MATHEMATICS 1302 (Applied Mathematics 2)

YEAR 2008–2009, TERM 2

HANDOUT #8: COUPLED OSCILLATIONS AND NORMAL MODES

1 Review of simple harmonic oscillator

In MATH 1301 you studied the simple harmonic oscillator : this is the name given to any
physical system (be it mechanical, electrical or some other kind) with one degree of freedom
(i.e. one dependent variable x) satisfying the equation of motion
mẍ = −kx , (1)
where m and k are constants (and the dot ˙ denotes d/dt as usual). For instance, if we have
a particle of mass m attached to a spring of spring constant k (with the other end of the
spring attached to a fixed wall), then the force on the particle is F = −kx where x is the
particle’s position (with x = 0 taken to be the equilibrium point of the spring), so Newton’s
Second Law F = ma is indeed (1).
Let us review briefly the solution of the harmonic-oscillator equation (1). Since this is
a one-dimensional problem with a position-dependent force, it can be solved by the energy
method, with potential energy U (x) = 21 kx2 . [This was done as a worked example in one
of the Kleppner & Kolenkow handouts.] But a simpler method is to recognize that (1) is
a homogeneous linear differential equation with constant coefficients, so its solutions can
be written (except in certain degenerate cases) as linear combinations of suitably chosen
exponentials, which we can write either as x(t) = eαt or as x(t) = eiωt . Let us use the latter
form (which is more convenient for oscillatory systems, because ω will come out to be a real
number). So the method is to guess a solution of the form
x(t) = eiωt (2)
and then choose ω so that this indeed solves (1). Inserting x(t) = eiωt into (1), we find
−mω 2 eiωt = −keiωt , (3)
p
which is a solution if (and only if) ω = ± k/m. We conclude that the general solution of
(1) is
x(t) = Aeiωt + Be−iωt (4)
p
with ω = k/m.1 This can equivalently be written in “sine-cosine” form as
x(t) = C1 cos ωt + C2 sin ωt (5)
or in “amplitude-phase” form as
x(t) = C cos(ωt + φ) . (6)
1
This is wrong when ω = 0, because then the solutions eiωt and e−iωt are not linearly independent. In
this case the linearly independent solutions are eiωt (i.e. 1) and teiωt (i.e. t). This is the “degenerate case”
I referred to earlier.

1
2 Coupled oscillations: A simple example
Now let us consider a simple situation with two degrees of freedom. Suppose we have
two particles, of masses m1 and m2 , respectively, connected as follows:

Let us assume for simplicity that all three springs have the same spring constant k, and that
the distance between the walls is exactly the sum of the equilibrium lengths of the three
springs. Then the equations of motion are

m1 ẍ1 = −kx1 + k(x2 − x1 ) (7a)

m2 ẍ2 = −kx2 + k(x1 − x2 ) (7b)

(You should check this carefully and make sure you understand the signs on all four forces.)
Let us now try a solution of the form

x1 (t) = A1 eiωt (8a)

x2 (t) = A2 eiωt (8b)

where A1 and A2 are constants. Substituting this into (7) yields

−m1 ω 2 A1 eiωt = −2kA1 eiωt + kA2 eiωt (9a)

−m2 ω 2 A2 eiωt = kA1 eiωt − 2kA2 eiωt (9b)

Extracting the common factor eiωt and moving everything to the right-hand side, we obtain

(m1 ω 2 − 2k)A1 + kA2 = 0 (10a)

kA1 + (m2 ω 2 − 2k)A2 = 0 (10b)

which is most conveniently written in matrix form as

    
m1 ω 2 − 2k k A1 0
2 = . (11)
k m2 ω − 2k A2 0
This is a homogeneous linear equation; it has a nonzero solution (i.e. a solution other than
A1 = A2 = 0) if and only if the matrix on the left-hand side is singular, i.e. has a zero
determinant. Setting the determinant equal to zero gives a quadratic equation for ω 2 , namely

(m1 ω 2 − 2k)(m2 ω 2 − 2k) − k 2 = 0 , (12)

which can be solved by the quadratic formula. Then, for each of the two possible  values
 for
A 1
ω 2 , we can go back to the linear equation (11) and solve for the “eigenvector” .
A2

