Chapter 1

Oscillations
David Morin, [email protected]

A wave is a correlated collection of oscillations. For example, in a transverse wave traveling

along a string, each point in the string oscillates back and forth in the transverse direc-
tion (not along the direction of the string). In sound waves, each air molecule oscillates
back and forth in the longitudinal direction (the direction in which the sound is traveling).
The molecules don’t have any net motion in the direction of the sound propagation. In
water waves, each water molecule also undergoes oscillatory motion, and again, there is no
overall net motion.1 So needless to say, an understanding of oscillations is required for an
understanding of waves.
The outline of this chapter is as follows. In Section 1.1 we discuss simple harmonic
motion, that is, motioned governed by a Hooke’s law force, where the restoring force is
proportional to the (negative of the) displacement. We discuss various ways to solve for the
position x(t), and we give a number of examples of such motion. In Section 1.2 we discuss
damped harmonic motion, where the damping force is proportional to the velocity, which
is a realistic damping force for a body moving through a fluid. We will find that there are
three basic types of damped harmonic motion. In Section 1.3 we discuss damped and driven
harmonic motion, where the driving force takes a sinusoidal form. (When we get to Fourier
analysis, we will see why this is actually a very general type of force to consider.) We present
three different methods of solving for the position x(t). In the special case where the driving
frequency equals the natural frequency of the spring, the amplitude becomes large. This is
called resonance, and we will discuss various examples.

1.1 Simple harmonic motion

1.1.1 Hooke’s law and small oscillations
Consider a Hooke’s-law force, F (x) = −kx. Or equivalently, consider the potential energy,
V (x) = (1/2)kx2 . An ideal spring satisfies this force law, although any spring will deviate
significantly from this law if it is stretched enough. We study this F (x) = −kx force because:
1 The ironic thing about water waves is that although they might be the first kind of wave that comes to

mind, they’re much more complicated than most other kinds. In particular, the oscillations of the molecules
are two dimensional instead of the normal one dimensional linear oscillations. Also, when waves “break”
near a shore, everything goes haywire (the approximations that we repeatedly use throughout this book
break down) and there ends up being some net forward motion. We’ll talk about water waves in Chapter
12.

1
 2 CHAPTER 1. OSCILLATIONS

• We can study it. That it, we can solve for the motion exactly. There are many
problems in physics that are extremely difficult or impossible to solve, so we might as
well take advantage of a problem we can actually get a handle on.
• It is ubiquitous in nature (at least approximately). It holds in an exact sense for
an idealized spring, and it holds in an approximate sense for a real-live spring, a
small-angle pendulum, a torsion oscillator, certain electrical circuits, sound vibrations,
molecular vibrations, and countless other setups. The reason why it applies to so many
V(x)
situations is the following.
Let’s consider an arbitrary potential, and let’s see what it looks like near a local min-
imum. This is a reasonable place to look, because particles generally hang out near a
minimum of whatever potential they’re in. An example of a potential V (x) is shown in
Fig. 1. The best tool for seeing what a function looks like in the vicinity of a given point
x is the Taylor series, so let’s expand V (x) in a Taylor series around x0 (the location of the
x0
minimum). We have
Figure 1 1 00 1
V (x) = V (x0 ) + V 0 (x0 )(x − x0 ) + V (x0 )(x − x0 )2 + V 000 (x0 )(x − x0 )3 + · · · (1)
2! 3!
On the righthand side, the first term is irrelevant because shifting a potential by a constant
amount doesn’t change the physics. (Equivalently, the force is the derivative of the potential,
and the derivative of a constant is zero.) And the second term is zero due to the fact that
we’re looking at a minimum of the potential, so the slope V 0 (x0 ) is zero at x0 . Furthermore,
the (x − x0 )3 term (and all higher order terms) is negligible compared with the (x − x0 )2
term if x is sufficiently close to x0 , which we will assume is the case.2 So we are left with

1 00
V (x) ≈ V (x0 )(x − x0 )2 (2)
2
parabola In other words, we have a potential of the form (1/2)kx2 , where k ≡ V 00 (x0 ), and where we
have shifted the origin of x so that it is located at x0 . Equivalently, we are just measuring
x relative to x0 .
We see that any potential looks basically like a Hooke’s-law spring, as long as we’re close
V(x) enough to a local minimum. In other words, the curve can be approximated by a parabola,
as shown in Fig. 2. This is why the harmonic oscillator is so important in physics.
We will find below in Eqs. (7)p and (11) that the (angular) frequency of the motion in
Figure 2 a Hooke’s-law potential is ω = k/m. So for a general potential V (x), the k ≡ V 00 (x0 )
equivalence implies that the frequency is
r
V 00 (x0 )
ω= . (3)
m

1.1.2 Solving for x(t)

The long way
The usual goal in a physics setup is to solve for x(t). There are (at least) two ways to do
this for the force F (x) = −kx. The straightforward but messy way is to solve the F = ma
differential equation. One way to write F = ma for a harmonic oscillator is −kx = m·dv/dt.
However, this isn’t so useful, because it contains three variables, x, v, and t. We therefore
2 The one exception occurs when V 00 (x) equals zero. However, there is essentially zero probability that

V 00 (x0 ) = 0 for any actual potential. And even if it does, the result in Eq. (3) below is still technically true;
they frequency is simply zero.
1.1. SIMPLE HARMONIC MOTION 3

can’t use the standard strategy of separating variables on the two sides of the equation
and then integrating. Equation have only two sides, after all. So let’s instead write the
acceleration as a = v · dv/dx. 3 This gives
µ ¶ Z Z
dv
F = ma =⇒ −kx = m v =⇒ − kx dx = mv dv. (4)
dx

Integration then gives (with E being the integration constant, which happens to be the
energy) r r
1 2 1 2 2 1 2
E − kx = mv =⇒ v = ± E− kx . (5)
2 2 m 2
Writing v as dx/dt here and separating variables one more time gives
r Z
dx 2
√ q kx2
=±
m
dt. (6)
E 1− 2E

A trig substitution turns the lefthand side into an arccos (or arcsin) function. The result is
(see Problem [to be added] for the details)
r
k
x(t) = A cos(ωt + φ) where ω= (7)
m

and where A and φ are arbitrary constants that are determined by the two initial conditions
(position
p and velocity); see the subsection below on initial conditions. A happens to be
2E/k, where E is the above constant of integration. The solution in Eq. (7) describes
simple harmonic motion, where x(t) is a simple sinusoidal function of time. When we discuss
damping in Section 1.2, we will find that the motion is somewhat sinusoidal, but with an
important modification.

The short way

F = ma gives
d2 x
−kx = m . (8)
dt2
This equation tells us that we want to find a function whose second derivative is proportional
to the negative of itself. But we already know some functions with this property, namely
sines, cosines, and exponentials. So let’s be fairly general and try a solution of the form,

x(t) = A cos(ωt + φ). (9)

A sine or an exponential function would work just as well. But a sine function is simply
a shifted cosine function, so it doesn’t really generate anything new; it just changes the
phase. We’ll talk about exponential solutions in the subsection below. Note that a phase φ
(which shifts the curve on the t axis), a scale factor of ω in front of the t (which expands or
contracts the curve on the t axis), and an overall constant A (which expands or contracts
the curve on the x axis) are the only ways to modify a cosine function if we want it to stay
a cosine. (Well, we could also add on a constant and shift the curve in the x direction, but
we want the motion to be centered around x = 0.)
3 This does indeed equal a, because v · dv/dx = dx/dt · dv/dx = dv/dt = a. And yes, it’s legal to cancel

the dx’s here (just imagine them to be small but not infinitesimal quantities, and then take a limit).
4 CHAPTER 1. OSCILLATIONS

If we plug Eq. (9) into Eq. (8), we obtain

³ ´ ³ ´
−k A cos(ωt + φ) = m − ω 2 A cos(ωt + φ)
³ ´
=⇒ (−k + mω 2 ) A cos(ωt + φ) = 0. (10)

Since this must be true for all t, we must have

r
2 k
k − mω = 0 =⇒ ω = , (11)
m

in agreement with Eq. (7). The constants φ and A don’t appear in Eq. p (11), so they can
be anything and the solution in Eq. (9) will still work, provided that ω = k/m. They are
determined by the initial conditions (position and velocity).
We have found one solution in Eq. (9), but how do we know that we haven’t missed any
other solutions to the F = ma equation? From the trig sum formula, we can write our one
solution as
A cos(ωt + φ) = A cos φ cos(ωt) − A sin φ sin(ωt), (12)
So we have actually found two solutions: a sin and a cosine, with arbitrary coefficients in
front of each (because φ can be anything). The solution in Eq. (9) is simply the sum of
these two individual solutions. The fact that the sum of two solutions is again a solution
is a consequence of the linearity our F = ma equation. By linear, we mean that x appears
only through its first power; the number of derivatives doesn’t matter.
We will now invoke the fact that an nth-order linear differential equation has n indepen-
dent solutions (see Section 1.1.4 below for some justification of this). Our F = ma equation
in Eq. (8) involves the second derivative of x, so it is a second-order equation. So we’ll
accept the fact that it has two independent solutions. Therefore, since we’ve found two, we
know that we’ve found them all.

The parameters
A few words on the various quantities that appear in the x(t) in Eq. (9).

