Chapter 3

Lagrangian Mechanics

3.1 Action Principle

In this chapter we will see how the familiar laws of mechanics can be expressed and understood
from a very different point of view, which is known as the Lagrangian formulation of mechanics.
This is in many ways more elegant than the Newtonian formulation, and it is particularly useful
when moving to quantum mechanics. For example, quantum field theories are usually studied in
a Lagrangian framework.
The idea is similar to Fermat’s principle in optics, according to which light follows the short-
est optical path, i.e., the path of shortest time to reach its destination. As a reminder, let us see
how Snell’s law
sin ✓2 n1
= , (3.1.1)
sin ✓1 n2
which tells how a light ray bends at the interface of two materials with refractive indices n1 and
n2 .
Consider a light ray from point (xa , ya ) to (xb , yb ). There is a horizontal interface at y, and
between ya and y, the refractive index is n1 and between y and yb it is n2 . We now assume that
the light follows a straight line from (xa , ya ) to a point (x, y) on the interface, and then from
(x, y) to (xb , yb ). The only unknown is therefore x. Because of the speed of light in medium is
c/n, the optical path length is
Z b Z b
n n1 p n2 p
S(x) = dt = dl = (x xa )2 + (y ya )2 + (xb x)2 + (yb y)2 . (3.1.2)
a a c c c
Note that we have made the simplification S(x, y) ! S(x) since we are assuming straight lines
i.e. y = x. According to Fermat’s principle, we need to find the minimum of this quantity. At
the minimum, the derivative with respect to x vanishes, so
@S n1 x xa n2 xb x n1 n2
0= = p p = sin ✓1 sin ✓2 ,
@x c (x xa )2 + (y ya ) 2 c (xb x)2 + (yb y)2 c c
(3.1.3)
from which Snell’s law (3.1.1) follows immediately.

34
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

Fermat had proposed his principle in 1657, and it motivated Maupertuis to suggest in 1746
that also matter particles would obey an analogous variational principle. He postulated that there
exists a quantity called action, which the trajectory of the matter particle would minimise. This
idea was later refined by Lagrange and Hamilton, who developed it into its current form, in which
the action S is defined as an integral over a function L(x, ẋ, t) known as the Lagrangian, as
Z tb
S[x] = L(x(t), ẋ(t), t)dt. (3.1.4)
ta

The Lagrangian is a function of the position x and velocity ẋ of the particle, and it may also have
some explicit time dependence. We will see later that for conservative systems, the Lagrangian
is simply the difference of the kinetic energy T and the potential energy V , i.e., L = T V .
Because the action S is given by an integral over time, it depends on the position and velocity
at all times, i.e., on the whole trajectory of the particle. It is therefore a function from the space
of functions x(t) to real numbers, and we indicate that by having the argument (i.e. function x)
in square brackets. Such function of functions are called functionals.
Given a Lagrangian L, the dynamics is determined by Hamilton’s principle (or action prin-
ciple), which states that to move from position xa at time ta to position xb at time tb , the particle
follows the trajectory that minimises the action S[x]. In other words, the actual physical trajec-
tory is the function x that minimises the action subject to the boundary conditions x(ta ) = xa
and x(tb ) = xb .
To find this minimising function x(t), we want to calculate the derivative of the action S[x]
with respect to the function x(t) and set it to zero. Functional derivatives such as this are studied
in the branch of mathematics known as functional analysis. However, here we adopt a slightly
simpler approach and consider small variations of the trajectory. This is known as variational
calculus.
Let us assume that x(t) is the function that minimises the action, and consider a slightly
perturbed trajectory
x̃(t) = x(t) + x(t), (3.1.5)
where we assume that the perturbation is infinitesimally small and vanishes at the endpoints,
x(ta ) = x(tb ) = 0. (3.1.6)
This perturbation changes the action by
S = S[x + x] S[x]
Z tb
= [L(x(t) + x(t), ẋ(t) + ẋ(t), t) L(x(t), ẋ(t), t)] dt
ta
Z tb 
@L @L
= x(t) + ẋ(t) dt
ta @x @ ẋ
Z tb 
@L @L d x(t)
= x(t) + dt
ta @x @ ẋ dt
Z tb  tb Z tb 
@L @L d @L
= x(t)dt + x(t) x(t)dt, (3.1.7)
ta @x @ ẋ ta ta dt @ ẋ

