Masses joined by springs can move in complicated ways, but some
collective motions are simple: the masses keep a fixed displacement
pattern while oscillating together. We call these motions modes.
For the spring chains we will study, every motion is a combination of
them. The equations have the form 𝑞″=−𝐾𝑞, and the eigenvectors of 𝐾
give the mode shapes; its eigenvalues give their squared angular frequencies.
10.1Coupled Springs
Consider two equal masses joined by three springs and stretched between
fixed supports. Each has mass 𝑚. At equilibrium the masses and supports
lie on a straight line, with spacing 𝑎, and each spring is under tension 𝑇0>0. We let
the masses move perpendicular to this line, neglecting gravity and
friction. This is called transverse motion. Write 𝑞1 and 𝑞2 for
their displacements, measured upwards in the picture.
Figure 10.1 Two masses displaced perpendicular to a row of stretched springs. The
gold arrows show the net transverse spring force on each mass in our
linear approximation; brackets mark the displacements 𝑞1 and 𝑞2.
Drag either mass and release it to start both masses from rest, or press
Play to release the configuration shown.
The middle spring has horizontal span 𝑎 and height difference
𝛿=𝑞2−𝑞1, so its length is √𝑎2+𝛿2. The length
changes only to second order in 𝛿, so to first order its tension
remains 𝑇0.
But the spring now pulls at an angle 𝜃 to the horizontal. Its
vertical rise divided by its length gives
sin𝜃=𝛿√𝑎2+𝛿2≈𝛿𝑎.
The transverse component of its pull on the first mass is therefore
𝐹⟂≈𝑇0sin𝜃≈𝑇0𝑎(𝑞2−𝑞1).
The same reasoning applies to the outside springs, whose other ends are
held at height zero. Keeping only these first-order forces, Newton's law
gives
𝑚𝑞″1=𝑇0𝑎[−𝑞1+(𝑞2−𝑞1)],𝑚𝑞″2=𝑇0𝑎[(𝑞1−𝑞2)−𝑞2].
For now, to keep things simple, let's set 𝑚=𝑎=𝑇0=1. With
𝑞=(𝑞1,𝑞2), our small-displacement model becomes
𝑞″=−𝐾𝑞,𝐾=(2−1−12).
The same reasoning applies to any number of masses. Consider 𝑛 equal
masses in a row between two fixed supports, joined by 𝑛+1 springs. At
equilibrium the masses lie on a straight line with spacing 𝑎, and every
spring is under tension 𝑇0. As before, the masses move perpendicular
to this line, and we set 𝑚=𝑎=𝑇0=1.
Mass 𝑗 is attached to two springs, one leading to each of its
neighbors. The small-displacement calculation from our two masses
applies to each spring: a neighbor at height 𝑞𝑗+1 contributes the
transverse force 𝑞𝑗+1−𝑞𝑗. The springs at the ends are attached to
supports, which stay at height zero. Adding the two forces, Newton's law
gives
𝑞″𝑗=(𝑞𝑗−1−𝑞𝑗)+(𝑞𝑗+1−𝑞𝑗)=𝑞𝑗−1−2𝑞𝑗+𝑞𝑗+1,
where for the first and last masses, the height of the missing neighbor
is the support's height, zero. Each mass is pulled toward the heights of
its two neighbors.
With 𝑞=(𝑞1,…,𝑞𝑛), these equations again have the form
𝑞″=−𝐾𝑞. Here is 𝐾 for chains of one, two, three, and four masses:
Each row comes from one mass. The diagonal entry 2 records its two
springs, and the two entries −1 beside it couple the mass to its
neighbors. Adding a mass to the chain just adds another row with the
same pattern. Every chain, however long, is described by 𝑞″=−𝐾𝑞 with
this one kind of matrix.
Figure 10.2 Chains of one to four equal masses between fixed supports. Drag any mass
and release it to start all the masses from rest. With one mass, every
release gives the same simple oscillation. With more masses, compare
releases that start from different shapes.
Try a few releases. With a single mass, the motion is always the same
simple oscillation. With two or more masses, the motion is harder to
describe. Pull one mass aside and let go: its neighbors are soon dragged
along, and the pattern of displacements keeps changing as the masses
pass motion back and forth.
10.2Eigenvectors and Solutions
Every chain of springs is described by 𝑞″=−𝐾𝑞. Let's solve it,
starting with our two masses.
We already know how to turn this into a first-order system. The state
records both displacements and both velocities:
The first two rows say that the derivatives of the displacements are the
velocities; the last two give the accelerations from our spring equations.
Grouping the first two coordinates into 𝑞 and the last two into 𝑞′,
we can write this more compactly:
𝑋=(𝑞𝑞′),𝑋′=(0𝐼−𝐾0)⏟__⏟__⏟𝐴𝑋.
Here each block is 2×2. A chain of 𝑛 masses gives the same
block matrix with 𝑛×𝑛 blocks, so its state space has dimension
2𝑛.
The matrix exponential gives 𝑋(𝑡)=𝑒𝑡𝐴𝑋(0), and Chapter 9 tells us
where to start: factor the characteristic polynomial. For our two
masses, expanding the 4×4 determinant gives
det(𝑟𝐼−𝐴)=𝑟4+4𝑟2+3=(𝑟2+1)(𝑟2+3).
Neither factor has real roots, so 𝐴 has no real eigenvectors. In
Chapter 9, a factor 𝑟2+𝜔2 meant an invariant plane rotating at
angular frequency 𝜔, with no growth or decay. So the
four-dimensional state space splits into two such planes, with angular
frequencies 1 and √3. Every motion of our springs combines two
pure oscillations.
Let's recover these oscillations by looking directly for patterns in the
matrix exponential 𝑒𝑡𝐴. In Chapter 7, we computed the spring's flow
𝑒𝑡𝐽 using 𝐽2=−𝐼. What happens when we square our matrix 𝐴?
Multiplying the blocks,
𝐴2=(−𝐾00−𝐾).
Squaring 𝐴 gives a block diagonal matrix, with −𝐾 in each diagonal
block where Chapter 7 had −1. Multiplying by 𝐴 again moves the
blocks off the diagonal:
𝐴3=(0−𝐾𝐾20),𝐴4=(𝐾200𝐾2).
In general, the even powers are block diagonal and the odd powers fill
the off-diagonal blocks:
𝐴2𝑗=(−1)𝑗(𝐾𝑗00𝐾𝑗),𝐴2𝑗+1=(−1)𝑗(0𝐾𝑗−𝐾𝑗+10).
Only powers of 𝐾 ever appear. Let's sort the terms of the exponential
series the same way, collecting the even powers and the odd powers
separately:
𝑒𝑡𝐴=(𝐼+𝑡2𝐴22!+𝑡4𝐴44!+⋯)+(𝑡𝐴+𝑡3𝐴33!+𝑡5𝐴55!+⋯).
Every even power is block diagonal, so the first group is too:
Look at the bottom-left block. Factoring out −𝐾 gives exactly the
series in the top-right block:
−𝑡𝐾+𝑡3𝐾23!−𝑡5𝐾35!+⋯=−𝐾(𝑡𝐼−𝑡3𝐾3!+𝑡5𝐾25!−⋯).
So only two different series appear. Let's give them names:
𝐶𝐾(𝑡)=𝐼−𝑡2𝐾2!+𝑡4𝐾24!−⋯,𝑆𝐾(𝑡)=𝑡𝐼−𝑡3𝐾3!+𝑡5𝐾25!−⋯.
The subscript records that these matrices are built from 𝐾. They
converge for every 𝑡, because they are blocks of the absolutely
convergent exponential series. In this notation, the two groups of terms
are
(𝐶𝐾(𝑡)00𝐶𝐾(𝑡))and(0𝑆𝐾(𝑡)−𝐾𝑆𝐾(𝑡)0).
Adding them gives the whole exponential:
𝑒𝑡𝐴=(𝐶𝐾(𝑡)𝑆𝐾(𝑡)−𝐾𝑆𝐾(𝑡)𝐶𝐾(𝑡)).
Reading the top row gives the displacement.
Theorem 10.1 (Second-order evolution). For any constant real square matrix 𝐾, the solution of
𝑞″=−𝐾𝑞,𝑞(0)=𝑞0,𝑞′(0)=𝑝0
is
𝑞(𝑡)=𝐶𝐾(𝑡)𝑞0+𝑆𝐾(𝑡)𝑝0.