2
 In particular, in the “symmetric case” m1 = m2 = m, the solutions are
p
ω1 = k/m (13a)
p
ω2 = 3k/m (13b)

and the corresponding eigenvectors are

 
1
e1 = (14a)
1
 
1
e2 = (14b)
−1

The normal modes — that is, the solutions of (7) that are pure oscillations at a single
frequency — are therefore
   
x1 (t) 1 p
= C1 cos( k/m t + φ1 ) (15)
x2 (t) 1

and    
x1 (t) 1 p
= C2 cos( 3k/m t + φ2 ) . (16)
x2 (t) −1
The general solution is a linear combination of these two normal modes. In the first (slower)
normal mode, the two particles are oscillating in phase, with thepsame amplitude; the middle
spring therefore exerts no force at all, and the frequency is k/m as it would be if the
middle spring were simply absent. In the second (faster) normal mode, the two particles are
oscillating 180◦ out of phase, with the same amplitude; therefore,peach particle feels a force
that is −3k times its displacement (why?), and the frequency is 3k/m.
This solution can be interpreted in another way. Let us build a matrix N whose columns
are the eigenvectors corresponding to the normal modes,
 

1 1
N = e1 e2 = , (17)
1 −1

and let us make the change of variables

   0
x1 x1
= N (18)
x2 x02

or equivalently

x1 = x01 + x02 (19a)

x2 = x01 − x02 (19b)

Then a solution with x01 6= 0, x02 = 0 corresponds to the first normal mode, while a solution
with x01 = 0, x02 6= 0 corresponds to the second normal mode. The point is that the change

3
of variables (18)/(19) decouples the system (7) [when m1 = m2 = m]: after a bit of algebra
we obtain
mẍ01 = −kx01 (20a)
mẍ02 = −3kx02 (20b)
(You should check this!) In other words, by a linear change of variables corresponding to
passage to the normal modes, the system (7) of coupled harmonic oscillators turns into a
system (20) of decoupled simple harmonic oscillators, each of which may be solved separately
by the elementary method reviewed in Section 1.

3 Coupled oscillations: The general case

We can now see how to handle the general case of coupled oscillators with an arbitrary
finite number n of degrees of freedom. We will have a system of homogeneous linear constant-
coefficient differential equations of the form
M ẍ + Kx = 0 (21)
where
 
x1
 x2 
• x =  ..  is a column vector of coordinates;
 
 . 
xn

• M (the so-called mass matrix) is a symmetric positive-definite n × n real matrix

(usually it will be a diagonal matrix, but it need not be);
• K (the so-called stiffness matrix) is a symmetric n × n real matrix (usually it too
will be positive-definite, but it need not be); and
• 0 denotes the zero vector.
We then try a solution of the form
x(t) = e eiωt (22)
where e is some fixed vector. This will solve (21) if (and only if)
(K − ω 2 M )e = 0 (23)
or equivalently
Ke = ω 2 M e . (24)
This is a generalized eigenvalue problem (it would be the ordinary eigenvalue problem
if M were the identity matrix). The eigenvalues — that is, the values of ω 2 for which there
exists a solution e 6= 0 — are the solutions of the nth-degree polynomial equation
det(K − ω 2 M ) = 0 , (25)

4
and then the corresponding vectors e are the eigenvectors. The solution x(t) = e e iωt [or
x(t) = e cos ωt] corresponding to such an eigenpair (ω, e) is a normal mode.
A generalized eigenvalue problem with a pair of real symmetric matrices, at least one of
which is positive-definite, always has a basis of eigenvectors — that is, we can always find
eigenvalues λ1 , . . . , λn and corresponding eigenvectors e1 , . . . , en such that

(a) Kej = λj M ej for j = 1, . . . , n; and

(b) {e1 , . . . , en } is a basis of Rn .

Let us prove this as follows:

Lemma 1 Every symmetric positive-definite real matrix has a symmetric positive-definite

square root. That is, if M is a symmetric positive-definite real n × n matrix, then there exists
another symmetric positive-definite real n × n matrix, which we shall denote M 1/2 , such that
M 1/2 M 1/2 = M . (In fact this matrix M 1/2 is unique, but we shall not need this fact.)

Proof. Basic linear algebra tells us that any symmetric real matrix can be diagonalized by
an orthogonal transformation: that is, there exists a matrix R satisfying R T R = RRT = I
(that is the definition of “orthogonal matrix”) and RT M R = D, where D is a diagonal ma-
trix. In fact, D = diag(m1 , . . . , mn ), where m1 , . . . , mn are the eigenvalues of M . In our case,
the matrix M is positive-definite, so all the eigenvalues m1 , . . . , mn are > 0; in particular,
1/2 1/2
they have square roots. We can therefore define the matrix D 1/2 = diag(m1 , . . . , mn ),
which manifestly satisfies D 1/2 D 1/2 = D.
Now define M 1/2 = RD 1/2 RT . We have

M 1/2 M 1/2 = RD 1/2 RT RD 1/2 RT

= RD 1/2 D 1/2 RT
= RDRT
= R(RT M R)RT
= M

where we used RT R = I, then D 1/2 D 1/2 = D, then D = RT M R, and finally RRT = I. (You
should check this carefully!)
I leave it to you to verify that M 1/2 = RD 1/2 RT is symmetric.
Finally, since M 1/2 = RD 1/2 RT with D 1/2 positive-definite and R nonsingular, it follows
that M 1/2 is positive-definite as well. (You should go back to the definition of “positive-
definite matrix” and verify this assertion too.) 

