11 Normal Modes -- Matrix Methods Fall 2003

Matrix methods provide a very powerful tool for analyzing normal modes of coupled
oscillator systems. We'll introduce this general approach by means of an example, the
same system we discussed in detail in Section 8 of these notes. Our development will
parallel the previous treatment, but we'll use matrix language.

Example

For the system discussed in Section 8, the equations of motion (from ΣF = ma) are
mx&&1 = − ( k + k ' ) x1 + k ' x 2 ,
(1)
mx&&2 = k ' x1 − ( k + k ' ) x2 .
The equations of motion can be written as a single matrix equation:

FG m 0 IJ FG &&x IJ
1
= −
FG k + k ' −k ' IJ FG x IJ .
1
H 0 mK H &&x K2 H −k ' k + k' K Hx K
2
(2)

Each side of this equation is a column matrix. You should verify that when the matrix
products are carried out, equating the first (top) elements on the two sides gives the first
of Eqs. (1), and equating the second (bottom) elements gives the second.
We define the matrices M, K, and x as follows:

M =
FG m 0 IJ , K =
FG k + k ' − k' IJ x=
FG x IJ .
1
H 0 mK H −k ' k + k'
,
K Hx K
2
(3)

Then Eq. (2) can be written simply as

Mx&& = −Kx. (4)

For other systems with different arrangements of masses and springs, the M and K
matrices will be different. But if the spring forces are linear functions of the coordinates,
the equations of motion can always be written in the form of Eq. (4). So the following
development is general, and is not restricted to the specific example cited above.

Note that the (1,1) element of K represents the negative of the force on mass 1 when it
is displaced a distance x 1 . The (1,2) element is the negative of the force on mass 1
when mass 2 is displaced a distance x 2 , and so on. This gives an alternative way to
determine the elements of K, instead of working out Eqs. (1) from ΣF = ma.

Caution: Some books define K with the opposite sign, so there is no minus in Eq. (4).
(See, for example, Edwards and Penney, p. 321.) We prefer to retain the minus sign, so
that Eq. (4) has the same form as the equation of motion mx&& = −kx for a single
particle.
11-2 11 Normal Modes -- Matrix Methods

General Formulation

For a normal mode, each x must vary sinusoidally, and all x's must have the same
frequency. Therefore we try to find a solution of Eq. (4) in the form

FG x IJ = C FG a IJ cosbω t + ϕg,
1 1
b g
x = Ca cos ω t + ϕ .
H x K Ha K
2 2
or (5)

In this expression, C is a scalar amplitude factor (determined by initial conditions), and

a is a column matrix (or vector) whose elements a1 , a2 give the ratios of the amplitudes
of the x's for each normal mode. For example, if a2 = 2a1 , the motion of the mass
with coordinate x 2 is in phase with that of x 1 , but with amplitude twice as great.

Taking the second time derivative of Eq. (5), we get

b g b
x&& = − ω 2 Ca cos ω t + ϕ = − λCa cos ω t + ϕ , g where λ = ω 2. (6)
To test whether Eq. (5) really is a solution of Eq. (4), we substitute Eqs. (5) and (6)
into Eq. (4). After dividing out the common factor C cos (ωt + ϕ), we get

− λMa = − Ka, or bK − λ Mga = 0. (7)

This is a set of simultaneous, homogeneous equations for the elements of the column
matrix a. (This should sound familiar!) Non-trivial equations exist if, and only if, the
determinant of the system is zero, that is, when
K − λ M = 0. (8)
Thus Eq. (5) is a solution of Eq. (4) if, and only if, the value of λ = ω2 satisfies Eq. (8).
This condition is called the secular equation for the system; it is satisfied only for certain
particular values of λ, the roots of the secular equation.

For our example, we use the M and K matrices from Eqs. (3). The corresponding
secular equation is:
k + k ' − mλ − k'
= 0. (9)
−k ' k + k ' − mλ
This agrees with Eq. (4) in Section 8 of these notes (page 8-2).

Equation (8) is an algebraic equation for λ. The degree of the equation (and hence the
number of roots) is equal to the number of coordinates of the system (i.e., the number of
degrees of freedom). In some problems two or more roots may be equal.