• ω is the angular frequency.4 Note that

µ ¶
2π ¡ ¢
x t+ = A cos ω(t + 2π/ω) + φ = A cos(ωt + φ + 2π)
ω
= A cos(ωt + φ)
= x(t). (13)

Also, using v(t) = dx/dt = −ωA sin(ωt + φ), we find that v(t + 2π/ω) = v(t). So
after a time of T ≡ 2π/ω, both the position and velocity are back to where they were
(and the force too, since it’s proportional to x). This time T isp
therefore the period.
The motion
p repeats after every time interval of T . Using ω = k/m, we can write
T = 2π m/k.
4 It is sometimes also called the angular speed or angular velocity. Although there are technically differ-

ences between these terms, we’ll generally be sloppy and use them interchangeably. Also, it gets to be a
pain to keep saying the word “angular,” so we’ll usually call ω simply the “frequency.” This causes some
ambiguity with the frequency, ν, as measured in Hertz (cycles per second); see Eq. (14). But since ω is a
much more natural quantity to use than ν, we will invariably work with ω. So “frequency” is understood
to mean ω in this book.
1.1. SIMPLE HARMONIC MOTION 5

The frequency in Hertz (cycles per second) is given by ν = 1/T . For example, if T =
0.1 s, then ν = 1/T = 10 s−1 , which means that the system undergoes 10 oscillations
per second. So we have r
1 ω 1 k
ν= = = . (14)
T 2π 2π m
To remember where the “2π” in ν = ω/2π goes, note that ω is larger than ν by a
factor of 2π, because one revolution has 2π radians in it, and ν is concerned with
revolutions whereas ω is concerned with radians.
Note the extremely important point that the frequency is independent of the ampli-
tude. You might think that the frequency should be smaller if the amplitude is larger,
because the mass has farther to travel. But on the other hand, you might think that
the frequency should be larger if the amplitude is larger, because the force on the
mass is larger which means that it is moving faster at certain points. It isn’t intu-
itively obvious which of these effects wins, although it does follow from dimensional
analysis (see Problem [to be added]). It turns out that the effects happen to exactly
x(t)
cancel, making the frequency independent of the amplitude. Of course, in any real-life
A
situation, the F (x) = −kx form of the force will break down if the amplitude is large
enough. But in the regime where F (x) = −kx is a valid approximation, the frequency t
is independent of the amplitude.
• A is the amplitude. The position ranges from A to −A, as shown in Fig. 3 -A

• φ is the phase. It gives a measure of what the position is a t = 0. φ is dependent on Figure 3

when you pick the t = 0 time to be. Two people who start their clocks at different
times will have different phases in their expressions for x(t). (But they will have the
same ω and A.) Two phases that differ by 2π are effectively the same phase.
Be careful with the sign of the phase. Fig. 4 shows plots of A cos(ωt + φ), for φ = 0,
±π/2, and π. Note that the plot for φ = +π/2 is shifted to the left of the plot for
φ = 0, whereas the plot for φ = −π/2 is shifted to the right of the plot for φ = 0.
These are due to the fact that, for example, the φ = −π/2 case requires a larger time
to achieve the same position as the φ = 0 case. So a given value of x occurs later in
the φ = −π/2 plot, which means that it is shifted to the right.

Acos(ωt)

Acos(ωt-π/2)
t
2 4 6 8 10 12
Acos(ωt+π/2)
Acos(ωt+π)

Figure 4

Various ways to write x(t)

We found above that x(t) can be expressed as x(t) = A cos(ωt + φ). However, this isn’t the
only way to write x(t). The following is a list of equivalent expressions.

x(t) = A cos(ωt + φ)
6 CHAPTER 1. OSCILLATIONS

= A sin(ωt + φ0 )
= Bc cos ωt + Bs sin ωt
= Ceiωt + C ∗ e−iωt
¡ ¢
= Re Deiωt . (15)
A, Bc , and Bs are real quantities here, but C and D are (possibly) complex. C ∗ denotes
the complex conjugate of C. See Section 1.1.5 below for a discussion of matters involving
complex quantities. Each of the above expressions for x(t) involves two parameters – for
example, A and φ, or the real and imaginary parts of C. This is consistent with the fact
that there are two initial conditions (position and velocity) that must be satisfied.
The two parameters in a given expression are related to the two parameters in each of
the other expressions. For example, φ0 = φ + π/2, and the various relations among the other
parameters can be summed up by
Bc = A cos φ = 2Re(C) = Re(D),
Bs = −A sin φ = −2Im(C) = −Im(D), (16)
and a quick corollary is that D = 2C. The task of Problem [to be added] is to verify these
relations. Depending on the situation at hand, one of the expressions in Eq. (15) might
work better than the others, as we’ll see in Section 1.1.7 below.

1.1.3 Linearity
As we mentioned right after Eq. (12), linear differential equations have the property that
the sum (or any linear combination) of two solutions is again a solution. For example, if
cos ωt and sin ωt are solutions, then A cos ωt + B sin ωt is also a solution, for any constants
A and B. This is consistent with the fact that the x(t) in Eq. (12) is a solution to our
Hooke’s-law mẍ = −kx equation.
This property of linear differential equations is easy to verify. Consider, for example, the
second order (although the property holds for any order) linear differential equation,
Aẍ + B ẋ + Cx = 0. (17)
Let’s say that we’ve found two solutions, x1 (t) and x2 (t). Then we have
Aẍ1 + B ẋ1 + Cx1 = 0,
Aẍ2 + B ẋ2 + Cx2 = 0. (18)
If we add these two equations, and switch from the dot notation to the d/dt notation, then
we have (using the fact that the sum of the derivatives is the derivative of the sum)
d2 (x1 + x2 ) d(x1 + x2 )
A +B + C(x1 + x2 ) = 0. (19)
dt2 dt
But this is just the statement that x1 + x2 is a solution to our differential equation, as we
wanted to show.
What if we have an equation that isn’t linear? For example, we might have
Aẍ + B ẋ2 + Cx = 0. (20)
If x1 and x2 are solutions to this equation, then if we add the differential equations applied
to each of them, we obtain
"µ ¶2 µ ¶2 #
d2 (x1 + x2 ) dx1 dx1
A +B + + C(x1 + x2 ) = 0. (21)
dt2 dt dt
1.1. SIMPLE HARMONIC MOTION 7

This is not the statement that x1 + x2 is a solution, which is instead

µ ¶2
d2 (x1 + x2 ) d(x1 + x2 )
A +B + C(x1 + x2 ) = 0. (22)
dt2 dt

The two preceding equations differ by the cross term in the square in the latter, namely
2B(dx1 /dt)(dx2 /dt). This is in general nonzero, so we conclude that x1 +x2 is not a solution.
No matter what the order if the differential equation is, we see that these cross terms will
arise if and only if the equation isn’t linear.
This property of linear differential equations – that the sum of two solutions is again a
solution – is extremely useful. It means that we can build up solutions from other solutions.
Systems that are governed by linear equations are much easier to deal with than systems
that are governed by nonlinear equations. In the latter, the various solutions aren’t related
in an obvious way. Each one sits in isolation, in a sense. General Relativity is an example
of a theory that is governed by nonlinear equations, and solutions are indeed very hard to
come by.

1.1.4 Solving nth-order linear differential equations

The “fundamental theorem of algebra” states that any nth-order polynomial,

an z n + an−1 z n−1 + · · · + a1 z + a0 , (23)

can be factored into

an (z − r1 )(z − r2 ) · · · (z − rn ). (24)
This is believable, but by no means obvious. The proof is a bit involved, so we’ll just accept
it here.
Now consider the nth-order linear differential equation,

dn x dn−1 x dx
an n
+ a n−1 n−1
+ · · · + a1 + a0 = 0. (25)
dt dt dt
Because differentiation by t commutes with multiplication by a constant, we can invoke the
equality of the expressions in Eqs. (23) and (24) to say that Eq. (25) can be rewritten as
µ ¶µ ¶ µ ¶
d d d
an − r1 − r2 · · · − rn x = 0. (26)
dt dt dt

In short, we can treat the d/dt derivatives here like the z’s in Eq. (24), so the relation
between Eqs. (26) and (25) is the same as the relation between Eqs. (24) and (23). And
because all the factors in Eq. (26) commute with each other, we can imagine making any of
the factors be the rightmost one. Therefore, any solution to the equation,
µ ¶
d dx
− ri x = 0 ⇐⇒ = ri x, (27)
dt dt

is a solution to the original equation, Eq. (25). The solutions to these n first-order equations
are simply the exponential functions, x(t) = Aeri t . We have therefore found n solutions,
so we’re done. (We’ll accept the fact that there are only n solutions.) So this is why our
strategy for solving differential equations is to always guess exponential solutions (or trig
solutions, as we’ll see in the following section).
8 CHAPTER 1. OSCILLATIONS

