741 words
4 minutes
工科数学分析-3.23

test

5.4.3 有约束的极值LagrangeLagrange乘数法#

1.一个约束条件#

二元函数z=f(x,y)z=f(x,y),目标函数,g(x,y)=0g(x,y)=0约束条件
设方程g(x,y)=0g(x,y)=0确定的隐函数y=y(x),gy(P0)≠0,P0(x0,y0)y=y(x),g_y(P_0)\neq 0,P_0(x_0,y_0)
若极值点在点P0(x0,y0)P_{0}(x_{0},y_{0})取到,则z=f(x,y(x))z=f(x,y(x))在x=x0x=x_{0}取到极值,于是
dzdx=fx+fydydx=0\frac{dz}{dx}=f_{x}+f_{y}\frac{dy}{dx}=0
ddxg(x,y(x))=gx+gydydx=0\frac{d}{dx}g(x,y(x))=g_{x}+g_{y}\frac{dy}{dx}=0
∇f,∇g⊥(1,dydx=T⃗),∇f//∇g\nabla f,\nabla g \perp(1,\frac{dy}{dx}=\vec{T}),\nabla f//\nabla g
因此,存在常数λ0\lambda_{0},使得∇f(P0)=λ0∇g(P0),g(p0)=0(∗2)\nabla f(P_{0})=\lambda_{0}\nabla g(P{0}),g(p_{0})=0(*2)
引入变量λ\lambda与函数L(x,y,λ)=f(x,y)−λg(x,y)(∗3)L(x,y,\lambda)=f(x,y)-\lambda g(x,y)(*3)
因为

Lx=fx−λgxLy=fy−λgyLλ=−g\begin{align} L_{x}&=f_{x}-\lambda g_{x}\\ L_{y}&=f_{y}-\lambda g_{y}\\ L_{\lambda}&=-g \end{align}

所以(∗2)(*2)等价于∇L(P0)=0(∗4)\nabla L(P_0)=0(*4)
若P0(x0,y0)P_{0}(x_{0},y_{0})为条件极值,则(x0,y0,λ0)(x_{0},y_{0},\lambda_{0})为L(x,y,λ)L(x,y,\lambda)的驻点,该方法称为拉格朗日乘数法
若约束集{(x,y)∣g(x,y)=0}\{(x,y)|g(x,y)=0\}是有限闭集,则条件极值存在
设gx,gyg_{x},g_{y}在DD上不全为零,求出拉格朗日函数(∗3)(*3)的全部驻点以及相应的函数值,再取最大或最小

ex.4 设x2+xy+y2=1x^{2}+xy+y^{2}=1,求2x+3y2x+3y的最大值#

解
令L(x,y,λ)=2x+3y−λ(x2+xy+y2−1)L(x,y,\lambda)=2x+3y-\lambda(x^{2}+xy+y^{2}-1),由∇L=0\nabla L=0,得

Lx=2−λ(2x+y)=0Ly=3−λ(x+2y)=0Lλ=−(x2+xy+y2−1)=0\begin{align} L_{x}&=2-\lambda(2x+y)=0\\ L_{y}&=3-\lambda(x+2y)=0\\ L_{\lambda}&=-(x^{2}+xy+y^{2}-1)=0 \end{align}

x=±121,y=±421,2x+3y=±1421x=\frac{\pm 1}{\sqrt{21}},y=\frac{\pm 4}{\sqrt{21}},2x+3y=\pm\frac{14}{\sqrt{21}}
三元函数,u=f(x,y,z),g(x,y,z)=0u=f(x,y,z),g(x,y,z)=0
若极值在点P0(x0,y0,z0)P_0(x_{0},y_{0},z_{0})处取到,设gz≠0g_{z}\neq0,确定z=z(x,y),z0=z(x0,y0)z=z(x,y),z_{0}=z(x_{0},y_{0})
gx+gzzx=0,∇g⊥(1,0,zx)=v1⃗g_{x}+g_{z}z_{x}=0,\nabla g\perp (1,0,z_{x})=\vec{v_{1}}
gy+gyzy=0,∇g⊥(0,1,zy)=v2⃗g_{y}+g_{y}z_{y}=0,\nabla g\perp(0,1,z_{y})=\vec{v_2}
∇g//v1⃗×v2⃗=(−zx,−zy,1)=v3⃗\nabla g//\vec{v_1}\times\vec{v_2}=(-z_{x},-z_{y},1)=\vec{v_3}
因为u=f(x,y,z(x,y))u=f(x,y,z(x,y))在Q0(x0,y0)Q_{0}(x_{0},y_{0})取极值,所以