In our example, there are two roots:

k k + 2k '
λ1 = , λ2 = . (10)
m m
11 Normal Modes -- Matrix Methods 11-3

The values of λ that satisfy Eq. (8) are called eigenvalues. There are as many values of
λ as the number of degrees of freedom of the system. For each value of λ, there is a
corresponding column matrix (or vector), giving the amplitude ratios for the x's for the
corresponding normal mode. We'll call the eigenvalues λ1 , λ2 , λ3 , L ; let a be the
column matrix corresponding to λ1 , let b be the column matrix corresponding to λ2 ,
and so on. The column matrices a, b, c, L are the eigenvectors corresponding to the
eigenvalues λ1 , λ2 , λ3 , L , respectively. Then

bK − λ Mga = 0,
1 bK − λ Mgb = 0,
2 bK − λ Mgc = 0,
3 L. (11)

The most general solution of Eq. (4) can now be written as

b
x = C1 a cos ω1 t + ϕ1 g b
+ C2 b cos ω 2 t + ϕ 2 g b g
+ C3 c cos ω 3 t + ϕ 3 + L . (12)

The constants C1 , C2 , C3 , L and ϕ1 , ϕ2 , ϕ3 L depend on the initial positions and

velocities of the two masses), and the normal-mode frequencies are
ω1 = λ1 , ω2 = λ2 , L .
As before, we define the normal coordinates for the system as

q1 = C1 cos(ω1 t + ϕ1 ), q2 = C2 cos(ω2t + ϕ2 ), q3 = C3 cos(ω3t + ϕ3 ), L . (13)

Then the normal-coordinate transformation can be written as

x = q1 a + q2 b + q3 c + L . (14)

Now we define q to be a column matrix whose elements are the normal coordinates
q1 , q2 , q3 , L , and we define a square matrix A whose columns are the eigenvectors
a, b, c, L . For two degrees of freedom,

q=
FG q IJ
1
A=
FG a1 IJ
b1
Hq K
2
and
Ha 2 b2K. (15)

The normal coordinate transformation then becomes simply

x = Aq. (16)
For our example,

x=
FG x IJ ,
1
q=
FG q IJ ,
1
A=
FG1 1 IJ ,
Hx K
2 Hq K
2 H1 −1K
which agrees with the results at the end of Section 8 (page 8-5). (But note that there is a
lot of arbitrariness in the definition of A, because each q can be multiplied by an
arbitrary numerical constant.)
11-4 11 Normal Modes -- Matrix Methods

This whole procedure is closely related to the discussion of eigenvalues and eigenvectors
in Section 10 of these notes. Suppose we multiply both sides of Eq. (7) on the left by
M −1 . The result is

(M −1 K − λΜ
Μ −1Μ )a = 0, or (M−1 K)a = λa. (17)

That is, a is an eigenvector of the matrix (M −1 K) , with eigenvalue λ1 . Similarly, from

Eq. (11), b is an eigenvector of (M −1 K) with eigenvalue λ2 , and so on.

If our objective is only to find the normal-mode frequencies and vibration patterns, we
are finished. Each eigenvalue λ is the square of a normal-mode frequency ω, and each
corresponding eigenvector gives the relative amplitudes of motion of the various masses
for the normal-mode motion with this frequency.

However, there are problems in which we need to determine the constants C1 , C2 , C3 ,

L and ϕ1, ϕ2, ϕ3, L in Eq. (12) in terms of the initial values of the coordinates and
velocities. This would be fairly simple if the eigenvectors a, b, c, L were all mutually
orthogonal. We could then multiply Eq. (12) successively by the transpose of each
eigenvector, and all but one term on the right would be zero. We stated in Section 10
(page 10-4) that the eigenvectors of a symmetric matrix are all mutually orthogonal if
they correspond to distinct eigenvalues.

In our example problem, the matrices M, M −1 , K, and M −1 K are all symmetric, and
the eigenvectors are orthogonal. But this is an accident resulting from the symmetry of
our physical system and our choice of coordinates. In general M and K are not
symmetric; even when they are symmetric, the product M −1 K in general is not
symmetric. So in general the eigenvectors a, b, c, L are not mutually orthogonal.
But it turns out that we can develop generalized concepts of normalization and
orthogonality that are useful for fitting initial conditions and in other situations.

