u=f(x)u = f(\bm x) 在区域 DD 上有定义,x0D\bm x_0 \in Dv=(cosθ1,cosθ2,,cosθn)\bm v = (\cos \theta_1, \cos \theta_2, \dots, \cos \theta_n) 为一方向,如果极限 limt0+0f(x0+tv)f(x0)t\lim_{t \to 0 + 0} \dfrac{f(\bm x_0 + t \bm v) - f(\bm x_0)}{t} 存在,则称为 f(x)f(\bm x)x0\bm x_0 处沿 v\bm v方向导数,记为 f(x0)v\dfrac{\partial f(\bm x_0)}{\partial \bm v}

一个二元函数在 (x0,y0)(x_0, y_0) 处的所有方向导数都存在也未必连续,例如 f(x,y)=[y=x2,x0]f(x, y) = [y = x^2, x \ne 0](0,0)(0, 0) 处。

Δx=(Δx1,Δx2,,Δxn)\Delta \bm x = (\Delta x_1, \Delta x_2, \cdots, \Delta x_n),并称它为自变量的全增量。若存在仅依赖于 x0\bm x_0 的常数 AiA_i,使得 Δf(x0)=f(x0+Δx)f(x0)=i=1nAiΔxi+o(Δx)\Delta f(\bm x_0) = f(\bm x_0 + \Delta \bm x) - f(\bm x_0) = \sum_{i=1}^{n} A_i \Delta x_i + o(|\Delta \bm x|),则称 f(x)f(\bm x)x0\bm x_0 处可微,并称 i=1nAiΔxi\sum_{i=1}^{n} A_i \Delta x_i全微分,记为 df(x0)\mathrm{d} f(\bm x_0)

可微一定连续,且必然有 df(x0)=i=1nf(x0)xidxi{\rm d}f(\bm x_0) = \sum_{i=1}^{n} \dfrac{\partial f(\bm x_0)}{\partial x_i} {\rm d} x_i

定理:如果 f(x)f(\bm x)U(x0,δ0)U(\bm x_0, \delta_0) 存在各个偏导数,并且这些偏导数在 x0\bm x_0 处连续,即 f(x)C1(D)f(\bm x) \in C^1(D),则 f(x)f(\bm x)x0\bm x_0 处可微。

nn 归纳。n=kn=k 成立,当 n=k+1n=k+1 时,由一元函数拉格朗日中值定理,f(x0+Δx)f(x10+Δx1,,xk0+Δxk,xk+10)=f(x0)xk+1Δxk+1+o(1)Δxk+1f(\bm x_0 + \Delta \bm x) - f(x_1^0 + \Delta x_1, \cdots, x_k^0 + \Delta x_k, x_{k+1}^0) = \dfrac{\partial f(\bm x_0)}{\partial x_{k+1}}\Delta x_{k+1} + o(1) |\Delta x_{k+1}|
根据归纳假设,f(x10+Δx1,,xk0+Δxk,xk+10)f(x0)=i=1kf(x0)xiΔxi+o(1)i=1k(Δxi)2f(x_1^0 + \Delta x_1, \cdots, x_k^0 + \Delta x_k, x_{k+1}^0) - f(\bm x_0) = \sum_{i=1}^{k} \dfrac{\partial f(\bm x_0)}{\partial x_i} \Delta x_i + o(1) \sqrt{\sum_{i=1}^{k}(|\Delta x_i|)^2}。相加得到 Δf(x0)=i=1k+1f(x0)xiΔxi+o(1)i=1k+1(Δxi)2\Delta f(\bm x_0) = \sum_{i=1}^{k+1} \dfrac{\partial f(\bm x_0)}{\partial x_i} \Delta x_i + o(1) \sqrt{\sum_{i=1}^{k+1}(|\Delta x_i|)^2}

定理:如果 f(x0)f(\bm x_0) 可微,对于 v=(cosθ1,,cosθn)\bm v = (\cos \theta_1, \cdots, \cos \theta_n),它的方向导数为 f(x0)v=i=1nf(x0)xicosθi\dfrac{\partial f(\bm x_0)}{\partial \bm v} = \sum_{i=1}^{n} \dfrac{\partial f(\bm x_0)}{\partial x_i} \cos \theta_i