ux=fx+fzzx=0,∇f⊥v1⃗uy=fy+fzzy=0,∇f⊥v2⃗\begin{align} u_{x}&=f_x+f_zz_x=0,\nabla f\perp \vec{v_1}\\ u_{y}&=f_y+f_zz_y=0,\nabla f\perp \vec{v_2}\\ \end{align}

说明∇f(p0)//∇g(P0)\nabla f(p_{0})//\nabla g(P_{0})
令L(x,y,z,λ)=f(x,y,z)−λg(x,y,z)(6)L(x,y,z,\lambda)=f(x,y,z)-\lambda g(x,y,z)(6)
则条件极值=>,∇L=0=>,\nabla L=0即∇f−λ∇g=0,g=0\nabla f-\lambda\nabla g=0,g=0

例4.8 设x,y,z>0,xyz=v,u=xy+2yz+2yzx,y,z>0,xyz=v,u=xy+2yz+2yz,求uminu_{min}#

解:
令L(x,y,z,λ)=xy+2yz+2xz−λ(xyz−v)L(x,y,z,\lambda)=xy+2yz+2xz-\lambda(xyz-v),由∇L=0\nabla L=0

Lx=y+2z−λ(yz)=0Ly=x+2z−λ(xz)=0Lz=2y+2x−λ(xy)=0Lλ=−(xyz−v)=0\begin{align} L_{x}&=y+2z-\lambda(yz)=0\\ L_{y}&=x+2z-\lambda(xz)=0\\ L_{z}&=2y+2x-\lambda(xy)=0\\ L_{\lambda}&=-(xyz-v)=0 \end{align}

z=v43,u=12z2=12(v4)23z=\sqrt[3]{\frac{v}{4}},u=12z^{2}=12(\frac{v}{4})^{\frac{2}{3}}

ex x,y,z>0x,y,z>0,证明xyz≤(x+y+z3)3xyz\le(\frac{x+y+z}{3})^{3}#

证明:
u=xyz,x+y+z=a>0u=xyz,x+y+z=a>0
令L(x,y,z,λ)=xyz−λ(x+y+z−a)L(x,y,z,\lambda)=xyz-\lambda(x+y+z-a),由∇L=0\nabla L=0

Lx=yz−λ=0Ly=zx−λ=0Lz=xy−λ=0Lλ=−(x+y+z−a)=0\begin{align} L_{x}&=yz-\lambda =0\\ L_{y}&=zx-\lambda=0\\ L_{z}&=xy-\lambda=0\\ L_{\lambda}&=-(x+y+z-a)=0 \end{align}

x=y=z=a3,u=(a3)3x=y=z=\frac{a}{3},u=(\frac{a}{3})^{3}
约束集S={(x,y,z)∣x+y+z=0,x,y,z≥0}S=\{(x,y,z)|x+y+z=0,x,y,z\ge 0\}
为△ABC\triangle ABC,它是有界闭域u=xyzu=xyz在SS上连续,可取到最大值与最小值,在∂S\partial S上,u=0u=0,
于是在大致在内部取到,因此最大值为(a3)3(\frac{a}{3})^3