35
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

where we Taylor expanded to linear order in x(t) and integrated the second term by parts.
Because of the boundary conditions (3.1.6), the substitution term vanishes, and we have
Z tb 
@L d @L
S= x(t)dt. (3.1.8)
ta @x dt @ ẋ

For x(t) to be minimum, the variation of the action (3.1.8) has to vanish for any function x(t).
You can see this by noting that if S < 0 for any perturbation x(t), then S[x + x] < S[x].
Correspondingly, if S > 0, then S[x x] < S[x]. In either case, we have found a function that
has lower action than x(t). Therefore x(t) can only be the minimum if S = 0.
The only way we can have S = 0 for every perturbation x(t) is that the expression inside
the brackets in Eq. (3.1.8) vanishes, i.e.,
d @L @L
= 0. (3.1.9)
dt @ ẋ @x
This is known as the Euler-Lagrange equation, and it is the equation of motion in the Lagrangian
formulation of mechanics. When using Eq. (3.1.9), it is very important to understand the differ-
ence between the partial (@) and total (d) derivatives.
To check that Eq. (3.1.9) really describes the same physics as Newtonian mechanics, let us
consider a simple example of a particle in a one-dimensional potential V (x). Because the system
is conservative, the Lagrangian is
1
L=T V = mẋ2 V (x), (3.1.10)
2
and the Euler-Lagrange equation is
d @L @L d @V @V
= (mẋ) + = mẍ + = 0. (3.1.11)
dt @ ẋ @x dt @x @x
This is nothing but Newton’s second law
@V
mẍ = . (3.1.12)
@x
It is interesting to note that although Newtonian and Lagrangian formulations of mechanics
are mathematically equivalent and describe the same physics, their starting point is very different.
Newton’s laws describe the evolution of the system as an initial value problem: We know the
position and velocity of the particle at the initial time, x(ta ) and ẋ(ta ), and we then use Newton’s
laws to determine the evolution x(t) at later times t > ta .
In contrast, the Lagrangian formulation describes the same physics as a boundary value prob-
lem. We know the initial and final positions of the particle x(ta ) and x(tb ), and we use the action
principle to determine x(t) for ta < t < tb , i.e., how the particle travels from one to the other. In
particular, we cannot choose the initial velocity because it is determined by the final destination
of the particle through the action principle. This may appear very non-local in time because the
behaviour of the particle in the far future determines its motion at the current time. However,

36
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

because because the two formulations are equivalent, this apparent non-locality in time does not
actually affect the physics. For example, it is not possible to use it to send information back in
time. In practice, it is usually easier to solve initial value problems, and therefore one usually
uses the Lagrangian formulation to set up the problem and derive the equations of motion but
then solves them as an initial value problem.
In many ways the Lagrangian formulation is closer to quantum mechanics, which does not
allow one to determine the initial position and velocity of the particle either. Furthermore, the
principle that the particle chooses one out of all possible trajectories resembles the double slit
experiment in quantum mechanics, with the key difference that in the quantum case one has to
sum over all possible trajectories rather than just selecting one. This correspondence turns out to
be fully accurate and becomes obvious in the path integral formulation in quantum mechanics.