The first term evolves an initial displacement released from rest; the
second evolves an initial velocity starting at equilibrium. We can read
those roles directly from the series:
𝐶𝐾(0)=𝐼,𝐶′𝐾(0)=0,𝑆𝐾(0)=0,𝑆′𝐾(0)=𝐼.
No symmetry or positivity was needed for this calculation.
The names come from the single spring. When 𝐾 is the 1×1 matrix
(𝜔2), the two series are
and our theorem gives the familiar solution
𝑞(𝑡)=𝑞0cos(𝜔𝑡)+𝑝0sin(𝜔𝑡)/𝜔. For a single mass in
our chain, 𝜔2=2.
For a general 𝐾, these series are hard to sum: every term involves a
new power of 𝐾. But for some vectors they are easy. Suppose
𝐾𝑣=𝜆𝑣. As in Chapter 9, applying 𝐾 again just multiplies by
𝜆 again, so 𝐾𝑚𝑣=𝜆𝑚𝑣 for every 𝑚. In our series, every
power of 𝐾 acting on 𝑣 becomes a power of the number 𝜆:
𝐶𝐾(𝑡)𝑣=(𝐼−𝑡2𝐾2!+𝑡4𝐾24!−⋯)𝑣=(1−𝑡2𝜆2!+𝑡4𝜆24!−⋯)𝑣.
When 𝜆>0, this is the cosine series from the single spring, with
𝜆 in the place of 𝜔2. The same happens to the sine
series. So
𝐶𝐾(𝑡)𝑣=cos(√𝜆𝑡)𝑣,𝑆𝐾(𝑡)𝑣=sin(√𝜆𝑡)√𝜆𝑣.
An eigenvector of 𝐾 is also an eigenvector of 𝐶𝐾(𝑡) and 𝑆𝐾(𝑡),
for every 𝑡, with eigenvalues cos(√𝜆𝑡) and
sin(√𝜆𝑡)/√𝜆, respectively.
Now use the theorem. Pull the masses into the shape 𝑣 and release them
from rest, so 𝑞0=𝑣 and 𝑝0=0. Then
𝑞(𝑡)=𝐶𝐾(𝑡)𝑣=cos(𝜔𝑡)𝑣,𝜔=√𝜆.
The vector 𝑣 never changes; only the number multiplying it oscillates.
The displacements stay proportional to the components of 𝑣: the
pattern scales, passes through zero, and reverses. Any mass whose
component is zero stays still; the others oscillate at the same angular
frequency 𝜔. This is what we mean by keeping the shape 𝑣.
Starting at equilibrium with velocity 𝑣 gives the same kind of motion,
𝑞(𝑡)=𝑆𝐾(𝑡)𝑣=sin(𝜔𝑡)𝑣/𝜔. Any initial displacement
𝑐𝑣 and velocity 𝑑𝑣 along the shape combine these:
𝑞(𝑡)=(𝑐cos(𝜔𝑡)+𝑑sin(𝜔𝑡)𝜔)𝑣.
The coefficient of 𝑣 is exactly the solution of the single spring
equation 𝑎″=−𝜆𝑎 with 𝑎(0)=𝑐 and 𝑎′(0)=𝑑. In fact, we can
check directly that this pairing works for any eigenvalue. If
𝑞(𝑡)=𝑎(𝑡)𝑣, then
𝑞″=𝑎″𝑣,−𝐾𝑞=−𝑎𝐾𝑣=−𝜆𝑎𝑣,
so 𝑞 solves our equation exactly when 𝑎″=−𝜆𝑎.
Definition 10.2 (Mode and mode shape). For the system 𝑞″=−𝐾𝑞, let 𝑣 be a nonzero real eigenvector of 𝐾,
with eigenvalue 𝜆. The family of solutions
𝑞(𝑡)=𝑎(𝑡)𝑣,𝑎″=−𝜆𝑎,
is called a mode, and 𝑣 is a mode shape. Multiplying 𝑣 by a
nonzero constant gives the same mode.
When 𝜆>0, the mode is oscillatory, with angular frequency
𝜔=√𝜆.
Each mode gives an invariant plane for the first-order system: the
states (𝑐𝑣,𝑑𝑣), with displacement and velocity along the same shape
𝑣, stay in that plane as the system evolves.
10.3Modes of Spring Chains
Let's find the modes of our chains, starting with two masses. The
characteristic polynomial of 𝐾 is
det(𝐾−𝜆𝐼)=(2−𝜆)2−1=(𝜆−1)(𝜆−3).
For 𝜆=1, the eigenvector equation says the two components are
equal; for 𝜆=3, it says they are opposite. We can therefore
choose the two shapes and write their solutions released from rest:
𝑣1=(11),𝑞(1)(𝑡)=cos𝑡(11),
𝑣2=(1−1),𝑞(2)(𝑡)=cos(√3𝑡)(1−1).
Both solutions start in their mode shape with zero velocity. Their
angular frequencies are 𝜔1=1 and 𝜔2=√3. These are
the same frequencies the characteristic polynomial of 𝐴 gave us in
section 10.2. Now we can see the motion in each invariant plane:
the masses moving together, or moving oppositely.
Figure 10.3 The two modes released from rest, using the same clock and displacement
scale. In the first, both masses move together; in the second, they move
oppositely. Start them together and watch how much sooner the second
mode completes an oscillation.
Each frequency has a physical explanation. When the masses move
together, their equal displacements keep the middle spring horizontal,
so each mass feels only the restoring force of its outside spring. When
they move oppositely, with displacements 1 and −1, the height
difference across the middle spring is −2. Its force adds to the
outside spring's force, giving the first mass a total restoring force of
−3. The extra tilt of the middle spring makes this motion faster.
The same reasoning applies to every mode shape. The force on the masses
is −𝐾𝑞, and when 𝑞 is a multiple of an eigenvector, this force is
−𝜆𝑞. Each mass feels a restoring force proportional to its own
displacement, with the same constant 𝜆 for every mass. So the
whole chain moves like a single spring with stiffness 𝜆.
For three masses, expanding along the first row gives
det(𝐾−𝜆𝐼)=(2−𝜆)[(2−𝜆)2−1]−(2−𝜆)=(2−𝜆)[(2−𝜆)2−2],
so the eigenvalues are 2 and 2±√2. For 𝜆=2, the middle
row of (𝐾−2𝐼)𝑣=0 says the first and third components are opposite, and
the first row says the middle component is zero. For 𝜆=2−√2,
the first row says the middle component is √2 times the first, and
the last row says the same about the third. Taking 𝜆=2+√2
changes the sign of √2. We list the modes in order of increasing
frequency:
Eigenvalue 𝜆
Angular frequency 𝜔
Mode shape
2−√2
≈0.765
(1,√2,1)
2
√2≈1.414
(1,0,−1)
2+√2
≈1.848
(1,−√2,1)
In the slowest mode, all three masses move to the same side together,
with the middle mass moving furthest. In the second, the middle mass
never moves at all: its neighbors are always displaced equally in
opposite directions, so their pulls on it cancel. In the fastest mode,
the middle mass moves opposite the outer two.
For four masses, the eigenvalues involve the golden ratio
𝜑=(1+√5)/2. Here are the modes, which you can check by
multiplying by 𝐾 and using 𝜑2=𝜑+1:
Eigenvalue 𝜆
Angular frequency 𝜔
Mode shape
2−𝜑≈0.382
≈0.618
(1,𝜑,𝜑,1)
3−𝜑≈1.382
≈1.176
(𝜑,1,−1,−𝜑)
1+𝜑≈2.618
≈1.618
(𝜑,−1,−1,𝜑)
2+𝜑≈3.618
≈1.902
(1,−𝜑,𝜑,−1)
A pattern is emerging. In each chain, the slowest mode moves every mass
to the same side, and each faster mode changes sign one more time along
the chain. More sign changes mean more neighbors pulling against each
other, giving a stiffer restoring force and a higher frequency.
Figure 10.4 Release a chain of two, three, or four masses in one of its mode shapes.
Every mass oscillates at the same frequency, and the chain keeps its
shape throughout the motion. Compare these with the releases in
figure 10.2, where the shape kept changing. For four masses,
𝜑=(1+√5)/2.
