂u∂xAv+∂v∂xATu\begin{aligned} \frac{\partial{(\mathbf{u} \cdot \mathbf{Av})}}{\partial{\mathbf{x}}} &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \mathbf{Av} + \frac{\partial{(\mathbf{Av})}}{\partial{\mathbf{x}}} \mathbf{u} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{\mathbf{x}}} \mathbf{Av} + \frac{\partial{\mathbf{v}}}{\partial{\mathbf{x}}} \mathbf{A}^T \mathbf{u} \end{aligned}

Back to table



  • Proof 2-3
    • ∂2f∂x∂xT\color{royalblue}{\LARGE \frac{\partial^2{f}}{\partial{\mathbf{x}}\partial{\mathbf{x}}^T}}

      x=(x1⋯xn)n×1T\mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂2f∂x∂xT=∂∂xT∂f∂x=∂∂xT(∂f∂x1⋯∂f∂xn)1×n=(∂f∂x1∂x1⋯∂f∂xn∂x1⋮⋱⋮∂f∂x1∂xn⋯∂f∂xn∂xn)n×n=HT\begin{aligned} \frac{\partial^2{f}}{\partial{\mathbf{x}}\partial{\mathbf{x}}^T} &= \frac{\partial}{\partial{\mathbf{x}^T}} \frac{\partial{f}}{\partial{\mathbf{x}}} \\[7mm] &= \frac{\partial}{\partial{\mathbf{x}^T}} \large \begin{pmatrix} \frac{\partial{f}}{\partial{x_1}} & \cdots & \frac{\partial{f}}{\partial{x_n}} \end{pmatrix}_{1 \times n} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{f}}{\partial{x_1}\partial{x_1}} & \cdots & \frac{\partial{f}}{\partial{x_n}\partial{x_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{f}}{\partial{x_1}\partial{x_n}} & \cdots & \frac{\partial{f}}{\partial{x_n}\partial{x_n}} \end{pmatrix}_{n \times n} \\[7mm] &= \mathbf{H}^T \end{aligned}



      Denominator layout:

      ∂2f∂x∂xT=∂∂xT∂f∂x=∂∂xT(∂f∂x1⋮∂f∂xn)n×1=(∂f∂x1∂x1⋯∂f∂x1∂xn⋮⋱⋮∂f∂xn∂x1⋯∂f∂xn∂xn)n×n=H\begin{aligned} \frac{\partial^2{f}}{\partial{\mathbf{x}}\partial{\mathbf{x}}^T} &= \frac{\partial}{\partial{\mathbf{x}^T}} \frac{\partial{f}}{\partial{\mathbf{x}}} \\[7mm] &= \frac{\partial}{\partial{\mathbf{x}^T}} \large \begin{pmatrix} \frac{\partial{f}}{\partial{x_1}} \\ \vdots \\ \frac{\partial{f}}{\partial{x_n}} \end{pmatrix}_{n \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{f}}{\partial{x_1}\partial{x_1}} & \cdots & \frac{\partial{f}}{\partial{x_1}\partial{x_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{f}}{\partial{x_n}\partial{x_1}} & \cdots & \frac{\partial{f}}{\partial{x_n}\partial{x_n}} \end{pmatrix}_{n \times n} \\[7mm] &= \mathbf{H} \end{aligned}

Back to table



  • Proof 2-4
    • ∂(a⋅x)∂x\color{royalblue}{\LARGE \frac{\partial{(\mathbf{a \cdot x})}}{\partial{\mathbf{x}}}} ,  a, \; \mathbf{a} is not a function of x\mathbf{x}

      a=(a1⋯an)n×1T,x=(x1⋯xn)n×1T\mathbf{a} = \begin{pmatrix} a_1 & \cdots & a_n \end{pmatrix}_{n \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂(a⋅x)∂x=aT∂x∂x+xT∂a∂x=aTI+xT0=aT\begin{aligned} \frac{\partial{(\mathbf{a \cdot x})}}{\partial{\mathbf{x}}} &= \mathbf{a}^T \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} + \mathbf{x}^T \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{I} + \mathbf{x}^T \mathbf{0} \\[7mm] &= \mathbf{a}^T \end{aligned}



      Denominator layout:

      ∂(a⋅x)∂x=∂x∂xa+∂a∂xx=Ia+0x=a\begin{aligned} \frac{\partial{(\mathbf{a \cdot x})}}{\partial{\mathbf{x}}} &= \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{a} + \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} \mathbf{x} \\[7mm] &= \mathbf{I} \mathbf{a} + \mathbf{0} \mathbf{x} \\[7mm] &= \mathbf{a} \end{aligned}