2.两个约束条件#

u=f(x,y,z),g(z,y,z)=0,h(x,y,z)=0(8)u=f(x,y,z),g(z,y,z)=0,h(x,y,z)=0(8)
设方程组g(x,y,z)=0,h(x,y,z)=0(9)g(x,y,z)=0,h(x,y,z)=0(9)
确定隐函数组x=x(z),y=y(z)(10)x=x(z),y=y(z)(10)
若u=f(x(z),y(z),z)u=f(x(z),y(z),z)取极值,则

dudz=fxdxdz+fydydz+fz=0=(fx,fy,fz)⋅(dxdz,dydz,1),∇f⊥T⃗\begin{align} \frac{du}{dz}&=f_x\frac{dx}{dz}+f_{y}\frac{dy}{dz}+f_{z}=0\\ &=(f_x,f_y,f_z)\cdot(\frac{dx}{dz},\frac{dy}{dz},1), \nabla f\perp \vec{T} \end{align}

将x=x(z),y=y(z)x=x(z),y=y(z)带入结束条件

gxdxdz+gydydz+gz=0,∇g⊥T⃗g_x\frac{dx}{dz}+g_{y}\frac{dy}{dz}+g_{z}=0,\nabla g\perp \vec{T}

同理

hxdxdz+hydydz+hz=0,∇h⊥T⃗h_x\frac{dx}{dz}+h_{y}\frac{dy}{dz}+h_{z}=0,\nabla h\perp \vec{T}

故存在λ,μ\lambda,\mu,使得
∇f+λ∇g+μ∇h=0\nabla f+\lambda\nabla g+\mu\nabla h=0
令L(x,y,z,λ,μ)=f(x,y,z)+λg(x,y,z)+μh(x,y,z)L(x,y,z,\lambda,\mu)=f(x,y,z)+\lambda g(x,y,z)+\mu h(x,y,z)

ex.5 求u=x2+y2+z2u=x^{2}+y^{2}+z^{2}在约束条件x24+y25+z225=1,z=x+y\frac{x^{2}}{4}+\frac{y^{2}}{5}+\frac{z^2}{25}=1,z=x+y 的极值#

解:
令L(x,y,z,λ,μ)=x2+y2+z2−λ(x24+y25+z225−1)−μ(x+y−z)L(x,y,z,\lambda,\mu)=x^{2}+y^{2}+z^{2}-\lambda(\frac{x^{2}}{4}+\frac{y^{2}}{5}+\frac{z^2}{25}-1)-\mu(x+y-z)
由∇L=0\nabla L=0

Lx=2x−12λx−μ=0(1)Ly=2y−25λy−μ=0(2)Lz=2z−225λz+μ=0(3)Lλ=−(x24+y25+z225−1)=0(4)Lμ=−(x+y−z)=0(5)\begin{align} L_{x}&=2x-\frac{1}{2}\lambda x-\mu=0&(1)\\ L_{y}&=2y-\frac{2}{5}\lambda y-\mu=0&(2)\\ L_{z}&=2z-\frac{2}{25}\lambda z+\mu=0&(3)\\ L_{\lambda}&=-(\frac{x^{2}}{4}+\frac{y^{2}}{5}+\frac{z^2}{25}-1)=0&(4)\\ L_{\mu}&=-(x+y-z)=0&(5)\\ \end{align}

(1)×x+(2)×y+(3)×z(1)\times x+(2)\times y+(3)\times z:
2(x2+y2+z2)−λ(x22+2y25+2z225)−μ(x+y−z02(x^{2}+y^{2}+z^{2})-\lambda(\frac{x^{2}}{2}+\frac{2y^{2}}{5}+\frac{2z^2}{25})-\mu(x+y-z0
结合(4)(5)(4)(5),x2+y2+z2=λx^{2}+y^2+z^2=\lambda
由(1)(2)(3)(1)(2)(3)
x=μ2−12λ,y=μ2−25λ,z=μ2−225λx=\frac{\mu}{2-\frac{1}{2}\lambda},y=\frac{\mu}{2-\frac{2}{5}\lambda},z=\frac{\mu}{2-\frac{2}{25}\lambda}
带入(5)(5),则λ=10,7517\lambda=10,\frac{75}{17}