We have found the modes of chains with two, three, and four masses one at
a time. Is there a pattern that works for every chain? Look again at the
slowest three-mass shape, (1,√2,1). Scaling it by √2/2 gives
(√2/2,1,√2/2). Including a zero at each wall, these heights
are
0,√22,1,√22,0,
which are familiar sine values:
sin0,sin𝜋4,sin2𝜋4,sin3𝜋4,sin4𝜋4.
The shape samples half a sine wave between the walls! The other two
shapes do the same with a full wave and one and a half waves:
(1,0,−1) lists the values sin(2𝑗𝜋/4), and (1,−√2,1) is a
multiple of sin(3𝑗𝜋/4), for 𝑗=1,2,3. You can check that our two-
and four-mass shapes follow the same pattern, sampling sin(𝑟𝑗𝜋/3)
and sin(𝑟𝑗𝜋/5).
Could the modes of every chain follow this pattern? Let's try
𝑣𝑗=sin(𝑗𝜃)
and see which values of 𝜃 make this an eigenvector of 𝐾.
Changing 𝜃 changes how much of the sine wave fits between the
walls. We will check the eigenvector equation for this whole family of
shapes, then use the fixed ends to determine which shapes are allowed.
For 𝑣 to be an eigenvector, each row of 𝐾𝑣=𝜆𝑣 must satisfy
2𝑣𝑗−𝑣𝑗−1−𝑣𝑗+1=𝜆𝑣𝑗,𝑗=1,…,𝑛.
The endpoint conditions 𝑣0=𝑣𝑛+1=0 make this equation agree with
the first and last rows as well. Substituting our guess, the sine
addition formula gives
So every sine pattern satisfies the same relation between neighboring
values, with 𝜆=2−2cos𝜃. To obtain an eigenvector of our
finite matrix, we still need a nonzero vector satisfying the endpoint
conditions. At the left wall, 𝑣0=sin0=0 automatically. At the right
wall, we need
sin((𝑛+1)𝜃)=0.
The sine must return to zero there: an integer number of half-waves must
fit between the walls. One half-wave gives 𝜃=𝜋/(𝑛+1), a full
wave gives 𝜃=2𝜋/(𝑛+1), and one and a half waves gives
𝜃=3𝜋/(𝑛+1). In general, taking
𝜃𝑟=𝑟𝜋𝑛+1,𝑟=1,…,𝑛,
gives the shapes
𝑣(𝑟)𝑗=sin𝑟𝑗𝜋𝑛+1,𝑗=1,…,𝑛,
with eigenvalues
𝜆𝑟=2−2cos𝑟𝜋𝑛+1=4sin2𝑟𝜋2(𝑛+1).
The index 𝑟 counts the half-waves in the sine curve we sample. In our
transverse chain, the masses themselves lie along this curve.
Have we found all the modes? Each of these 𝑛 vectors is nonzero, since
its first component is sin𝜃𝑟>0. Their eigenvalues are positive
and strictly increasing: the angle in the last sine runs through values
between 0 and 𝜋/2. Eigenvectors with distinct eigenvalues are
linearly independent, so these 𝑛 vectors form a basis of ℝ𝑛.
This accounts for all the eigenvalues, and every eigenvector is a nonzero
multiple of one of these shapes. Our guess has found a complete basis of
mode shapes!
The angular frequencies are
𝜔𝑟=√𝜆𝑟=2sin𝑟𝜋2(𝑛+1).
Every mode oscillates, and higher mode numbers give higher frequencies,
as our tables suggested.
These shapes also resemble the standing waves of a stretched string.
To compare chains of the same total length, we now restore the physical
constants. The matrix is multiplied by 𝑇0/(𝑚𝑎), so the shapes stay the
same and the angular frequencies become
𝜔𝑟=2√𝑇0𝑚𝑎sin𝑟𝜋2(𝑛+1).
In the figure, increase the number of masses while keeping the length,
tension, and mass per unit length fixed. The spacing and each mass get
smaller, and the masses sample the same sine-shaped curve more finely.
This differs from our earlier chains, where 𝑚=𝑎=𝑇0=1 for every size.
Figure 10.5 A transverse mode of a chain held under tension. Each mass moves
perpendicular to the resting line, and the fixed endpoints remain still.
Increase the number of masses to see the chain fill out the same
standing-wave shape. Length, tension, and mass per unit length are held
fixed; transverse motion is enlarged.
10.4Combining Modes
A single mode keeps its shape. But in figure 10.2, most releases
did not: the pattern kept changing. Let's see why, starting again with
two masses. Adding their two mode solutions gives another solution,
by superposition. With both modes present, the combined motion need not
keep a fixed shape: its two components oscillate at different frequencies.
Figure 10.6 The first two systems show the two modes released from rest,
𝑐1cos(𝑡)𝑣1 and 𝑐2cos(√3𝑡)𝑣2. The third adds their
displacements, mass by mass. All three use the same clock and scale.
Change the coefficients to vary the contribution of each mode, or set
one to zero to recover a pure mode. With both present, the combined
displacement pattern keeps changing, because the two components
oscillate at different frequencies.
Do these combinations give every motion? Our two shapes form a basis of
ℝ2, so we can write any initial displacement and velocity in
that basis. The same reasoning works for any number of masses whenever
we have a basis of eigenvectors, as we now do for every chain.
Theorem 10.3 (Complete solution by modes). Suppose the 𝑛×𝑛 matrix 𝐾 has a basis of eigenvectors
𝑣1,…,𝑣𝑛 with positive eigenvalues:
𝐾𝑣𝑗=𝜆𝑗𝑣𝑗,𝜔𝑗=√𝜆𝑗.
Write the initial displacement and velocity in this basis:
𝑞0=𝑛∑𝑗=1𝑐𝑗𝑣𝑗,𝑝0=𝑛∑𝑗=1𝑑𝑗𝑣𝑗.
Then the solution of 𝑞″=−𝐾𝑞 with 𝑞(0)=𝑞0 and 𝑞′(0)=𝑝0 is
𝑞(𝑡)=𝑛∑𝑗=1(𝑐𝑗cos(𝜔𝑗𝑡)+𝑑𝑗sin(𝜔𝑗𝑡)𝜔𝑗)𝑣𝑗.
In particular, the 2𝑛 mode solutions cos(𝜔𝑗𝑡)𝑣𝑗 and
sin(𝜔𝑗𝑡)𝑣𝑗/𝜔𝑗 form a basis of the solution space.
Proof. By theorem 10.1, 𝑞(𝑡)=𝐶𝐾(𝑡)𝑞0+𝑆𝐾(𝑡)𝑝0. The matrices
𝐶𝐾(𝑡) and 𝑆𝐾(𝑡) act on each term of the two expansions separately,
and on each eigenvector 𝑣𝑗 they act by multiplication by
cos(𝜔𝑗𝑡) and sin(𝜔𝑗𝑡)/𝜔𝑗. This gives the formula.
The mode solutions have initial states (𝑣𝑗,0) and (0,𝑣𝑗). Since
the 𝑣𝑗 form a basis of ℝ𝑛, these 2𝑛 states form a basis
of the state space, so by
Chapter 6 the corresponding solutions form a basis of the solution space. ∎
So once we know the eigenvectors and eigenvalues of 𝐾, solving any
initial value problem comes down to writing 𝑞0 and 𝑝0 in the
eigenvector basis. The eigenvectors give the collective shapes, and the
eigenvalues give their squared angular frequencies.
Example 10.4 (Releasing one mass). In figure 10.2, pulling the first of three masses aside and
releasing it produced a motion whose shape kept changing. Let's find
that motion. Take 𝑞0=(1,0,0) and 𝑝0=0, so every 𝑑𝑗 is zero.
Using the three shapes from section 10.3, with the same scaling,
𝑣1=(1,√2,1),𝑣2=(1,0,−1),𝑣3=(1,−√2,1),
the equation 𝑞0=𝑐1𝑣1+𝑐2𝑣2+𝑐3𝑣3 says
𝑐1+𝑐2+𝑐3=1,√2𝑐1−√2𝑐3=0,𝑐1−𝑐2+𝑐3=0.
The middle equation gives 𝑐1=𝑐3. Subtracting the third equation from
the first gives 2𝑐2=1, so 𝑐2=12 and then 𝑐1=𝑐3=14.
