自用的变分法与最优控制推导

经典变分法

推导1

推导步骤来自 理论力学3—变分法的核心，欧拉-拉格朗日方程
$$\begin{array}{rl}\delta J& =\delta {\int }_a^{b}f(y,y^{\prime })\text{d}x\\ & ={\int }_a^b\delta f(y,y^{\prime })\text{d}x\\ & ={\int }_a^b\left(\frac{\partial f}{\partial y}\delta y+\frac{\partial f}{\partial y^{\prime }}\delta y^{\prime }\right)\text{d}x\\ & ={\int }_a^b\frac{\partial f}{\partial y}\delta y\text{d}x+{\int }_a^b\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}}{\text{d}x}\delta y\text{d}x\\ {\int }_a^b\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}}{\text{d}x}\delta y\text{d}x& ={\int }_a^b\frac{\partial f}{\partial y^{\prime }}\text{d}\delta y={\left[\frac{\partial f}{\partial y^{\prime }}\delta y\right]}_a^b-{\int }_a^b\delta y\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\text{d}x\\ \delta J& ={\left[\frac{\partial f}{\partial y^{\prime }}\delta y\right]}_a^b+{\int }_a^b\frac{\partial f}{\partial y}\delta y\text{d}x-{\int }_a^b\delta y\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\text{d}x\\ & ={\int }_a^b\left(\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\right)\delta y\text{d}x\end{array}$$
即
$$\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}=0$$
另
$$\begin{array}{rl}\frac{\text{d}f}{\text{d}x}& =\frac{\partial f}{\partial y}\frac{\text{d}y}{\text{d}x}+\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}{y}^{\prime }}{\text{d}x}\\ & =\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}y}{\text{d}x}+\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}{y}^{\prime }}{\text{d}x}\end{array}$$
即
$$\begin{gathered} \frac{\text{d}}{\text{d}x}\left(y^{\prime }\frac{\partial f}{\partial y^{\prime }}-f\right)=0 \\ {y}^{\prime }\frac{\partial f}{\partial y^{\prime }}-f=C \end{gathered}$$

推导2

推导步骤来自 两小时搞定变分法
设$y_0(x)$为最优函数，$y(x)$为其它的函数，令$y(x)-y_0(x)=ϵ\eta (x)=\delta y(x)$，则

$$\begin{array}{rl}J(ϵ)& ={\int }_a^{b}f(y(ϵ),y^{\prime }(ϵ),x)\text{d}x\\ & ={\int }_a^{b}f(y_0+ϵ\eta (x),y_0^{\prime }+ϵ{\eta }^{\prime }(x),x)\text{d}x\\ \frac{\text{d}J}{\text{d}ϵ}& ={\int }_a^b\left(\frac{\partial f}{\partial y}\frac{\text{d}y}{\text{d}ϵ}+\frac{\partial f}{\partial y^{\prime }}\frac{\text{d}{y}^{\prime }}{\text{d}ϵ}+\frac{\partial f}{\partial x}\frac{\text{d}x}{\text{d}ϵ}\right)\text{d}x\\ & ={\int }_a^b\left(\frac{\partial f}{\partial y}\eta +\frac{\partial f}{\partial y^{\prime }}{\eta }^{\prime }\right)\text{d}x\\ & ={\int }_a^b\frac{\partial f}{\partial y}\eta \text{d}x+{\int }_a^b\frac{\partial f}{\partial y^{\prime }}\text{d}\eta \\ & ={\int }_a^b\frac{\partial f}{\partial y}\eta \text{d}x+{\left[\frac{\partial f}{\partial y^{\prime }(x)}\eta (x)\right]}_a^b+{\int }_a^b\eta (x)\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\text{d}x\\ & ={\int }_a^b\left(\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\right)\eta (x)\text{d}x+{\left[\frac{\partial f}{\partial y^{\prime }(x)}\eta (x)\right]}_a^b\end{array}$$

终端时刻固定、终端状态自由

$$\begin{array}{rl}\delta J& ={\int }_a^b\left(\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\right)\delta y\text{d}x+\frac{\partial f}{\partial y^{\prime }(b)}\delta y(b)=0\end{array}$$
所以（等号两边都必须等于0的证明见下面的变分预备定理）
$$\begin{array}{r}{\int }_a^b\left(\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\right)\delta y\text{d}x=\frac{\partial f}{\partial y^{\prime }(b)}\delta y(b)=0\end{array}$$
因为$\delta y(b)$不受限，所以
$$\frac{\partial f}{\partial y^{\prime }(b)}=0$$

