Hamiltons Principle

Ravitej Uppu

Introduction
In this talk, I will assume the principle of virtual work and related topics. The talk consists of introducing and later analyzing the principle of least action axiomatically and semi-rigorously.
Lets consider a holonomic1 system of particles with n degrees of freedom. Its evolution can be
considered as a sequence of equilibrium states (`a la dAlembert) under the action of all the forces. On
applying the principle of virual work, the equations of dynamics can be rewritten as
d T
T

= Qh
dt qh qh
where Qh are the generalized forces. In our systems of interest, i.e. conservative systems, Qh is
expressed in term of the potential energy as
Qh =

U
d U

dt qh qh

(1)

This leads to the Euler-Lagrange equation of motion by the definition of the Lagrangian function
L = T U as
d L
L

=0
h 1, . . . , n
dt qh qh

An axiomatic approach to the Hamiltons Principle

The state of a system is completely specified by specifying the aptitude2 of the system.
The evolution is a sequence of such states. This sequence is completely defined by specifying the
function L(q, q,
t) (which is defined on the set of states) and two different configurations qA and qB
at tA and tB .
The Principle of Least Action says that among all the curves qh = qh (t) joining the position A and
B, the one for which the integral S[q] is minimized3 represents the path of evolution. The integral
S[q] is given as
Z tB
S[q] =
L(q, q,
t)dt
(2)
tA
1

Refer Talk 1 : page 3

Refer Talk 1: page 3
3
The form of S[q] holds on every small part of the trajectory. On the whole trajectory S[q] may just have an extremum.
This is sufficient to get to the equations of motion ( S = 0, for an extremum).
2

Hamiltons Principle

Introducing the concept of variation

Let u = u0 (x) extremize I[u] among all the curves in the plane joining points A and B.
Analyzing the properties of I[u] implies we needs to compare the value of I[u0 ] taken by the
integral on this curve with other curve v that has the value I[v].
The difference v(x) u0 (x) is the variation (for a fixed x) in passing from curve u0 to v.
Analogous to two close points being represented by dx, we represent the variation between two
close curves as u.
Corresponding to u, the variations that occur in I[u] is I[v] I[u0 ]. This can be decomposed in
terms of I, 2 I, . . .
The condition for I[u] to take an extremum is I = 0.

The next obvious extension that we would expect is identification of the type of extremum from
the the sign of 2 I. But, this is not correct. Because, unlike two points, two curves can be located
arbitrarily close in space, yet at close points have tangents that are totally different. This effects
even more when the integral is a function as u0 which is precisely the case that we are dealing with
here.
The clarification comes from the principles of the functional calculus which I will discuss in brief
in the following section.

3
3.1

Building up functional calculus

Continuity

If we consider the obvious extension from calculus of one-variable to our case and say
Consider a functional F [u]. It is said to be continuous, if for a small change of u(x), F [u]
undergoes a small change.
But, this definition is very restricted as a generic functional F [u] can also have an explicit
dependance on u0 . So the question is How do we define close curves?
The definition should include the information of how close should our curves be. So,
curves can be defined to be close to various orders as follows :
Zero-order proximity close : u(x) and v(x) are zero-order proximity close, if |u(x)v(x)|
is small.
First-order proximity close : u(x) and v(x) are first-order proximity close, if |u(x)v(x)|
and |u0 (x) v 0 (x)| are small.
In general, u(x) and v(x) are kth-order proximity close, if |u(x)v(x)|, |u0 (x)v 0 (x)|, . . . ,
|u(k) (x) v (k) (x)| are small.
From this we can define a notion of distance () between the two curves as follows
Considering the existence of continuous derivatives of u and v upto order k, the distance of order k
is given by
k
X
k (u, v) :=
max |u(h) (x) v (h) (x)|
h=1

x0 xx1

For close curves, k is small.