Back to table



  • Proof 2-5
    • ∂(aTAx)∂x\color{royalblue}{\LARGE \frac{\partial{(\mathbf{a}^T \mathbf{Ax})}}{\partial{\mathbf{x}}}} ,  A, \; \mathbf{A}, a\mathbf{a} are not functions of x\mathbf{x}

      a=(a1⋯am)m×1T,A=(a11⋯a1n⋮⋱⋮am1⋯amn)m×n,x=(x1⋯xn)n×1T\mathbf{a} = \begin{pmatrix} a_1 & \cdots & a_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix}_{m \times n} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂(aTAx)∂x=∂(a⋅Ax)∂x=aTA∂x∂x+xTAT∂a∂x=aTAI+xTAT0=aTA\begin{aligned} \frac{\partial{(\mathbf{a}^T \mathbf{Ax})}}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{a} \cdot \mathbf{Ax})}{\partial{\mathbf{x}}}\\[7mm] &= \mathbf{a}^T \mathbf{A} \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} + \mathbf{x}^T \mathbf{A}^T \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{AI} + \mathbf{x}^T \mathbf{A}^T \mathbf{0} \\[7mm] &= \mathbf{a}^T \mathbf{A} \end{aligned}



      Denominator layout:

      ∂(aTAx)∂x=∂(a⋅Ax)∂x=∂a∂xAx+∂x∂xATa=0Ax+IATa=ATa\begin{aligned} \frac{\partial{(\mathbf{a}^T \mathbf{Ax})}}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{a} \cdot \mathbf{Ax})}{\partial{\mathbf{x}}}\\[7mm] &= \frac{\partial{\mathbf{a}}}{\partial{\mathbf{x}}} \mathbf{Ax} + \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{A}^T \mathbf{a} \\[7mm] &= \mathbf{0} \mathbf{Ax} + \mathbf{I} \mathbf{A}^T \mathbf{a} \\[7mm] &= \mathbf{A}^T \mathbf{a} \end{aligned}

Back to table



  • Proof 2-6
    • ∂(xTAx)∂x\color{royalblue}{\LARGE \frac{\partial({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}}}},   A\; \mathbf{A} is not a function of x\mathbf{x}

      A=(a11⋯a1n⋮⋱⋮an1⋯ann)n×n,x=(x1⋯xn)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{n1} & \cdots & a_{nn} \end{pmatrix}_{n \times n} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂(xTAx)∂x=∂(x⋅Ax)∂x=xTA∂x∂x+xTAT∂x∂x=xTAI+xTATI=xT(A+AT)\begin{aligned} \frac{\partial({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{x} \cdot \mathbf{Ax})}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{x}^T \mathbf{A} \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} + \mathbf{x}^T \mathbf{A}^T \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{x}^T \mathbf{A} \mathbf{I} + \mathbf{x}^T \mathbf{A}^T \mathbf{I} \\[7mm] &= \mathbf{x}^T (\mathbf{A} + \mathbf{A}^T) \end{aligned}



      Denominator layout:

      ∂(xTAx)∂x=∂(x⋅Ax)∂x=∂x∂xAx+∂x∂xATx=IAx+IATx=(A+AT)x\begin{aligned} \frac{\partial({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{x} \cdot \mathbf{Ax})}{\partial{\mathbf{x}}} \\[7mm] &= \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{A} \mathbf{x} + \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{A}^T \mathbf{x} \\[7mm] &= \mathbf{IAx} + \mathbf{I} \mathbf{A}^T \mathbf{x} \\[7mm] &= (\mathbf{A} + \mathbf{A}^T) \mathbf{x} \end{aligned}

Back to table



  • Proof 2-7
    • ∂2(xTAx)∂x∂xT\color{royalblue}{\LARGE \frac{\partial^2({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}\partial{\mathbf{x}^T}}}},   A\; \mathbf{A} is not a function of x\mathbf{x}

      A=(a11⋯a1n⋮⋱⋮an1⋯ann)n×n,x=(x1⋯xn)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{n1} & \cdots & a_{nn} \end{pmatrix}_{n \times n} , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂2(xTAx)∂x∂xT=∂∂xT{xT(A+AT)}={∂∂x[xT(A+AT)]}T={(A+AT)T}T=A+AT\begin{aligned} \frac{\partial^2({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}\partial{\mathbf{x}^T}}} &= \frac{\partial}{\partial{\mathbf{x}^T}} \{\mathbf{x}^T (\mathbf{A} + \mathbf{A}^T)\} \\[7mm] &= \left\{\frac{\partial}{\partial{\mathbf{x}}} [\mathbf{x}^T (\mathbf{A} + \mathbf{A}^T)] \right\}^T \\[7mm] &= \left\{(\mathbf{A} + \mathbf{A}^T)^T\right\}^T \\[7mm] &= \mathbf{A} + \mathbf{A}^T \end{aligned}



      Denominator layout:

      ∂2(xTAx)∂x∂xT=∂∂xT{xT(A+AT)}={∂∂x[xT(A+AT)]}T=(A+AT)T=A+AT\begin{aligned} \frac{\partial^2({\mathbf{x}^T\mathbf{Ax}})}{\partial{\mathbf{x}\partial{\mathbf{x}^T}}} &= \frac{\partial}{\partial{\mathbf{x}^T}} \{\mathbf{x}^T (\mathbf{A} + \mathbf{A}^T)\} \\[7mm] &= \left\{\frac{\partial}{\partial{\mathbf{x}}} [\mathbf{x}^T (\mathbf{A} + \mathbf{A}^T)] \right\}^T \\[7mm] &= (\mathbf{A} + \mathbf{A}^T)^T \\[7mm] &= \mathbf{A} + \mathbf{A}^T \end{aligned}

Back to table



  • Proof 2-8
    • ∂∥x∥2∂x\color{royalblue}{\LARGE \frac{\partial{\Vert \mathbf{x} \Vert^2}}{\partial{\mathbf{x}}}}

      x=(x1⋯xn)n×1T\mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂∥x∥2∂x=∂(x⋅x)∂x=xT∂x∂x+xT∂x∂x=xTI+xTI=2xT\begin{aligned} \frac{\partial{\Vert \mathbf{x} \Vert^2}}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{x} \cdot \mathbf{x})}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{x}^T \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} + \mathbf{x}^T \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{x}^T \mathbf{I} + \mathbf{x}^T \mathbf{I} \\[7mm] &= 2\mathbf{x}^T \end{aligned}



      Denominator layout:

      ∂∥x∥2∂x=∂(x⋅x)∂x=∂x∂xx+∂x∂xx=Ix+Ix=2x\begin{aligned} \frac{\partial{\Vert \mathbf{x} \Vert^2}}{\partial{\mathbf{x}}} &= \frac{\partial(\mathbf{x} \cdot \mathbf{x})}{\partial{\mathbf{x}}} \\[7mm] &= \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{x} + \frac{\partial{\mathbf{x}}}{\partial{\mathbf{x}}} \mathbf{x} \\[7mm] &= \mathbf{I} \mathbf{x} + \mathbf{I} \mathbf{x} \\[7mm] &= 2\mathbf{x} \end{aligned}

Back to table



  • Proof 2-9
    • ∂(aTxxTb)∂x\color{royalblue}{\LARGE \frac{\partial({\mathbf{a}^T \mathbf{xx}^T \mathbf{b}})}{\partial{\mathbf{x}}}},   a,b\; \mathbf{a}, \mathbf{b} are not functions of x\mathbf{x}

      a=(a1⋯an)n×1T,b=(b1⋯bn)n×1T,x=(x1⋯xn)n×1T\mathbf{a} = \begin{pmatrix} a_1 & \cdots & a_n \end{pmatrix}_{n \times 1}^T , \qquad \mathbf{b} = \begin{pmatrix} b_1 & \cdots & b_n \end{pmatrix}_{n \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂(aTxxTb)∂x=∂[(aTx)(xTb)]∂x=aTx∂(xTb)∂x+xTb∂(aTx)∂x=aTxbT+xTbaT=xTabT+xTbaT=xT(abT+baT)\begin{aligned} \frac{\partial(\mathbf{a}^T \mathbf{xx}^T \mathbf{b})}{\partial{\mathbf{x}}} &= \frac{\partial[(\mathbf{a}^T \mathbf{x}) (\mathbf{x}^T \mathbf{b})]}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{x} \frac{\partial(\mathbf{x}^T \mathbf{b})}{\partial{\mathbf{x}}} + \mathbf{x}^T \mathbf{b} \frac{\partial(\mathbf{a}^T \mathbf{x})}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{x} \mathbf{b}^T + \mathbf{x}^T \mathbf{b} \mathbf{a}^T \\[7mm] &= \mathbf{x}^T \mathbf{a} \mathbf{b}^T + \mathbf{x}^T \mathbf{b} \mathbf{a}^T \\[7mm] &= \mathbf{x}^T (\mathbf{a} \mathbf{b}^T + \mathbf{b} \mathbf{a}^T) \end{aligned}



      Denominator layout:

      ∂(aTxxTb)∂x=∂[(aTx)(xTb)]∂x=aTx∂(xTb)∂x+xTb∂(aTx)∂x=aTxb+xTba=aTxb+bTxa=baTx+abTx=(abT+baT)x\begin{aligned} \frac{\partial(\mathbf{a}^T \mathbf{xx}^T \mathbf{b})}{\partial{\mathbf{x}}} &= \frac{\partial[(\mathbf{a}^T \mathbf{x}) (\mathbf{x}^T \mathbf{b})]}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{x} \frac{\partial(\mathbf{x}^T \mathbf{b})}{\partial{\mathbf{x}}} + \mathbf{x}^T \mathbf{b} \frac{\partial(\mathbf{a}^T \mathbf{x})}{\partial{\mathbf{x}}} \\[7mm] &= \mathbf{a}^T \mathbf{x} \mathbf{b} + \mathbf{x}^T \mathbf{b} \mathbf{a} \\[7mm] &= \mathbf{a}^T \mathbf{x} \mathbf{b} + \mathbf{b}^T \mathbf{x} \mathbf{a} \\[7mm] &= \mathbf{ba}^T \mathbf{x} + \mathbf{ab}^T \mathbf{x} \\[7mm] &= (\mathbf{ab}^T + \mathbf{ba}^T) \mathbf{x} \end{aligned}