1.1.5 Taking the real part

In the second (short) derivation of x(t) we presented above, we guessed a solution of the
form, x(t) = A cos(ωt + φ). However, anything that can be written in terms of trig functions
can also be written in terms of exponentials. This fact follows from one of the nicest formulas
in mathematics:
eiθ = cos θ + i sin θ (28)
This can be proved in various ways, the quickest of which is to write down the Taylor series
for both sides and then note that they are equal. If we replace θ with −θ in this relation,
we obtain e−iθ = cos θ − i sin θ, because cos θ and sin θ are even and odd functions of θ,
respectively. Adding and subtracting this equation from Eq. (28) allows us to solve for the
trig functions in terms of the exponentials:

eiθ + e−iθ eiθ − e−iθ

cos θ = , and sin θ = . (29)
2 2i
So as we claimed above, anything that can be written in terms of trig functions can also be
written in terns of exponentials (and vice versa). We can therefore alternatively guess an
exponential solution to our −kx = mẍ differential equation. Plugging in x(t) = Ceαt gives
r
αt 2 αt 2 k k
−kCe = mα Ce =⇒ α = − =⇒ α = ±iω, where ω = . (30)
m m
We have therefore found two solutions, x1 (t) = C1 eiωt , and x2 (t) = C2 e−iωt . The C1
coefficient here need not have anything to do with the C2 coefficient. Due to linearity, the
most general solution is the sum of these two solutions,

x(t) = C1 eiωt + C2 e−iωt (31)

This expression for x(t) satisfies the −kx = mẍ equation for any (possibly complex) values
of C1 and C2 . However, x(t) must of course be real, because an object can’t be at a position
of, say, 3+7i meters (at least in this world). This implies that the two terms in Eq. (31) must
be complex conjugates of each other, which in turn implies that C2 must be the complex
conjugate of C1 . This is the reasoning that leads to the fourth expression in Eq. (15).
There are two ways to write any complex number: either as the sum of a real and
imaginary part, or as the product of a magnitude and a phase eiφ . The equivalence of these
is a consequence of Eq. (28). Basically, if we plot the complex number in the complex plane,
we can write it in either Cartesian or polar coordinates. If we choose the magnitude-phase
way and write C1 as C0 eiφ , then the complex conjugate is C2 = C0 e−iφ . Eq. (31) then
becomes

x(t) = C0 eiφ eiωt + C0 e−iφ e−iωt

= 2C0 cos(ωt + φ), (32)

where we have used Eq. (29). We therefore end up with the trig solution that we had
originally obtained by guessing, so everything is consistent.
Note that by adding the two complex conjugate solutions together in Eq. (32), we ba-
sically just took the real part of the C0 eiφ eiωt solution (and then multiplied by 2, but that
can be absorbed in a redefinition of the coefficient). So we will often simply work with the
exponential solution, with the understanding that we must take the real part in the end to
get the actual physical solution.
If this strategy of working with an exponential solution and then taking the real part
seems suspect or mysterious to you, then for the first few problems you encounter, you
1.1. SIMPLE HARMONIC MOTION 9

should do things the formal way. That is, you should add on the second solution and then
demand that x(t) (or whatever the relevant variable is in a given setup) is real. This will
result in the sum of two complex conjugates. After doing this a few of times, you will realize
that you always end up with (twice) the real part of the exponential solutions (either of
them). Once you’re comfortable with this fact, you can take a shortcut and forget about
adding on the second solution and the other intermediate steps. But that’s what you’re
really doing.

Remark: The original general solution for x(t) in Eq. (31) contains four parameters, namely the
real and imaginary parts of C1 , and likewise for C2 (ω is determined by k and m). Or equivalently,
the four parameters are the magnitude and phase of C1 , and likewise for C2 . These four parameters
are all independent, because we haven’t yet invoked the fact that x(t) must be real. If we do invoke
this, it cuts down the number of parameters from four to two. These two parameters are then
determined by the two initial conditions (position and velocity).
However, although there’s no arguing with the “x(t) must be real” reasoning, it does come a
little out of the blue. It would be nice to work entirely in terms of initial conditions. But how
can we solve for four parameters with only two initial conditions? Well, we can’t. But the point
is that there are actually four initial conditions, namely the real and imaginary parts of the initial
position, and the real and imaginary parts of the initial velocity. That is, x(0) = x0 + 0 · i, and
v(0) = v0 + 0 · i. It takes four quantities (x0 , 0, v0 , and 0 here) to specify these two (possibly
complex) quantities. (Once we start introducing complex numbers into x(t), we of course have to
allow the initial position and velocity to be complex.) These four given quantities allow us to solve
for the four parameters in x(t). And in the end, this process gives (see Problem [to be added]) the
same result as simply demanding that x(t) is real. ♣

1.1.6 Phase relations and phasor diagrams

Let’s now derive the phase relation between x(t), v(t), and a(t). We have

x(t) = A cos(ωt + φ),

dx ³ π´
=⇒ v(t) = = −ωA sin(ωt + φ) = ωA cos ωt + φ + ,
dt 2 ωA A
dv
=⇒ a(t) = = −ω 2 A cos(ωt + φ) = ω 2 A cos (ωt + φ + π) . (33) ωt+φ
dt
x,v,a
We see that a leads v by π/2, and v leads x by π/2. These phase relations can be conveniently
expressed in the phasor diagram in Fig. 5. The values of x, v, and a are represented by
vectors, where it is understood that to get their actual values, we must take the projection ω2 A
of the vectors onto the horizontal axis. The whole set of three vectors swings around
counterclockwise with angular speed ω as time goes on. The initial angle between the x horizontal projections
phasor and the horizontal axis is picked to be φ. So the angle of the x phasor as a function give x,v, and a
of time is ωt + φ, the angle of the v phasor is ωt + φ + π/2, and the angle of the a phasor
Figure 5
is ωt + φ + π. Since taking the horizontal projection simply brings in a factor of the cosine
of the angle, we do indeed obtain the expressions in Eq. (33), as desired.
The units of x, v, and a are different, so technically we shouldn’t be drawing all three
phasors in the same diagram, but it helps in visualizing what’s going on. Since the phasors
swing around counterclockwise, the diagram makes it clear that a(t) is π/2 ahead of v(t),
which is π/2 ahead of x(t). So the actual cosine forms of x, v, and a look like the plots
shown in Fig. 6 (we’ve chosen φ = 0 here).
 10 CHAPTER 1. OSCILLATIONS

a(t) v(t) x(t)

x,v,a
1.0

0.5

t
2 4 6 8 10 12

- 0.5

- 1.0

equilibrium Figure 6

a(t) reaches its maximum before v(t) (that is, a(t) is ahead of v(t)). And v(t) reaches its
(max a) maximum before x(t) (that is, v(t) is ahead of x(t)). So the plot of a(t) is shifted to the left
from v(t), which is shifted to the left from x(t). If we look at what an actual spring-mass
system is doing, we have the three successive pictures shown in Fig. 7. Figures 5, 6, and 7
are three different ways of saying the same thing about the relative phases.
(then max v)

1.1.7 Initial conditions

(then max x) As we mentioned above, each of the expressions for x(t) in Eq. (15) contains two parameters,
and these two parameters are determined from the initial conditions. These initial conditions
Figure 7 are invariably stated as the initial position and initial velocity. In solving for the subsequent
motion, any of the forms in Eq. (15) will work, but the

x(t) = Bc cos ωt + Bs sin ωt (34)

form is the easiest one to work with when given x(0) and v(0). Using

dx
v(t) = = −ωBc sin ωt + ωBs cos ωt, (35)
dt
the conditions x(0) = x0 and v0 = v0 yield
v0
x0 = x(0) = Bc , and v0 = v(0) = ωBs =⇒ Bs = . (36)
ω
Therefore,
v0
x(t) = x0 cos ωt + sin ωt (37)
ω
If you wanted to use the x(t) = A cos(ωt + φ) form instead, then v(t) = −ωA sin(ωt + φ).
The initial conditions now give x0 = x(0) = A cos φ and v0 = −ωA sin pφ. Dividing gives
tan φ = −v0 /ωx0 . Squaring and adding (after dividing by ω) gives A = x20 + (v0 /ω)2 . We
have chosen the positive root for A; the negative root would simply add π on to φ. So we
have r µ
³ v ´2 ³ −v ´¶
0 0
x(t) = x20 + cos ωt + arctan . (38)
ω ωx0
The correct choice from the two possibilities for the arctan angle is determined by either
cos φ = x0 /A or sin φ = −v0 /ωA. The result in Eq. (38) is much messier than the result in
Eq. (37), so it is clearly advantageous to use the form of x(t) given in Eq. (34).
1.1. SIMPLE HARMONIC MOTION 11

All of the expressions for x(t) in Eq. (15) contain two parameters. If someone proposed
a solution with only one parameter, then there is no way that it could be a general solution,
because we don’t have the freedom to satisfy the two initial conditions. Likewise, if someone
proposed a solution with three parameters, then there is no way we could ever determine
all three parameters, because we have only two initial conditions. So it is good that the
expressions in Eq. (15) contain two parameters. And this is no accident. It follows from the
fact that our F = ma equation is a second-order differential equation; the general solution
to such an equation always contains two free parameters.

We therefore see that the fact that two initial conditions completely specify the motion
of the system is intricately related to the fact that the F = ma equation is a second-
order differential equation. If instead of F = mẍ, Newton’s second law was the first-order
equation, F = mẋ, then we wouldn’t have the freedom of throwing a ball with an initial
velocity of our choosing; everything would be determined by the initial position only. This is
clearly not how things work. On the other hand, if Newton’s second law was the third-order
equation, F = md3 x/dt3 , then the motion of a ball wouldn’t be determined by an initial
position and velocity (along with the forces in the setup at hand); we would also have to
state the initial acceleration. And again, this is not how things work (in this world, at least).

