853 words
4 minutes
工科数学分析-3.19

5.4.2 无约束条件极值与最值#

1.无约束极值#

设函数ff定义在一个点p0p_{0}的某个领域Up0U_{p_{0}}内,若对于任何pUp0p\in U_{p_{0}},有f(p)fp0f(p)\le f_{p_{0}},此时称ffp0p_{0}取到极大值,若不等号反向,则称ffp0p_{0}取到极小值。
f(x,y)f(x,y)在点P(x0,y0)P(x_{0},y_{0})取到极值,则对任何实数hh,考虑两个一元函数φ(t)=f(x0+th,y0),ψ(t)=f(x0,y0+th)\varphi(t)=f(x_{0}+th,y_{0}),\psi(t)=f(x_{0},y_{0}+th)(要求偏导数存在)
均在t=0t=0取极值,有FermatFermat引理
φ(0)=fx(x0,y0)h=0,ψ(0)=fy(x0,y0)h=0\varphi'(0)=f_{x}(x_{0},y_{0})h=0,\psi'(0)=f_{y}(x_{0},y_{0})h=0
由于hh的任意性,fx(x0,y0)=fy(x0,y0)=0f_{x}(x_{0},y_{0})=f_{y}(x_{0},y_{0})=0

定理4.3(极值的必要条件)#

设函数ff在点X0X_{0}的一阶偏导数都存在,若ffX0X_{0}取极值,则f(X0)=0\nabla f(X_{0})=0
Φ(t)=f(x0+tk,y0+th)\Phi(t)=f(x_{0}+tk,y_{0}+th)需要可微
满足f(X0)=0\nabla f(X_{0})=0的点X0X_{0}称为驻点
极值点一定是驻点(函数一阶偏导数存在时),反过来不成立,例如鞍点
TaylorTaylor公式
f(X0+ΔX)f(X0)=f(X0),ΔX+12(ΔX)THf(X0)ΔX+o(ΔX2)f(X_{0}+\Delta X)-f(X_{0})=\langle\nabla f(X_{0}),\Delta X\rangle+\frac{1}{2}(\Delta X)^{T}H_{f}(X_{0})\Delta X+o(||\Delta X||^{2})

X0X_{0}是驻点,则
f(X0+ΔX)f(X0)=12(ΔX)THf(X0)ΔX+o(ΔX2)f(X_{0}+\Delta X)-f(X_{0})=\frac{1}{2}(\Delta X)^{T}H_{f}(X_{0})\Delta X+o(||\Delta X||^{2})

定理4.4(极值充分条件)设f(X)f(X)U(X0)U(X_{0})内二阶偏导数连续,f(X0)=0\nabla f(X_{0})=0Hf(X0)H_{f}(X_{0})ff在点X0X_{0}HesseHesse矩阵(2fXiXj)(X0)n×n(\frac{\partial^{2}f}{\partial X_{i}\partial X_{j}})(X_{0})_{n\times n},#

Hf(X0)H_{f}(X_{0})是正定(负定)矩阵,则f(X0)f(X_{0})为极小(大)值,
因为Hf(X0)H_{f}(X_{0})是实对称矩阵,存在实正交矩阵QQ,使得
QTHf(X0)Q=diag{λ1,...,λn}Q^{T}H_{f}(X_{0})Q=diag\{\lambda_{1},...,\lambda_{n}\}
其中λ1,...,λn\lambda_{1},...,\lambda_{n}Hf(X0)H_{f}(X_{0})的实特征值,令ΔX=Qy\Delta X=Q\cdot y,则

Δz=12yTQTHf(X0)Qy+o(y2),y=(y1,..,yn)T=12yTdiag{λ1,...,λn}y+o(y2)=12(λ1y12+...+λnyn2)+o(y12+...,yn2)\begin{align} \Delta z&=\frac{1}{2}y^{T}Q^{T}H_{f}(X_{0})Qy+o(||y||^{2}),y=(y_{1},..,y_{n})^{T}\\ &=\frac{1}{2}y^{T}diag\{\lambda_{1},...,\lambda_{n}\}y+o(||y||^{2})\\ &=\frac{1}{2}(\lambda_{1}y_{1}^{2}+...+\lambda_{n}y_{n}^{2})+o(y_{1}^{2}+...,y_{n}^{2}) \end{align}