Let us now show that a pair of real quadratic forms, one of which is positive-definite, can
be simultaneously diagonalized:

Proposition 2 Let M and K be a real symmetric n × n matrices, with M positive-definite.

Then there exists a nonsingular real n × n matrix N such that N T M N = I (the identity
matrix) and N T KN = Λ, where Λ is a diagonal matrix.

5
Proof. Let M 1/2 be the symmetric positive-definite square root of M whose existence
is guaranteed by the Lemma. Since M 1/2 is positive-definite, it is invertible. Then L =
(M 1/2 )−1 K(M 1/2 )−1 is a real symmetric matrix, so it can be diagonalized by an orthogonal
transformation: that is, there exists a matrix R satisfying R T R = RRT = I and RT LR = Λ,
where Λ is a diagonal matrix. Now define N = (M 1/2 )−1 R. We have N T = RT (M 1/2 )−1
(why?). Then N T KN = Λ by construction (why?), and

N T M N = RT (M 1/2 )−1 M (M 1/2 )−1 R

= RT (M 1/2 )−1 M 1/2 M 1/2 (M 1/2 )−1 R
= RT R
= I.

Corollary 3 Let M and K be a real symmetric n × n matrices, with M positive-definite.

Then there exist real numbers λ1 , . . . , λn and vectors e1 , . . . , en ∈ Rn such that

(a) Kej = λj M ej for j = 1, . . . , n; and

(b) {e1 , . . . , en } is a basis of Rn .

Proof. Let N be the matrix whose existence is guaranteed by the Proposition, and let
e1 , . . . , en be its columns. Since N is nonsingular, its columns are linearly independent,
hence form a basis of Rn . Obviously N T is also nonsingular, hence invertible, and the
Proposition tells us that

M N = (N T )−1
KN = (N T )−1 Λ

(why?). Here Λ is a diagonal matrix; let λ1 , . . . , λn be its diagonal entries. Now, the jth
column of M N is M ej (why?), so the jth column of (N T )−1 is M ej . It follows that the jth
column of (N T )−1 Λ is λj M ej (why?). Since the jth column of KN is Kej , this proves that
Kej = λj M ej . 

This manipulation of matrices is quick but perhaps a bit abstract. Here is a more direct
proof of the Corollary that analyzes directly the generalized eigenvalue problem:

Alternate proof of Corollary 3. Let M 1/2 be the symmetric positive-definite square

root of M whose existence is guaranteed by the Lemma. We can then rewrite

Ke = λM e (26)

as
Ke = λM 1/2 M 1/2 e . (27)

6
Defining f = M 1/2 e, we have e = (M 1/2 )−1 f [note that M 1/2 is invertible because it is
positive-definite] and hence the equation can be rewritten as
K(M 1/2 )−1 f = λM 1/2 f . (28)
And we can left-multiply both sides by (M 1/2 )−1 to obtain
(M 1/2 )−1 K(M 1/2 )−1 f = λf (29)
[note that this operation is reversible because (M 1/2 )−1 is invertible]. So we now have an
ordinary eigenvalue problem for the symmetric real matrix (M 1/2 )−1 K(M 1/2 )−1 . This matrix
has eigenvalues λ1 , . . . , λn and a corresponding basis of linearly independent eigenvectors
f1 , . . . , fn . Defining ej = (M 1/2 )−1 fj , a simple calculation shows that
Kej = λj M ej for j = 1, . . . , n . (30)
And {e1 , . . . , en } is a basis of Rn because {f1 , . . . , fn } is a basis of Rn and the matrix (M 1/2 )−1
is nonsingular. 

4 Another example: n masses with springs

Now let us generalize the example of Section 2 by considering a chain of n particles, each
of mass m, joined by n + 1 springs, each of spring constant k, between a pair of walls. Then
the equations of motion are
mẍ1 = −kx1 + k(x2 − x1 ) (31a)
mẍ2 = k(x1 − x2 ) + k(x3 − x2 ) (31b)
mẍ3 = k(x2 − x3 ) + k(x4 − x3 ) (31c)
..
.
mẍn−1 = k(xn−2 − xn−1 ) + k(xn − xn−1 ) (31d)
mẍn = k(xn−1 − xn ) − kxn (31e)
(You should check this carefully and make sure you understand this, including all the signs!)
This system of differential equations can be written compactly in the form
M ẍ + Kx = 0 (32)
where M = mI and the matrix K has entries 2k on the diagonal and −k just above and
below the diagonal (and zero entries everywhere else): that is, K = kL where
 
2 −1
 −1 2 −1
 

 
 −1 2 −1 
L =  (33)
 
.. .. .. 