Therefore
𝑞(𝑡)=14cos(𝜔1𝑡)𝑣1+12cos(𝜔2𝑡)𝑣2+14cos(𝜔3𝑡)𝑣3,
with the frequencies 𝜔1≈0.765, 𝜔2=√2, and
𝜔3≈1.848 from our table. Reading off each mass,
Each mode keeps its own shape and frequency. The changing pattern in the
figure comes from adding three motions with different frequencies.
Our chains are one particular arrangement of springs. Equal masses can be
joined in many other patterns, such as a grid or the floors of a
building. To use theorem 10.3 on any of them, we need two things:
the matrix 𝐾, and a basis of its eigenvectors. Let's see how the
springs determine 𝐾, and what that tells us about its eigenvectors.
Consider a system of equal masses connected by springs, with each spring
responding to the relative displacement of the masses it joins. A spring
of stiffness 𝑘 connecting masses 𝑖 and 𝑗 contributes
𝑚𝑞″𝑖=⋯+𝑘(𝑞𝑗−𝑞𝑖),𝑚𝑞″𝑗=⋯+𝑘(𝑞𝑖−𝑞𝑗).
After dividing by the common mass, this spring adds 𝑘/𝑚 to the
diagonal entries 𝐾𝑖𝑖 and 𝐾𝑗𝑗 of 𝐾 in 𝑞″=−𝐾𝑞. It adds the
same coefficient −𝑘/𝑚 to both 𝐾𝑖𝑗 and 𝐾𝑗𝑖. A spring joining
mass 𝑖 to a fixed support adds only 𝑘/𝑚 to 𝐾𝑖𝑖. To build 𝐾,
start from the zero matrix and add the contributions of every spring.
For our transverse chain, each spring has 𝑘=𝑇0/𝑎 and each mass has two
springs. With 𝑚=𝑎=𝑇0=1, this gives the 2 on the diagonal and the
−1 beside it.
Adding more connections preserves the equality of 𝐾𝑖𝑗 and 𝐾𝑗𝑖,
even when the springs have different stiffnesses. Springs attached to
fixed supports contribute only diagonal entries. The resulting matrix is
therefore symmetric:
𝐾𝑖𝑗=𝐾𝑗𝑖,orequivalently𝐾𝑇=𝐾.
Every real symmetric matrix has a basis of real eigenvectors. This fact
comes from linear algebra; we will not prove it here. So every system of
equal masses joined by springs has a basis of mode shapes. When the
eigenvalues are also positive, as they were for our chains,
theorem 10.3 expresses every motion as a combination of
oscillating modes.
Symmetry gives us more. In example 10.4, we found the
coefficients of 𝑞0 by solving three equations. Look at the three-mass
shapes again:
(1,√2,1),(1,0,−1),(1,−√2,1).
Every pair has dot product zero! The two-mass shapes (1,1) and
(1,−1) are perpendicular too. This is no accident.
Theorem 10.5 (Orthogonality and initial conditions). Eigenvectors of a real symmetric matrix with distinct eigenvalues are
orthogonal, and its eigenbasis can be chosen orthogonal.
For a symmetric matrix with positive eigenvalues, choose such an
orthogonal eigenbasis. In theorem 10.3, the solution with
𝑞(0)=𝑞0 and 𝑞′(0)=𝑝0 has coefficients
𝑐𝑗=𝑣𝑗⋅𝑞0𝑣𝑗⋅𝑣𝑗,𝑑𝑗=𝑣𝑗⋅𝑝0𝑣𝑗⋅𝑣𝑗.
Proof. Suppose 𝐾𝑢=𝜆𝑢 and 𝐾𝑣=𝜇𝑣, with 𝜆≠𝜇. Since 𝐾𝑇=𝐾,
(𝐾𝑢)⋅𝑣=𝑢⋅(𝐾𝑇𝑣)=𝑢⋅(𝐾𝑣).
Substituting the eigenvector equations gives
𝜆(𝑢⋅𝑣)=𝜇(𝑢⋅𝑣).
Thus (𝜆−𝜇)(𝑢⋅𝑣)=0, so 𝑢⋅𝑣=0.
Different eigenspaces are therefore orthogonal. Within each eigenspace
we can also choose an orthogonal basis; this is another fact from linear
algebra that we won't prove here. Often the eigenspace is just a line,
so any nonzero vector along it will do.
Recall how to find coordinates in an orthogonal basis. If
𝑥=𝑎1𝑣1+⋯+𝑎𝑛𝑣𝑛,
taking the dot product with 𝑣𝑗 removes every term except one:
𝑣𝑗⋅𝑥=𝑎𝑗(𝑣𝑗⋅𝑣𝑗).
Thus 𝑎𝑗=(𝑣𝑗⋅𝑥)/(𝑣𝑗⋅𝑣𝑗). We can find each coefficient
separately, without solving a system of equations. In
theorem 10.3, the 𝑐𝑗 and 𝑑𝑗 are exactly the coordinates of
𝑞0 and 𝑝0 in the eigenvector basis, so this formula gives them. ∎
For our three-mass release, the first shape gives 𝑣1⋅𝑞0=1 and
𝑣1⋅𝑣1=4, so 𝑐1=14 at once. In the same way,
𝑐2=12 and 𝑐3=14, with no system to solve.
For a symmetric 𝐾 with positive eigenvalues, we can now find an
orthogonal basis of mode shapes, project the initial displacement and
velocity onto them, and add their motions. The shapes need not have
length one. Let's apply this to our two coupled masses, now with an
initial velocity.
Example 10.6 (Motion from initial conditions). Our two-mass system has the orthogonal mode shapes
𝑣1=(11),𝑣2=(1−1),
with angular frequencies 1 and √3.
Suppose the first mass starts displaced by one unit and at rest, while
the second starts at equilibrium with velocity one:
𝑞0=(10),𝑝0=(01).
Both shapes have squared length 2. Since 𝑣1⋅𝑞0=1 and
𝑣2⋅𝑞0=1, the initial displacement is
𝑞0=12𝑣1+12𝑣2.
Since 𝑣1⋅𝑝0=1 and 𝑣2⋅𝑝0=−1, the initial velocity is
Each mode carries part of the starting displacement and part of the
starting velocity, and the two modes oscillate at different frequencies.
10.5Patterns on a Grid
Now let's connect the masses in two directions. Arrange twenty-five equal
masses in a square grid, with taut springs joining neighboring masses
and attaching the outer rows and columns to a fixed frame. Each mass
moves perpendicular to the grid's resting plane. As before, we consider
small displacements and neglect gravity and friction.
Figure 10.7 Twenty-five masses are joined in a square grid with a fixed boundary.
Each mass moves perpendicular to the grid's resting plane. Drag a mass
and release it: its motion pulls on neighboring masses, which in turn
pull on their neighbors, spreading the disturbance across the grid.
This time we keep the constants general. Let each mass be 𝑚, the
resting distance between connected points be 𝑎, and the equilibrium
tension in each spring be 𝑇0. Number the masses across each row as in
figure 10.8, and let 𝑞𝑗(𝑡) be the displacement of mass 𝑗,
positive above the resting plane.
Figure 10.8 The grid viewed from above. Numbers label the moving masses; crosses
mark fixed attachment points. The masses move perpendicular to the page.
Consider mass 7. Its four neighbors are masses 2, 6, 8, and 12.
Each spring pulls it toward the height of the mass at the other end.
The small-displacement calculation from our two-mass system applies
to each connection: to first order, a neighbor at height 𝑞𝑗
contributes the transverse force (𝑇0/𝑎)(𝑞𝑗−𝑞7). Adding the four
forces gives
The masses next to the frame obey the same rule, except that a fixed
attachment point has displacement zero. For example, mass 1 has two
moving neighbors and two fixed connections, so
We can now assemble the whole system. With
𝑞=(𝑞1,…,𝑞25), write
𝑞″=−𝑇0𝑚𝑎𝐿grid𝑞.
Each row of 𝐿grid has a 4 on the diagonal and a −1
for each neighboring mass. All other entries are zero. This is exactly
what our recipe from section 10.4 produces, with
𝑘=𝑇0/𝑎 for each spring. Here is the
entire matrix, with lines separating the five rows of the physical grid:
The entries next to the main diagonal connect masses within a row;
the entries five places away connect masses in neighboring rows.