1.1.8 Energy

F (x) = −kx is a conservative force. That is, the work done by the spring is path-
independent. Or equivalently, the work done depends only on R the initial position xi and
the final position xf . You can quickly show that that work is (−kx) dx = kx2i /2 − kx2f /2.
Therefore, since the force is conservative, the energy is conserved. Let’s check that this is
indeed the case. We have

1 2 1
E = kx + mv 2
2 2
1 ¡ ¢2 1 ¡ ¢2
= k A cos(ωt + φ) + m − ωA sin(ωt + φ)
2 2
1 2³ ´
= A k cos2 (ωt + φ) + mω 2 sin2 (ωt + φ)
2
1 2³ ´
= kA cos2 (ωt + φ) + sin2 (ωt + φ) (using ω 2 ≡ k/m)
2
1 2
= kA . (39)
2

This makes sense, because kA2 /2 is the potential energy of the spring when it is stretched
the maximum amount (and so the mass is instantaneously at rest). Fig. 8 shows how the
energy breaks up into kinetic and potential, as a function of time. We have arbitrarily
chosen φ to be zero. The energy sloshes back and forth between kinetic and potential. It is
all potential at the points of maximum stretching, and it is all kinetic when the mass passes
through the origin.
 12 CHAPTER 1. OSCILLATIONS

x, PE, KE
1.5 x(t) = Acos ωt
1_ 1_
1.0 PE = kx2 = kA2cos2 ωt
2 2
0.5 1_ 2 1_
KE = mv = mω2A2sin2 ωt
2 2
t
2 4 6 8 10 12 1_
- 0.5 = kA2sin2 ωt
2
- 1.0
- 1.5

Figure 8

1.1.9 Examples
Let’s now look at some examples of simple harmonic motion. Countless examples exist in
the real word, due to the Taylor-series argument in Section 1.1. Technically, all of these
examples are just approximations, because a force never takes exactly the F (x) = −kx form;
there are always slight modifications. But these modifications are negligible if the amplitude
is small enough. So it is understood that we will always work in the approximation of small
amplitudes. Of course, the word “small” here is meaningless by itself. The correct statement
is that the amplitude must be small compared with some other quantity that is inherent
to the system and that has the same units as the amplitude. We generally won’t worry
about exactly what this quantity is; we’ll just assume that we’ve picked an amplitude that
is sufficiently small.
The moral of the examples below is that whenever you arrive at an equation of the
form√z̈ + (something)z = 0, you know that z undergoes simple harmonic motion with
ω = something.
θ
l
Simple pendulum
Consider the simple pendulum shown in Fig. 9. (The word “simple” refers to the fact that
m the mass is a point mass, as opposed to an extended mass in the “physical ” pendulum
below.) The mass hangs on a massless string and swings in a vertical plane. Let ` be
Figure 9 the length of the string, and let θ(t) be the angle the string makes with the vertical. The
gravitational force on the mass in the tangential direction is −mg sin θ. So F = ma in the
tangential direction gives
−mg sin θ = m(`θ̈) (40)
The tension in the string combines with the radial component of gravity to produce the
radial acceleration, so the radial F = ma equation serves only to tell us the tension, which
we won’t need here.
Eq. (40) isn’t solvable in closed form. But for small oscillations, we can use the sin θ ≈ θ
approximation to obtain
r
g
θ̈ + ω 2 θ = 0, where ω ≡ . (41)
`

This looks exactly like the ẍ+ω 2 x equation for the Hooke’s-law
p spring, sop
all of our previous
results carry over. The only difference is that ω is now g/` instead of k/m. Therefore,
we have
θ(t) = A cos(ωt + φ), (42)
1.1. SIMPLE HARMONIC MOTION 13

where A and φ are determined by the initial

p conditions. So the pendulum undergoes simple
p
harmonic motion with a frequency of g/`. The period is therefore T = 2π/ω = 2π `/g.
The true motion is arbitrarily close to this, for sufficiently small amplitudes. Problem [to be
added] deals with the higher-order corrections to the motion in the case where the amplitude
is not small.
Note that the angle θ bounces back and forth between ±A, where A is small. But the
phase angle ωt + φ increases without bound, or equivalently keeps running from 0 to 2π
repeatedly. pivot
θ d
Physical pendulum CM
Consider the “physical” pendulum shown in Fig. 10. The planar object swings back and
forth in the vertical plane that it lies in. The pivot is inside the object in this case, but it
need not be. You could imagine a stick (massive or massless) that is glued to the object
and attached to the pivot at its upper end.
We can solve for the motion by looking at the torque on the object around the pivot. mass m,
If the moment of inertia of the object around the pivot is I, and if the object’s CM is a moment of inertia I
distance d from the pivot, then τ = Iα gives (using the approximation sin θ ≈ θ)
Figure 10
r
mgd
−mgd sin θ = I θ̈ =⇒ θ̈ + ω 2 θ = 0, where ω ≡ . (43)
I

So we again have simple harmonic motion. Note that if the object

p is actually a point mass,
then we have I = md2 , and the frequency becomes ω = g/d. This agrees with the
simple-pendulum result in Eq. (41) with ` → d.
If you wanted, you could also solve this problem by using τ = Iα around the CM, but
then you would need to also use the Fx = max and Fy = may equations. All of these
equations involve messy forces at the pivot. The point is that by using τ = Iα around the
pivot, you can sweep these forces under the rug.
Q
I L C
-Q
LC circuit

Consider the LC circuit shown in Fig. 11. Kirchhoff’s rule (which says that the net voltage
drop around a closed loop must be zero) applied counterclockwise yields Figure 11

dI Q
−L − = 0. (44)
dt C

But I = dQ/dt, so we have

r
Q 2 1
−LQ̈ − =0 =⇒ Q̈ + ω Q = 0, where ω ≡ . (45)
C LC

So we again have simple harmonic motion. In comparing this LQ̈ + (1/C)Q equation with
the simple-harmonic mẍ + kx = 0 equation, we see that L is the analog of m, and 1/C is
the analog of k. L gives a measure of the inertia of the system; the larger L is, the more
the inductor resists changes in the current (just as a large m makes it hard to change the
velocity). And 1/C gives a measure of the restoring force; the smaller C is, the smaller the
charge is that the capacitor wants to hold, so the larger the restoring force is that tries to
keep Q from getting larger (just as a large k makes it hard for x to become large).
14 CHAPTER 1. OSCILLATIONS

1.2 Damped oscillations

1.2.1 Solving the differential equation
Let’s now consider damped harmonic motion, where in addition to the spring force we also
have a damping force. We’ll assume that this damping force takes the form,

Fdamping = −bẋ. (46)

Note that this is not the force from sliding friction on a table. That would be a force with
constant magnitude µk N . The −bẋ force here pertains to a body moving through a fluid,
provided that the velocity isn’t too large. So it is in fact a realistic force. The F = ma
equation for the mass is

Fspring + Fdamping = mẍ

=⇒ −kx − bẋ = mẍ
r
k b
=⇒ ẍ + γ ẋ + ω02 x = 0, where ω0 ≡ , γ≡ . (47)
m m
p
We’ll use ω0 to denote k/m here instead of the ω we used in Section 1.1, because there
will be a couple frequencies floating around in this section, so it will be helpful to distinguish
them.
In view of the discussion in Section 1.1.4, the method we will always use to solve a linear
differential equation like this is to try an exponential solution,

x(t) = Ceαt . (48)

Plugging this into Eq. (47) gives

α2 Ceαt + γαCeαt + ω02 Ceαt = 0

=⇒ α2 + γα + ω02 = 0
p
−γ ± γ 2 − 4ω02
=⇒ α = . (49)
2
We now have three cases to consider, depending on the sign of γ 2 − 4ω02 . These cases are
called underdamping, overdamping, and critical damping.

1.2.2 Underdamping (γ < 2ω0 )

The first of the three possible cases is the case of light damping, which holds if γ < 2ω0 . In
this case, the discriminant in Eq. (49) is negative, so α has a complex part.5 Let’s define
the real quantity ωu (where the “u” stands for underdamped) by

q r ³ γ ´2
1
ωu ≡ 4ω02 − γ2 =⇒ ωu = ω0 1− (50)
2 2ω0

Then the α in Eq. (49) becomes α = −γ/2 ± iωu . So we have the two solutions,

x1 (t) = C1 e(−γ/2+iωu )t and x2 (t) = C2 e(−γ/2−iωu )t . (51)

5 This reminds me of a joke: The reason why life is complex is because it has both a real part and an

imaginary part.
1.2. DAMPED OSCILLATIONS 15

We’ll accept here the fact that a second-order differential equation (which is what Eq. (47)
is) has at most two linearly independent solutions. Therefore, the most general solution is
the sum of the above two solutions, which is
³ ´
x(t) = e−γt/2 C1 eiωu t + C2 e−iωu t . (52)

However, as we discussed in Section 1.1.5, x(t) must of course be real. So the two terms
here must be complex conjugates of each other, to make the imaginary parts cancel. This
implies that C2 = C1∗ , where the star denotes complex conjugation. If we let C1 = Ceiφ
then C2 = C1∗ = Ce−iφ , and so x(t) becomes
³ ´
xunderdamped (t) = e−γt/2 C ei(ωu t+φ) + e−i(ωu t+φ)
= e−γt/2 C · 2 cos(ωu t + φ)
≡ Ae−γt/2 cos(ωu t + φ) (where A ≡ 2C). (53)

As we mentioned in Section 1.1.5, we’ve basically just taken the real part of either of the
two solutions that appear in Eq. (52).
We see that we have sinusoidal motion that slowly decreases in amplitude due to the x
e−γt/2 factor. In other words, the curves ±Ae−γt/2 form the envelope of the sinusoidal 1.0

motion. The constants A and φ are determined by the initial conditions, whereas the
t
constants γ and ωu (which is given by Eq. (50)) are determined by the setup. The task of 5 10 15 20 25 30
Problem [to be added] is to find A and φ for the initial conditions, x(0) = x0 and v(0) = 0.
-1.0
Fig. 12 shows a plot of x(t) for γ = 0.2 and ω0 = 1 s−1 , with the initial conditions of x(0) = 1
and v(0) = 0. p Figure 12
Note that the frequency ωu = ω0 1 − (γ/2ω0 )2 is smaller than the natural frequency,
ω0 . However, this distinction is generally irrelevant, because if γ is large enough to make ωu
differ appreciably from ω0 , then the motion becomes negligible after a few cycles anyway.
For example, if ωu differs from ω0 by even just 20% (so that ωu = (0.8)ω0 ), then you can
show that this implies that γ = (1.2)ω0 . So after just two cycles (that is, when ωu t = 4π),
the damping factor equals

e−(γ/2)t = e−(0.6)ω0 t = e−(0.6/0.8)ωu t = e−(3/4)(4π) = e−3π ≈ 1 · 10−4 , (54)

which is quite small.