Normalization of Eigenvectors

We first generalize the concept of normalization. Because the eigenvectors a and b

serve only to give the ratios of the amplitudes of motion of the various masses, for each
normal mode, we are free to multiply each eigenvector by any constant. It turns out to be
useful to multiply each eigenvector by a constant such that

aTMa = 1, bTMb = 1, cTMc = 1, and so on. (18)

(The more usual normalization condition is simply to require that aT a = 1, bT b = 1, and

so on, or, a1 2 + a 2 2 + a3 2 + L = 1 , and so on.) When the eigenvectors are multiplied by
factors such that they satisfy Eqs. (18), they are said to be normalized, in this
generalized sense.
11 Normal Modes -- Matrix Methods 11-5

In our example, where

a =
FG1IJ M =
FG m 0 IJ ,
H1K and
H 0 mK
the un-normalized a gives aT Ma = 2m. We multiply a by 1 2m; then aT Ma = 1.
So in our example problem, the normalized eigenvector a corresponding to the
eigenvalue λ1 = k m is

a =
1 1 FG IJ (19)
2m 1
,
HK
and the normalized eigenvector b corresponding to eigenvalue λ2 = bk + 2k 'g m is

b=
1 1 FG IJ 1 −1FG IJ
. (20)
2 m −1 H K or b=
2m 1 H K
(Note that even with the normalization conditions, the eigenvectors are not determined
uniquely; each one can always be multiplied by −1.)

Orthogonality of Eigenvectors

In the general case, if λ1 , λ2 , λ3 , L are all unequal, and if M and K are symmetric
matrices (as in our example), then it can be shown that
aTMb = 0, aTMc = 0, bTMc = 0, L . (21)
Note that the usual definition of orthogonality of two vectors a and b is aT b = 0, i.e.,
the scalar product of a and b is zero. So Eq. (21) represents a generalization of the
concepts of the scalar product and of orthogonality of two vectors. We invite you to
verify that the eigenvectors a and b in our example problem do satisfy this condition.

If the system has three or more degrees of freedom, then any two eigenvectors that
correspond to unequal eigenvalues are orthogonal in the sense of Eq. (21).

To prove Eqs. (21), we start with Eqs. (11):

Ka = λ1Ma and Kb = λ 2 Mb. (22)
Multiply the first equation on the left by bT , and the second on the left by aT .

bT Ka = λ 1bT Ma , a T Kb = λ 2a T Mb. (23)

Take the transpose of the entire first of Eqs. (23):

eb Kaj e j
T T
T
= λ1 bT Ma . (24)

The transpose of a product of matrices equals the product of the transposes in the reverse
order; also, (bT )T = b, so Eq. (24) can be written
11-6 11 Normal Modes -- Matrix Methods

a T K Tb = λ1a TM Tb. (25)

If K and M are symmetric, as in our example, then KT = K and M T = M. (We'll
discuss this property later.) Then Eq. (25) can be written.
a T Kb = λ1a T Mb. (26)
Finally, we subtract Eq. (26) from the second of Eqs. (23):

bλ 2 g
− λ1 a T Mb = 0. (27)

Thus if λ 2 ≠ λ1 , then a T Mb = 0 , and the eigenvectors a and b are orthogonal in

the generalized sense we have defined. Similarly, any pair of eigenvectors with unequal
eigenvalues are orthogonal in this sense.

Initial Conditions

We can now use the orthogonality properties of the normalized eigenvectors to work out
relations to find the constants C1 , C2 , C3 , L and ϕ1 , ϕ2 , ϕ3 , L in Eq. (12) if we
are given the initial values of the coordinates and velocities, which we denote by
x( 0) and x& ( 0) , respectively. At time t = 0, Eq. (12) becomes

x( 0) = C1 a cos ϕ1 + C2 b cos ϕ 2 + C3 c cos ϕ 3 + L . (28)

Similarly, taking the time derivative of Eq. (12) and setting t = 0, we find

x& ( 0) = − C1ω 1a sinϕ1 − C2 ω 2 b sinϕ 2 − C3 ω 3 c sinϕ 3 − L . (29)

Now see what happens when we multiply Eq. (28) on the left by aT M. We get

a T Mx( 0) = C1a T Ma cos ϕ1 + C2 a T Mb cos ϕ 2 + C3 a T Mc cos ϕ3 + L . (30)