Notice the gaps at each row break: masses 5 and 6 are not connected,
nor are masses 10 and 11.
The repeated five-by-five groups make both kinds of connection visible.
Our grid contains chains in two directions. Let's try to build its
mode shapes from the chain modes we already know.
Take a chain mode 𝑣=(𝑣1,…,𝑣5) and use it as the shape along
each row. Allow the rows to have different amplitudes: multiply the
first row by 𝑤1, the second by 𝑤2, and so on. The twenty-five
displacements, shown in their positions on the grid, are
Every row is a multiple of 𝑣. Now look down the columns: each is a
multiple of 𝑤=(𝑤1,…,𝑤5). If we choose 𝑤 to be another chain
mode, our shape has a chain mode in both directions! The displacement
at column 𝑖, row 𝑗 is the product 𝑣𝑖𝑤𝑗.
Figure 10.9 Two chain modes combine to make a grid mode. If their components are
𝑣𝑖 and 𝑤𝑗, the height at position (𝑖,𝑗) in the resulting shape
is 𝑣𝑖𝑤𝑗. The highlighted row shows a copy of 𝑣, scaled by one
component of 𝑤. Change either mode to see how it changes the whole
shape, then press Play to watch that shape oscillate.
Let's check that this is an eigenvector of the grid matrix. Suppose
𝐾𝑣=𝜆𝑣 and 𝐾𝑤=𝜇𝑤, where 𝐾 is the chain matrix for five masses
from section 10.3.
At column 𝑖, row 𝑗, the grid matrix takes four times the displacement
and subtracts the four neighboring displacements. For our product shape,
this gives
We use 𝑣0=𝑣6=𝑤0=𝑤6=0 at the fixed boundary, so the calculation
works at the edges too. Every component has been multiplied by
𝜆+𝜇. Thus, reading these heights row by row gives an
eigenvector of 𝐿grid with eigenvalue 𝜆+𝜇.
The mode shapes multiply, and the eigenvalues add. Since our system is
𝑞″=−(𝑇0/(𝑚𝑎))𝐿grid𝑞, this product shape oscillates with
angular frequency
𝜔=√𝑇0𝑚𝑎(𝜆+𝜇).
There are five chain modes to choose from in each direction, giving
5⋅5=25 product shapes. Do these give a basis for all twenty-five
displacements? They do, because the products inherit the orthogonality
of the chain modes. Take two product shapes 𝑣𝑖𝑤𝑗 and
̃𝑣𝑖̃𝑤𝑗. Their dot product is
The five chain modes have different eigenvalues, so by
theorem 10.5 they are mutually orthogonal. If we chose
different chain modes in either direction, the corresponding factor is
zero. Thus our twenty-five product shapes are nonzero and mutually
orthogonal. If a combination of them is zero, taking the dot product with
any one of them shows that its coefficient is zero. So they are linearly
independent, and twenty-five of them form a basis of ℝ25.
Every motion of the grid is a combination of these oscillating shapes.
The same construction works for a rectangular grid with 𝑁 masses
across each row and 𝑀 rows, keeping the same spacing 𝑎, mass 𝑚,
and equilibrium tension 𝑇0. We already know the chain modes for
any number of masses, so we can write the grid modes explicitly.
Choosing chain mode 𝑝 across each row (the 𝑥 direction) and chain
mode 𝑟 down each column (the 𝑦 direction) gives the shape
𝑊(𝑝,𝑟)𝑖𝑗=sin𝑝𝜋𝑖𝑁+1sin𝑟𝜋𝑗𝑀+1,1≤𝑝≤𝑁,1≤𝑟≤𝑀.
Here 𝑖 labels the column and 𝑗 labels the row. Adding the two
chain eigenvalues gives the squared angular frequency
𝜔2𝑝,𝑟=4𝑇0𝑚𝑎(sin2𝑝𝜋2(𝑁+1)+sin2𝑟𝜋2(𝑀+1)).
The same dot-product calculation shows that these 𝑁𝑀 shapes form
an orthogonal basis. We have found a complete set of modes for grids
of any size!
The indices 𝑝 and 𝑟 count the half-waves sampled in the two
directions. In the (1,1) mode, all masses move to the same side of
the resting plane together. In the (2,1) mode, the two halves in
the 𝑥 direction move to opposite sides. Try changing one mode index
at a time in the figure to see how each factor changes the whole shape.
Figure 10.10 Explore the product modes on larger grids. The x-mode and y-mode
determine the pattern in the two directions; neighboring lobes move
in opposite directions. Increase the number of masses to see the shapes
sampled more finely. Fabric view draws a surface through the same moving
points, making the hills, valleys, and stationary lines easier to see.
10.6Applications
Let's put modes to work on buildings and molecules. In each, a mode shape
shows which parts move together, which move in opposite directions, and
which stay nearly still, and its frequency tells us how quickly that
collective motion happens.
10.6.1Buildings
Tall buildings sway. Let's build the simplest model of that motion.
Treat each floor as a rigid mass 𝑚 that moves sideways, in one
horizontal direction. The columns between two floors resist being
tilted, like a spring: the story between floors 𝑗−1 and 𝑗 pushes back
in proportion to the drift𝑞𝑗−𝑞𝑗−1, the difference between the
two floors' displacements. We give every story the same stiffness 𝑘.
The ground does not move, so the lowest story's drift is just 𝑞1,
and nothing lies above the roof.
Engineers call this idealization a shear building.
Our recipe from section 10.4 assembles 𝐾 directly.
Each story is a spring joining two neighboring floors, and the lowest
story joins the first floor to the fixed ground. Every floor except the
roof has a story below it and a story above it. The roof has only the
story below. For three floors,
𝐾=𝑘𝑚⎛⎜
⎜
⎜⎝2−10−12−10−11⎞⎟
⎟
⎟⎠.
With more floors, the pattern continues: 2 on the diagonal and −1
beside it, except for the final 1. This is our chain matrix with one
entry changed! The building is a chain with its bottom end fixed and its
top end free.
Figure 10.11 An idealized building on fixed ground, swaying in one of its modes.
Choose the number of floors and the mode. The floors move sideways while
the columns between them bend. The gold columns mark the story with the
largest drift, the greatest difference in displacement between
neighboring floors. Playback is slowed so that each building's slowest
mode takes the same time; the readout compares each mode's frequency
with the slowest.
The modes look much like the chain's. In the slowest mode every floor
sways the same way, with the roof moving furthest. Each faster mode
changes direction one more time on the way up. As for the chain, a guess
of the form 𝑣𝑗=sin(𝑗𝜃) finds them all. Changing the last entry
changes which shapes fit: now
𝜃𝑟=(2𝑟−1)𝜋2𝑛+1,𝜔𝑟=2sin𝜃𝑟2√𝑘𝑚,𝑟=1,…,𝑛.
You can check this in the exercises. For the lowest modes of a tall
building, 𝜔𝑟 is nearly proportional to 2𝑟−1, so their frequencies
are close to the ratio 1:3:5. For a long chain with both ends fixed, they are close
to 1:2:3. Freeing one end changes the whole pattern of frequencies.
The slowest frequency also tells us how a building's sway depends on its
height. For many floors, 𝜔1≈𝜋2𝑛+1√𝑘/𝑚, so
the period of the slowest sway is
𝑇1=2𝜋𝜔1≈2(2𝑛+1)√𝑚𝑘.
The period grows in proportion to the number of floors. Engineers use
exactly this kind of rule of thumb: in the US building code (ASCE 7), a
moment-frame building's slowest period is estimated as roughly 0.1
seconds per story, for buildings up to about twelve stories. Our crude model can't supply the constant, which depends on
real masses and stiffnesses, but it predicts the proportionality.
The modes also separate two different questions. Which floor moves the
most, and which story deforms the most? In the slowest mode of a
three-floor building, the floor displacements are in the ratio
0.45:0.80:1, so the roof moves furthest. The drifts are in the ratio
0.45:0.36:0.20, so the bottom story is tilted the most. Engineers
watch drift closely, because it measures how much the columns themselves
are bent.
This model leaves a lot out. Real buildings also bend like a beam as a
whole, twist, rest on ground that gives a little, and lose energy to
friction. It also leaves out the ground's own motion during an
earthquake, which pushes on the building from below.