Method 2#

把x,y,z,μx,y,z,\mu当做未知数,λ\lambda为系数,线性方程组有解

∣2−12λ00−102−25λ0−1002−225λ111−10∣=(2−12λ)(2−25λ+2−225λ)−(−1)(2−25λ)(2−225λ)=12−9825λ+34125λ2\begin{align} \begin{vmatrix} &2-\frac{1}{2}\lambda &0 &0 &-1 \\ &0 &2-\frac{2}{5}\lambda &0 &-1 \\ &0 &0 &2-\frac{2}{25}\lambda &1 \\ &1 &1 &-1 &0 \\ \end{vmatrix} &=(2-\frac{1}{2}\lambda)(2-\frac{2}{5}\lambda+2-\frac{2}{25}\lambda)-(-1)(2-\frac{2}{5}\lambda)(2-\frac{2}{25}\lambda)\\&=12-\frac{98}{25}\lambda+\frac{34}{125}\lambda^2 \end{align}

3.一般形式#

u=f(X)X=(x1,x2,...,xn)∈Rnφk(X)=01≤k≤m<n\begin{align} u&=f(X) &X=(x_1,x_2,...,x_n)\in R^n\\ \varphi_k(X)&=0 &1\le k\le m<n\\ \end{align}

L(X,λ)=f(X)+λ1φ1(X)+...+λmφm(X),λ=(λ1,...,λm)∈RmL(X,\lambda)=f(X)+\lambda_1\varphi_1(X)+...+\lambda_m\varphi_m(X),\lambda=(\lambda_1,...,\lambda_m)\in R^m
由∇L=0\nabla L=0,求出全部驻点(X0,λ0)(X_0,\lambda_0)
X0=(x0,1,x0,2,...,x0,n),λ0=(λ0,1,...,λ0,m)X_0=(x_{0,1},x_{0,2},...,x_{0,n}),\lambda_0=(\lambda_{0,1},...,\lambda_{0,m})
并要求rank(∂φk∂xi)m×n=mrank(\frac{\partial\varphi_k}{\partial x_i})_{m\times n}=m

∂φ1∂x1dx1+...+∂φ1∂xmdxm+∂φ1∂xm+1dxm+1+...+∂φ1∂xndxn=0...∂φm∂x1dx1+...+∂φm∂xmdxm+∂φm∂xm+1dxm+1+...+∂φm∂xndxn=0\begin{align} \frac{\partial \varphi_{1}}{\partial x_{1}}d_{x_{1}}+...+\frac{\partial \varphi_{1}}{\partial x_{m}}d_{x_{m}}+\frac{\partial \varphi_{1}}{\partial x_{m+1}}d_{x_{m+1}}+...+\frac{\partial \varphi_{1}}{\partial x_{n}}d_{x_{n}}=0\\ ...\\ \frac{\partial \varphi_{m}}{\partial x_{1}}d_{x_{1}}+...+\frac{\partial \varphi_{m}}{\partial x_{m}}d_{x_{m}}+\frac{\partial \varphi_{m}}{\partial x_{m+1}}d_{x_{m+1}}+...+\frac{\partial \varphi_{m}}{\partial x_{n}}d_{x_{n}}=0\\ \end{align}

设X‾=(xm+1,...,xn)\overline{X}=(x_{m+1},...,x_{n})且优越数据确定的隐函数组
x1=x1(X‾),...,xm=xm(X‾)x_{1}=x_{1}(\overline{X}),...,x_{m}=x_{m}(\overline {X})

条件极值的充分条件#

对LL的每个驻点(X0,λ0)(X_{0},\lambda_{0})
计算Φ(X)=L(X,λ)\Phi(X)=L(X,\lambda)在点X0X_0的HesseHesse矩阵,