 . . . 

−1 2 −1
 
 
−1 2

7
is called the one-dimensional discrete Laplace matrix (with Dirichlet boundary condi-
tions at the endpoints).2 The eigenvalues λ = ω 2 of our generalized eigenvalue problem are
simply k/m times the eigenvalues of the matrix L.
How can we find the eigenvalues and eigenvectors of L? This is not so obvious; it requires
a bitof cleverness.
 Let us start by observing that the equation Lf = µf for the eigenvector
f1
 f2 
f =  ..  can be written in the simple form
 
 . 
fn

−fs−1 + 2fs − fs+1 = µfs for s = 1, . . . , n (34)

if we simply define f0 = 0 and fn+1 = 0. (Why?) Now this is a linear constant-coefficient

difference equation; and by analogy with linear constant-coefficient differential equations,
we might expect the solutions of (34) to be linear combinations of (complex) exponentials,
e.g.
fs = eiαs (35)
where we can always take −π < α ≤ π (why?). Plugging the guess (35) into (34), we see
that this guess indeed solves (34) provided that α and µ are related by

2 − 2 cos α = µ . (36)

(You should check this carefully!) In particular we must have 0 ≤ µ ≤ 4. Note that to each
allowed value of µ there corresponds a pair of allowed values of α — namely, a value α > 0
and its negative — because cos is an even function.3 So any linear combination of the two
solutions eiαs and e−iαs is also a solution for the given value of µ; in particular, any linear
combination of sin(αs) and cos(αs) is a solution.
But we are not done yet: we have solved the difference equation (34), but we have not
yet dealt with the “boundary conditions” f0 = 0 and fn+1 = 0. The condition f0 = 0 can be
satisfied simply by choosing the solution

fs = sin(αs) (37)

(why?). And the condition fn+1 = 0 can then be satisfied by making sure that (n + 1)α is a
multiple of π, i.e.
jπ
α = for some integer j . (38)
n+1
2
L is called the discrete Laplace matrix because it is the discrete analogue of the Laplacian operator
∂2 ∂2
−d2 /dx2 in one dimension or − 2 − · · · − in n dimensions. Indeed, if you were to try to solve on
∂x1 ∂x2n
the computer a differential equation involving the operator −d2 /dx2 , you would probably discretize space
(i.e. replace continuous space by a mesh of closely spaced points) and replace the operator −d 2 /dx2 by the
matrix L or something similar.
3
There are two exceptions: µ = 0 corresponds only to α = 0, and µ = 4 corresponds only to α = π.

8
We can’t take j = 0, because that would make f identically zero; but we can take any integer
j from 1 up to n. We have thus obtained eigenvectors f1 , . . . , fn for the matrix L, given by
 πjs 
(fj )s = sin , (39)
n+1
with corresponding eigenvalues
 πj 
µj = 2 − 2 cos (40a)
n+1
 πj 
2
= 4 sin . (40b)
2(n + 1)
Since there are n of these and they are linearly independent (they must be because the values
µj are all different!), we conclude that we have found a complete set of eigenvectors for the
matrix L.4
Physically, these eigenvectors are standing waves. To see this, let us make some plots
for n = 5 and j = 1, 2, 3, 4, 5. For each value of j, we first plot the function
 πjs 
fj (s) = sin (41)
n+1
for real values of s in the interval 0 ≤ s ≤ n + 1 (this shows most clearly the “standing
wave”); then we indicate the points corresponding to s = 1, 2, . . . , n, which are the entries
in the eigenvector fj .
n=5, j=1
1.0

0.5

0.0
1 2 3 4 5 6

-0.5

-1.0

n=5, j=2
1.0

0.5

0.0
1 2 3 4 5 6

-0.5

-1.0

4
It follows, in particular, that nothing new is obtained by going outside the range 1 ≤ j ≤ n. For instance,
j = n + 1 again yields the zero function, j = n + 2 yields a multiple of what j = n yields, j = n + 3 yields a
multiple of what j = n − 1 yields, and so forth.

9
 n=5, j=3
1.0

0.5

0.0
1 2 3 4 5 6

-0.5

-1.0

n=5, j=4
1.0

0.5

0.0
1 2 3 4 5 6

-0.5

-1.0

n=5, j=5
1.0

0.5

0.0
1 2 3 4 5 6

-0.5

-1.0

Next week we will look at the limit n → ∞ of this problem — namely, waves on a string
of length L — and we will find standing-wave solutions corresponding to the functions
 πjs 
fj (s) = sin (42)
L
where s is now a real number satisfying 0 ≤ s ≤ L, and j is now an arbitrarily large positive
integer.

10