2

3.2

Differentiability

3.2

Hamiltons Principle

Differentiability

Analogous to the total and partial derivatives that we know of, even here there exists
two kinds of derivatives : Strong or the Frechet derivative and the Weak or the Gateaux
derivative.
3.2.1

Strong derivative

Consider U and V (normed vector spaces). As a clarification to the question raised by

Amaresh, U and V can be similar. Only possible distinguishing feature is the sense
of distance that is considered on the space. For U , we define the distance as defined
above. Whereas, the distance measure on V is the standard measure of separation between points.
A is an open subset in U . Consider a map F from A to U .
F :AU V
F is said to be differentiable at u U , if F 0 , such that
F [u + h] F [u] = F 0 h + (u, h)

(3)

where, is infinitesimal with respect to the distance measure in V as given by the norm
(i.e. lim||h||0 ||(u,h)||
= 0).
||h||
F 0 h is called the strong differential and F 0 is the strong derivative of F at u. Composition
is straightforward. Uniqueness can be checked in the same way as total derivatives in
calculus.
3.2.2

Weak derivative

Consider a map G that is defined as

G:AU V
Consider the limit,
DG(u, h) =

d
G[u + th] G[u]
G[u + th]|t=0 = lim
t0
dt
t

(4)

The convergence is only with respect to the norm on V . This is the general definition for
the weak derivative. In general DG can be non-linear with respect to h. In the case of DG
being linear, we can write
DG(u, h) = Gu h
where Gu is the weak derivative of G at u.
Composition does not hold good for weak derivatives similar to the case of partial derivatives of functions. At this point few remarks will clarify the differences between the two
kinds of derivatives.
If strong derivative of F exists, weak derivative exists and they will coincide.

3.3

Gradient of a functional

Hamiltons Principle

Proof : If F is strongly derivable, then

F [u + th] F [u] = tF 0 [u]h + (u, th)
F [u + th] F [u]
= F 0 [u]h
t
where the RHS is nothing but Fu , from the definition of weak derivative.
Q.E.D.
lim

t0

Converse of the above statement does not hold. Proof was left as an excercise.
Analogous to the situation where existence of partial derivates that are continuous,
implies the existence of total derivate, we can say
Given a map F , if weak derivative Fu exists in the neighbourhood of u0 and
represents a continuous operator at u0 , then F 0 exists at uo and coincides with
Fu .

3.3

Gradient of a functional

Consider U (has a scalar product defined on it), a space of numerical functions u(x). Let
F be a functional defined on U as
F : u U F [u] R
Gradient of F or the functional derivative of F with respect to the scalar product is G =
F
u , defined by the relation


d
F
F [u + ]|=0 =:
,
(5)
d
u
Second functional derivative is the weak derviative Gu of G i.e.
d2
F [u +  + ]|= =0 =: (Gu , )
dd
It is easy to check that Gu is symmteric with respect to scalar product.

A step away - an example

Let U be the space of all C functions1 u(x), defined on the interval I = [a, b) R. Let the functions
and all their derivatives go to zero at a and b i.e.
u(k) (a) = 0 & u(k) (b) = 0

k [1, )

Let F [u] be a functional defined as

Z
F [u] =

f (u, ux , uxx , . . . , unx , . . .)dx

a

1
These are also called smooth functions, because neither the functions nor their derivatives have any corners, as all of
them are continuous

Hamiltons Principle

where un x is the nth-dervative of u with respect to x. From the definition of gradient, we get
F

d
F [u + h]|=0
d

Z b
f
f
f
f
hx +
hxx + . . . +
hnx + . . . dx
h+
=
u
ux
uxx
unx
a

Z b
f
dn f
d f
h
=
+ . . . + h(1)n n
+ . . . dx
h
u
dx ux
dx unx
a
=

It can now be easily checked that

F
dn f
f
d f
+ . . . + (1)n n
+ ...
=

u
u dx ux
dx unx

(6)

Towards Hamiltons Principle

Consider the special case when f is only a function of u and ux . Then the Eqn.(6) simplifies to
F
f
d f
=

u
u dx ux
This can be generalized to n functions u1 , . . . , un such that


F
F
F

,...,
u
u1
un


d f
d f
f
f

,...,
=
u1 dx u1x
un dx unx

(7)
(8)

Now, if we recognize x 7 t and u 7 q (which implies ux 7 q,

we can attribute f 7 L and F 7 S,
which says that
d L
L

= S
(9)
dt qi qi
Extremal variation of S as defined in Eqn. (2) gives us the Euler-Lagrange equations of motion
from the above result.

Additional remark
Actually, starting from Eqn. (6), and recognizing x 7 t and u 7 q, we can also get the generalized
expression Eqn. (1) for the Potential energy U in consevative systems by recognizing f 7 U(q, q).

This is just a consequence of the definition of gradient of a functional and the fact that for conservative
systems gradient vanishes.