对二元函数f(x,y)f(x,y),记
A=fxx(X0),B=fxy(X0),C=fyy(X0),Hf(X0)=(ABBC)A=f_{xx}(X_{0}),B=f_{xy}(X_{0}),C=f_{yy}(X_{0}),H_{f}(X_{0})=\begin{pmatrix}A&B\\B&C\end{pmatrix}

(1)A>0,ACB2>0,X0>min(2)A<0,ACB2>0,X0>max(3)ACB2<0,X0>nope(4)ACB2=0,X0>uncertain\begin{align} &(1)A>0,AC-B^{2}>0,X_{0}->min\\ &(2)A<0,AC-B^{2}>0,X_{0}->max\\ &(3)AC-B^{2}<0,X_{0}->nope\\ &(4)AC-B^{2}=0,X_{0}->uncertain\\ \end{align}

f(x,y)=x2,(0,0)f(x,y)=x^{2},(0,0)是极小值点,但A=2,B=C=0A=2,B=C=0
g(x,y)=x2+y,(0,0)g(x,y)=x^{2}+y,(0,0)不是极值点,但A=2,B=C=0A=2,B=C=0

例4.2 求f(x,y)=x3+y3+3xyf(x,y)=x^{3}+y^{3}+3xy的极值#

解:
先求驻点
fx=3x2+3y=0f_{x}=3x^{2}+3y=0
fy=3y2+3x=0f_{y}=3y^{2}+3x=0
P1(0,0),P2(1,1)P_{1}(0,0),P_{2}(-1,-1)为两个驻点
再求二阶偏导数
fxx=6xf_{xx}=6x
fxy=3f_{xy}=3
fyy=6yf_{yy}=6y

于是
Hf(P1)=(0330),Hf(P2)=(6336)H_f(P_{1})=\begin{pmatrix}0&3\\3&0\end{pmatrix},H_f(P_{2})=\begin{pmatrix}-6&3\\3&-6\end{pmatrix}
因为Hf(P1)H_{f}(P_{1})不定,Hf(P2)H_f(P_{2})负定,则P1P_{1}不是极值点,P2P_{2}为极大值点,f(P2)=1f(P_{2})=1为极大值

例4.3 f(x,y)=2x23xy2+y4f(x,y)=2x^{2}-3xy^{2}+y^{4}的极值点#

解:
fx=4x3y2=0f_{x}=4x-3y^{2}=0
fy=6xy+4y3=0f_{y}=-6xy+4y^{3}=0
驻点P(0,0)P(0,0)
fxx=4f_{xx}=4
fxy=6yf_{xy}=-6y
fyy=6x+12y2f_{yy}=-6x+12y^{2}
fxx(0,0)=4,fxy(0,0)=0,fyy(0,0)=0f_{xx}(0,0)=4,f_{xy}(0,0)=0,f_{yy}(0,0)=0
此时ACB2=0AC-B^{2}=0
有因为f(x,y)f(0,0)=(2xy2)(xy2)f(x,y)-f(0,0)=(2x-y^{2})(x-y^{2})
x>y2,2x<y2x>y^{2},2x<y^{2}时大于零,12y2<x<y2\frac{1}{2}y^{2}<x<y^{2}时小于零
因此,f(0,0)=0f(0,0)=0不是极值

ex.2 方程2x2+y2+z2+2xy2x2y4z+4=0()2x^{2}+y^{2}+z^{2}+2xy-2x-2y-4z+4=0(*)所确定的隐函数z=Z(x,y)z=Z(x,y)的极值#

(下面的求导是我自己算的,解法在解后面)
zx=4x+2zzx+2y24zx=4x+2y2=0z_{x}=4x+2z\frac{\partial z}{\partial x}+2y-2-4\frac{\partial z}{\partial x}=4x+2y-2=0
zy=2y+2zzy+2x24zy=2y+2x2=0z_{y}=2y+2z\frac{\partial z}{\partial y}+2x-2-4\frac{\partial z}{\partial y}=2y+2x-2=0
zxx=4+2(zx)2+(2z4)2zx2z_{xx}=4+2(\frac{\partial z}{\partial x})^{2}+(2z-4)\frac{\partial^{2}z}{\partial x^{2}}
zxy=2zxzy+242zxyz_{xy}=2\frac{\partial z}{\partial x}\frac{\partial z}{\partial y}+2-4\frac{\partial^{2}z}{\partial x\partial y}
zyy=2+2(zx)2+(2z4)2zx2z_{yy}=2+2(\frac{\partial z}{\partial x})^{2}+(2z-4)\frac{\partial^{2}z}{\partial x^{2}}
解:
()(*)取微分