10.6.2Molecules
Molecules vibrate too. Their atoms are held together by chemical bonds,
which act like stiff springs when the atoms move slightly from their
resting positions. Unlike our chains and buildings, though, a molecule's
atoms have different masses. Carbon dioxide, for example, has an oxygen
atom on either side of a carbon atom, and each oxygen is about 4/3 as
heavy as the carbon.
Let's set up the simplest model of carbon dioxide. Place the three atoms
on a line, oxygen–carbon–oxygen, and let 𝑞1,𝑞2,𝑞3 be their
displacements along the line. Each bond is a spring of stiffness 𝑘,
pushing back in proportion to its change in length. Nothing holds the
molecule in place. Newton's law gives 𝑀𝑞″=−𝐾𝑞, where 𝑀 records the
masses and 𝐾 is assembled from the springs without dividing by them:
Each oxygen has only one bond, which gives the 1s at the ends of the
diagonal.
Multiplying by 𝑀−1 gives 𝑞″=−(𝑀−1𝐾)𝑞. This is our familiar
equation with 𝑀−1𝐾 in place of 𝐾. The same setup works for any
number of masses, with 𝑀=diag(𝑚1,…,𝑚𝑛).
Neither theorem 10.1 nor theorem 10.3 required
symmetry: eigenvectors of 𝑀−1𝐾 give modes, and a basis of them lets
us evolve each component separately. To find them, notice that
𝑀−1𝐾𝑣=𝜆𝑣 is the same as 𝐾𝑣=𝜆𝑀𝑣. So we solve
det(𝐾−𝜆𝑀)=0,(𝐾−𝜆𝑀)𝑣=0:
the familiar calculation, with 𝑀 in place of 𝐼. We can still find
the coefficients by writing the initial displacement and velocity in
the eigenvector basis. But 𝑀−1𝐾 is usually not symmetric, so the
ordinary dot-product formulas from section 10.4 do not
automatically apply.
Something new happens because nothing holds the molecule in place:
𝐾(1,1,1)=0. The shape (1,1,1) has eigenvalue 0. In this mode all
three atoms move together and the bonds never stretch; the molecule
drifts as a whole. Our series still handle it: built from 𝑀−1𝐾, they
give 𝐶𝑀−1𝐾(𝑡)𝑣=𝑣 and 𝑆𝑀−1𝐾(𝑡)𝑣=𝑡𝑣, so a molecule given
a push drifts steadily. Every free molecule has motions like this, called
zero modes. They are not vibrations. Our theorem theorem 10.3
assumed positive eigenvalues, but its argument still works with zero
modes: each one contributes (𝑐𝑗+𝑑𝑗𝑡)𝑣𝑗 in place of an oscillation.
The other two modes are vibrations. In the symmetric stretch(1,0,−1), the carbon stays still while the two oxygens move in and out
together: both bonds lengthen together and shorten together. The name
refers to these bond-length changes, even though the oxygens move in
opposite directions. In the antisymmetric stretch(1,−2𝑚𝑂/𝑚𝐶,1),
the carbon moves against both oxygens. In both, 𝑚𝑂𝑣1+𝑚𝐶𝑣2+𝑚𝑂𝑣3=0: the
vibrations carry no total momentum, so the molecule's center of mass
stays still while it vibrates.
Remark 10.7. The matrix 𝑀−1𝐾 is usually not symmetric, but we can still
guarantee a basis of modes. Substituting 𝑦=𝑀1/2𝑞, where 𝑀1/2
has diagonal entries √𝑚𝑖, gives
𝑦″=−(𝑀−1/2𝐾𝑀−1/2)𝑦. This new matrix is symmetric. By
section 10.4, it has an orthogonal basis of eigenvectors
𝑤𝑗, and 𝑣𝑗=𝑀−1/2𝑤𝑗 are mode shapes for the original problem.
You can work through an example in the exercises.
You can find the two frequencies in the exercises. Their ratio is
𝜔anti𝜔sym=√1+2𝑚𝑂𝑚𝐶≈1.91.
Standard tables
list these frequencies for carbon dioxide as 1333 and
2349 inverse centimeters, a unit proportional to frequency. Their
ratio is 1.76. Our model is off by about 9%.
What's missing? In a real molecule, stretching one bond changes how the
other bond pulls. One simple way to capture this is to add a weak spring
of stiffness 𝑘′ joining the two oxygens directly. Writing 𝑘𝑏 for the
bond stiffness, our recipe gives
𝐾=⎛⎜
⎜
⎜⎝𝑘𝑏+𝑘′−𝑘𝑏−𝑘′−𝑘𝑏2𝑘𝑏−𝑘𝑏−𝑘′−𝑘𝑏𝑘𝑏+𝑘′⎞⎟
⎟
⎟⎠.
Now 𝐾 is no longer tridiagonal, because the two ends are connected
directly. With 𝑘′ about 9% of 𝑘𝑏, this model reproduces both
frequencies. But we chose 𝑘′ to make that happen, so the match is not
evidence that the model is right. A real test needs a new prediction.
For example: how do the frequencies change when the carbon is replaced
by the heavier isotope carbon-13? That makes a good project.
Water is harder, because the molecule is bent: the two hydrogen atoms
sit on either side of the oxygen, at an angle of about 104.5∘. Its
atoms can move in two directions within the plane of the molecule. So a
water molecule has six coordinates, and 𝑀 is 6×6, with each
atom's mass appearing twice. The restoring forces are shown in
figure 10.12.
Each bond acts as a spring along its own direction. To first order, a
bond responds only to the part of the motion that stretches it. Sliding
one atom sideways past the other changes the bond length only to second
order, by the same calculation as in section 10.1. Our springs
there pulled sideways only because they were already under tension. A
third restoring force resists changes in the angle between the bonds.
In this planar model, three of water's six modes are zero modes: the molecule can drift in
two directions and rotate in its plane. The other three are vibrations,
a bend, a symmetric stretch, and an antisymmetric stretch, with
frequencies 1595, 3657, and 3756 inverse centimeters. Setting up
the 6×6 matrix with its constants, and comparing its modes with
these measurements, is another good project.
Figure 10.12 A spring model of a water molecule. Each oxygen–hydrogen bond acts as a
spring along its own direction, and a third restoring force, drawn as an
arc at the oxygen, resists changes in the angle between the bonds. The
atoms move within the plane of the page.
10.7Beyond Masses and Springs
Modes appear wherever a system makes small oscillations about an
equilibrium. A guitar
string and a drumhead vibrate in patterns very like our chain and grid
modes, and engineers compute the modes of a bridge or an airplane wing
before it is built. Even the whole Earth has modes. After a great
earthquake, the planet rings like a bell for weeks, and sensitive
instruments can measure its
slowest modes. In the "breathing" mode, the
whole Earth swells and shrinks with a period of about 20 minutes. In
the "football" mode, it stretches into a slightly elongated shape and
back, with a period of about 54 minutes.
Our chains point toward continuous systems like these. A chain has
finitely many masses, so a mode shape is a list of numbers and 𝐾 is a
matrix. A string has its mass spread along its whole length, so a mode
shape is a function: a whole curve. In section 10.3, as we divided
a string into more and more small masses, the chain's mode shapes
sampled sine curves more and more finely, keeping length, tension, and
mass per unit length fixed. In the limit, these become the continuous
modes of the string. This passage is called a continuum limit.
If we measure position along a
string of length ℓ by 𝑥, the eigenvector equation 𝐾𝑣=𝜆𝑣,
suitably rescaled, turns in the limit into an eigenvalue equation for
functions,
−𝑑2𝑣𝑑𝑥2=𝜇𝑣,𝑣(0)=𝑣(ℓ)=0,
whose solutions are the curves sin(𝑟𝜋𝑥/ℓ). A grid becomes a drumhead
in the same way.
Every motion of a chain is a combination of its modes, and the same is
true for a string. Writing an initial shape of the string as a sum of its
sine modes is called a Fourier series. Its coefficients come from the
same projections we used in section 10.4, with the dot
product replaced by an integral. Vibrating strings, drumheads, and
Fourier series belong to the study of partial differential equations.
The modes we have found here are the first step.