3.2 Generalised Coordinates

One attractive aspect of the Lagrangian formulation is that it is independent of the variables that
are used to describe the state of the system. This is because the minimum of the function does
not depend on the coordinate system, and the same applies to a functional such as the action S.
Therefore, in contrast with Newtonian mechanics, we do not have to use the Cartesian position
coordinates, and the Euler-Lagrange equation still has the same form (3.1.9). Instead, we are
free to choose whichever set of variables we want to parameterise the state of the system, and
which are then called generalised coordinates and usually denoted by q. They can be position
coordinates, but also angles etc.
Usually we need more than one generalised coordinate, which we label by index i, so that we
have some number N generalised coordinates qi , with i = 1, . . . , N . Each coordinate satisfies
the corresponding Euler-Lagrange equation
d @L @L
= 0. (3.2.1)
dt @ q̇i @qi
As the first example of generalised coordinates, let us consider a simple pendulum that has a
mass m at the end of a light rod of fixed length l. The angle of the pendulum from the vertical
position is ✓, which we choose as the generalised coordinate q = ✓. The Lagrangian is
1
L=T V = ml2 ✓˙2 mgl(1 cos ✓). (3.2.2)
2
The Euler-Lagrange equation is
d @L @L d ⇣ 2 ˙⌘
= ml ✓ + mgl sin ✓ = ml2 ✓¨ + mgl sin ✓ = 0, (3.2.3)
dt @ ✓˙ @✓ dt
from which we find
g
✓¨ = sin ✓. (3.2.4)
l

37
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

As a slightly more complex example, let us consider the motion of a particle in a central
potential V (r) in three dimensions. We use the spherical coordinates (r, ✓, ) as the generalised
coordinates. The Lagrangian is
1 ⇣ ⌘
L=T V = m ṙ2 + r2 ✓˙2 + r2 ˙2 sin2 ✓ V (r). (3.2.5)
2
The Euler-Lagrange equation for r is
d @L @L d ⇣ ⌘ dV
= (mṙ) mr ✓˙2 + ˙ 2 sin2 ✓ + = 0, (3.2.6)
dt @ ṙ @r dt dr
which gives
⇣ ⌘
dV
mr̈ = mr ✓˙2 + ˙ 2 sin2 ✓ . (3.2.7)
dr
The second term on the right hand side is the force due to the potential, and the first term is the
centrifugal force.
The Euler-Lagrange equation for ✓ is
d @L @L d ⇣ 2 ˙⌘
= mr ✓ mr2 ˙ 2 sin ✓ cos ✓ = 0. (3.2.8)
dt @ ✓˙ @✓ dt

Finally, because does not appear in the Lagrangian (except as a time derivative ˙ ), the Euler-
Lagrange equation for is
d @L @L d @L
= = 0. (3.2.9)
dt @ ˙ @ dt @ ˙
This means that the quantity
@L
= mr2 sin2 ✓ ˙ (3.2.10)
@ ˙
is conserved. We note that this is simply Lz , the z component of the angular momentum vector
L = mr ⇥ ṙ.
We can easily see that this is, in fact, a very general result. For any generalised coordinate qi ,
we define the generalised momentum pi by
@L
pi = . (3.2.11)
@ q̇i
The Euler-Lagrange equation implies that whenever the Lagrangian does not depend on qi , then
the corresponding generalised momentum pi is conserved. This is a simple example of a more
general result known as Noether’s theorem, which we will come back to later.
As a very simple example, let us consider the Cartesian position coordinate x as our gener-
alised coordinate. With the Lagrangian L = 12 mẋ2 V (x), the generalised momentum is simply
the conventional momentum p = @L/@ ẋ = mẋ.