终端时刻自由、终端状态固定

终端时刻自由的一个公式
$$\delta y(b+\text{d}b)=\delta y(b)+y^{\prime }(b)\text{d}b(1-1)$$
$$\begin{array}{rl}\delta J& ={\int }_a^{b+\text{d}b}f(y+\delta y,y^{\prime }+\delta y^{\prime })\text{d}x-{\int }_a^{b}f(y,y^{\prime })\text{d}x\\ & ={\int }_a^b\delta f(y,y^{\prime })\text{d}x+{\int }_b^{\text{d}b}f(y+\delta y,y^{\prime }+\delta y^{\prime })\text{d}x\\ & ={\int }_a^b\left(\frac{\partial f}{\partial y}-\frac{\text{d}}{\text{d}x}\frac{\partial f}{\partial y^{\prime }}\right)\delta y\text{d}x+\frac{\partial f}{\partial y^{\prime }(b)}\delta y(b)+f(y(b),y^{\prime }(b))\text{d}b\end{array}$$
(补充说明:${\int }_x^{x+\text{d}x}f(t)\text{d}t=f(x)\text{d}x$)
即
$$\frac{\partial f}{\partial y^{\prime }(b)}\delta y(b)+f(y(b),y^{\prime }(b))\text{d}b=0(1-2)$$
又因为$\delta y(b+\text{d}b)=\delta y(b)+y^{\prime }(b)\text{d}b=0$，所以
$$\frac{\partial f}{\partial y^{\prime }(b)}{y}^{\prime }(b)=f(y(b),y^{\prime }(b))$$

终端时刻自由、终端状态自由

将式(1-1)代入(1-2)
$$\begin{array}{rl}& \frac{\partial f}{\partial y^{\prime }(b)}(\delta y(b+\text{d}b)-y^{\prime }(b)\text{d}b)+f(y(b),y^{\prime }(b))\text{d}b=0\\ & \frac{\partial f}{\partial y^{\prime }(b)}\delta y(b+\text{d}b)+\left(f(y(b),y^{\prime }(b))-\frac{\partial f}{\partial y^{\prime }(b)}\right)\text{d}b=0\end{array}$$
得到
$$\begin{array}{rl}& \frac{\partial f}{\partial y^{\prime }(b)}=f(y(b),y^{\prime }(b))=0\end{array}$$

终端时刻自由、终端状态约束

设约束方程$c(x)=0$，则约束时满足$\delta y(b+\text{d}b)=\delta y(b)+y^{\prime }(b)\text{d}b=c^{\prime }(b)\text{d}b$，代入式(1-2)得
$$\begin{array}{rl}& \frac{\partial f}{\partial y^{\prime }(b)}(c^{\prime }(b)-y^{\prime }(b))\text{d}b+f(y(b),y^{\prime }(b))\text{d}b=0\\ & {\left[\frac{\partial f}{\partial y^{\prime }}(c^{\prime }-y^{\prime })+f(y,y^{\prime })\right]}_{x=b}=0\end{array}$$

极小值原理

下面的推导均可推广到$x$、$\lambda$等为向量形式。

哈密顿函数

设$$H(x,u,\lambda ,t)=f(x,u,t)+\lambda (t)g(x,u,t)$$
系统状态方程为$x^{\prime }=g(x,u,t)$。定义泛函
$$\begin{array}{rl}J& ={\int }_a^{b}f(x,u,t)+\lambda (t)(g(x,u,t)-x^{\prime }(t))\text{d}t\\ & ={\int }_a^{b}H(x,u,\lambda ,t)\text{d}t-{\int }_a^b\lambda (t)\text{d}x(t)\\ & ={\int }_a^{b}H(x,u,\lambda ,t)\text{d}t+{\int }_a^b{\lambda }^{\prime }(t)x(t)\text{d}t-[\lambda (t)x(t){]}_a^b\\ \delta J& ={\int }_a^b\delta H(x,u,\lambda ,t)\text{d}t+{\int }_a^b{\lambda }^{\prime }(t)\delta x(t)\text{d}t-\lambda (b)\delta x(b)\\ & ={\int }_a^b\left(\frac{\partial H}{\partial x}+{\lambda }^{\prime }\right)\delta x+\frac{\partial H}{\partial u}\delta u\text{d}t-\lambda (b)\delta x(b)\end{array}$$
由变分预备定理得到
$$\begin{array}{rl}& \frac{\partial H}{\partial \lambda }=x^{\prime }\\ & \frac{\partial H}{\partial x}=-{\lambda }^{\prime }\\ & \frac{\partial H}{\partial u}=0\end{array}$$