Very light damping (γ ¿ ω0 )

In the limit
√ of very light damping (that is, γ ¿ ω0 ), we can use the Taylor-series approxi-
mation 1 + ² ≈ 1 + ²/2 in the expression for ωu in Eq. (50) to write
r ³ γ ´2 µ ¶
1 ³ γ ´2 γ2
ωu = ω0 1 − ≈ ω0 1 − = ω0 − . (55)
2ω0 2 2ω0 8ω02
So ωu essentially equals ω0 , at least to first order in γ.

Energy
Let’s find the energy of an underdamped oscillator, E = mẋ2 /2 + kx2 /2, as a function of
time. To keep things from getting too messy, we’ll let the phase φ in the expression for x(t)
in Eq. (53) be zero. The associated velocity is then
dx ³ γ ´
v= = Ae−γt/2 − cos ωu t − ωu sin ωu t . (56)
dt 2
16 CHAPTER 1. OSCILLATIONS

The energy is therefore

1 1
E = mẋ2 + kx2
2 2
1 ³ γ ´2 1
= mA2 e−γt − cos ωu t − ωu sin ωu t + kA2 e−γt cos2 ωu t. (57)
2 2 2

Using the definition of ωu from Eq. (50), and using k = mω02 , this becomes
µ 2 ³ ¶
1 2 −γt γ 2 2 γ2 ´ 2 2 2
E = mA e cos ωu t + γωu cos ωu t sin ωu t + ω0 − sin ωu t + ω0 cos ωu t
2 4 4
µ 2³ ´ ³ ´¶
1 2 −γt γ 2 2 2 2 2
= mA e cos ωu t − sin ωu t + γωu cos ωu t sin ωu t + ω0 cos ωu t + sin ωu t
2 4
µ 2 ¶
1 2 −γt γ γωu 2
= mA e cos 2ωu t + sin 2ωu t + ω0 . (58)
2 4 2

As a double check, when γ = 0 this reduces to the correct value of E = mω02 A2 /2 = kA2 /2.
For nonzero γ, the energy decreases in time due to the e−γt factor. The lost energy goes
into heat that arises from the damping force.
Note that the oscillating parts of E have frequency 2ωu . This is due to the fact that the
forward and backward motions in each cycle are equivalent as far as the energy loss due to
damping goes.
Eq. (58) is an exact result, but let’s now work in the approximation where γ is very
small, so that the e−γt factor decays very slowly. In this approximation, the motion looks
essentially sinusoidal with a roughly constant amplitude over a few oscillations. So if we take
the average of the energy over a few cycles (or even just exactly one cycle), the oscillatory
terms in Eq. (58) average to zero, so we’re left with

1 1
hEi = mω02 A2 e−γt = kA2 e−γt , (59)
2 2
where the brackets denote the average over a few cycles. In retrospect, we could have
obtained this small-γ result without going through the calculation in Eq. (58). If γ is very
small, then Eq. (53) tells us that at any given time we essentially have a simple harmonic
oscillator with amplitude Ae−γt/2 , which is roughly constant. The energy of this oscillator
is the usual (k/2)(amplitude)2 , which gives Eq. (59).

Energy decay
What is the rate of change of the average energy in Eq. (59)? Taking the derivative gives

dhEi
= −γhEi (60)
dt

This tells us that the fractional rate of change of hEi is γ. That is, in one unit of time, hEi
loses a fraction γ of its value. However, this result holds only for small γ, and furthermore
it holds only in an average sense. Let’s now look at the exact rate of change of the energy
as a function of time, for any value of γ, not just small γ.
The energy of the oscillator is E = mẋ2 /2 + kx2 /2. So the rate of change is

dE
= mẋẍ + kxẋ = (mẍ + kx)ẋ. (61)
dt
1.2. DAMPED OSCILLATIONS 17

Since the original F = ma equation was mẍ + bẋ + kx = 0, we have mẍ + kx = −bẋ,
Therefore,
dE dE
= (−bẋ)ẋ =⇒ = −bẋ2 (62)
dt dt
This is always negative, which makes sense because energy is always being lost to the
damping force. In the limit where b = 0, we have dE/dt = 0, which makes sense because we
simply have undamped simple harmonic motion, for which we already know that the energy
is conserved.
We could actually have obtained this −bẋ2 result with the following quicker reason-
ing. The damping force is Fdamping = −bẋ, so the power (which is dE/dt) it produces is
Fdamping v = (−bẋ)ẋ = −bẋ2 .
Having derived the exact dE/dt = −bẋ2 result, we can give another derivation of the
result for hEi in Eq. (60). If we take the average of dE/dt = −bẋ2 over a few cycles, we
obtain (using the fact that the average of the rate of change equals the rate of change of the
average)
dhEi
= −bhẋ2 i. (63)
dt
We now note that the average energy over a few cycles equals twice the average kinetic
energy (and also twice the average potential energy), because the averages of the kinetic
and potential energies are equal (see Fig. 8). Therefore,

hEi = mhẋ2 i. (64)

Comparing Eqs. (63) and (64) yields

dhEi b
= − hEi ≡ −γhEi, (65)
dt m
in agreement with Eq. (60). Basically, the averages of both the damping power and the
kinetic energy are proportional to hẋ2 i. And the ratio of the proportionality constants is
−b/m ≡ −γ.

Q value
The ratio γ/ω0 (or its inverse, ω0 /γ) comes up often (for example, in Eq. (50)), so let’s
define
ω0
Q≡ (66)
γ
Q is dimensionless, so it is simply a number. Small damping means large Q. The Q stands
for “quality,” so an oscillator with small damping has a high quality, which is a reasonable
word to use. A given damped-oscillator system has particular values of γ and ω0 (see Eq.
(47)), so it therefore has a particular value of Q. Since Q is simply a number, a reasonable
question to ask is: By what factor has the amplitude decreased after Q cycles? If we consider
the case of very small damping (which is reasonable, because if the damping isn’t small, the
oscillations die out quickly, so there’s not much to look at), it turns out that the answer is
independent of Q. This can be seen from the following reasoning.
The time to complete Q cycles is given by ωu t = Q(2π) =⇒ t = 2πQ/ωu . In the case
of very light damping (γ ¿ ω0 ), Eq. (50) gives ωu ≈ ω0 , so we have t ≈ 2πQ/ω0 . But since
we defined Q ≡ ω0 /γ, this time equals t ≈ 2π(ω0 /γ)/ω0 = 2π/γ. Eq. (53) then tells us that
at this time, the amplitude has decreased by a factor of

e−γt/2 = e−(γ/2)(2π/γ) = e−π ≈ 0.043, (67)

 18 CHAPTER 1. OSCILLATIONS

which is a nice answer if there ever was one! This result provides an easy way to determine
Q, and hence γ. Just count the number of cycles until the amplitude is about 4.3% of the
original value. This number equals Q, and then Eq. (66) yields γ, assuming that we know
the value of ω0 .

1.2.3 Overdamping (γ > 2ω0 )

If γ > 2ω0 , then the two solutions for α in Eq. (49) are both real. Let’s define µ1 and µ2 by
r r
γ γ2 2 γ γ2
µ1 ≡ + − ω0 , and µ2 ≡ − − ω02 . (68)
2 4 2 4
The most general solution for x(t) can then be written as

xoverdamped (t) = C1 e−µ1 t + C2 e−µ2 t (69)

x
1.5 where C1 and C2 are determined by the initial conditions. Note that both µ1 and µ2 are
1.0 positive, so both of these terms undergo exponential decay, and not exponential growth
0.5 (which would be a problem physically). Since µ1 > µ2 , the first term decays more quickly
0.0 t than the second, so for large t we are essentially left with only the C2 e−µ2 t term.
2 4 6 8
- 0.5
The plot of x(t) might look like any of the plots in Fig. 13, depending on whether you
Figure 13 throw the mass away from the origin, release it from rest, or throw it back (fairly quickly)
toward the origin. In any event, the curve can cross the origin at most once, because if we
set x(t) = 0, we find
µ ¶
−µ1 t −µ2 t C1 (µ1 −µ2 )t 1 −C1
C1 e + C2 e = 0 =⇒ − =e =⇒ t = ln . (70)
C2 µ1 − µ2 C2

We have found at most one solution for t, as we wanted to show. In a little more detail, the
various cases are: There is one positive solution for t if −C1 /C2 > 1; one zero solution if
−C1 /C2 = 1; one negative solution if 0 < −C1 /C2 < 1; and no (real) solution if −C1 /C2 < 0.
Only in the first of these four cases does that mass cross the origin at some later time after
you release/throw it (assuming that this moment corresponds to t = 0).