The matrix product in the first term on the right is unity, because of the normalization of
the eigenvectors, and all the other matrix products on the right are zero because of
orthogonality . So Eq. (30) becomes simply

a T Mx( 0) = C1 cos ϕ1 . (31)

The same thing happens when we multiply Eq. (29) on the left by aT M

a T Mx& ( 0) = − C1 ω1 a T Ma sinϕ1 − C2 ω 2 a T Mb sin ϕ2 − C3 ω 3 a T Mc sin ϕ 3 + L. (32)

Using the normalization and orthogonality conditions, we get

a T Mx& ( 0) = − C1 ω1 sin ϕ1 . (33)

11 Normal Modes -- Matrix Methods 11-7

Now it is straightforward to solve Eqs. (31) and (33) simultaneously to obtain C1 and
ϕ1 . And all the other C's and ϕ's can be obtained the same way by multiplying Eqs.
(28) and (29) by b, c, L .

Diagonalization of K and M Matrices

The eigenvectors and eigenvalues can be used to construct the normal-coordinate

transformation; this transformation converts both K and M into diagonal matrices.
From Eqs. (11),
Ka = λ1Ma , Kb = λ 2 Mb, Kc = λ 3 Mc, L . (34)
We can combine these equations into a single matrix equation. Let a1 , a2 , a3 , ... be the
elements of the normalized eigenvector a, and so on; then define a matrix A consisting
of the normalized eigenvectors as columns; that is.

A =
FG a 1 b1 IJ
Ha 2 b2 K (for two degrees of freedom), (35)

Fa b1 c1 I
A = Ga JJ
1

GH a 2 b2
b3
c2
c3 K
(for three degrees of freedom), (36)
3

and so on. (From here on we assume two degrees of freedom; the generalization to three
or more degrees of freedom will be clear.)

We can combine Eqs. (34) as

FG a1 b1 IJ = MFG λ a 1 1 λ 2b1IJ
K
Ha 2 b2 K Hλ a 1 2 λ 2b2 K
. (37)

Also note that

FG λ a λ b IJ = FG a b IJ FG λ
1 1 2 1 1 1 1 0IJ
Hλ a λ b K Ha b K H 0
1 2 2 2 2 2 λ2 K
. (38)

Now we define a diagonal matrix Λ as

Λ=G
F λ 0 IJ .1

H0 λ K 2
(39)

Then
FG λ a
1 1 λ 2 b1 IJ = AΛ
Λ,
Hλ a
1 2 λ 2b2 K
and Eq. (37) can be written as

Λ.
KA = MAΛ (40)
11-8 11 Normal Modes -- Matrix Methods

Now we multiply this equation on the left by AT :

Λ.
AT KA = ATMAΛ (41)
T
Consider the product A MA. In general, this is equal to

FG a
1 a2 IJ FG m
11 m12 IJ FG a 1 b1IJ
Hb1 b2 K Hm 21 m22 K Ha 2 b2 K
.

We see that the first row of AT and the first column of A give the normalization
condition for the eigenvector a. The first row of AT and the second column of A give
the orthogonality relation for a and b, and so on. Thus the diagonal elements of the
product AT MA are all 1, and the off-diagonal elements are all zero. That is, AT MA is
equal to the identity (unit) matrix, denoted by I.

It follows from this result that

eA MAjΛ = Λ = FGH λ0
T 1 0
λ2
IJ
K
, (42)

and, from Eq. (41),

AT KA = Λ. (43)
(Note that Λ is always diagonal, and that the diagonal elements are the eigenvalues.)

Thus we have the fairly astonishing result that the transformation AT MA turns M into
the identity matrix I, and it turns K into a matrix AT KA = Λ that is always diagonal,
with diagonal elements equal to the eigenvalues.