加入限制条件

下面开始最优控制，将积分上下限由$(a,b)$改为$(t_0,t_f)$。
末端时刻固定、末端状态约束时，加入末端性能指标$\varphi (x_f)$极小和末端状态约束条件$\psi (x_f)=0$。
$$\begin{array}{rl}J& =\varphi (x)+\gamma \psi (x)+{\int }_{t_0}^{t_f}f(x,u,t)+\lambda (t)(g(x,u,t)-x^{\prime }(t))\text{d}t\\ \delta J& =(\frac{\partial \varphi }{\partial x}+\gamma \frac{\partial \psi }{\partial x}-\lambda (t_f))\delta x(t_f)+{\int }_{t_0}^{t_f}\left(\frac{\partial H}{\partial x}+{\lambda }^{\prime }\right)\delta x+\frac{\partial H}{\partial u}\delta u\text{d}t\end{array}$$
横截条件
$$\lambda (t_f)=\frac{\partial \varphi }{\partial x(t_f)}+\gamma \frac{\partial \psi }{\partial x(t_f)}$$
的矢量形式为
$$\left[\begin{array}{c}{\lambda }_1(t_f)\\ {\lambda }_2(t_f)\end{array}\right]=\left[\begin{array}{c}\frac{\partial \varphi }{\partial x_1(t_f)}\\ \frac{\partial \varphi }{\partial x_2(t_f)}\end{array}\right]+\left[\begin{array}{cc}\frac{\partial {\psi }_1}{\partial x_1(t_f)}& \frac{\partial {\psi }_2}{\partial x_1(t_f)}\\ \frac{\partial {\psi }_1}{\partial x_2(t_f)}& \frac{\partial {\psi }_2}{\partial x_2(t_f)}\end{array}\right]\left[\begin{array}{c}{\gamma }_1\\ {\gamma }_2\end{array}\right]$$
如果末端状态自由则删除约束条件$\psi (x_f)=0$。如果末端状态固定，则删除横截条件$\lambda (t_f)$。
（因为末端状态固定时，$\delta x(t_f)$任意导致横截条件$\lambda (t_f)$不成立）
如果末端时刻自由，则增加哈密顿函数变化律公式
$$H(t_f)=-\frac{\partial \varphi }{\partial t_f}-\gamma \frac{\partial \psi }{\partial t_f}$$

限制输入

几个前置证明

变分预备定理证明

• 证明1：若$h(x)$为任意函数时均有${\int }_a^{b}M(x)h(x)\text{d}x=0$则$M(x)\equiv 0$。

反证法，若$M(x)\not\equiv 0$，因为$h(x)$任意，令$h(x)=-M(x)(x-a)(x-b)$，则
$$M(x)h(x)=-M^2(x)(x-a)(x-b)$$
$\because x\in [a,b]$，$\therefore M(x)h(x)\ge 0$，且$\exists x\in (a,b)$使得$M(x)h(x)>0$，从而${\int }_a^{b}M(x)h(x)\text{d}x>0$，假设不成立。

• 证明2：不存在$M(x)$使得若$h(x)$为任意函数时均有${\int }_a^{b}M(x)h(x)\text{d}x=C\ne 0$。

同理令$h(x)=k/M(x)$，则
$${\int }_a^{b}M(x)h(x)\text{d}x=k(b-a)\not\equiv C$$
假设不成立。

可交换性证明

• 变分和微分

$$\begin{array}{rl}& \frac{\text{d}}{\text{d}x}\delta y=\frac{\text{d}}{\text{d}x}(ϵ\eta (x))=ϵ\frac{\text{d}}{\text{d}x}\eta (x)\\ & \delta \frac{\text{d}}{\text{d}x}y=y^{\prime }(x)-y_0^{\prime }(x)={\eta }^{\prime }(x)\end{array}$$

• 变分和积分

$$\begin{array}{rl}& \delta {\int }_a^{b}y(x)\text{d}x={\int }_a^{b}y(x)\text{d}x-{\int }_a^b{y}_0(x)\text{d}x={\int }_a^{b}y(x)-y_0(x)\text{d}x={\int }_a^b\delta y(x)\text{d}x\end{array}$$

泛函的拉格朗日乘子法证明

$$J={\int }_a^{b}f(y,y^{\prime })\text{d}x$$
若满足约束$g(y,y^{\prime })=0$，则
$$\begin{array}{rl}& \delta g=\frac{\partial g}{\partial y}\delta y+\frac{\partial g}{\partial y^{\prime }}\delta y^{\prime }=0(2-1)\\ & \delta J={\int }_a^b\delta f(y,y^{\prime })\text{d}x+\lambda \delta g(y,y^{\prime })=0(2-2)\\ & \delta f(y,y^{\prime })=\frac{\partial f}{\partial y}\delta y+\frac{\partial f}{\partial y^{\prime }}\delta y^{\prime }=0(2-3)\end{array}$$
由(2-1)(2-3)式可得
$$\begin{array}{rl}\frac{\partial f}{\partial y}& =-\lambda \frac{\partial g}{\partial y}\\ \frac{\partial f}{\partial y^{\prime }}& =-\lambda \frac{\partial g}{\partial y^{\prime }}\\ \delta f(y,y^{\prime })& =\frac{\partial f}{\partial y}\delta y+\frac{\partial f}{\partial y^{\prime }}\delta y^{\prime }\\ & =\frac{\partial f}{\partial y}\delta y+\frac{\partial f}{\partial y^{\prime }}(-\frac{\frac{\partial g}{\partial y}}{\frac{\partial g}{\partial y^{\prime }}})\\ & =(\frac{\partial f}{\partial y}+\lambda \frac{\partial g}{\partial y})\delta y\end{array}$$
于是可以构造拉格朗日函数
$$\begin{array}{rl}& L(y,y^{\prime })=f(y,y^{\prime })+\lambda g(y,y^{\prime })\\ & \frac{\partial L}{\partial y}=\frac{\partial f}{\partial y}+\lambda \frac{\partial g}{\partial y}\end{array}$$
需要注意的是，多元函数中$z=f(x_0,y_0)$是个点，所以$\lambda$是个常数；而$\delta y$是个函数，所以泛函中的$\lambda$也是个函数。