38
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

3.3 Precession of a Symmetric Top

Let us now consider an example of the use of Lagrangian mechanics to solve a real problem: a
symmetric top. The kinetic energy is given in terms of Euler angles by Eq. (2.8.4) whereas the
gravitational potential energy is V = M gR cos ✓. This leaves us with a Lagrangian
⇣ ⌘2
L = 12 I1 ˙ 2 sin2 ✓ + 12 I1 ✓˙2 + 12 I3 ˙ + ˙ cos ✓ M gR cos ✓ . (3.3.1)

The Euler-Lagrange equation (3.2.1) for ✓ is

d ⇣ ˙⌘ ⇣ ⌘
I1 ✓ = I1 ˙ 2 sin ✓ cos ✓ I3 ˙ + ˙ cos ✓ ˙ sin ✓ + M gR sin ✓ . (3.3.2)
dt
The Lagrangian function (3.3.1) does not contain the other two Euler angles and so the
generalised momenta p = @L/@ ˙ and p = @L/@ ˙ are constant
d h ˙ 2 ⇣
˙ ˙
⌘ i
I1 sin ✓ + I3 + cos ✓ cos ✓ = 0 (3.3.3)
dt
d h ⇣ ⌘i
I3 ˙ + ˙ cos ✓ = 0. (3.3.4)
dt
Note that comparison of Eqs. (3.3.4) and (2.8.2) tells us that

!3 = ˙ + ˙ cos ✓ = constant. (3.3.5)

We are interested in the situation of steady precession at a constant angle ✓. In this case
we conclude from Eqs. (3.3.3) and (3.3.4) that ˙ and ˙ are constant. Hence the axis of the top
precesses around the vertical with a constant angular velocity, which we denote by ⌦, i.e., ˙ = ⌦.
Because we are looking for a solution with fixed ✓, the left side of Eq. (3.3.2) must vanish, and
we obtain
I1 ⌦2 cos ✓ I3 !3 ⌦ + M gR = 0 , (3.3.6)
which we can solve for ⌦. The general solution is
p
I3 !3 ± I32 !32 4I1 M gR cos ✓
⌦= . (3.3.7)
2I1 cos ✓
This only has real roots if
4I1 M gR cos ✓
!32 !c2 ⌘ . (3.3.8)
I32
If the top is spinning more slowly than this, there is no solution with constant ✓. Instead, the top
starts to wobble.
For a rapidly spinning top, !3 !c , we can expand the square root in Eq. (3.3.7) to obtain
⇣ ⌘
I1 M gR cos ✓
I3 !3 ± I3 !3 1 2 I 2 !2 ⇢
M gR/I3 !3 sign
⌦⇡ 3 3
! (3.3.9)
2I1 cos ✓ I3 !3 /I1 cos ✓ +sign

39
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

The first of these is the precession frequency calculated in (2.6.6), while the second is the preces-
sion of a free system discussed in section 2.8. Note the absence of any contribution from gravity
in the second expression.
Note that if ✓ > 12 ⇡, the top is hanging below its point of support, and there is no limit on
!3 . In particular, for !3 = 0, we find the possible angular velocities of a compound pendulum
swinging in a circle s
M gR
⌦=± (3.3.10)
I1 |cos ✓|

3.4 Constraints
Consider a system of N particles in three dimensions. To specify the position of each particle,
you need 3N generalised coordinates. However, in many cases the coordinates are not all inde-
pendent but subject to some constraints, such as the rigidity conditions (see Section 2.2), which
reduce the number of generalised coordinates required. For a rigid body, the original 3N coordi-
nates may be reduced to six generalised coordinates: three translational, such as the coordinates
X, Y, Z of the centre of mass, and three rotational, such as the Euler angles, , ✓, .
We will assume that the constraints can be written in the form f (x1 , . . . , x3N , t) = 0. Con-
straints like that are called holonomic. For a rigid body, the constraints are of this form: The
distance between each pair of particles i and j is fixed to a constant dij , and therefore one has

(ri rj ) 2 d2ij = 0. (3.4.1)

Another example is motion on the surface of a sphere of radius R, for which the constraint is

f (x, y, z) = x2 + y 2 + z 2 R2 = 0. (3.4.2)