0=4xdx+2ydy+2zdz+2ydx+2xdy2dx2dy4dz=(4x+2y2)dx+(2y+2x2)dy+(2z4)dz(2)\begin{align} 0&=4xdx+2ydy+2zdz+2ydx+2xdy-2dx-2dy-4dz\\ &=(4x+2y-2)dx+(2y+2x-2)dy+(2z-4)dz&(*2) \end{align}

先求驻点,令dz=0dz=0
(4x+2y2)dx+(2y2x2)dy=0(4x+2y-2)dx+(2y-2x-2)dy=0
由于dx,dydx,dy的任意性
4x+2y2=0,2y2x2=04x+2y-2=0,2y-2x-2=0
x=0,y=1x=0,y=1带入()(*)z24z+3=0z^{2}-4z+3=0
z=1z=1z=3z=3,说明()(*)(0,1,1),(0,1,3)(0,1,1),(0,1,3)附近分别确定一个隐函数z=Z1(x,y),z=Z2(x,y)z=Z_{1}(x,y),z=Z_{2}(x,y)(2)(*2)取微分,得

0=(4dx+2dy)dz+(2dy+2dx)dy+2(dz)2+(2z4)d2z=4(dx)2+4dxdy+2(dy)2+(2zd)d2zdz=0\begin{align} 0&=(4dx+2dy)dz+(2dy+2dx)dy+2(dz)^{2}+(2z-4)d^{2}z\\ &=4(dx)^{2}+4dxdy+2(dy)^{2}+(2z-d)d^{2}z&dz=0 \end{align}


d2z=12z(2(dx)2+2dxdy+(dy)2)d^{2}z=\frac{1}{2-z}(2(dx)^{2}+2dxdy+(dy)^{2})
x=0,y=1,z=1x=0,y=1,z=1时,d2z=2(dx)2+2dxdy+(dy)2d^{2}z=2(dx)^{2}+2dxdy+(dy)^{2}
A=2,B=2,C=1,Hz1(0,1)=1,Z1(0,1)=1A=2,B=2,C=1,H_{z_{1}}(0,1)=1,Z_{1}(0,1)=1为极小值
x=0,y=1,z=3x=0,y=1,z=3时,d2z=2(dx)22dxdy(dy)2d^{2}z=-2(dx)^{2}-2dxdy-(dy)^{2}
A=2,B=2,C=1,Hz2(0,1)=1,Z1(0,1)=3A=-2,B=-2,C=-1,H_{z_{2}}(0,1)=1,Z_{1}(0,1)=3为极大值

2.最大值,最小值 有界闭区域上的连续函数可取到#

最大值,最小值,他们可能在区域内部取到(此时他们是极值),也能在边界上取到
例4.4 f(x,y)=x2+2x2y+y2f(x,y)=x^{2}+2x^{2}y+y^{2}在圆域D={(x,y)x2+y21}D=\{(x,y)|x^{2}+y^{2}\le 1\}上的最值
解:
fx=2x+4xy=0f_{x}=2x+4xy=0
fy=2x2+2y=0f_{y}=2x^{2}+2y=0
驻点P1(0,0),P2(12,12),P3(12,12)P_{1}(0,0),P_{2}(\frac{1}{\sqrt{2}},-\frac{1}{2}),P_{3}(-\frac{1}{\sqrt{2}},-\frac{1}{2}),以及f(P1)=0,f(P2)=f(P3)=14f(P_{1})=0,f(P_{2})=f(P_{3})=\frac{1}{4}
DD的边界上x2+y2=1,f(x,y)=1+2(1y2)y=φ(y),y1x^{2}+y^{2}=1,f(x,y)=1+2(1-y^{2})y=\varphi(y),|y|\le1
φ(y)=26y2=0\varphi'(y)=2-6y^{2}=0y=±13,φ(±13)=1±439y=\pm\frac{1}{\sqrt{3}},\varphi(\pm\frac{1}{\sqrt{3}})=1\pm\frac{4\sqrt{3}}{9}
比较这几个值即可
maxDf=1+439,minDf=0max_{D}f=1+\frac{4\sqrt{3}}{9},min_{D}f=0