从上式看出,可微函数的方向导数最大的方向向量为 1i=1n(f(x0)xi)2(f(x0)x1,f(x0)x2,,f(x0)xn)\dfrac{1}{\sqrt{\sum_{i=1}^{n}(\dfrac{\partial f(\bm x_0)}{\partial x_i})^2}}(\dfrac{\partial f(\bm x_0)}{\partial x_1}, \dfrac{\partial f(\bm x_0)}{\partial x_2}, \cdots, \dfrac{\partial f(\bm x_0)}{\partial x_n}),同时它的模就是方向导数的值。称 (f(x0)x1,f(x0)x2,,f(x0)xn)(\dfrac{\partial f(\bm x_0)}{\partial x_1}, \dfrac{\partial f(\bm x_0)}{\partial x_2}, \cdots, \dfrac{\partial f(\bm x_0)}{\partial x_n})x0\bm x_0 处的梯度,记为 gradf(x0){\bf grad} f(\bm x_0)。所以 f(x0)v=gradf(x0)v\dfrac{\partial f(\bm x_0)}{\partial \bm v} = {\bf grad} f(\bm x_0) \cdot \bm v

后文所有向量默认为列向量

设向量函数 f(x)=(f1(x),f2(x),,fm(x))T\bm f(\bm x) = (f_1(\bm x), f_2(\bm x), \cdots, f_m(\bm x))^T 在区域 DD 上有定义,Δx=(Δx1,Δx2,,Δxn)T\Delta \bm x = (\Delta x_1, \Delta x_2, \cdots, \Delta x_n)^Tx\bm xx0\bm x_0 处的全增量,如果存在只与 x0\bm x_0 有关的 m×nm \times n 矩阵 A\bm A 使得 Δx0|\Delta \bm x| \to 0 时,Δf(x0)=(Δf1(x0),Δf2(x0),,Δfm(x0))T=AΔx+α(Δx)\Delta \bm f(\bm x_0) = (\Delta f_1(\bm x_0), \Delta f_2(\bm x_0), \cdots, \Delta f_m(\bm x_0))^T = \bm A \Delta \bm x + \bm \alpha(|\Delta \bm x|),其中 α(Δx)=(α1(Δx),α2(Δx),,αm(Δx))T\bm \alpha(|\Delta \bm x|) = (\alpha_1(|\Delta \bm x|), \alpha_2(|\Delta \bm x|), \cdots, \alpha_m(|\Delta \bm x|))^Tαj\alpha_j 依赖 Δx\Delta \bm xlimΔx0αj(Δx)Δx=0\lim_{|\Delta \bm x|\to 0} \dfrac{\alpha_j(|\Delta \bm x|)}{|\Delta \bm x|} = 0,则称 f(x)\bm f(\bm x)x0\bm x_0可导/可微A\bm A 称为 Frechet 导数,记作 f(x0)\bm f'(\bm x_0) 或者 Df(x0)\text D\bm f(\bm x_0)AΔx\bm A \Delta \bm x 称为全微分,记作 df(x0)\text d \bm f(\bm x_0),即 df(x0)=Df(x0)Δx\text d \bm f(\bm x_0) = \text D \bm f(\bm x_0) \Delta \bm x

定理:向量函数可微的充要条件是它包含的所有函数可微,此时 f(x0)=(f1(x0)x1f1(x0)x2f1(x0)xnf2(x0)x1f2(x0)x2f2(x0)xnfm(x0)x1fm(x0)x2fm(x0)xn)\bm f' (\bm x_0) = \begin{pmatrix} \dfrac{\partial f_1(\bm x_0)}{\partial x_1} & \dfrac{\partial f_1(\bm x_0)}{\partial x_2} & \cdots & \dfrac{\partial f_1(\bm x_0)}{\partial x_n} \\ \dfrac{\partial f_2(\bm x_0)}{\partial x_1} & \dfrac{\partial f_2(\bm x_0)}{\partial x_2} & \cdots & \dfrac{\partial f_2(\bm x_0)}{\partial x_n} \\ \vdots\\ \dfrac{\partial f_m(\bm x_0)}{\partial x_1} & \dfrac{\partial f_m(\bm x_0)}{\partial x_2} & \cdots & \dfrac{\partial f_m(\bm x_0)}{\partial x_n} \end{pmatrix} 称为雅可比 Jacobi 矩阵,记为 Jf(x0)\bm J_{\bm f}(\bm x_0),特别的,如果是方阵,那么它的行列式称为雅可比行列式,记作 (f1,f2,,fn)(x1,x2,,xn)x0\dfrac{\partial(f_1, f_2,\cdots, f_n)}{\partial(x_1, x_2, \cdots, x_n)}\Big| _{\bm x_0}

如果 fj(x)f_j(\bm x) 的各个偏导数在区域 DD 上连续,我们称 f(x)\bm f(\bm x)DD 上是 C1C^1 的,记作 f(x)C1(D)\bm f(\bm x) \in C^1(D)。特别的,如果区域 DDΩ\it \Omega 的变换 f(x)C1(D)\bm f(\bm x) \in C^1(D)f1C1(Ω)\bm f^{-1} \in C^1(\it \Omega),则说这个变换是 C1C^1 的。

如果 f(x):RnRf(\bm x) : \R^n \to \Rg(x):RnRm\bm g(\bm x) : \R^n \to \R^m 列向量函数,则 (f(x)g(x))=f(x)g(x)+g(x)f(x)(f(\bm x) \bm g(\bm x))' = f(\bm x) \bm g'(\bm x) + \bm g(\bm x) f'(\bm x)

(f(x)g(x))(i;j)=(f(x)gi(x))xj=f(x)gi(x)xj+gi(x)f(x)xj=f(x)g(x)(i;j)+g(x)f(x)(i;j)(f(\bm x) \bm g(\bm x))'(i; j) = \dfrac{\partial (f(\bm x) g_i(\bm x))}{\partial x_j} = f(\bm x) \dfrac{\partial g_i(\bm x)} {\partial x_j} + g_i(\bm x) \dfrac{\partial f(\bm x)} {\partial x_j} = f(\bm x) \bm g'(\bm x)(i; j) + \bm g(\bm x) f'(\bm x)(i; j)

f(u)f(\bm u)u0\bm u_0 处可微,u0=u(x0)\bm u_0 = \bm u(\bm x_0)x0\bm x_0 处可微,则 f(u(x))f(\bm u(\bm x))x0\bm x_0 处可微,且 f(u(x0))=f(u0)u(x0)f'(\bm u(\bm x_0)) = f'(\bm u_0) \bm u'(\bm x_0)

Δf(u(x))=f(u0)Δu(x)+β(Δu(x))=f(u0)(u(x0)Δx+α(Δx))+β(Δu(x))\Delta f(\bm u(\bm x)) = f'(\bm u_0) \Delta \bm u(\bm x) + \beta(|\Delta \bm u(\bm x)|) = f'(\bm u_0) (\bm u'(\bm x_0) \Delta \bm x + \bm \alpha(|\Delta \bm x|)) + \beta(|\Delta \bm u(\bm x)|)
下面说明 limΔx0f(u0)α(Δx)+β(u(x0)Δx+α(Δx))Δx=0\lim_{|\Delta \bm x \to \bm 0|} \dfrac{f'(\bm u_0) \bm \alpha(|\Delta \bm x|) + \beta(|\bm u'(\bm x_0) \Delta \bm x + \bm \alpha(|\Delta \bm x|)|)}{|\Delta \bm x|} = 0 即可。前一半是显然的,后一半由于 u(x0)Δx+α(Δx)u(x0)Δx+ΔxΔx(u(x0)+1)0|\bm u'(\bm x_0) \Delta \bm x + \bm \alpha(|\Delta \bm x|)| \le ||\bm u'(\bm x_0)|| | \Delta \bm x| + |\Delta \bm x| \le |\Delta \bm x|(||\bm u'(\bm x_0)|| + 1) \to 0,故同样也是 00

推论f,u\bm f, \bm u 均可微,df(u(x))=f(u(x))u(x)dx\text d \bm f(\bm u(\bm x)) = \bm f'(\bm u(\bm x)) \bm u'(\bm x) \text d \bm x

推论y=f(x)\bm y = \bm f(\bm x)DDΩ\it \OmegaC1C^1 变换,则 (f1)(y)=[f(x)]1(\bm f^{-1})'(\bm y) = [\bm f'(\bm x)]^{-1},从而 (y1,y2,,yn)(x1,x2,,xn)(x1,x2,,xn)(y1,y2,,yn)=1\dfrac{\partial(y_1, y_2,\cdots, y_n)}{\partial(x_1, x_2, \cdots, x_n)} \cdot \dfrac{\partial(x_1, x_2, \cdots, x_n)}{\partial(y_1, y_2,\cdots, y_n)} = 1

推论(链锁法则)f(u)f(\bm u)u0\bm u_0 处可微,u(x)=(u1(x),u2(x),,um(x))T\bm u(\bm x) = (u_1(\bm x), u_2(\bm x), \cdots, u_m(\bm x))^Tx0\bm x_0 处可微(可以减弱为存在各个偏导数),u0=u(x0)\bm u_0 = \bm u(\bm x_0),则 f(u(x0))xi=j=1m(f(u0)ujuj(x0)xi)\dfrac{\partial f(\bm u(\bm x_0))}{\partial x_i} = \sum_{j=1}^{m} (\dfrac{\partial f(\bm u_0)}{\partial u_j} \cdot \dfrac{\partial u_j(\bm x_0)}{\partial x_i})

(f(x)xi)xk\dfrac{\partial(\frac{\partial f(\bm x)}{\partial x_i})}{\partial x_k} 存在时,可以将其记为 2f(x)xkxi\dfrac{\partial^2 f(\bm x)}{\partial x_k \partial x_i}fxkxi(x)f''_{x_k x_i}(\bm x)

定理:对于 j,kj, k,如果 fjk(x)f_{jk}''(\bm x)fkj(x)f_{kj}''(\bm x)U(x0,δ)U(\bm x_0, \delta) 内存在,且在 x0\bm x_0 处连续,则 fjk(x0)=fjk(x0)f_{jk}''(\bm x_0) = f_{jk}''(\bm x_0)

考虑 I=f(x0+Δx,y0+Δy)f(x0+Δx,y0)f(x0,y0+Δy)+f(x0,y0)ΔxΔyI = \dfrac{f(x_0 + \Delta x, y_0 + \Delta y) - f(x_0 + \Delta x, y_0) - f(x_0, y_0 + \Delta y) + f(x_0, y_0)}{\Delta x\Delta y},一方面,I=fx(x0+θΔx,y0+Δy)fx(x0+θΔx,y0)Δy=fyx(x0+θΔx,y0+μΔy)I = \dfrac{f'_x(x_0 + \theta\Delta x, y_0 + \Delta y) - f'_x(x_0 + \theta \Delta x, y_0)}{\Delta y} = f'_{yx}(x_0 + \theta\Delta x, y_0 + \mu \Delta y),另一方面可以写作 fxy(x0+θΔx,y0+μΔy)f'_{xy}(x_0 + \theta'\Delta x, y_0 + \mu' \Delta y),由连续性得到相等。

对于高阶微分可以形式化的记 dkf(x)=(i=1ndxixi)kf(x)\text d^k f(\bm x) = (\sum_{i=1}^{n} \text d x_i \dfrac{\partial}{\partial x_i})^k f(\bm x)

特别的,对于二元函数 f(x,y)f(x, y)dkf(x,y)=j=0kCkjkf(x,y)xkjyjdxkjdyj\text d^k f(x, y) = \sum_{j=0}^k C_k^j \dfrac{\partial^k f(x, y)}{\partial x^{k-j} \partial y^j} \text d x^{k-j} \text d y^j

定理(拉格朗日余项泰勒公式):设函数 f(x)f(\bm x)U(x0,δ0)U(\bm x_0, \delta_0) 内具有连续 K+1K+1 阶偏导数,则 x0+hU(x0,δ0)\forall \bm x_0 + \bm h \in U(\bm x_0, \delta_0),存在 0<θ<10 < \theta < 1f(x0+h)=f(x0)+k=1K1k!(i=1nhixi)kf(x0)+1(K+1)!(i=1nhixi)K+1f(x0+θh)f(\bm x_0 + \bm h) = f(\bm x_0) + \sum_{k=1}^{K} \dfrac{1}{k!} (\sum_{i=1}^{n} h_i \dfrac{\partial}{\partial x_i})^k f(\bm x_0) + \dfrac{1}{(K+1)!}(\sum_{i=1}^{n} h_i \dfrac{\partial}{\partial x_i})^{K+1} f(\bm x_0 + \theta \bm h)

构造 φ(t)=f(x0+th)\varphi(t) =f(\bm x_0 + t \bm h),则 φ(t)\varphi(t)K+1K+1 阶连续导数,使用一元函数泰勒公式即可证明。

推论(皮亚诺余项泰勒公式)f(x)CK(U(x0,δ0))f(\bm x) \in C^K(U(\bm x_0, \delta_0)),则 f(x0+h)=f(x0)+k=1K1k!(i=1nhixi)kf(x0)+o(hK)f(\bm x_0 + \bm h) = f(\bm x_0) + \sum_{k=1}^{K} \dfrac{1}{k!} (\sum_{i=1}^{n} h_i \dfrac{\partial}{\partial x_i})^k f(\bm x_0) + o(|\bm h|^K)

推论(拉格朗日微分中值定理)f(x)C1(D)f(\bm x) \in C^1(D),则 t[0,1]\forall t \in [0, 1]x0+t(xx0)D\bm x_0 + t(\bm x - \bm x_0) \in D,有 f(x)f(x0)=i=1nf(x0+θ(xx0))xi(xixi0)=f(x0+θ(xx0))(xx0)f(\bm x) - f(\bm x_0) = \sum_{i=1}^{n} \dfrac{\partial f(\bm x_0 + \theta(\bm x - \bm x_0))}{\partial x_i}(x_i - x_i^0) = f'(\bm x_0 + \theta (\bm x - \bm x_0)) (\bm x - \bm x_0)