10.8Optional: Matrix Functions and Evolution
In section 10.2, we wrote every solution of 𝑞″=−𝐾𝑞 as
𝑞(𝑡)=𝐶𝐾(𝑡)𝑞0+𝑆𝐾(𝑡)𝑝0. Compare the first series
𝐶𝐾(𝑡)=𝐼−𝑡2𝐾2!+𝑡4𝐾24!−⋯
with the cosine series
cos(𝑡𝜔)=1−𝑡2𝜔22!+𝑡4𝜔44!−⋯.
The matrix 𝐾 occupies the place of 𝜔2. To express our series
as a cosine, we need a matrix playing the role of 𝜔: a square root
of 𝐾.
Return now to the symmetric matrices with positive eigenvalues studied
in section 10.4. Write
𝐾=𝑃diag(𝜆1,…,𝜆𝑛)𝑃−1,
where the columns of 𝑃 are eigenvectors. For a diagonal matrix, taking
a square root is straightforward: take the positive square root of each
diagonal entry. Converting back to the original coordinates suggests
Ω=𝑃diag(√𝜆1,…,√𝜆𝑛)𝑃−1.
Squaring confirms that it works:
Ω2=𝑃diag(𝜆1,…,𝜆𝑛)𝑃−1=𝐾.
We write this matrix as √𝐾. As with the matrix exponential,
√𝐾 does not mean taking square roots of the individual entries of
𝐾; we built it from the diagonalized matrix.
We define matrix cosine and sine using their familiar power series.
Substituting 𝐾=Ω2 into our first series gives
It has exactly the form of the single spring solution, with the scalar
frequency replaced by a matrix.
In section 10.2, we saw that 𝐶𝐾(𝑡) multiplies each
eigenvector 𝑣𝑗 by cos(𝜔𝑗𝑡), and 𝑆𝐾(𝑡) multiplies it by
sin(𝜔𝑗𝑡)/𝜔𝑗. This is the same rule we found for the
exponential: on an eigenvector, a matrix function acts by applying the
scalar function to its eigenvalue. For a diagonalizable matrix, these
actions determine the whole matrix function. Here, for example,
cos(𝑡Ω)=𝑃diag(cos(𝑡𝜔1),…,cos(𝑡𝜔𝑛))𝑃−1.
The change of coordinates, separate evolution of each mode, and return
to physical coordinates are all contained in this one matrix.
Finally, suppose 𝐾 is symmetric with nonnegative eigenvalues, allowing
some to be zero. The same construction still gives Ω=√𝐾.
If any eigenvalue is zero, Ω−1 does not exist. The original
series remain valid.
On a zero mode 𝐾𝑣=0, they give
𝐶𝐾(𝑡)𝑣=𝑣,𝑆𝐾(𝑡)𝑣=𝑡𝑣,
so that mode carries the familiar rest or drift motion, as for the
drifting molecule in section 10.6.2. This agrees with
lim𝜔→0sin(𝑡𝜔)𝜔=𝑡.
An expression that includes both positive and zero frequencies is
𝑆𝐾(𝑡)=∫𝑡0cos(𝑠Ω)𝑑𝑠.
Along each mode, the integral supplies precisely its scalar
initial-velocity solution.
Problems
1Why the Springs Must Be Stretched
In section 10.1, every spring was already under tension
𝑇0>0 at equilibrium. Here we see why that matters. A spring of
stiffness 𝑘 and natural length ℓ0 joins two points whose
horizontal separation is 𝑎, with ℓ0<𝑎. One end moves sideways by
𝛿 relative to the other.
(a)
Show that the spring's length is
ℓ=√𝑎2+𝛿2=𝑎+𝛿22𝑎+⋯, so it changes
only to second order in 𝛿. (Use √1+𝑥≈1+𝑥/2.)
(b)
The tension is 𝑇=𝑘(ℓ−ℓ0). Show that
𝑇=𝑇0+(second-orderterms), where 𝑇0=𝑘(𝑎−ℓ0).
(c)
The sideways component of the spring's force is 𝑇𝛿/ℓ. Show
that, to first order, it is (𝑇0/𝑎)𝛿, as in
section 10.1.
(d)
Now drop the assumption ℓ0<𝑎, and suppose the spring is
unstretched at equilibrium: ℓ0=𝑎, so
𝑇0=0. Show that the sideways force is approximately 𝑘𝛿3/(2𝑎2).
Why doesn't our theory of modes apply to an unstretched chain moving
sideways?
2Launching the Masses
Our two masses from section 10.1 start at equilibrium,
𝑞(0)=0, with velocities 𝑞′(0)=(1,−1).
(a)
Explain why this motion is a single mode, and write 𝑞(𝑡).
(b)
Check by differentiating that your 𝑞(𝑡) satisfies 𝑞″=−𝐾𝑞 and both
initial conditions.
(c)
Each mass starts with speed 1. How far does it travel before turning
around? Why is this distance less than 1?
3Masses on a Ring
Place 𝑛 equal masses at the corners of a regular 𝑛-sided polygon,
joined in a closed loop by springs under tension. Each mass slides on a
guide that lets it move only perpendicular to the plane of the polygon.
Let 𝑞𝑗 be the displacement of mass 𝑗, and choose units in which
𝑇0/(𝑚𝑎)=1, where 𝑎 is the side length.
(a)
Using the calculation in section 10.1, explain why
𝑞″𝑗=𝑞𝑗−1−2𝑞𝑗+𝑞𝑗+1,
where the indices wrap around: the neighbors of mass 1 are masses 𝑛
and 2, and the neighbors of mass 𝑛 are masses 𝑛−1 and 1. Write 𝐾
for 𝑛=4. How does it differ from the chain's matrix?
(b)
Show that (1,1,…,1) is an eigenvector with eigenvalue 0. What
motion is this?
(c)
Following section 10.3, show that both 𝑣𝑗=cos(𝑗𝜃) and
𝑣𝑗=sin(𝑗𝜃) satisfy every row equation with
𝜆=2−2cos𝜃, provided they are consistent with the wrapping
around. Which values of 𝜃 are allowed?
(d)
Take 𝑛=4. Find all four modes and their frequencies. Show that the
eigenvalue 2 has two shapes, (0,−1,0,1) and (1,0,−1,0). Show that
any combination of them, such as (1,−1,−1,1), is also a mode with the
same frequency.
(e)
Check that your four shapes are orthogonal. Why does this mean that
every motion of the ring is a combination of them?
4Repeated Frequencies on a Square Grid
For the 5×5 grid of section 10.5, choosing chain mode 𝑝
across the rows and 𝑟 down the columns gives the shape
𝑊(𝑝,𝑟)𝑖𝑗=sin𝑝𝜋𝑖6sin𝑟𝜋𝑗6.
(a)
Explain why 𝑊(1,2) and 𝑊(2,1) have the same frequency. Why does
any combination of them also give a mode with that frequency?
(b)
Show that 𝑊(1,2)−𝑊(2,1) is zero along the diagonal 𝑖=𝑗, and
that 𝑊(1,2)+𝑊(2,1) is zero along the other diagonal 𝑖+𝑗=6. (For the sketch, it helps to factor:
𝑊(1,2)±𝑊(2,1)=2sin𝜋𝑖6sin𝜋𝑗6(cos𝜋𝑗6±cos𝜋𝑖6).) Sketch both shapes. How do their still lines
differ from those of a single product mode?
(c)
Show that whenever 𝑝+𝑟=6, the frequency formula gives the same value,
because sin2𝑝𝜋12+sin2𝑟𝜋12=1. How many modes
share this frequency?
5Modes of a Shear Building
In section 10.6.1, we stated the modes of a building with 𝑛
equal floors and equal stories. Take 𝑘=𝑚=1, so 𝐾 has 2 on the
diagonal and −1 beside it, except that its last diagonal entry is 1.
In this problem, you will derive those modes by following the chain
calculation in section 10.3.
(a)
Write out the last row of 𝐾𝑣=𝜆𝑣. Show that it is the same as
the chain's row equation
2𝑣𝑛−𝑣𝑛−1−𝑣𝑛+1=𝜆𝑣𝑛
if we define 𝑣𝑛+1=𝑣𝑛. Explain this in physical terms: imagine an
extra floor above the roof, moving exactly with the roof. Why does the
spring joining them exert no force?
(b)
Guess 𝑣𝑗=sin(𝑗𝜃), so that 𝑣0=0 at the ground. Using the
calculation in section 10.3, explain why every row of
𝐾𝑣=𝜆𝑣 holds with 𝜆=2−2cos𝜃, provided that
sin((𝑛+1)𝜃)=sin(𝑛𝜃).