Sometimes one has to deal with non-holonomic constraints, for example if the constraint depends
on velocities, but these are more complicated to handle, and we will not discuss them in this
course.
In principle, solving each constraint equation eliminates one coordinate. If one has initially
N coordinates xi (with i = 1, . . . , N ) and C constraints, solving them will allow one to express
the original coordinates in terms of M = N C generalised coordinates qj , with j = 1, . . . , M ,

xi = xi (q1 , . . . , qM , t), (3.4.3)

with possibly explicit time-dependence if the constraints are time-dependent. In that case the
system is called forced, otherwise it is natural.
If one can solve the constraints and find the explicit relations (3.4.3), one can then write the
Lagrangian in terms of the generalised coordinates qj and solve the Euler-Lagrange equation.
Substituting this solution to Eq. (3.4.3) then gives the solution in terms of the original coordi-
nates.

40
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

An alternative approach, which is sometimes useful, is to implement the constraints us-

ing Lagrange multipliers. Starting with a Lagrangian L(x1 , . . . , xN ) and a constraint function
f (x1 , . . . , xN ), we define a new Lagrangian L0

L0 (x1 , . . . , xN , ) = L(x1 , . . . , xN ) + f (x1 , . . . , xN ), (3.4.4)

which is a function of the original coordinates and a Lagrange multiplier . If we now treat as
the (N + 1)th coordinate, its Euler-Lagrange equation is
d @L0 @L0
= f (x1 , . . . , xN ) = 0, (3.4.5)
dt @ ˙ @
and therefore it satisfies the constraint automatically. The Euler-Lagrange equations for the orig-
inal coordinates are
d @L @L @f
(t) = 0, (3.4.6)
dt @ ẋi @xi @xi
where the extra term can be interpreted as the (generalised) force that has to be applied to the
system to enforce the constraint.
As an example, consider a mass m hanging from a rope that is wrapped around a pulley of
radius R and moment of inertia I. Using the vertical position z of the mass, and the angle ✓ of
the pulley as the coordinates, the Lagrangian is
1 1
L=T V = mż 2 + I ✓˙2 mzg. (3.4.7)
2 2
If the rope does not slip, the pulley has to rotate as the mass moves, and this imposes a constraint
R✓˙ = ż. Choosing the origin appropriately, we can write this as a holonomic constraint

f (✓, z) = R✓ + z = 0. (3.4.8)

Introducing the Lagrange multiplier , the new Lagrangian is

1 1
L0 = L + (R✓ + z) = mż 2 + I ✓˙2 mzg + (R✓ + z). (3.4.9)
2 2
The Euler-Lagrange equations are
d
mż + mg = 0 for z,
dt
d ˙
I✓ R = 0 for ✓,
dt
(R✓ + z) = 0 for . (3.4.10)
The third equation implements the constraint ✓ = z/R, and substituting this to the first two
gives
mz̈ + mg = 0,
I
z̈ R = 0. (3.4.11)
R
41
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

Solving this pair of equations for z̈ and , we obtain

✓ ◆
I
m + 2 z̈ + mg = 0,
R
mg
= . (3.4.12)
1 + mR2 /I
The first line shows that the moment of inertia I of the pulley gives the mass extra inertia. The
second line gives the force that rope has to apply in order to enforce the constraint. This is just
the tension of the rope.

3.5 Normal Modes

3.5.1 Orthogonal Coordinates
Instead of rigid constraints, let us now consider a situation where the constraints are flexible so
that the particles can move around their equilibrium positions. We assume that the system is
described by N generalised coordinates qi . We also assume that it is natural, which means that
the kinetic energy is a quadratic homogeneous function of the generalised velocities. We can
then write it as
1X
T = aij (q1 , . . . , qN )q̇i q̇j , (3.5.1)
2 ij
where the coefficients aij can depend on the coordinates qi but not on velocities q̇i . They can
chosen to be symmetric (aji = aij ) without any loss of generality.
The coordinates are said to be orthogonal if there are no cross terms, i.e., aij = 0 if i 6= j.
Then the kinetic energy is simply
1X
T = aii (q1 , . . . , qN )q̇i2 . (3.5.2)
2 i