海色 Hessi 矩阵Hf(x0)=(2f(x0)x122f(x0)x1x22f(x0)x1xn2f(x0)x2x12f(x0)x222f(x0)x2xn2f(x0)xnx12f(x0)xnx22f(x0)xn2)\bm H_f(\bm x_0) = \begin{pmatrix} \dfrac{\partial^2 f(\bm x_0)}{\partial x_1^2} & \dfrac{\partial^2 f(\bm x_0)}{\partial x_1 \partial x_2} & \cdots & \dfrac{\partial^2 f(\bm x_0)}{\partial x_1 \partial x_n}\\ \dfrac{\partial^2 f(\bm x_0)}{\partial x_2 \partial x_1} & \dfrac{\partial^2 f(\bm x_0)}{\partial x_2^2} & \cdots & \dfrac{\partial^2 f(\bm x_0)}{\partial x_2 \partial x_n} \\ \vdots \\ \dfrac{\partial^2 f(\bm x_0)}{\partial x_n \partial x_1} & \dfrac{\partial^2 f(\bm x_0)}{\partial x_n \partial x_2} & \cdots & \dfrac{\partial^2 f(\bm x_0)}{\partial x_n^2} \end{pmatrix}

在 Taylor 展式中令 K=1K = 1,则有 f(x0+h)=f(x0)+f(x0)h+12hTHf(x0)h+o(h2)f(\bm x_0 + \bm h) = f(\bm x_0) + f'(\bm x_0) \bm h + \dfrac{1}{2} \bm h^T \bm H_f(\bm x_0) \bm h + o(|\bm h|^2)

隐函数存在定理:设二元函数 F(x,y)F(x, y)U((x0,y0),δ)U((x_0, y_0), \delta) 内满足:

  1. F(x0,y0)=0F(x_0, y_0) = 0
  2. F(x,y),Fy(x,y)F(x, y), F'_y(x, y)U((x0,y0),δ)U((x_0, y_0), \delta) 内连续。
  3. Fy(x0,y0)0F_y'(x_0, y_0) \ne 0

则存在 0<δ0<δ0 < \delta_0 < \delta,使得在 U(x0,δ0)U(x_0, \delta_0) 内存在唯一满足下述条件的连续函数 y=f(x)y = f(x)

不妨 Fy(x0,y0)>0F'_y(x_0, y_0) > 0,则存在 δ0\delta_0(x,y)U((x0,y0),δ0)\forall (x, y) \in U((x_0, y_0), \delta_0),有 Fy(x,y)>0F'_y(x, y) > 0。故 F(x0,y0δ0)<0,F(x0,y0+δ0)>0F(x_0, y_0 - \delta_0) < 0, F(x_0, y_0 + \delta_0) > 0,进而存在 δ1<δ0\delta_1 < \delta_0xU(x0,δ1)\forall x \in U(x_0, \delta_1)F(x,y0δ0)<0F(x, y_0 - \delta_0) < 0F(x,y0+δ0)>0F(x, y_0 + \delta_0) > 0,故由单调性有唯一的 F(x,f(x))=0F(x, f(x)) = 0
至于连续性,则由于 ε>0\forall \varepsilon > 0,对于 y=f(x)\overline y = f(\overline x),有 F(x,yε)<0,F(x,y+ε)>0F(\overline x, \overline y - \varepsilon) < 0, F(\overline x, \overline y + \varepsilon) > 0,故存在 δ\delta 使得 xU(x,δ),F(x,yε)<0,F(x,y+ε)>0\forall x \in U(\overline x, \delta), F(x, \overline y - \varepsilon) < 0, F(x, \overline y + \varepsilon) > 0,从而 f(x)(yε,y+ε)f(x) \in (\overline y - \varepsilon, \overline y + \varepsilon),也即 f(x)f(x)<ε|f(x) - f(\overline x)| < \varepsilon
如果 xx 的偏导数连续,那么设 Δy=f(x+Δx)f(x)\Delta y = f(\overline x + \Delta x) - f(\overline x)。由微分中值定理,0=F(x+Δx,y+Δy)F(x,y)=Fx(x+θΔx,y+θΔy)Δx+Fy(x+θΔx,y+θΔy)Δy0 = F(\overline x + \Delta x, \overline y + \Delta y) - F(\overline x, \overline y) = F'_x(\overline x + \theta \Delta x, \overline y + \theta \Delta y) \Delta x + F'_y(\overline x + \theta \Delta x, \overline y + \theta \Delta y) \Delta y,移项后令 Δx0\Delta x \to 0 由连续性得证。

推广隐函数存在定理:设函数 F(x,y)F(\bm x, y)U((x0,y0),δ)U((\bm x_0, y_0), \delta) 内满足:

  1. F(x0,y0)=0F(\bm x_0, y_0) = 0
  2. F(x,y),Fy(x,y)F(\bm x, y), F'_y(\bm x, y)U((x0,y0),δ)U((\bm x_0, y_0), \delta) 内连续。
  3. Fy(x0,y0)0F_y'(\bm x_0, y_0) \ne 0

则存在 0<δ0<δ0 < \delta_0 < \delta,使得在 U(x0,δ0)U(\bm x_0, \delta_0) 内存在唯一满足下述条件的连续函数 y=f(x)y = f(\bm x)

隐函数组存在定理:设 F(x,u)\bm F(\bm x, \bm u)U((x0,u0),δ)U((\bm x_0, \bm u_0), \delta) 内有定义,且满足

  1. F(x0,u0)=0\bm F(\bm x_0, \bm u_0) = \bm 0
  2. Fj(x,u)F_j(\bm x, \bm u) 及它的各个偏导数在 U((x0,u0),δ)U((\bm x_0, \bm u_0), \delta) 内连续。
  3. (F1,F2,,Fm)(u1,u2,,um)(x0,u0)0\dfrac{\partial (F_1, F_2, \cdots, F_m)}{\partial (u_1, u_2, \cdots, u_m)} \Big|_{(\bm x_0, \bm u_0)} \ne 0

则存在 0<δ0<δ0 < \delta_0 < \delta,使得在 U(x0,δ0)U(\bm x_0, \delta_0) 内存在唯一满足下述条件的 mmnn 元向量函数 f(x)\bm f(\bm x)

逆映射存在定理:设 y=f(x)\bm y = \bm f(\bm x) 是区域 DDΩ\it \Omega 的一个 C1C^1 映射,并且在 x0D\bm x_0 \in D 处有 (f1,f2,,fn)(x1,x2,,xn)x00\dfrac{\partial(f_1, f_2, \cdots, f_n)}{\partial (x_1, x_2, \cdots, x_n)} \Big|_{\bm x_0} \ne 0,记 y0=f(x0)\bm y_0 = f(\bm x_0),则存在 U(x0,δ0)DU(\bm x_0, \delta_0) \sub D,使得 y=f(x)\bm y = \bm f(\bm x)U(x0,δ0)U(\bm x_0, \delta_0)f(U(x0,δ0))\bm f(U(\bm x_0, \delta_0))C1C^1 同胚映射。

Fj(x,y)=yjfj(x)F_j(\bm x, \bm y) = y_j - f_j(\bm x),考虑 F(x,y)=0\bm F(\bm x, \bm y) = \bm 0 设个方程组,各个子函数的各偏导数都连续,并且 (F1,F2,,Fn)(x1,x2,,xn)(x0,y0)=(1)n(f1,f2,,fn)(x1,x2,,xn)x00\dfrac{\partial (F_1, F_2, \cdots, F_n)}{\partial (x_1, x_2, \cdots, x_n)} \Big|_{(\bm x_0, \bm y_0)} = (-1)^n \dfrac{\partial (f_1, f_2, \cdots, f_n)}{\partial (x_1, x_2, \cdots, x_n)} \Big|_{\bm x_0} \ne 0,因此存在隐函数 x=g(y)\bm x = \bm g(\bm y),且有各个连续偏导数。

定理:设 f(x)f(\bm x)x0\bm x_0 取极值且 f(x)f(\bm x) 关于 xix_i 可偏导,则有 f(x0)xi=0\dfrac{\partial f(\bm x_0)}{\partial x_i} = 0,特别的,若可微,则 f(x0)=0f'(\bm x_0) = \bm 0