Very heavy damping (γ À ω0 )

Consider the limit where γ À ω0 . This corresponds to a very weak spring (small ω0 )
immersed in a very thick fluid (large γ), such a molasses. If γ À ω0 , then Eq. (68) gives
µ1 ≈ γ. And if we use a Taylor series for µ2 we find
s µ ¶
γ γ 4ω02 γ γ 1 4ω02 ω02
µ2 = − 1− 2 ≈ − 1− = ¿ γ. (71)
2 2 γ 2 2 2 γ2 γ

x We therefore see that µ1 À µ2 , which means that the e−µ1 t term goes to zero much faster
1.2 than the e−µ2 t term. So if we ignore the quickly-decaying e−µ1 t term, Eq. (69) becomes
1.0
0.8 2 γ
0.6 x(t) ≈ C2 e−µ2 t ≈ C2 e−(ω0 /γ)t ≡ C2 e−t/T , where T ≡ . (72)
0.4
0.2
ω02
0.0 t
- 0.2 2 4 6 8 A plot of a very heavily damped oscillator is shown in Fig. 14. We have chosen ω0 = 1 s−1
and γ = 3 s−1 . The initial conditions are x0 = 1 and v0 = 0. The two separate exponential
Figure 14
decays are shown, along with their sum. This figure makes it clear that γ doesn’t have to be
much larger than ω0 for the heavy-damping approximation to hold. Even with γ/ω0 only
1.2. DAMPED OSCILLATIONS 19

equal to 3, the fast-decay term dies out on a time scale of 1/µ1 ≈ 1/γ = (1/3) s, and the
slow-decay term dies out on a time scale of 1/µ2 ≈ γ/ω02 = 3 s.
T = γ/ω02 is called the “relaxation time.” The displacement decreases by a factor of 1/e
for every increase of T in the time. If γ À ω0 , we have T ≡ γ/ω02 À 1/ω0 . In other words,
T is much larger than the natural period of the spring, 2π/ω0 . The mass slowly creeps back
toward the origin, as you would expect for a weak spring in molasses.
Note that T ≡ γ/ω02 ≡ (b/m)(k/m) = b/k. So Eq. (72) becomes x(t) ≈ C2 e−(k/b)t .
What is the damping force is associated with this x(t)? We have
µ ¶ ³ ´
k −(k/b)t
Fdamping = −bẋ = −b − C2 e = k C2 e−(k/b)t = k · x(t) = −Fspring . (73)
b
This makes sense. If we have a weak spring immersed in a thick fluid, the mass is hardly
moving (or more relevantly, hardly accelerating). So the drag force and the spring force must
essentially cancel each other. This also makes it clear why the relaxation time, T = b/k, is
independent of the mass. Since the mass is hardly moving, its inertia (that is, its mass) is
irrelevant. The only things that matter are the (essentially) equal and opposite spring and
damping forces. So the only quantities that matter are b and k.

1.2.4 Critical damping (γ = 2ω0 )

If γ = 2ω0 , then we have a problem with our method of solving for x(t), because the two α’s
in Eq. (49) are equal, since the discriminant is zero. Both solutions are equal to −γ/2, which
equals ω0 because we’re assuming γ = 2ω0 . So we haven’t actually found two independent
solutions, which means that we won’t be able to satisfy arbitrary initial conditions for x(0)
and v(0). This isn’t good. We must somehow find another solution, in addition to the e−ω0 t
one.
It turns out that te−ω0 t is also a solution to the F = ma equation, ẍ + 2ω0 ẋ + ω02 x = 0
(we have used γ = 2ω0 here). Let’s verify this. First note that
d ¡ −ω0 t ¢
ẋ = te = e−ω0 t (1 − ω0 t),
dt
d ¡ −ω0 t ¢ ¡ ¢
=⇒ ẍ = e (1 − ω0 t) = e−ω0 t − ω0 − ω0 (1 − ω0 t) . (74)
dt
Therefore,
³ ´
ẍ + 2ω0 ẋ + ω02 x = e−ω0 t (−2ω0 + ω02 t) + 2ω0 (1 − ω0 t) + ω02 t = 0, (75)

as desired. Why did we consider the function te−ω0 t as a possible solution? Well, it’s a
general result from the theory of differential equations that if a root α of the characteristic
equation is repeated k times, then
eαt , teαt , t2 eαt , · · · , tk−1 eαt (76)
are all solutions to the original differential equation. But you actually don’t need to invoke
this result. You can just take the limit, as γ → 2ω0 , of either of the underdamped or x
overdamped solutions. You will find that you end up with a e−ω0 t and a te−ω0 t solution 1.5
(see Problem [to be added]). So we have 1.0
0.5
xcritical (t) = (A + Bt)e−ω0 t (77) 0.0 t
2 4 6 8
- 0.5
A plot of this is shown in Fig. 15. It looks basically the same as the overdamped plot
in Fig. 13. But there is an important difference. The critically damped motion has the Figure 15
property that it converges to the origin in the quickest manner, that is, quicker than both
the overdamped or underdamped motions. This can be seen as follows.
20 CHAPTER 1. OSCILLATIONS

• Quicker than overdamped: From Eq. (77), the critically damped motion goes to
zero like e−ω0 t (the Bt term is inconsequential compared with the exponential term).
And from Eq. (69), the overdamped motion goes to zero like e−µ2 t (since µ1 > µ2 ).
But from the definition of µ2 in Eq. (68), you can show that µ2 < ω0 (see Problem
[to be added]). Therefore, e−ω0 t < e−µ2 t , so xcritical (t) < xoverdamped (t) for large t, as
desired.

• Quicker than underdamped: As above, the critically damped motion goes to zero
like e−ω0 t . And from Eq. (53), the envelope of the underdamped motion goes to zero
like e−(γ/2)t . But the assumption of underdamping is that γ < 2ω0 , which means that
(envelope)
γ/2 < ω0 . Therefore, e−ω0 t < e−(γ/2)t , so xcritical (t) < xunderdamped (t) for large t, as
desired. The underdamped motion reaches the origin first, of course, but it doesn’t
stay there. It overshoots and oscillates back and forth. The critically damped oscillator
has the property that it converges to zero quickest without overshooting. This is very
relevant when dealing with, for example, screen doors or car shock absorbers. After
going over a bump in a car, you want the car to settle down to equilibrium as quickly
as possible without bouncing around.

1.3 Driven and damped oscillations

1.3.1 Solving for x(t)
Having looked at damped oscillators, let’s now look at damped and driven oscillators. We’ll
take the driving force to be Fdriving (t) = Fd cos ωt. The driving
p frequency ω is in general
equal to neither the natural frequency of the oscillator, ω0 = k/m, nor the frequency of
the underdamped oscillator, ωu . However, we’ll find that interesting things happen when
ω0 = ω. To avoid any confusion, let’s explicitly list the various frequencies:
p
• ω0 : the natural frequency of a simple oscillator, k/m.
p
• ωu : the frequency of an underdamped oscillator, ω02 − γ 2 /4.

• ω: the frequency of the driving force, which you are free to pick.

There are two reasons why we choose to consider a force of the form cos ωt (a sin ωt form
would work just as well). The first is due to the form of our F = ma equation:

Fspring + Fdamping + Fdriving = ma

=⇒ −kx − bẋ + Fd cos ωt = mẍ. (78)

This is simply Eq. (47) with the additional driving force tacked on. The crucial property
of Eq. (78) is that it is linear in x. So if we solve the equation and produce the function
x1 (t) for one driving force F1 (t), and then solve it again and produce the function x2 (t) for
another driving force F2 (t), then the sum of the x’s is the solution to the situation where
both forces are present. To see this, simply write down Eq. (78) for x1 (t), and then again
for x2 (t), and then add the equations. The result is

−k(x1 + x2 ) − b(ẋ1 + ẋ2 ) + (F1 + F2 ) = m(ẍ1 + ẍ2 ). (79)

In other words, x1 (t) + x2 (t) is the solution for the force F1 (t) + F2 (t). It’s easy to see that
this procedure works for any number of functions, not just two. It even works for an infinite
number of functions.
1.3. DRIVEN AND DAMPED OSCILLATIONS 21

The reason why this “superposition” result is so important is that when we get to Fourier
analysis in Chapter 3, we’ll see that any general function (well, as long as it’s reasonably
well behaved, which will be the case for any function we’ll be interested in) can be written
as the sum (or integral) of cos ωt and sin ωt terms with various values of ω. So if we use
this fact to write an arbitrary force in terms of sines and cosines, then from the preceding
paragraph, if we can solve the special case where the force is proportional to cos ωt (or
sin ωt), then we can add up the solutions (with appropriate coefficients that depend on the
details of Fourier analysis) to obtain the solution for the original arbitrary force. To sum
up, the combination of linearity and Fourier analysis tells us that it is sufficient to figure
out how to solve Eq. (78) in the specific case where the force takes the form of Fd cos ωt.
The second reason why we choose to consider a force of the form cos ωt is that F (t) =
Fd cos ωt is in fact a very realistic force. Consider a spring that has one end attached to
a mass and the other end attached to a support. If the support is vibrating with position
xend (t) = Aend cos ωt (which often happens in real life), then the spring force is

Fspring (x) = −k(x − xend ) = −kx + kAend cos ωt. (80)

This is exactly the same as a non-vibrating support, with the addition of someone exerting
a force Fd cos ωt directly on the mass, with Fd = kAend .
We’ll now solve for x(t) in the case of damped and driven motion. That is, we’ll solve
Eq. (78), which we’ll write in the form,
r
2 b k Fd
ẍ + γ ẋ + ω0 x = F cos ωt, where γ ≡ , ω0 ≡ , F ≡ . (81)
m m m
There are (at least) three ways to solve this, all of which involve guessing a sinusoidal or
exponential solution.