We can always make our coordinates orthogonal by using the Gram-Schmidt procedure. For
example, if N = 2, the general form of Eq. (3.5.1) is

T = 12 a11 q̇12 + a12 q̇1 q̇2 + 12 a22 q̇22 . (3.5.3)

Defining a new coordinate

a12
q10 = q1 + q2 , (3.5.4)
a11
the kinetic energy becomes
a212
T = 12 a11 q̇102 + 12 a022 q̇22 with a022 = a22 . (3.5.5)
a11
Furthermore, if we rescale the coordinates in Eq. (3.5.2) by
p
qi0 = aii qi , (3.5.6)

42
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

θ R

M
r
φ m

Figure 3.1: Double pendulum

the kinetic energy becomes

1 X 02
T = q̇ . (3.5.7)
2 i i
Therefore we can always assume that the kinetic energy has this form.
As an example, consider a double pendulum, with a second pendulum hanging from the first.
The kinetic energy is
1 2 1 ⇣ ⌘2 h i
T = M Ṙ + m Ṙ + ṙ = 12 M R2 ✓˙2 + 12 m R2 ✓˙2 + r2 ˙ 2 + 2Rr✓˙ ˙ cos(✓ ) (3.5.8)
2 2
where M , R and ✓ refer to the upper pendulum and m, r and to the lower. Note that the
kinetic energy of the lower pendulum depends not only on but also on the motion of the upper
pendulum to which it is attached. For small values of ✓ and we can set the cosine term to one,
so that the kinetic energy is

T = 12 (M + m)R2 ✓˙2 + 12 mr2 ˙ 2 + mRr✓˙ ˙ . (3.5.9)

However, because of the last term, the coordinates are not orthogonal. In this case it is obvious
that we could get an orthogonal set by simply considering the displacement of the 2 bobs. For
small angles this gives
x = R✓ y = R✓ + r , (3.5.10)
and the kinetic energy, T , becomes

T = 12 M ẋ2 + 12 mẏ 2 . (3.5.11)

Finally, we can reduce this to the standard form (3.5.7) using

p p
q1 = M x q2 = my . (3.5.12)

43
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

3.5.2 Small Oscillations

We now consider the potential energy, V . If T is given by (3.5.7), the Lagrangian is
1 X 02
L=T V = q̇ V (q1 , . . . , qN ). (3.5.13)
2 i i

The Euler-Lagrange equations are then simply

@V
q̈i = , (3.5.14)
@qi
for all i.
If we now assume that the amplitude of the oscillations is small, we can Taylor expand
the potential around the origin, which was chosen to correspond to the equilibrium state. To
quadratic order we have
X X
V (q1 , . . . , qN ) = V0 + bi qi + 12 kij qi qj + O(q 3 ), (3.5.15)
i ij

where V0 , bi and kij are constants.

Because we can subtract a constant from the potential without changing the equations of
motion (3.5.14), we are free to choose V0 = 0. Because we are assuming that the equilibrium
state is qi = 0 for all i, we find that bi = 0. Therefore, we only need to consider the quadratic
term X
V = 12 kij qi qj , (3.5.16)
ij

Note that we can also choose kij to be symmetric, i.e., kji = kij without any loss of generality.
With the potential (3.5.16), the Euler-Lagrange equation (3.5.14) is simply
X
q̈i = kij qj , (3.5.17)
j

which can be written in matrix form as

0 1 0 10 1
q1 k11 k12 · · · k1N q1
2 B q C B CB C
d B 2 C B k21 k22 · · · k2N CB q2 C
B . C = B .. .. .. .. CB .. C. (3.5.18)
dt2 @ .. A @ . . . . A@ . A
qN kN 1 kN 2 · · · kN N qN

By defining an N -dimensional coordinate vector q = (q1 , . . . , qN ) and an N ⇥ N matrix k with

elements kij , we can also write it more compactly as

q̈ = k · q. (3.5.19)

