变分法(欧拉

您所在的位置:网站首页 梯度法求最优解 变分法(欧拉

变分法(欧拉

2024-07-14 16:03| 来源: 网络整理| 查看: 265

泛函的简单理解:

xu的变量,F(x,u,{u}') 这样的就叫泛函 .

加个积分,\int_{a}^{b} F(x,u,{u}') dx这样的就叫积分泛函 .

欧拉 - 拉格朗日 (E - L) 公式:

定义一个能量泛函如下:

\large E(u) = \int_{a}^{b}F(x,u,{u}')dx

\large u = u(x)

我们的目的是找到能使 \large E(u) 取到极值的时候 \large u 的取值,所以我们就假设 \large u 就是当前能使\large E(u)取极值时的一个函数 :\large u = argminE(u)

所以一定有:

\large E(u)\leqslant E(u+t v)

\large E(u+t v) =\int_{a}^{b}F(x,u+tv,{(u+tv)}')dx

其中 \large u\large v 都是关于 \large x 的函数,也就是 \large u(x)\large v(x)  . 且\large v(a)=v(b)=0{\color{Blue} }{\color{DarkRed} }

我们把 \large E(u+t v) 看成一个 \large t 为变量的函数 \large \o (t) ,当 \large t \to 0\large \large E(u)= E(u+t v) ,所以可以推断出  \large \frac{\partial \o }{\partial t} = 0  .

原式:\frac{\partial \o }{\partial t}

= \frac{\partial E(u+t v) }{\partial t}

=\int_{a}^{b}[\frac{\partial F(x,u+tv,{(u+tv)}' )}{\partial x}*\frac{\partial x}{\partial t}+\frac{\partial F(x,u+tv,{(u+tv)}' )}{\partial (u+tv)}*\frac{\partial (u+tv)}{\partial t}+\frac{\partial F(x,u+tv,{(u+tv)}' )}{\partial {(u+tv)}'}*\frac{\partial ({(u+tv)}')}{\partial t}]dx

=\int_{a}^{b}[\frac{\partial F(x,u+tv,{(u+tv)}' }{\partial (u+tv)}* v+\frac{\partial F(x,u+tv,{(u+tv)}' }{\partial {(u+tv)}'}*{v}']dx

由于 \large t \to 0

 =\int_{a}^{b}[\frac{\partial F }{\partial u}* v+\frac{\partial F }{\partial {u}'}*{v}']dx

=\int_{a}^{b}\frac{\partial F }{\partial u}vdx+\int_{a}^{b}\frac{\partial F }{\partial {u}'}{v}'dx

=\int_{a}^{b}\frac{\partial F }{\partial u}vdx+\int_{a}^{b}\frac{\partial F }{\partial {u}'}dv

 对后一项使用分部积分公式:

 =\int_{a}^{b}\frac{\partial F }{\partial u}vdx+\frac{\partial F}{\partial {u}'}v-\int_{a}^{b}vd(\frac{\partial F}{\partial {u}'})

=\int_{a}^{b}\frac{\partial F }{\partial u}vdx+\frac{\partial F}{\partial {u}'}v-\int_{a}^{b}v\frac{d}{dx}(\frac{\partial F}{\partial {u}'})dx

合并第一项和第三项:

=\int_{a}^{b}\frac{\partial F }{\partial u}vdx-\frac{d}{dx}(\frac{\partial F}{\partial {u}'})vdx+\frac{\partial F}{\partial {u}'}v|_{a}^{b}

=\int_{a}^{b}[\frac{\partial F}{\partial u}-\frac{d}{dx}(\frac{\partial F}{\partial {u}'})]vdx+\frac{\partial F}{\partial {u}'}v|_{a}^{b}

文章开头我们提过 \large \large v(a)=v(b)=0,所以  \frac{\partial F}{\partial {u}'}v|_{a}^{b} = 0,若原式\frac{\partial \o }{\partial t}=0 ,则:

 \large \frac{\partial F}{\partial u}-\frac{d}{dx}(\frac{\partial F}{\partial {u}'}) = 0{\color{Red} }

\large -\frac{\partial F}{\partial u}+\frac{d}{dx}(\frac{\partial F}{\partial {u}'}) = 0

这就是变分法引理,也就是大名鼎鼎的欧拉 - 拉格朗日 (E - L) 公式

 我们利用相同的方法,可以推出几个衍生公式:

1 .  积分泛函为 : \large \large E(u) = \int_{a}^{b}F(x,u,{u}')dx  时,利用上述变分法得到E-L公式为

\large \large \frac{\partial F}{\partial u}-\frac{d}{dx}(\frac{\partial F}{\partial {u}'})+\frac{d^{2}}{dx}(\frac{\partial F}{\partial {u}''}) = 0

\large -\frac{\partial F}{\partial u}+\frac{d}{dx}(\frac{\partial F}{\partial {u}'})-\frac{d^{2}}{dx}(\frac{\partial F}{\partial {u}''}) = 0

 2 .  积分泛函为:\large \large \large E(u) = \int_{a}^{b}F(x,y,u,u_{x},u_{y})dx时,利用上述变分法得到E-L公式为

 \large \frac{\partial F}{\partial u}-\frac{d}{dx}(\frac{\partial F}{\partial {u_{x}}'})-\frac{d}{dy}(\frac{\partial F}{\partial {u_{y}}'}) = 0

\large -\large \frac{\partial F}{\partial u}+\frac{d}{dx}(\frac{\partial F}{\partial {u_{x}}'})+\frac{d}{dy}(\frac{\partial F}{\partial {u_{y}}'}) = 0

意义及应用:

文章的标题已经说过了,在关于水平集图像分割中,其公式的意义主要是用于梯度下降求泛函最优解,那为什么要引用一个新变量 \large t ?根据水平集和曲线演化的概念,曲线是随着时间慢慢演化的,我们就可以理解为 \large t 取一个趋近于 \large 0 很小的数,所以 \large \frac{\partial F}{\partial u}-\frac{d}{dx}(\frac{\partial F}{\partial {u}'}) 也就会很小,把 \large t 当成时间变量,我们利用时间一点点向前,一点一点演化曲线,最终达到目的。



【本文地址】


今日新闻


推荐新闻


CopyRight 2018-2019 办公设备维修网 版权所有 豫ICP备15022753号-3