ex.3证明:圆的外切三角形中,正三角形的面积最小#

Proof
假设圆的半径为aa,则三角形面积为
S=a2(tan(α2)+tan(β2)+tan(γ2))=a2(tan(α2)+tan(β2)tan(α+β2))S=a^{2}(tan(\frac{\alpha}{2})+tan(\frac{\beta}{2})+tan(\frac{\gamma}{2}))=a^{2}(tan(\frac{\alpha}{2})+tan(\frac{\beta}{2})-tan(\frac{\alpha+\beta}{2}))
α+β+γ=2π,D:0<α,βπ,α+β>π\alpha+\beta+\gamma=2\pi,D:0<\alpha,\beta\pi,\alpha+\beta>\pi
ABC\triangle ABC,不含三边
先求驻点
Sα=a2(sec2(α2)sec2(α+β2))=0S_{\alpha}=a^{2}(sec^{2}(\frac{\alpha}{2})-sec^{2}(\frac{\alpha+\beta}{2}))=0
Sβ=a2(sec2(β2)sec2(α+β2))=0S_{\beta}=a^{2}(sec^{2}(\frac{\beta}{2})-sec^{2}(\frac{\alpha+\beta}{2}))=0
cos(α2)=cos(β2)=cos(α+β2)cos(\frac{\alpha}{2})=cos(\frac{\beta}{2})=|cos(\frac{\alpha+\beta}{2})|
有唯一解α=β=23π\alpha=\beta=\frac{2}{3}\pi
A=Sαα=43a2A=S_{\alpha\alpha}=4\sqrt{3}a^{2}
B=Sαβ=23a2B=S_{\alpha\beta}=2\sqrt{3}a^{2}
B=Sαβ=43a2B=S_{\alpha\beta}=4\sqrt{3}a^{2}
A>0,ACB2>0A>0,AC-B^{2}>0
因此SS(23π,23π)(\frac{2}{3}\pi,\frac{2}{3}\pi)取到极小值33a23\sqrt{3}a^{2}
下面证明:minDS=33a2min_{D}S=3\sqrt{3}a^{2}
其一,SSDD上连续,但DD不是有解闭域,SSDD上处处有偏导数
其二,SSABC\triangle ABC的边界附近的值远大于33a23\sqrt{3}a^{2}
α+β>π+0\alpha+\beta->\pi+0时,S>a22(tan(α+β2))S>\frac{a^{2}}{2}(-tan(\frac{\alpha+\beta}{2}))趋向正无穷
α>π0\alpha->\pi-0时,S>a22tan(α2)S>\frac{a^{2}}{2}tan(\frac{\alpha}{2})趋向正无穷
β\beta类似
其三,直线α+β=πδ,α=πδ,β=πδ\alpha+\beta=\pi-\delta,\alpha=\pi-\delta,\beta=\pi-\delta围成A1B1C1\triangle A_{1}B_{1}C_{1}
D=A1B1C1(intABCA1B1C1)=D1(DD1)D=\triangle A_{1}B_{1}C_{1}\cup(int\triangle ABC-\triangle A_{1}B_{1}C_{1})=D_{1}\cup(D-D_{1})
其中D1=A1B1C1D_{1}=\triangle A_{1}B_{1}C_{1}时有界闭域且存在δ>0\delta>0使得SDD1>63a2S|_{D-D_{1}}>6\sqrt{3}a^{2}
于是minDS=minD1Smin_{D}S=min_{D_{1}}S
D1D_{1}的内点(23π,23π)(\frac{2}{3}\pi,\frac{2}{3}\pi)取到
因此minD=S(23π,23π)=33a2min_{D}=S(\frac{2}{3}\pi,\frac{2}{3}\pi)=3\sqrt{3}a^{2}

5.4.3 有约束的极值LagrangeLagrange乘数法#

问: