Functional & Calculus of variations

Fcuntional

普通形式:

定积分形式:

Functional derivatives[2]

  • 先从普通函数讲起:
    对于普通函数,我们通过对移动一小步,可以定义在处的differential:

    由此,可以推广到多元函数处的differential:

    所以说,导数derivative(以及偏导)其实就是微分计算中,每一个微分偏量前的系数
    我们将这个式子可以用sum重写成:

  • 泛函导数的定义:
    (当然,其实我们不能保证所有functional的导数都存在)
    由此式子(泛函的differential或者是variation of the functional F)定义泛函在处的导数,其中为任意函数:

    可以类比下普通函数在点的total differential [2]:

  • 普通函数导数类比泛函导数的理解方式:
    两者都可以理解为对的variation()等于在每“分量”处的变化率,即偏导数乘以其变化量。当时,即有无穷多个independent variables。类比泛函导数中,即为无限的连续的index
    同时泛函由于有一个项,所以可以类比普通函数的方向导数[2],即在方向上的导数(投影?)

  • 其他相关概念理解:
    其实泛函变分(variation)的概念即“variation of a functional”,与普通函数中微分(differential)的概念是相似的。同理"small change in input"在泛函中的普通函数中的也是类似的。

  • 泛函导数的推导:
    本身是EL equation的一种泛化,首先,针对EL的形式设:

    根据differential的定义,在处的:

    (1)
    (2)
    = \int [\frac {\partial f}{\partial x} \frac {\partial x}{\partial \epsilon} + \frac {\partial f}{\partial \rho} \frac {\partial (\rho + \epsilon \phi)}{\partial \epsilon} + \frac {\partial f}{\partial \rho'} \frac {\partial (\rho' + \epsilon \phi')}{\partial \epsilon}] dx(3)
    (4)
    (5)
    (6)
    对比(1)和(6),我们可以得到:
    (7)
    注,这里是对求导的意思,即


    解读:
    (2):通过定义,其实是就是转化成了求的极限值。注意,这里两边同时对求导,为常数,为的函数,所以得:
    (3):对求导,由于跟无关,直接对积分内的运用chain rule即可
    (4):化简,由于所以第一项对的偏导可以消除
    (5):这一步主要想消除由两边同时积分再移项:。注意,是关于的函数,所以这里是对积分求导。
    (6):由于在边界上vanish,所以,再合并积分项
    (7):为针对EL equation的推导,该导数的可以推广到更泛化的形式[5],可以通过多元的泰勒一阶展开推导[5.Doc]

  • 泛函导数的性质[6]:
    如下几个性质与普通函数导数相同
    Linearity
    Product Rule
    Chain Rule

泛函极值条件 Weak Extrema of Functional

函数中最小值的条件:一阶导数为0是必要非充分的,且极值点二阶导数非负,此时是充要的。
类似地,可以通过二阶变分[7] 得到泛函极小值的条件:
Necessary:
Sufficient:且在该点处, strongly positive[8]
相关条件以及证明见Refer[8]

边界条件 Natural boundary condition

Essential boundary conditions are imposed explicitly on the solution but natural boundary conditions are automatically satisfied after solution of the problem.
在没有端点固定的问题中。
1、需要满足除了EL-equation:一阶变分为0条件见上述式子(7)。这个逻辑非常天然符合直觉,即使得问题取得极值的话,肯定也满足在其边界固定的问题中的极值条件。
2、还需要满足:以及。
否则,泛函不可能取得极值。由于这个条件是在求变分极值过程中得到的条件,所以是自然边界条件。

Other

微积分演变
https://www.zhihu.com/question/27926053/answer/1017772036?utm_source=wechat_session

E-L equation
https://en.wikipedia.org/wiki/Euler%E2%80%93Lagrange_equation
推导:
https://math.stackexchange.com/questions/1885316/functional-derivative-how-to-obtain-delta-f-int-frac-delta-f-delta-f-de
https://www.cnblogs.com/bigmonkey/p/9519387.html

变分法(Calculus of Variations(variational method)):
https://baike.baidu.com/item/%E5%8F%98%E5%88%86%E6%B3%95/83603
https://en.wikipedia.org/wiki/Calculus_of_variations
https://zhuanlan.zhihu.com/p/20718489

思考,Functionals 与Cost Function
https://stats.stackexchange.com/questions/158348/can-a-neural-network-learn-a-functional-and-its-functional-derivative
以及Functional Gradient Descent
cost function本身,也可以当作是一个functional?我们用它来最优化得到最终的function。

Refer:

[1]
与的差异:
https://math.stackexchange.com/questions/317338/differentiation-using-d-or-delta
stands for the exact differential 一般用在math中
refers to an inexact differential 一般用在physics中,inexact differentials

[2]
泛函
Functionals and the Functional Derivative
泛函导数,定义与计算方法:
见Doc:Functional Derivative
见:http://julian.tau.ac.il/bqs/functionals/node1.html
理解为方向导数见:https://bjlkeng.github.io/posts/the-calculus-of-variations/

[3]
https://en.wikipedia.org/wiki/Differential_of_a_function#Differentials_in_several_variables
involving the [partial derivative] of y with respect to x, The sum of the partial differentials with respect to all of the independent variables is the total differential

[4]
TODO(有关Radon-Nikodym定理,测度理论)

[5]
https://en.wikipedia.org/wiki/Functional_derivative
见 Determining functional derivatives中“An analogous application of the definition of the functional derivative yields”

[6]
https://en.wikipedia.org/wiki/Functional_derivative
见Properties

[7]:
Second Variation二阶变分。
https://encyclopediaofmath.org/wiki/Second_variation#:~:text=a%20sufficient%2C%20condition%20(under%20certain,at%20the%20point%20x0.&text=(the%20derivatives%20are%20evaluated%20at,x0(t)).

[8]
见:
1、https://en.wikipedia.org/wiki/Calculus_of_variations#Variations_and_sufficient_condition_for_a_minimum中的Variations and sufficient condition for a minimum
2、Notes on Sufficient Conditions for Extrema
3、https://math.stackexchange.com/questions/3155772/difficulty-understanding-sufficient-conditions-for-weak-extrema-in-calculus-of-v

[9]
1、见:https://math.stackexchange.com/questions/3315027/what-are-natural-boundary-conditions-in-the-calculus-of-variations
2、变分原理(Doc)中的自然边界条件。

你可能感兴趣的:(Functional & Calculus of variations)