Method 1
Let’s try a solution of the form,

x(t) = A cos(ωt + φ), (82)

where the ω here is the same as the driving frequency. If we tried a different frequency, then
the lefthand side of Eq. (81) would have this different frequency (the derivatives don’t affect
the frequency), so it would have no chance of equaling the F cos ωt term on the righthand
side.
Note how the strategy of guessing Eq. (82) differs from the strategy of guessing Eq. (48)
in the damped case. The goal there was to find the frequency of the motion, whereas in the
present case we’re assuming that it equals the driving frequency ω. It might very well be
the case that there doesn’t exist a solution with this frequency, but we have nothing to lose
by trying. Another difference between the present case and the damped case is that we will
actually be able to solve for A and φ, whereas in the damped case these parameters could
take on any values, until the initial conditions are specified.
If we plug x(t) = A cos(ωt + φ) into Eq. (81), we obtain

−ω 2 A cos(ωt + φ) − γωA sin(ωt + φ) + ω02 A cos(ωt + φ) = F cos ωt. (83)

We can cleverly rewrite this as

ω 2 A cos(ωt + φ + π) + γωA cos(ωt + φ + π/2) + ω02 A cos(ωt + φ) = F cos ωt, (84)

 ω2A
22 CHAPTER 1. OSCILLATIONS

γωA which just says that a is 90◦ ahead of v, which itself is 90◦ ahead of x. There happens to
be a slick geometrical interpretation of this equation that allows us to quickly solve for A
and φ. Consider the diagram in Fig. 16. The quantities ω0 , γ, ω, and F are given. We have
ω02 A picked an arbitrary value of A and formed a vector with length ω02 A pointing to the right.
Then we’ve added on a vector with length γωA pointing up, and then another vector with
Figure 16 length ω 2 A point to the left. The sum is the dotted-line vector shown. We can make the
magnitude of this vector be as small or as large as we want by scaling the diagram down or
up with an arbitrary choice of A.
F
If we pick A so that the magnitude of the vector sum equals F , and if we rotate the
φ whole diagram through the angle φ that makes the sum horizontal (φ is negative here), then
ω2A we end up with the diagram in Fig. 17. The benefit of forming this diagram is the following.
Consider the horizontal projections of all the vectors. The fact that the sum of the three
ω02 A tilted vectors equals the horizontal vector implies that the sum of their three horizontal
components equals F . That is (remember that φ is negative),
γωA ω02 A cos φ + γωA cos(φ + π/2) + ω 2 A cos(φ + π) = F cos(0). (85)

Figure 17 This is just the statement that Eq. (84) holds when t = 0. However, if A and φ are chosen
so that it holds at t = 0, then it holds at any other time too, because we can imagine
rotating the entire figure counterclockwise through the angle ωt. This simply increases the
arguments of all the cosines by ωt. The statement that the x components of the rotated
vectors add up correctly (which they do, because the figure keeps the same shape as it is
rotated, so the sum of the three originally-tilted vectors still equals the originally-horizontal
vector) is then
ω02 A cos(ωt + φ) + γωA cos(ωt + φ + π/2) + ω 2 A cos(ωt + φ + π) = F cos ωt, (86)
which is the same as Eq. (84), with the terms on the left in reverse order. Our task therefore
reduces to determining the values of A and φ that generate the quadrilateral in Fig. 17.

The phase φ
If we look at the right triangle formed by drawing the dotted line shown, we can quickly
read off
−γωA −γω
tan φ = 2 2
=⇒ tan φ = 2 (87)
F (ω0 − ω )A ω0 − ω 2
ω02 A We’ve put the minus sign in by hand here because φ is negative. This follows from the fact
φ that we made a “left turn” in going from the ω02 A vector to the γωA vector. And then
another left turn in going from the γωA vector to the ω 2 A vector. So no matter what the
ω2A value of ω is, φ must lie somewhere in the range −π ≤ φ ≤ 0. Angles less than −90◦ arise
γωA when ω > ω0 , as shown in Fig. 18. φ ≈ 0 is obtained when ω ≈ 0, and φ ≈ −π is obtained
when ω ≈ ∞. Plots of φ(ω) for a few different values of γ are shown in Fig. 19.
Figure 18
φ(ω)
ω (in units of ω0)
0.5 1.0 1.5 2.0 2.5 3.0
- 0.5
- 1.0 γ (in units of ω0)
- 1.5 50
- 2.0 5
2
- 2.5 1
.5
- 3.0 .1
.01
1.3. DRIVEN AND DAMPED OSCILLATIONS 23

Figure 19

If γ is small, then the φ curve starts out with the roughly constant value of zero, and
then jumps quickly to the roughly constant value of −π. The jump takes place in a small
range of ω near ω0 . Problem [to be added] addresses this phenomenon. If γ is large, then
the φ curve quickly drops to −π/2, and then very slowly decreases to π. Nothing interesting
happens at ω = ω0 in this case. See Problem [to be added].
Note that for small γ, the φ curve has an inflection point (which is point where the second
derivative is zero), but
√ for large γ it doesn’t. The value of γ that is the cutoff between these
two regimes is γ = 3ω0 (see Problem [to be added]). This is just a fact of curiosity; I don’t
think it has any useful consequence.

The amplitude A
To find A, we can apply the Pythagorean theorem to the right triangle in Fig. 17. This
gives
³ ´2 F
(ω02 − ω 2 )A + (γωA)2 = F 2 =⇒ A = p 2 (88)
(ω0 − ω 2 )2 + γ 2 ω 2

Our solution for x(t) is therefore

x(t) = A cos(ωt + φ) (89)

where φ and A are given by Eqs. (87) and (88). Plots of the amplitude A(ω) for a few
different values of γ are shown in Fig. 20.

A(ω) (in units of Fd /m)

10

6 γ (in units of ω0)

.1
4
.2
2 .5
1
0
3 ω (in units of ω0)
0.0 0.5 1.0 1.5 2.0

Figure 20

At what value of ω does the maximum of A(ω) occur? A(ω) is maximum when the
denominator of the expression in Eq. (88) is minimum, so setting the derivative (with respect
to ω) of the quantity under the square root equal to zero gives
q
2(ω02 − ω 2 )(−2ω) + γ 2 (2ω) = 0 =⇒ ω = ω02 − γ 2 /2. (90)
√
For small γ (more precisely, for γ ¿ ω0 ), this √ yields ω ≈ ω0 . If γ = 2ω0 , then the
maximum occurs at ω = 0. If γ is larger than 2ω0 , then the maximum occurs at ω = 0,
and the curve monotonically decreases as ω increases. These facts are consistent with Fig.
20. For small γ, which is the case we’ll usually be concerned with, the maximum value of
A(ω) is essentially equal to the value at ω0 , which is A(ω0 ) = F/γω0 .
24 CHAPTER 1. OSCILLATIONS

A(ω) goes to zero as ω → ∞. The value of A(ω) at ω = 0 is

F Fd /m Fd
A(0) = ≡ = . (91)
ω02 k/m k
This is independent of γ, consistent with Fig. 20. The reason why A(0) takes the very simple
form of Fd /k will become clear in Section 1.3.2 below.
Using the same techniques that we’ll use below to obtain in Eq. (128) the width of the
power curve, you can show (see Problem [to be added]) that the width of the A(ω) curve
(defined to be the width at half max) is
√
width = 3γ (92)

So the curves get narrower (at half height) as γ decreases.

However, the curves don’t get narrower in an absolute sense. By this we mean that for
a given value of ω, say ω = (0.9)ω0 , the value of A in Fig. 20 increases as γ decreases.
Equivalently, for a given value of A, the width of the curve at this value increases as γ
decreases. These facts follow from the fact that as γ → 0, the A in Eq. (88) approaches the
function F/|ω02 − ω 2 |. This function is the envelope of all the different A(ω) functions for
different values of γ. If we factor the denominator in F/|ω02 − ω 2 |, we see that near ω0 (but
not right at ω0 ), A behaves like (F/2ω0 )/|ω0 − ω|.
Remember that both A and φ are completely determined by the quantities ω0 , γ, ω,
and F . The initial conditions have nothing to do with A and φ. How, then, can we satisfy
arbitrary initial conditions with no free parameters at our disposal? We’ll address this
question after we discuss the other two methods for solving for x(t).