44
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

In the same notation, the Lagrangian is

1 1
L = q̇ · q̇ q · k · q. (3.5.20)
2 2
In our double pendulum example, the potential is

V (✓, ) = M gR(1 cos ✓) + mg [R(1 cos ✓) + r(1 cos )]

1 1
⇡ (M + m)gR✓2 + mgr 2
2 2
M + m 2 mg
= gx + (y x)2
✓2R 2r◆ r
1 m R+r g 2 g 2 mg
= 1+ q1 + q2 q1 q 2 . (3.5.21)
2 M r R 2r Mr
The coefficient matrix is therefore
✓ m R+r g p m g◆
1+p
k= M r
mg
R
g
M r . (3.5.22)
M r r

3.5.3 Eigenvalue Problem

Eq. (3.5.19) is a set of N coupled linear second-order equation. Therefore we should find 2N
linearly independent solutions.
We look for solutions of the form

q(t) = Aei!t , (3.5.23)

where A and ! are constants. We will check later that we have found all 2N solutions.
Substituting the Ansatz (3.5.23) into Eq. (3.5.19) gives

! 2 Aei!t = k · Aei!t , (3.5.24)

which is equivalent to
k · A = ! 2 A. (3.5.25)
This has the form of the eigenvalue equation: It shows that A is an eigenvector of the matrix k
with eigenvalue ! 2 .
We know from linear algebra that the eigenvalues of a matrix are given by solutions of the
characteristic equation

k11 ! 2 k12 ··· k1N

k21 k22 ! 2 · · · k2N
det(k ! 2 1) ⌘ .. .. .. .. = 0. (3.5.26)
. . . .
kN 1 kN 2 ··· kN N !2

45
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

Once we have found the eigenvalue ! 2 , we can substitute it to Eq. (3.5.25) to find the eigenvector.
Because k is a symmetric N ⇥ N matrix, it has N real eigenvalues !↵2 , where ↵ = 1, . . . , N ,
corresponding to N eigenvectors A↵ , which we can choose to be orthonormal,

A↵ · A = ↵ . (3.5.27)

For each eigenvalue !↵2 , there are two independent solutions,

q+
↵ (t) = A↵ e
i!↵ t
and q↵ (t) = A↵ e i!↵ t
. (3.5.28)

For N eigenvalues, this takes the total number of linearly independent solutions to 2N , proving
that we have the complete solution.
The frequencies !↵ are real because the eigenvalues !↵2 have to be non-negative: If we had
!↵ < 0, then for the corresponding eigenvector A↵ and small ✏ we would have
2

1 1
V (✏A↵ ) = ✏2 A↵ · k · A↵ = ✏2 !↵2 < 0, (3.5.29)
2 2
meaning that q = 0 could not be the minimum of the potential as we assumed.
The individual solutions q±
↵ are complex, so physical solutions have to be real linear combi-
nations of them,
q↵ (t) = a+ q+
↵ (t) + a q↵ (t) = a+ e
i!↵ t
+ a e i!↵ t A↵
= [(a+ + a ) cos !↵ t + i(a+ a ) sin !↵ t] A↵ . (3.5.30)
Because we want a real solution, the coefficients a1 ⌘ (a+ + a ) and a2 = i(a+ a ) have to
be real, and we can equally well write

q↵ (t) = (a1 cos !↵ t + a2 sin !↵ t) A↵ , (3.5.31)

which we can also write a pure cosine term with a phase shift,

q↵ (t) = c cos(!↵ t + )A↵ , (3.5.32)

where c and are real constants that have to be determined from the initial conditions.
Finally, the general solution is a linear combination of solutions of the form (3.5.32),
N
X
q(t) = c↵ cos(!↵ t + ↵ )A↵ . (3.5.33)
↵=1