H(X0)=(∂2L∂xi∂yj)n×nH(X_0)=\begin{pmatrix} \frac{\partial^2 L}{\partial x_i\partial y_j} \end{pmatrix}_{n\times n}

若H(X0)H(X_0)正(负)定在X0X_{0}为条件极小(大)值点
证明:
注意到
∇Φ(X)=∇f(X)+λ0,1∇φ1(X)+...+λ0,m∇φm(X),∇Φ(X0)=0\nabla\Phi(X)=\nabla f(X)+\lambda_{0,1}\nabla\varphi_{1}(X)+...+\lambda_{0,m}\nabla\varphi_{m}(X),\nabla\Phi(X_{0})=0
定义约束集S={X∈Rn∣φk(X)=0,1≤k≤m}S=\{X\in R^n|\varphi_{k}(X)=0,1\le k\le m\}
在SS上f(X)=Φ(X)f(X)=\Phi(X),设X,X+h∈SX,X+h\in S,则

f(X+h)−f(X)=Φ(X+h)−Φ(X)=∇Φ(X)h+12hTH(X)h+o(∣∣h∣∣2)=12hTH(X)h+o(∣∣h∣∣2)\begin{align} f(X+h)-f(X)&=\Phi(X+h)-\Phi(X)\\ &=\nabla \Phi(X)h+\frac{1}{2}h^{T}H(X)h+o(||h||^2)\\ &=\frac{1}{2}h^{T}H(X)h+o(||h||^2) \end{align}

当H(X0)H(X_0)正定时,a=min∣∣h∣∣=1hTH(X0)h>0a=min_{||h||=1}h^{T}H(X_0)h>0
此时hTH(X0)h≥a∣∣h∣∣2h^TH(X_0)h\ge a||h||^2,于是
f(X0+h)−f(X0)>∣∣h∣∣2(fraca2+o(1))≥0f(X_{0}+h)-f(X_{0})>||h||^2(\\frac{a}{2}+o(1))\ge 0

ex.6 g(x,y,z)=2x2+y2+z2+2xy−2x−2y−4z+4=0g(x,y,z)=2x^{2}+y^2+z^2+2xy-2x-2y-4z+4=0确定的隐函数z=z(x,y)z=z(x,y)的极值#

解:

u=f(x,y,z)=z,g(x,y,z)=0L(x,y,z,λ)=z+λg(x,y,z),∇L=0Lx=λ(4x+2y−2)=0Ly=λ(2y+2x−2)=0Lz=1+λ(2z−4)=0\begin{align} u&=f(x,y,z)=z,g(x,y,z)=0\\ L(x,y,z,\lambda)&=z+\lambda g(x,y,z),\nabla L=0\\ L_x&=\lambda(4x+2y-2)=0\\ L_y&=\lambda(2y+2x-2)=0\\ L_z&=1+\lambda(2z-4)=0\\ \end{align}

x=0,y=0x=0,y=0则
z1=1,λ1=12z_1=1,\lambda_1=\frac{1}{2}或z2=3,λ2=−12z_2=3,\lambda_2=-\frac{1}{2}
Lxx=4λL_{xx}=4\lambda
Lyy=2λL_{yy}=2\lambda
Lzz=2λL_{zz}=2\lambda
Lxy=2λL_{xy}=2\lambda
Lxz=0L_{xz}=0
Lyz=0L_{yz}=0

H1=(210220001)H_{1}=\begin{pmatrix} 2&1&0\\ 2&2&0\\ 0&0&1 \end{pmatrix}

正定,z1(0,1)=1z_1(0,1)=1为极小值

H2=(−2−10−2−2000−1)H_{2}=\begin{pmatrix} -2&-1&0\\ -2&-2&0\\ 0&0&-1 \end{pmatrix}

负定,z2(0,1)=3z_2(0,1)=3为极大值

蜂房问题#

容积相同,一个六棱柱用怎样的三个全等菱形为底,表面积最小