f(x0)=0f'(\bm x_0) = 0,则称 x0\bm x_0f(x)f(\bm x) 的一个驻点/临界点,如果一个驻点不是极值点,则称为鞍点

定理 f(x)f(\bm x) 有二阶连续偏导数,f(x0)=0f'(\bm x_0) = 0,设 Hf(x0)\bm H_f(\bm x_0) 为满秩矩阵,则

  1. 正定 - 极小值
  2. 负定 - 极大值
  3. 不定 - 不取极值

由于 hTHf(x0)h\bm h^T \bm H_f(\bm x_0) \bm h{h:h=1}\{ \bm h : |\bm h| = 1\} 上的一个连续函数,又由于这是一个紧集,故有最大值 MM 最小值 mm。下面只证明正定,即 m>0m > 0
f(x)=f(x0)+f(x0)(xx0)+12(xx0)THf(x0)(xx0)+o(xx02)>f(x0)+12(xx0)Txx0Hf(x0)xx0xx0xx02+o(xx02)>f(x0)+m4xx02f(\bm x) = f(\bm x_0) + f'(\bm x_0)(\bm x - \bm x_0) + \dfrac{1}{2}(\bm x - \bm x_0)^T \bm H_f(\bm x_0)(\bm x - \bm x_0) + o(|\bm x - \bm x_0|^2) > f(\bm x_0) + \dfrac{1}{2}\dfrac{(\bm x - \bm x_0)^T}{|\bm x - \bm x_0|} \bm H_f(\bm x_0)\dfrac{\bm x - \bm x_0}{|\bm x - \bm x_0|} |\bm x - \bm x_0|^2 + o(|\bm x - \bm x_0|^2) > f(\bm x_0) + \dfrac{m}{4}|\bm x - \bm x_0|^2

定理:设函数 f(x),φ(x)=(φ1(x),,φm(x))Tf(\bm x), \bm \varphi(\bm x) = (\varphi_1(\bm x), \cdots, \varphi_m(\bm x))^TDRnD \sub \R^n 有各个连续偏导数,m<nm < nx0\bm x_0f(x)f(\bm x) 在约束条件 φ(x)=0\bm \varphi(\bm x) = \bm 0 下的极值点,φ(x0)\bm \varphi'(\bm x_0) 的秩为 mm,则存在 λ1,λ2,,λmR\lambda_1, \lambda_2, \cdots, \lambda_m \in \R,使得 x0\bm x_0 处成立下述等式:f(x0)xi+j=1mλjφj(x0)xi=0i=1,2,,n\dfrac{\partial f(\bm x_0)}{\partial x_i} + \sum_{j=1}^{m} \lambda_j \dfrac{\partial \varphi_j(\bm x_0)}{\partial x_i} = 0 \qquad i = 1, 2, \cdots, n

证明 n=2,m=1n = 2, m = 1 的情况。φ(x0,y0)0,φ(x0,y0)=0\varphi'(x_0, y_0) \ne \bm 0, \varphi(x_0, y_0) = 0,因此不妨存在隐函数 y=g(x)y = g(x)(或者反过来),x0x_0f(x,g(x))f(x, g(x)) 的一个通常极值点,因此 f(x0,y0)x+f(x0,y0)yg(x0)=0\dfrac{\partial f(x_0, y_0)}{\partial x} + \dfrac{\partial f(x_0, y_0)}{\partial y} g'(x_0) = 0,结合 g(x0)=φx(x0,y0)φy(x0,y0)g'(x_0) = - \dfrac{\varphi'_x(x_0, y_0)}{\varphi'_y(x_0, y_0)},原命题得证。

这种求极值点的必要条件的方法称为拉格朗日乘数法

设曲线 Γ\it \Gamma 由连续映射 h(t)\bm h(t)t[α,β]t \in [\alpha, \beta] 所确定,且是单射,那么称为简单曲线。如果 h(α)=h(β)\bm h(\alpha) = \bm h(\beta)h\bm h[α,β)[\alpha, \beta) 是单射,那么称为简单闭曲线 / Jordan 曲线

对于三维空间参数方程曲线 x=x(t),y=y(t),z=z(t)x = x(t), y = y(t), z = z(t),曲线在 x(t0)\bm x(t_0) 处的切向量为 x(t0)\bm x'(t_0),法平面用向量内积记为 x(t0)(xx(t0))=0\bm x'(t_0) \cdot (\bm x - \bm x(t_0)) = 0。而对于由 F1(x,y,z)=F2(x,y,z)=0F_1(x, y, z) = F_2(x, y, z) = 0 确定的曲线方程,如果 F(x0,y0,z0)\bm F'(x_0, y_0, z_0) 的秩为 22,则由隐函数存在定理,在邻域内存在参数方程形式,故 F1F_1tt 求偏导得 F1(x0)xx(t0)+F1(x0)yy(t0)+F1(x0)zz(t0)=0\dfrac{\partial F_1(\bm x_0)}{\partial x} x'(t_0) + \dfrac{\partial F_1(\bm x_0)}{\partial y} y'(t_0) + \dfrac{\partial F_1(\bm x_0)}{\partial z} z'(t_0) = 0,对于 F2F_2 同理,因此切向量与 F1(x0)×F2(x0)=Ai+Bj+CkF_1'(\bm x_0) \times F_2'(\bm x_0) = A \bm i + B \bm j + C \bm k 平行,从而切向量为 xx0A=yy0B=zz0C\dfrac{x - x_0}{A} = \dfrac{y - y_0}{B} = \dfrac{z - z_0}{C},而法平面为 A(xx0)+B(yy0)+C(zz0)=0A(x - x_0) + B(y - y_0) + C(z - z_0) = 0