The modes of vibration of the system, i.e., the individual solutions (3.5.32) are known as normal
modes. For a forced system there are resonances at the frequencies of the normal modes.
It is often useful to use the normal modes to define a set of generalised coordinates known
as the normal coordinates, which we denote by q̃↵ . They are defined by expressing the original
coordinates qi in terms of the eigenvectors A↵ as
X
q= q̃↵ A↵ . (3.5.34)
↵

46
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

Substituting this to Eq. (3.5.20) we find that the Lagrangian becomes simply
X ✓1 1 2 2
◆
L= ˙
q̃ 2
! q̃ . (3.5.35)
↵
2 ↵ 2 ↵ ↵

The Euler-Lagrange equations are

q̃¨↵ + !↵2 q̃↵ = 0, (3.5.36)
which means that each normal coordinate q̃↵ oscillates independently of all others with its own
normal frequency !↵ .

3.6 Continuous Systems

In addition to mechanical systems consisting of a finite number of degrees of freedom, one is
often also interested in continuous systems, for example waves propagation in continuous media
or field theories describing particle physics or electromagnetism.
To see how continuous systems are described in the Lagrangian formulation, consider a
stretched string. We assume that in equilibrium the string is stretched to length `0 and has tension
k. The displacement of the string from its equilibrium position is given by the continuous func-
tion y(x, t), where x 2 {0, `0 }. To describe the time and space derivatives, we use the notation
@y @y
ẏ = , y0 = . (3.6.1)
@t @x
The Lagrangian is still given by the difference of the kinetic and potential energies, L =
T V . The kinetic energy can be calculated by considering an infinitesimal segment of length
dx. The mass of such a segment is dm = µdx where the constant µ is the mass per unit length.
The velocity of the segment is simply ẏ, and therefore the kinetic energy of the infinitesimal
segment is
1
dT = µ dx ẏ 2 . (3.6.2)
2
Integrating over the who distance `0 , we find the total kinetic energy
Z `0
1
T = dx µ ẏ 2 . (3.6.3)
0 2
The potential energy is V of the string is due to its tension k,

V = k(` `0 ), (3.6.4)

where ` is the length of the displaced string. Again, this can be calculated by considering an
infinitesimal segment of length dx. According to Pythagoras theorem, the length of the segment
is ✓ ◆
p p 1 02
d` = dx2 + dy 2 = 1 + y 02 dx ⇡ 1 + y dx, (3.6.5)
2

47
Advanced Classical Physics, Autumn 2016 Lagrangian Mechanics

where we have assumed that the displacement is small and smooth so that y 0 ⌧ 1, and Taylor
expanded to quadratic order. The length of the string is then obtained by summing over all the
infinitesimal segments
Z `0 ✓ ◆ Z `0
1 02 1
`= dx 1 + y = `0 + dx y 02 , (3.6.6)
0 2 0 2

and therefore the potential energy is

Z `0
1
V = dx ky 02 . (3.6.7)
0 2
We can now write the whole Lagrangian,
Z `0 ✓ ◆
1 2 1 02
L=T V = dx µẏ ky . (3.6.8)
0 2 2
R
The integrand is called the Lagrangian density and denoted by L, i.e., L = dxL, where

1 1 02
L = µẏ 2 ky . (3.6.9)
2 2
The action is then given by an integral over both time and space,
Z Z
S = dt dxL(ẏ, y 0 , y). (3.6.10)

Variation of the action is

Z Z  Z Z 
@L @L 0 @L d @L d @L @L
S = dt dx ẏ + 0 y + y = dt dx + y.
@ ẏ @y @y dt @ ẏ dx @y 0 @y
(3.6.11)
The action principle S = 0 therefore leads to the Euler-Lagrange equation
d @L d @L @L
+ = 0. (3.6.12)
dt @ ẏ dx @y 0 @y
Substituting the Lagrangian density (3.6.9) for the string, we find the equation of motion
d d
(µẏ) + ( ky 0 ) = µÿ ky 00 = 0, (3.6.13)
dt dx
which is the wave equation, as one would expect.

48