Back to table



  • Proof 2-10
    • ∂∥x−a∥∂x\color{royalblue}{\LARGE \frac{\partial{\Vert \mathbf{x - a} \Vert}}{\partial{\mathbf{x}}}},   a\; \mathbf{a} is not a function of x\mathbf{x}

      a=(a1⋯an)n×1T,x=(x1⋯xn)n×1T\mathbf{a} = \begin{pmatrix} a_1 & \cdots & a_n \end{pmatrix}_{n \times 1}^T , \qquad \mathbf{x} = \begin{pmatrix} x_1 & \cdots & x_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂∥x−a∥∂x=∂∑i=1n(xi−ai)2∂x=(∂∑i=1n(xi−ai)2∂x1⋯∂∑i=1n(xi−ai)2∂xn)1×n=(2(x1−a1)2∑i=1n(xi−ai)2⋯2(xn−an)2∑i=1n(xi−ai)2)1×n=1∥x−a∥(x1−a1⋮xn−an)1×nT=1∥x−a∥((x1⋮xn)1×nT−(a1⋮an)1×nT)=(x−a)T∥x−a∥\begin{aligned} \frac{\partial{\Vert \mathbf{x - a} \Vert}}{\partial{\mathbf{x}}} &= \large \frac{\partial \sqrt{\sum_{i=1}^n (x_i - a_i)^2}}{\partial{\mathbf{x}}} \\[7mm] &= \Large \begin{pmatrix} \frac{\partial \sqrt{\sum_{i=1}^n (x_i - a_i)^2}}{\partial{x_1}} & \cdots & \frac{\partial \sqrt{\sum_{i=1}^n (x_i - a_i)^2}}{\partial{x_n}} \end{pmatrix}_{1 \times n} \\[7mm] &= \Large \begin{pmatrix} \frac{2(x_1 - a_1)}{2\sqrt{\sum_{i=1}^n (x_i - a_i)^2}} & \cdots & \frac{2(x_n - a_n)}{2\sqrt{\sum_{i=1}^n (x_i - a_i)^2}} \end{pmatrix}_{1 \times n} \\[7mm] &= \frac{1}{\Vert \mathbf{x - a} \Vert} \begin{pmatrix} x_1 - a_1 \\ \vdots \\ x_n - a_n \end{pmatrix}_{1 \times n}^T \\[7mm] &= \frac{1}{\Vert \mathbf{x - a} \Vert} ( \begin{pmatrix} x_1 \\ \vdots \\ x_n \end{pmatrix}_{1 \times n}^T - \begin{pmatrix} a_1 \\ \vdots \\ a_n \end{pmatrix}_{1 \times n}^T ) \\[7mm] &= \frac{(\mathbf{x} - \mathbf{a})^T}{\Vert \mathbf{x - a} \Vert} \end{aligned}



      优化:

      ∂∥x−a∥∂x=∂(x−a)⋅(x−a)∂x=12(x−a)⋅(x−a)∂(x−a)⋅(x−a)∂x=12(x−a)⋅(x−a){(x−a)T∂(x−a)∂x+(x−a)T∂(x−a)∂x}=12(x−a)⋅(x−a){(x−a)TI+(x−a)TI}=2(x−a)T2(x−a)⋅(x−a)=(x−a)T∥x−a∥\begin{aligned} \frac{\partial{\Vert \mathbf{x - a} \Vert}}{\partial{\mathbf{x}}} &= \frac{\partial\sqrt{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}}{\partial{\mathbf{x}}} \\[7mm] &= \frac{1}{2 \sqrt{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}} \frac{\partial{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}}{\partial{\mathbf{x}}} \\[7mm] &= \frac{1}{2 \sqrt{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}} \left\{(\mathbf{x} - \mathbf{a})^T \frac{\partial(\mathbf{x} - \mathbf{a})}{\partial{\mathbf{x}}} + (\mathbf{x} - \mathbf{a})^T \frac{\partial(\mathbf{x} - \mathbf{a})}{\partial{\mathbf{x}}} \right\} \\[7mm] &= \frac{1}{2 \sqrt{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}} \left\{(\mathbf{x} - \mathbf{a})^T \mathbf{I} + (\mathbf{x} - \mathbf{a})^T \mathbf{I} \right\} \\[7mm] &= \frac{2 (\mathbf{x} - \mathbf{a})^T}{2 \sqrt{(\mathbf{x} - \mathbf{a}) \cdot (\mathbf{x} - \mathbf{a})}} \\[7mm] &= \frac{(\mathbf{x} - \mathbf{a})^T}{\Vert \mathbf{x} - \mathbf{a} \Vert} \end{aligned}



