[go: up one dir, main page]
Scribd
Upload
91 views3 pages

Calculus of Variations-1

Calculus of variations is used to find functions that minimize energy functionals, which are integrals involving an unknown function and its derivatives. It extends vector calculus to evaluate derivatives of functionals. The gradient of a functional is found by taking its directional derivative, yielding the Euler-Lagrange equation, a necessary condition for a minimizing function. This can be solved directly or via gradient descent, resulting in a partial differential equation. The methodology can be extended to problems with multiple variables or vector-valued functions.

Uploaded by

saleemnasir2k7154
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
91 views3 pages

Calculus of Variations-1

Calculus of variations is used to find functions that minimize energy functionals, which are integrals involving an unknown function and its derivatives. It extends vector calculus to evaluate derivatives of functionals. The gradient of a functional is found by taking its directional derivative, yielding the Euler-Lagrange equation, a necessary condition for a minimizing function. This can be solved directly or via gradient descent, resulting in a partial differential equation. The methodology can be extended to problems with multiple variables or vector-valued functions.

Uploaded by

saleemnasir2k7154
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

1

Calculus of Variations
Namrata Vaswani, namrata@iastate.edu

These notes are still under preparation. Please email me if you find any mistakes and typos.
These notes are based on Chapter 1 of [1] and some web sources.
Consider the problem of minimizing an energy functional E(u) which is an integral of a function of an
unknown function u(x) and its derivatives w.r.t. x. u and its derivatives are only known at the boundaries
of the integration domain.
Calculus of variations is used to find the gradient of a functional (here E(u)) w.r.t. a function (here
u(x)), which we denote by u E . Setting u E = 0 gives the Euler-Lagrange equation and this is

a necessary condition for the minimizing function to satisfy. In some cases the Euler-Lagrange can be
solved directly in closed form. For other cases one uses numerical techniques for gradient descent, which
gives rise to a Partial Differential Equation (PDE). In effect, Calculus of Variations extends vector calculus
to enable us to evaluate derivatives of functionals.

A. Evaluating u E
I explain here the simplest case: how to evaluate u E when E can be written as a definite integral
of u and its first partial derivative ux ,

u
x

and x is a scalar, i.e.


Z b
E(u) =
L(u, ux )dx

(1)

and u(a) and u(b) are known (fixed), while u(x), x (a, b) is variable. Here L(u, ux ) is referred to as
the Lagrangian.
1) Defining u E : From vector calculus (if u were a vector u = [u1 , u2 , ..un ]), then the directional
derivative in a direction is
E(u + ) E(u)
E
= lim
= u E
0

(2)

where the dot product expands as


u E =

n
X
(u E)i i
i=1

(3)

One way to evaluate u E is to write out a first order Taylor series expansion of E(u + ) and define
(u E)i by comparison as the term multiplying i .

We use this same methodology for calculus of variations, but now u is a continuous function of a
Rb
variable x and is also a continuous function of x with unit norm (||||2 = a (x)2 dx = 1). The dot
product is now defined as
Z
u E =

(u E)(x)(x)dx.

(4)

The boundary conditions, u(a) and u(b) are fixed and so (a) = (b) = 0.
2) The solution method: Expand E(u + ) using first order Taylor series as
Z b
E(u + ) E(u) + u E = E(u) +
(u E)(x)(x)dx

(5)

and (u E)(x) is the term multiplying (x) in this expansion.


Applying this to (1), we get
Z b
E(u + ) =
L(u + , ux + x )dx
a

L(u, ux )dx +

L
( )(x)(x)dx +
u

(
a

L
)(x)x (x)dx
ux

= E(u) + (T1 + T2 )

(6)

Now T1 is already in the form of a dot product with . We need to bring T2 =

Rb

L
a ux x dx

form. We do this using integration by parts. Recall that


Z b
Z b
f (x)gx (x)dx = f (b)g(b) f (a)g(a)
fx (x)g(x)dx
a

T2 =

(7)

We apply this to T2 with f =


Z

also in this

L
ux ,

g = and (a) = (b) = 0, so that the first two terms of (7) vanish.

L
L
L
x dx = (
)(b)(b) (
(a))(a)
ux
ux
ux
Z b
L
= 00
dx
a x ux

L
dx
x ux

(8)

Thus combining (6) with (8), we get


Z
E(u + ) = E(u) +

[(
a

L
L
)
]dx
u
x ux

(9)

and thus by comparison with (5), we have


(u E)(x) = [(

L
L
)
](x)
u
x ux

(10)

Thus the Euler Lagrange equation (necessary condition for a minimizer) is


[(

L
L
)
]=0
u
x ux

(11)

This can be either solved directly or using gradient descent. When using gradient descent to find the
minimizing u, we get a PDE with an artificial time variable t as
u
L
L
= (u E) = [( )
]
t
u
x ux

(12)

B. Extensions
The solution methodology can be easily extended to cases where (i) u is a function of multiple
variables, i.e. x is a vector (x = [x1 , x2 , ..xk ]T ) or (ii) when u itself is a vector of functions, i.e. u(x) =
[u1 (x), u2 (x)...um (x)]T or (iii) when E depends on higher order derivatives of u. Please UNDERSTAND

the basic idea in the above derivation carefully, some of these extensions may be Exam questions.
Exercise: Show that if u = u(x, y) i.e. u is a function of two scalar variables x and y , then
(u E) = (

L
L
L
)

u
x ux y uy

(13)

A trivial extension of this shows that if E is a function of two functions u(x, y) and v(x, y), then the
Euler-Lagrange equation is given by
L
L
L
)

=0
u
x ux y uy
L
L
L
(v E) = ( )

=0
v
x vx y vy

(u E) = (

(14)

One application of this is in estimating Optical Flow using Horn and Schunks method [2] (see Optical
flow handout). More applications will be seen in Segmentation problems, which attempt to find the
object contour C(p) = [C x (p), C y (p)] (where p is a parameter that goes from 0 to 1 over the contour
and C(0) = C(1) for closed contour) that minimizes an image dependent energy functional E(C).
R EFERENCES
[1] G.Sapiro, Geometric Partial Differential Equations and Image Analysis, Cambridge University Press, 2001.
[2] B.K.P. Horn and B.G.Schunk, Determining optical flow, Artificial Intelligence, vol. 17, pp. 185203, 1981.

You might also like