Example

For the same example we've been discussing all along,

a =
1 1 FG IJ b =
1FG IJ A = 1 FG1 1 IJ .
1
2m 1
,
HK H K
2m −1
,
2m H1 −1K
(44)

AT MA =
1 1FG 1 I Fm 0I 1 F1 1 I
J G J G J F1 0IJ
= G
2m 1H −1K H 0 mK 2m H 1 −1K H0 1K (45)

Fk I
1 F1 1 I F k + k ' − k ' I 1 F1 1 I
= GG m
G J G J G J
0
J . (46)
k + 2k'J
AT KA =
2m H1 −1K H − k ' k + k 'K 2m H 1 −1K GH 0 J
m K
You should verify Eqs. (45) and (46) by carrying out the matrix multiplication, both by
hand and using Maple.
11 Normal Modes -- Matrix Methods 11-9

Normal Coordinate Transformation

The normal coordinate transformation is now simply x = Aq. We substitute this into the
equations of motion, Eq. (4):
Mx&& = −Kx , && = −KAq.
MAq (47)
Multiply on the left by AT :
&& = − AT KAq.
AT MAq (48)
But, as we showed above,
AT MA = I and AT KA = Λ . (49)
So Eq. (48) becomes

&& = − Λq,
FG q&& IJ
1
= −
FG λ 1 0 IJ FG q IJ
1
= −
FG λ q IJ .
1 1
q or
H q&& K
2 H0 λ2 K Hq K
2 Hλ q K
1 2
(50)

We see that the equations for the q's are decoupled; each equation contains only one q:
q&&1 = − λ1 q1 = − ω1 2 q1 , q&&2 = − λ 2 q2 = − ω 2 2 q 2 , (51)
and so forth. Each q corresponds to a single normal mode.

Symmetry of M and K Matrices

In our example problem, both M and K happen to be symmetric. When they are not,
our orthogonality proof is not valid, and Eqs. (26) and (27) aren't correct. When the
equations of motion are obtained from Lagrange's equations, it can be shown that M and
K are always symmetric. But when we start with ΣF = ma, the resulting matrices may
or may not be symmetric, depending on our choice of coordinates.

Suppose, for instance, that for our example system we let x 1 be the displacement of the
left mass from equilibrium and let x 2 be the elongation of the spring k '. Then the
displacement of the right mass from equilibrium is x 1 + x 2 and its acceleration is x&&1 + x&&2
You can verify that the equations of motion, in terms of these coordinates, are
mx&&1 = − kx1 + k ' x2 ,
b
x1 + &&
m && g
x2 = −kx1 − k + k ' x2 . b g (52)

In matrix form,

FG m 0 IJ FG &&x IJ = − FG k
1 −k ' IJ FG x IJ .
1
H m mK H &&x K H k
2 k + k' K Hx K 2
(53)

The M and K matrices are

M=
FG m 0 IJ , K=
FG k −k ' IJ
H m mK Hk k + k'
.
K (54)
11-10 11 Normal Modes -- Matrix Methods

Note that neither of these is symmetric. The secular equation is

k − λm −k '
K − λM = = 0. (55)
k − λm k + k ' − λm
We leave it as an exercise for the reader to show that the eigenvalues are again
k k + 2k '
λ1 = , λ2 = , (56)
m m
and the normalized eigenvectors are

1 1 FG IJ 1 FG IJ F I
GH JK
1 1 1 1 3
a= b= A =
m 0
,
HK 3m −2 H K
,
m 0 −2 3
. (57)

You should check that again the operation AT MA reduces M to the identity matrix I,
and that

AT KA = Λ =
FG λ
1 0 IJ
H0 λ2
.
K (58)

You can also check the orthogonality of the eigenvectors; in this case they are not
orthogonal because M and K aren't symmetric. However, we can reformulate the
problem by adding and subtracting the equations of motion, Eqs. (52), to obtain M and
K matrices that are symmetric. Here's how we do it. We can add the two original
equations of motion, Eqs. (52), to obtain
bx1 + &&
m 2 && g
x2 = −2kx1 − kx2 . (59)
Now we take this equation and the second of Eqs. (52) to be the be the basic equations
of motion:
b g
x1 + &&
m 2 && x2 = −2kx1 − kx2 ,
mb && x g = −kx − b k + k 'g x .
(60)
x + &&
1 2 1 2

For these equations, the M and K matrices are

M =
FG 2m mIJ , K =
FG 2k k IJ
H m mK Hk k + k'K. (61)

Now both M and K are symmetric. The secular determinant is

2 k − 2λm k − λm
= 0. (62)
k − λm k + k ' − λm
You are invited to complete this analysis by (1) showing that the roots of the secular
equation are the same as before, (2) obtaining the normalized eigenvector for each
eigenvalue, and (3) showing that the eigenvectors are orthogonal. Have fun!