      Denominator layout:

      ∂∥x−a∥∂x={(x−a)T∥x−a∥}T=x−a∥x−a∥\begin{aligned} \frac{\partial{\Vert \mathbf{x - a} \Vert}}{\partial{\mathbf{x}}} &= \left\{ \frac{(\mathbf{x} - \mathbf{a})^T}{\Vert \mathbf{x - a} \Vert} \right\}^T \\[7mm] &= \frac{\mathbf{x} - \mathbf{a}}{\Vert \mathbf{x - a} \Vert} \end{aligned}

Back to table

向量对标量 (Vector-by-scalar)



  • Proof 3-1
    • ∂Au∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{Au}}}{\partial{x}}},   A\; \mathbf{A} is not a function of x\large x

      A=(a11⋯a1n⋮⋱⋮am1⋯amn)m×n,u=(u1⋯un)n×1T\mathbf{A} = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix}_{m \times n} , \qquad \mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂Au∂x=∂∂x(∑i=1na1iui⋮∑i=1namiui)m×1=(∂∂x∑i=1na1iui⋮∂∂x∑i=1namiui)m×1=(∑i=1na1i∂ui∂x⋮∑i=1nami∂ui∂x)m×1=(a11⋯a1n⋮⋱⋮am1⋯amn)m×n(∂u1∂x⋮∂un∂x)n×1=A∂u∂x\begin{aligned} \frac{\partial{\mathbf{Au}}}{\partial{x}} &= \large \frac{\partial}{\partial{x}} \begin{pmatrix} \sum_{i=1}^n a_{1i} u_i \\ \vdots \\ \sum_{i=1}^n a_{mi} u_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial}{\partial{x}} \sum_{i=1}^n a_{1i} u_i \\ \vdots \\ \frac{\partial}{\partial{x}} \sum_{i=1}^n a_{mi} u_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n a_{1i} \frac{\partial{u_i}}{\partial{x}}\\ \vdots \\ \sum_{i=1}^n a_{mi} \frac{\partial{u_i}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix}_{m \times n} \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{n \times 1} \\[7mm] &= \mathbf{A} \frac{\partial{\mathbf{u}}}{\partial{x}} \end{aligned}



      Denominator layout:

      ∂Au∂x=∂∂x(∑i=1na1iui⋮∑i=1namiui)m×1=(∂∂x∑i=1na1iui⋮∂∂x∑i=1namiui)m×1=(∑i=1na1i∂ui∂x⋮∑i=1nami∂ui∂x)m×1=(∂u1∂x⋯∂un∂x)1×n(a11⋯am1⋮⋱⋮a1n⋯amn)n×m=∂u∂xAT\begin{aligned} \frac{\partial{\mathbf{Au}}}{\partial{x}} &= \large \frac{\partial}{\partial{x}} \begin{pmatrix} \sum_{i=1}^n a_{1i} u_i \\ \vdots \\ \sum_{i=1}^n a_{mi} u_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial}{\partial{x}} \sum_{i=1}^n a_{1i} u_i \\ \vdots \\ \frac{\partial}{\partial{x}} \sum_{i=1}^n a_{mi} u_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n a_{1i} \frac{\partial{u_i}}{\partial{x}}\\ \vdots \\ \sum_{i=1}^n a_{mi} \frac{\partial{u_i}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{1 \times n} \normalsize \begin{pmatrix} a_{11} & \cdots & a_{m1} \\ \vdots & \ddots & \vdots \\ a_{1n} & \cdots & a_{mn} \end{pmatrix}_{n \times m} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{x}} \mathbf{A}^T \end{aligned}

Back to table



  • Proof 3-2
    • ∂(uTv)∂x\color{royalblue}{\LARGE \frac{\partial(\mathbf{u}^T \mathbf{v})}{\partial{x}}},   u=u(x),v=v(x)\; \mathbf{u} = \mathbf{u}(x), \mathbf{v} = \mathbf{v}(x)

      u=(u1⋯un)n×1T,v=(v1⋯vn)n×1T\mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_n \end{pmatrix}_{n \times 1}^T , \qquad \mathbf{v} = \begin{pmatrix} v_1 & \cdots & v_n \end{pmatrix}_{n \times 1}^T
      ∂u∂xnumerator=(∂u1∂x⋮∂un∂x)n×1,∂u∂xdenominator=(∂u1∂x⋯∂un∂x)1×n\underset{\bf{numerator}}{\frac{\partial{\mathbf{u}}}{\partial{x}}} = \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{n \times 1} , \qquad \normalsize \underset{\bf{denominator}}{\frac{\partial{\mathbf{u}}}{\partial{x}}} = \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{1 \times n}



      Numerator layout:

      ∂(uTv)∂x=∂∂x∑i=1nuivi=∑i=1n∂∂xuivi=∑i=1n(∂ui∂xvi+ui∂vi∂x)=∑i=1n∂ui∂xvi+∑i=1nui∂vi∂x=(∂u1∂x⋯∂un∂x)1×n(v1⋮vn)n×1+(u1⋯un)1×n(∂v1∂x⋮∂vn∂x)n×1=(∂u∂x)Tv+uT∂v∂x\begin{aligned} \frac{\partial(\mathbf{u}^T \mathbf{v})}{\partial{x}} &= \frac{\partial}{\partial{x}} \sum_{i=1}^n u_i v_i \\[7mm] &= \sum_{i=1}^n \frac{\partial}{\partial{x}} u_i v_i \\[7mm] &= \sum_{i=1}^n \left(\frac{\partial{u_i}}{\partial{x}} v_i + u_i \frac{\partial{v_i}}{\partial{x}}\right) \\[7mm] &= \sum_{i=1}^n \frac{\partial{u_i}}{\partial{x}} v_i + \sum_{i=1}^n u_i \frac{\partial{v_i}}{\partial{x}} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{1 \times n} \normalsize \begin{pmatrix} v_1 \\ \vdots \\ v_n \end{pmatrix}_{n \times 1} + \begin{pmatrix} u_1 & \cdots & u_n \end{pmatrix}_{1 \times n} \large \begin{pmatrix} \frac{\partial{v_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{v_n}}{\partial{x}} \end{pmatrix}_{n \times 1} \\[7mm] &= \left(\frac{\partial{\mathbf{u}}}{\partial{x}}\right)^T \mathbf{v} + \mathbf{u}^T \frac{\partial{\mathbf{v}}}{\partial{x}} \end{aligned}



      Denominator layout:

      ∂(uTv)∂x=(∂u1∂x⋯∂un∂x)1×n(v1⋮vn)n×1+(u1⋯un)1×n(∂v1∂x⋮∂vn∂x)n×1=∂u∂xv+uT(∂v∂x)T\begin{aligned} \frac{\partial(\mathbf{u}^T \mathbf{v})}{\partial{x}} &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{1 \times n} \normalsize \begin{pmatrix} v_1 \\ \vdots \\ v_n \end{pmatrix}_{n \times 1} + \begin{pmatrix} u_1 & \cdots & u_n \end{pmatrix}_{1 \times n} \large \begin{pmatrix} \frac{\partial{v_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{v_n}}{\partial{x}} \end{pmatrix}_{n \times 1} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{x}} \mathbf{v} + \mathbf{u}^T \left(\frac{\partial{\mathbf{v}}}{\partial{x}}\right)^T \end{aligned}

Back to table



  • Proof 3-3
    • ∂g(u)∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{x}}},   u=u(x)\; \mathbf{u} = \mathbf{u}(x)

      g=(g1⋯gm)m×1T,u=(u1⋯un)n×1T\mathbf{g} = \begin{pmatrix} g_1 & \cdots & g_m \end{pmatrix}_{m \times 1}^T , \qquad \mathbf{u} = \begin{pmatrix} u_1 & \cdots & u_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂g(u)∂x=(∂g1∂x⋮∂gm∂x)m×1=(∂g1∂u∂u∂x⋮∂gm∂u∂u∂x)m×1=((∂g1∂u1⋯∂g1∂un)(∂u1∂x⋮∂un∂x)⋮(∂gm∂u1⋯∂gm∂un)(∂u1∂x⋮∂un∂x))m×1=(∑i=1n∂g1∂ui∂ui∂x⋮∑i=1n∂gm∂ui∂ui∂x)m×1=(∂g1∂u1⋯∂g1∂un⋮⋱⋮∂gm∂u1⋯∂gm∂un)m×n(∂u1∂x⋮∂un∂x)n×1=∂g(u)∂u∂u∂x\begin{aligned} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{x}} &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{x}}\\ \vdots \\ \frac{\partial{g_m}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}} \\ \vdots \\ \frac{\partial{g_m}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_1}}{\partial{u_n}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{u_n}}{\partial{x}} \end{pmatrix} \\ \vdots \\ \begin{pmatrix} \frac{\partial{g_m}}{\partial{u_1}} & \cdots & \frac{\partial{g_m}}{\partial{u_n}} \end{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{u_n}}{\partial{x}} \end{pmatrix} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n \frac{\partial{g_1}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x}} \\ \vdots \\ \sum_{i=1}^n \frac{\partial{g_m}}{\partial{u_i}} \frac{\partial{u_i}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_1}}{\partial{u_n}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_m}}{\partial{u_1}} & \cdots & \frac{\partial{g_m}}{\partial{u_n}} \end{pmatrix}_{m \times n} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{n \times 1} \\[7mm] &= \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}} \end{aligned}



      Denominator layout:

      ∂g∂x=(∂g1∂x⋯∂gm∂x)1×m=(∂u∂x∂g1∂u⋯∂u∂x∂gm∂u)1×m=((∂u1∂x⋯∂un∂x)(∂g1∂u1⋮∂g1∂un)⋯(∂u1∂x⋯∂un∂x)(∂gm∂u1⋮∂gm∂un))1×m=(∑i=1n∂ui∂x∂g1∂ui⋯∑i=1n∂ui∂x∂gm∂ui)1×m=(∂u1∂x⋯∂un∂x)1×n(∂g1∂u1⋯∂gm∂u1⋮⋱⋮∂g1∂un⋯∂gm∂un)n×m=∂u∂x∂g(u)∂u\begin{aligned} \frac{\partial{\mathbf{g}}}{\partial{x}} &= \large \begin{pmatrix} \frac{\partial{g_1}}{\partial{x}} & \cdots & \frac{\partial{g_m}}{\partial{x}} \end{pmatrix}_{1 \times m} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{g_1}}{\partial{\mathbf{u}}} & \cdots & \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{g_m}}{\partial{\mathbf{u}}} \end{pmatrix}_{1 \times m} \\[7mm] &= \large \begin{pmatrix} \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_1}}{\partial{u_n}} \end{pmatrix} & \cdots & \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix} \begin{pmatrix} \frac{\partial{g_m}}{\partial{u_1}} \\ \vdots \\ \frac{\partial{g_m}}{\partial{u_n}} \end{pmatrix} \end{pmatrix}_{1 \times m} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n \frac{\partial{u_i}}{\partial{x}} \frac{\partial{g_1}}{\partial{u_i}} & \cdots & \sum_{i=1}^n \frac{\partial{u_i}}{\partial{x}} \frac{\partial{g_m}}{\partial{u_i}} \end{pmatrix}_{1 \times m} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{u_1}}{\partial{x}} & \cdots & \frac{\partial{u_n}}{\partial{x}} \end{pmatrix}_{1 \times n} \begin{pmatrix} \frac{\partial{g_1}}{\partial{u_1}} & \cdots & \frac{\partial{g_m}}{\partial{u_1}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{g_1}}{\partial{u_n}} & \cdots & \frac{\partial{g_m}}{\partial{u_n}} \end{pmatrix}_{n \times m} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \end{aligned}

Back to table



  • Proof 3-4
    • ∂f(g(u))∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{x}}},   u=u(x)\; \mathbf{u} = \mathbf{u}(x)



      Numerator:

      ∂f(g(u))∂x=∂f(g(u))∂g(u)∂g(u)∂x=∂f(g)∂g∂g(u)∂u∂u∂x\begin{aligned} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{x}} &= \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{g}(\mathbf{u})}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{x}} \\[7mm] &= \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{u}}}{\partial{x}} \end{aligned}



      Denominator layout:

      ∂f(g(u))∂x=∂g(u)∂x∂f(g(u))∂g(u)=∂u∂x∂g(u)∂u∂f(g)∂g\begin{aligned} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{x}} &= \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{x}} \frac{\partial{\mathbf{f}(\mathbf{g}(\mathbf{u}))}}{\partial{\mathbf{g}(\mathbf{u})}} \\[7mm] &= \frac{\partial{\mathbf{u}}}{\partial{x}} \frac{\partial{\mathbf{g}(\mathbf{u})}}{\partial{\mathbf{u}}} \frac{\partial{\mathbf{f}(\mathbf{g})}}{\partial{\mathbf{g}}} \end{aligned}

Back to table



  • Proof 3-5
    • ∂Uv∂x\color{royalblue}{\LARGE \frac{\partial{\mathbf{U}\mathbf{v}}}{\partial{x}}},   U=U(x),  v=v(x)\; \mathbf{U} = \mathbf{U}(x), \; \mathbf{v} = \mathbf{v}(x)

      U=(u11⋯u1n⋮⋱⋮um1⋯umn)m×n,v=(v1⋯vn)n×1T\mathbf{U} = \begin{pmatrix} u_{11} & \cdots & u_{1n} \\ \vdots & \ddots & \vdots \\ u_{m1} & \cdots & u_{mn} \end{pmatrix}_{m \times n} , \qquad \mathbf{v} = \begin{pmatrix} v_1 & \cdots & v_n \end{pmatrix}_{n \times 1}^T



      Numerator layout:

      ∂Uv∂x=∂∂x{(u11⋯u1n⋮⋱⋮um1⋯umn)m×n(v1⋮vn)n×1}=∂∂x(∑i=1nu1ivi⋮∑i=1numivi)m×1=(∑i=1n∂∂xu1ivi⋮∑i=1n∂∂xumivi)m×1=(∑i=1n(∂u1i∂xvi+u1i∂vi∂x)⋮∑i=1n(∂umi∂xvi+umi∂vi∂x))m×1=(∑i=1n∂u1i∂xvi⋮∑i=1n∂umi∂xvi)m×1+(∑i=1nu1i∂vi∂x⋮∑i=1numi∂vi∂x)m×1=(∂u11∂x⋯∂u1n∂x⋮⋱⋮∂um1∂x⋱∂umn∂x)m×n(v1⋮vn)n×1+(u11⋯u1n⋮⋱⋮um1⋯umn)m×n(∂v1∂x⋮∂vn∂x)n×1=∂U∂xv+U∂v∂x\begin{aligned} \frac{\partial{\mathbf{U}\mathbf{v}}}{\partial{x}} &= \frac{\partial}{\partial{x}} \left\{ \begin{pmatrix} u_{11} & \cdots & u_{1n} \\ \vdots & \ddots & \vdots \\ u_{m1} & \cdots & u_{mn} \end{pmatrix}_{m \times n} \begin{pmatrix} v_1 \\ \vdots \\ v_n \end{pmatrix}_{n \times 1} \right\} \\[7mm] &= \frac{\partial}{\partial{x}} \begin{pmatrix} \sum_{i=1}^n u_{1i} v_i \\ \vdots \\ \sum_{i=1}^n u_{mi} v_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n \frac{\partial}{\partial{x}} \normalsize u_{1i} v_i \\ \vdots \\ \sum_{i=1}^n \frac{\partial}{\partial{x}} \normalsize u_{mi} v_i \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n (\frac{\partial{u_{1i}}}{\partial{x}} \normalsize v_i + u_{1i} \large \frac{\partial{v_i}}{\partial{x}}) \\ \vdots \\ \sum_{i=1}^n (\frac{\partial{u_{mi}}}{\partial{x}} \normalsize v_i + u_{mi} \large \frac{\partial{v_i}}{\partial{x}}) \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \sum_{i=1}^n \frac{\partial{u_{1i}}}{\partial{x}} \normalsize v_i \\ \vdots \\ \sum_{i=1}^n \frac{\partial{u_{mi}}}{\partial{x}} \normalsize v_i \end{pmatrix}_{m \times 1} \normalsize + \large \begin{pmatrix} \sum_{i=1}^n \normalsize u_{1i} \large \frac{\partial{v_i}}{\partial{x}} \\ \vdots \\ \sum_{i=1}^n \normalsize u_{mi} \large \frac{\partial{v_i}}{\partial{x}} \end{pmatrix}_{m \times 1} \\[7mm] &= \large \begin{pmatrix} \frac{\partial{u_{11}}}{\partial{x}} & \cdots & \frac{\partial{u_{1n}}}{\partial{x}} \\ \vdots & \ddots & \vdots \\ \frac{\partial{u_{m1}}}{\partial{x}} & \ddots & \frac{\partial{u_{mn}}}{\partial{x}} \end{pmatrix}_{m \times n} \begin{pmatrix} v_1 \\ \vdots \\ v_n \end{pmatrix}_{n \times 1} + \normalsize \begin{pmatrix} u_{11} & \cdots & u_{1n} \\ \vdots & \ddots & \vdots \\ u_{m1} & \cdots & u_{mn} \end{pmatrix}_{m \times n} \large \begin{pmatrix} \frac{\partial{v_1}}{\partial{x}} \\ \vdots \\ \frac{\partial{v_n}}{\partial{x}} \end{pmatrix}_{n \times 1} \\[7mm] &= \frac{\partial{\mathbf{U}}}{\partial{x}} \mathbf{v} + \mathbf{U} \frac{\partial{\mathbf{v}}}{\partial{x}} \end{aligned}



      Denominator layout:

      ∂Uv∂x=(∑i=1n∂∂xu1ivi⋯∑i=1n∂∂xumivi)1×m={(∂U∂x)numeratorv+U(∂v∂x)numerator}T=vT(∂U∂x)denominator+(∂v∂x)denominatorUT=vT∂U∂x+∂v∂xUT\begin{aligned} \frac{\partial{\mathbf{U}\mathbf{v}}}{\partial{x}} &= \large \begin{pmatrix} \sum_{i=1}^n \frac{\partial}{\partial{x}} \normalsize u_{1i} v_i & \cdots & \sum_{i=1}^n \frac{\partial}{\partial{x}} \normalsize u_{mi} v_i \end{pmatrix}_{1 \times m} \\[7mm] &= \left\{\underset{\bf{numerator}}{\left(\frac{\partial{\mathbf{U}}}{\partial{x}}\right)} \mathbf{v} + \mathbf{U} \underset{\bf{numerator}}{\left(\frac{\partial{\mathbf{v}}}{\partial{x}}\right)}\right\}^T \\[7mm] &= \mathbf{v}^T \underset{\bf{denominator}}{\left(\frac{\partial{\mathbf{U}}}{\partial{x}}\right)} + \underset{\bf{denominator}}{\left(\frac{\partial{\mathbf{v}}}{\partial{x}}\right)} \mathbf{U}^T \\[7mm] &= \mathbf{v}^T \frac{\partial{\mathbf{U}}}{\partial{x}} + \frac{\partial{\mathbf{v}}}{\partial{x}} \mathbf{U}^T \end{aligned}

Back to table

参考

评论COMMENTS

◂ 返回索引