对于三维空间 C1C^1 曲面 F(x,y,z)=0F(x, y, z) = 0,在 (x0,y0,z0)(x_0, y_0, z_0) 处任取一条曲面上的光滑曲线 (x(t),y(t),z(t))(x(t), y(t), z(t)),则 F1(x0)xx(t0)+F1(x0)yy(t0)+F1(x0)zz(t0)=0\dfrac{\partial F_1(\bm x_0)}{\partial x} x'(t_0) + \dfrac{\partial F_1(\bm x_0)}{\partial y} y'(t_0) + \dfrac{\partial F_1(\bm x_0)}{\partial z} z'(t_0) = 0,因此可以推出切面方程为 F1(x0)x(xx0)+F1(x0)y(yy0)+F1(x0)z(zz0)=0\dfrac{\partial F_1(\bm x_0)}{\partial x} (x - x_0) + \dfrac{\partial F_1(\bm x_0)}{\partial y} (y - y_0) + \dfrac{\partial F_1(\bm x_0)}{\partial z} (z - z_0)=0,于是直线 xx0Fx(x0)=yy0Fy(x0)=zz0Fz(x0)\dfrac{x - x_0}{F'_x(\bm x_0)} = \dfrac{y - y_0}{F'_y(\bm x_0)} = \dfrac{z - z_0}{F'_z(\bm x_0)} 为法线方程。反之如果平面由参数方程 x=x(u,v),y=y(u,v),z=z(u,v)x = x(u, v), y = y(u, v), z = z(u, v) 定义,对于平面上的曲线 x=x(u0,v),y=y(u0,v),z=z(u0,v)x = x(u_0, v), y = y(u_0, v), z = z(u_0, v),切向量为 (x0,y0,z0)Tu=(xu(u0,v0),yu(u0,v0),zu(u0,v0))\dfrac{\partial(x_0, y_0, z_0)^T}{\partial u} = (x'_u(u_0, v_0), y'_u(u_0, v_0), z'_u(u_0, v_0)),同理可知,法向量为 n=(x0,y0,z0)Tu×(x0,y0,z0)Tv=Ai+Bj+Ck\bm n = \dfrac{\partial(x_0, y_0, z_0)^T}{\partial u} \times \dfrac{\partial(x_0, y_0, z_0)^T}{\partial v} = A \bm i + B \bm j + C \bm k,立刻得到法线方程和切平面方程。

DD 是一个凸域,f(x)f(\bm x)DD 内有定义,如果 x0,x1D\forall \bm x_0, \bm x_1 \in Dt(0,1)\forall t \in (0, 1),有 f(tx1+(1t)x2)tf(x1)+(1t)f(x2)f(t \bm x_1 + (1 - t) \bm x_2) \le t f(\bm x_1) + (1 - t) f (\bm x_2),则称 ffDD 内是凸函数,如果成立严格不等式则是严格凸函数

定理:如果 f(x)f(\bm x)DD 内有二阶连续偏导数,则 ff 凸与下面两条均等价:

  1. x0,x\forall \bm x_0, \bm x,成立 f(x)f(x0)+f(x0)(xx0)f(\bm x) \ge f(\bm x_0) + f'(\bm x_0)(\bm x - \bm x_0)
  2. x0\forall \bm x_0Hf(x0)\bm H_f(\bm x_0) 半正定。

f(x0+tΔx)=f(x0)+tf(x0)Δx+t22ΔxTHf(x0)Δx+o(tΔx2)()f(\bm x_0 + t\Delta \bm x) = f(\bm x_0) + t f'(\bm x_0)\Delta \bm x + \dfrac{t^2}{2} \Delta \bm x^T \bm H_f(\bm x_0) \Delta\bm x + o(|t\Delta\bm x|^2) \qquad (*)
由凸,f(x0+tΔx)tf(x0+Δx)+(1t)f(x0)f(\bm x_0 + t\Delta \bm x) \le tf(\bm x_0 + \Delta \bm x)+ (1 - t) f(\bm x_0),带入 ()(*) 的一阶形式则 1 得证。反过来,tf(x1)+(1t)f(x2)f(x0)+f(x0)(t(x1x0)+(1t)(x2x0))t f(\bm x_1) + (1 - t) f(\bm x_2) \ge f(\bm x_0) + f'(\bm x_0)(t(\bm x_1 - \bm x_0) + (1 - t)(\bm x_2 - \bm x_0)),令 x0=tx1+(1t)x2\bm x_0 = t \bm x_1 + (1 - t) \bm x_2 则得到凸的定义式。
把 1 和 ()(*) 联立,令 t0t \to 0 立刻得到 2。而 2 结合 ()(*) 自然说明 1。