(c)
Use the identity sin𝐴−sin𝐵=2cos𝐴+𝐵2sin𝐴−𝐵2 to
show that, for 0<𝜃<𝜋, this condition holds exactly when
cos((𝑛+12)𝜃)=0. Conclude that the allowed values
are
𝜃𝑟=(2𝑟−1)𝜋2𝑛+1,𝑟=1,…,𝑛.
(d)
Show that the eigenvalues 𝜆𝑟=2−2cos𝜃𝑟=4sin2(𝜃𝑟/2)
are positive and distinct. Explain why the 𝑛 shapes form a basis of
ℝ𝑛, as in section 10.3.
(e)
Check your answer for a building with two floors. Find the eigenvalues
of
𝐾=(2−1−11)
directly, and compare them with 𝜆1 and 𝜆2.
6A First Model of Carbon Dioxide
In section 10.6.2, we modeled carbon dioxide as three atoms on a
line, oxygen–carbon–oxygen, joined by two bond springs of stiffness 𝑘.
Let 𝑞1,𝑞2,𝑞3 be the displacements of the atoms along the line.
(a)
Write down the force on each atom, and show that Newton's law gives
𝑀𝑞″=−𝐾𝑞 with
For each eigenvalue, solve (𝐾−𝜆𝑀)𝑣=0 to find a mode shape.
Describe how the three atoms move in each mode. Which mode is not a
vibration?
(d)
Find the ratio 𝜔anti/𝜔sym of the two
vibration frequencies. Using 𝑚𝑂=16 and 𝑚𝐶=12, compare your answer
with the ratio 2349/1333 of the frequencies listed in standard tables.
(e)
Check that both vibration shapes satisfy 𝑚𝑂𝑣1+𝑚𝐶𝑣2+𝑚𝑂𝑣3=0.
Explain why this means that the molecule's center of mass does not move
during a vibration.
7Symmetrizing Unequal Masses
Two masses, 𝑚1=1 and 𝑚2=2, move along a line between two walls. A
spring of stiffness 1 joins the first mass to the left wall, a spring
of stiffness 2 joins the two masses, and a spring of stiffness 2 joins
the second mass to the right wall.
(a)
Show that the equations of motion are 𝑀𝑞″=−𝐾𝑞 with
𝐾=(3−2−24),𝑀=(1002).
(b)
Solve det(𝐾−𝜆𝑀)=0, and find a mode shape for each eigenvalue;
call them 𝑣1 and 𝑣2.
(c)
Show that your two shapes are not orthogonal, but that 𝑣1⋅𝑀𝑣2=0.
(d)
Compute 𝑀−1/2𝐾𝑀−1/2 and check that it is symmetric. Find its
eigenvalues and a pair of orthogonal eigenvectors 𝑤1,𝑤2. Check that
each 𝑀−1/2𝑤𝑗 is a multiple of your shape 𝑣𝑗. Explain, as in the
remark in section 10.6.2, why this always works.
8From springs to motion
Two masses, each of mass 1, slide along a line between two walls. Each is
attached to its wall by a spring of stiffness 1, and a spring of stiffness
2 joins them. Let 𝑞1 and 𝑞2 be their displacements from equilibrium.
(a)
Show that 𝑞″=−𝐾𝑞 with 𝐾=(3−2−23).
(b)
Find the mode shapes and their angular frequencies. Explain physically why
the mode in which the masses move together has angular frequency 1.
(c)
The first mass is pulled one unit to the right, and both masses are released
from rest. Use theorem 10.5 to find 𝑞(𝑡).
(d)
Does this motion ever return exactly to its starting state?
9Striking the middle mass
The three-mass chain of section 10.3 is at rest at equilibrium when its
middle mass is struck, so 𝑞(0)=0 and 𝑞′(0)=(0,1,0).
One mode is missing from your answer. Which one, and why? Use the shape of
that mode and the symmetry of the initial condition.
(c)
Show that 𝑞1(𝑡)=𝑞3(𝑡) for all 𝑡. Find initial conditions that excite
only the missing mode.
10Each mode is an invariant plane
Write 𝑞″=−𝐾𝑞 as the first-order system 𝑋′=𝐴𝑋, with 𝑋=(𝑞,𝑞′) and
𝐴=(0𝐼−𝐾0). Suppose 𝐾𝑣=𝜆𝑣 with
𝑣≠0.
(a)
Compute 𝐴 applied to the states (𝑣,0) and (0,𝑣). Explain why the plane
spanned by these two states is invariant.
(b)
Show that, in the basis (𝑣,0),(0,𝑣) of this plane, 𝐴 has matrix
(01−𝜆0). Which single-spring equation
has this matrix?
(c)
For 𝜆>0, what kind of flow does this give in the plane, in the
language of Chapter 8? What would happen if 𝜆<0?
(d)
For the two-mass chain, write down the two invariant planes, and explain why
together they fill the four-dimensional state space.
11The evolution matrices of two masses
For the two-mass chain, 𝐾=(2−1−12), with mode
shapes (1,1) and (1,−1) and angular frequencies 1 and √3.
(a)
Each eigenvector satisfies 𝐶𝐾(𝑡)𝑣=cos(√𝜆𝑡)𝑣 and
𝑆𝐾(𝑡)𝑣=sin(√𝜆𝑡)𝑣/√𝜆. Use this to find 𝐶𝐾(𝑡)
and 𝑆𝐾(𝑡) as 2×2 matrices.
(b)
Use theorem 10.1 to write the solution for any 𝑞0 and
𝑝0. Check that it agrees with example 10.6.
(c)
(Uses section 10.8.) Find √𝐾, and check that its square
is 𝐾.
12Using the chain formula
In section 10.3 we found that a chain of 𝑛 masses has mode shapes
𝑣(𝑟)𝑗=sin𝑟𝑗𝜋𝑛+1 and angular frequencies
𝜔𝑟=2sin𝑟𝜋2(𝑛+1).
(a)
Check that the formula gives the frequencies 1 and √3 for two
masses.
(b)
For a chain of five masses, list the five angular frequencies. In each mode,
which masses never move?
(c)
With 𝑚=𝑎=𝑇0=1, how many masses must a chain have before its lowest
angular frequency is below 0.1? Explain physically why adding masses
lowers it.
13Holding a mass still
In some modes, a mass never moves. This problem asks what happens if we hold
such a mass still.
(a)
In the three-mass chain of section 10.3, a clamp holds the middle mass
fixed at equilibrium. The two outer masses can still move. What angular
frequency do they oscillate with? Which mode of the unclamped chain has the
same frequency, and why does that mode not notice the clamp?
(b)
Now clamp the middle mass of the chain of five masses. Which modes of the
unclamped chain keep moving exactly as before? Compare their frequencies
with those of two separate two-mass chains.
(c)
Guitarists play a "harmonic" by lightly touching a string at its midpoint,
which damps any mode that moves there. Consider a long chain with an odd
number of masses, and hold its middle mass still. Which modes can keep
moving? For a long chain, 𝜔2≈2𝜔1. How does the lowest
remaining frequency compare with the chain's usual lowest frequency?
Python: Real Buildings
14A Four-Story Building
Mark Austin's
four-story shear building example
has unequal floor masses and story stiffnesses. From the roof down, the
floor masses are 1500, 3000, 3000, and 4500 kg, and the story
stiffnesses are 0.8, 1.6, 2.4, and 3.2 MN/m. The last story joins
the lowest floor to the ground.
(a)
Number the floors from the roof down. Use our recipe from
section 10.4 to assemble 𝐾, without dividing by the
masses as in section 10.6.2, and write down 𝑀. Each
story spring joins two neighboring floors, or the lowest floor to the
ground.
(b)
Use Python to solve 𝐾𝑣=𝜆𝑀𝑣. (The function scipy.linalg.eigh(K, M)
solves this problem directly.) Find the angular frequencies and the
periods 𝑇=2𝜋/𝜔.
(c)
Compare your periods with Austin's: 0.579, 0.260, 0.187, and
0.138 seconds. Describe the four mode shapes. Where does each move the
most, and how many times does it change direction?
(d)
For the slowest mode, compute the drift in each story. Which story
deforms the most? Compare with the equal-story building in
section 10.6.1, and explain the difference.