Method 2
Let’s try a solution of the form,

x(t) = A cos ωt + B sin ωt. (93)

(This A isn’t the same as the A in Method 1.) As above, the frequency here must be the
same as the driving frequency if this solution is to have any chance of working. If we plug
this expression into the F = ma equation in Eq. (81), we get a fairly large mess. But if we
group the terms according to whether they involve a cos ωt or sin ωt, we obtain (you should
verify this)

(−ω 2 B − γωA + ω02 B) sin ωt + (−ω 2 A + γωB + ω02 A) cos ωt = F cos ωt. (94)

We have two unknowns here, A and B, but only one equation. However, this equation is
actually two equations. The point is that we want it to hold for all values of t. But sin ωt
and cos ωt are linearly independent functions, so the only way this equation can hold for all t
is if the coefficients of sin ωt and cos ωt on each side of the equation match up independently.
That is,

−ω 2 B − γωA + ω02 B = 0,
−ω 2 A + γωB + ω02 A = F. (95)

We now have two unknowns and two equations. Solving for either A or B in the first
equation and plugging the result into the second one gives
(ω02 − ω 2 )F γωF
A= and B= . (96)
(ω02 − ω 2 )2 + γ 2 ω 2 (ω02 − ω 2 )2 + γ 2 ω 2
1.3. DRIVEN AND DAMPED OSCILLATIONS 25

The solution for x(t) is therefore

(ω02 − ω 2 )F γωF
x(t) = 2 2 2 2 2
cos ωt + 2 sin ωt (97)
(ω0 − ω ) + γ ω (ω0 − ω 2 )2 + γ 2 ω 2

We’ll see below in Method 3 that this solution is equivalent to the x(t) = A cos(ωt + φ)
solution from Method 1, given that A and φ take on the particular values we found.

Method 3
First, consider the equation,
ÿ + γ ẏ + ω02 y = F eiωt . (98)
This equation isn’t actually physical, because the driving “force” is the complex quantity
F eiωt . Forces need to be real, of course. And likewise the solution for y(t) will complex,
so it can’t actually represent an actual position. But as we’ll shortly see, we’ll be able to
extract a physical result by taking the real part of Eq. (98).
Let’s guess a solution to Eq. (98) of the form y(t) = Ceiωt . When we get to Fourier
analysis in Chapter 3, we’ll see that this is the only function that has any possible chance
of working. Plugging in y(t) = Ceiωt gives

F
−ω 2 Ceiωt + iγωCeiωt + ω02 Ceiωt = F · Ceiωt =⇒ C = . (99)
ω02 − ω 2 + iγω

What does this solution have to do with our original scenario involving a driving force
proportional to cos ωt? Well, consider what happens when we take the real part of Eq. (98).
Using the fact that differentiation commutes with the act of taking the real part, which is
true because µ ¶
d da d³ ´
Re (a + ib) = = Re(a + ib) , (100)
dt dt dt
we obtain

Re(ÿ) + Re(γ ẏ) + Re(ω02 y) = Re(F eiωt )

d2 ³ ´ d³ ´ ³ ´
2
Re(y) + γ Re(y) + ω02 Re(y) = F cos ωt. (101)
dt dt
In other words, if y is a solution to Eq. (98), then the real part of y is a solution to our
original (physical) equation, Eq. (81), with the F cos ωt driving force. So we just need to
take the real part of the solution for y that we found, and the result will be the desired
position x. That is,
³ ´ ³ ´ µ ¶
iωt F iωt
x(t) = Re y(t) = Re Ce = Re e . (102)
ω02 − ω 2 + iγω

Note that the quantity Re(Ceiωt ) is what matters here, and not Re(C)Re(eiωt ). The equiv-
alence of this solution for x(t) with the previous ones in Eqs. (89) and (97) can be seen as
follows. Let’s consider Eq. (97) first.

• Agreement with Eq. (97):

Any complex number can be written in either the Cartesian a + bi way, or the polar
(magnitude and phase) Aeiφ way. If we choose to write the C in Eq. (99) in the
26 CHAPTER 1. OSCILLATIONS

Cartesian way, we need to get the i out of the denominator. If we “rationalize” the
denominator of C and expand the eiωt term in Eq. (102), then x(t) becomes
à ¡ ¢ !
F (ω02 − ω 2 ) − iγω
x(t) = Re (cos ωt + i sin ωt) . (103)
(ω02 − ω 2 )2 + γ 2 ω 2

The real part comes from the product of the real parts and the product of the imaginary
parts, so we obtain
(ω02 − ω 2 )F γωF
x(t) = 2 2 2 2 2
cos ωt + 2 sin ωt, (104)
(ω0 − ω ) + γ ω (ω0 − ω 2 )2 + γ 2 ω 2
in agreement with Eq. (97) in Method 2.
• Agreement with Eq. (89):
If we choose to write the C in Eq. (99) in the polar Aeiφ way, we have
√ F
A= C · C∗ = p 2 , (105)
(ω0 − ω 2 )2 + γ 2 ω 2
and
Im(C) −γω
tan φ = = 2 , (106)
Re(C) ω0 − ω 2
where we have obtained the real and imaginary parts from the expression for C that
we used in Eq. (103). (This expression for tan φ comes from the fact that the ratio of
the imaginary and real parts of eiφ = cos φ + i sin φ equals tan φ.) So the x(t) in Eq.
(102) becomes
³ ´ ³ ´ ³ ´
x(t) = Re y(t) = Re Ceiωt = Re Aeiφ eiωt = A cos(ωt + φ). (107)

This agrees with the result obtained in Eq. (89) in Method 1, because the A and φ in
Eqs. (105) and (106) agree with the A and φ in Eqs. (88) and (87).

The complete solution

Having derived the solution for x(t) in three different ways, let’s look at what we’ve found.
We’ll use the x(t) = A cos(ωt + φ) form of the solution in the following discussion.
p
As noted above, A and φ have definite values, given the values of ω0 ≡ k/m, γ ≡ b/m,
F ≡ Fd /m, and ω. There is no freedom to impose initial conditions. The solution in
Eq. (107) therefore cannot be the most general solution, because the most general solution
must allow us to be able to satisfy arbitrary initial conditions. So what is the most general
solution? It is the sum of the A cos(ωt + φ) solution we just found (this is called the
“particular” solution) and the solution from Section 1.2.1 (the “homogeneous” solution)
that arose when there was no driving force, and thus a zero on the righthand side of the
F = ma equation in Eq. (47). This sum is indeed a solution to the F = ma equation in
Eq. (81) because this equation is linear in x. The homogeneous solution simply produces a
zero on the righthand side, which doesn’t mess up the equality generated by the particular
solution. In equations, we have (with the sum of the particular and homogeneous solutions
written as x = xp + xh )

ẍ + γ ẋ + ω02 x = (ẍp + ẍh ) + γ(ẋp + ẋh ) + ω02 (xp + xh )

¡ ¢ ¡ ¢
= ẍp + γ ẋp + ω02 xp + ẍh + γ ẋh + ω02 xh
= F cos ωt + 0, (108)
1.3. DRIVEN AND DAMPED OSCILLATIONS 27

which means that x = xp + xh is a solution to the F = ma equation, as desired. The

two unknown constants in the homogeneous solution yield the freedom to impose arbitrary
initial conditions. For concreteness, if we assume that we have an underdamped driven
oscillator, which has the homogeneous solution given by Eq. (53), then the complete solution,
x = xp + xh , is

x(t) = Ap cos(ωt + φ) + Ah e−γt/2 cos(ωh t + θ) (underdamped ) (109)

A word on the various parameters in this result:

• ω is the driving frequency, which can be chosen arbitrarily.

p
• Ap and φ are functions of ω0 ≡ k/m, γ ≡ b/m, F ≡ Fd /m, and ω. They are given
in Eqs. (88) and (87).

• ωh is a function of ω0 and γ. It is given in Eq. (50).

• Ah and θ are determined by the initial conditions.

However, having said all this, we should note the following very important point. For large t
(more precisely, for t À 1/γ), the homogeneous solution goes to zero due to the e−γt/2 term.
So no matter what the initial conditions are, we’re left with essentially the same particular
solution for large t. In other words, if two different oscillators are subject to exactly the same
driving force, then even if they start with wildly different initial conditions, the motions will
essentially be the same for large t. All memory of the initial conditions is lost.6 Therefore,
since the particular solution is the one that survives, let’s examine it more closely and discuss
some special cases.

1.3.2 Special cases for ω

Slow driving (ω ¿ ω0 )
If ω is very small compared with ω0 , then we can simplify the expressions for φ and A in
Eqs. (87) and (88). Assuming that γ isn’t excessively large (more precisely, assuming that
γω ¿ ω02 ), we find
F Fd /m Fd
φ ≈ 0, and A≈ 2 ≡ = . (110)
ω0 k/m k
Therefore, the position takes the form (again, we’re just looking at the particular solution
here),
Fd
x(t) = A cos(ωt + φ) = cos ωt. (111)
k
Note that the spring force is then Fspring = −kx = −Fd cos ωt, which is simply the negative
of the driving force. In other words, the driving force essentially balances the spring force.
This makes sense: The very small frequency, ω, of the motion implies that the mass is hardly
moving (or more relevantly, hardly accelerating), which in turn implies that the net force
must be essentially zero. The damping force is irrelevant here because the small velocity
(due to the small ω) makes it negligible. So the spring force must balance the driving force.
The mass and the damping force play no role in this small-frequency motion. So the effect
of the driving force is to simply balance the spring force.
6 The one exception is the case where there is no damping whatsoever, so that γ is exactly zero. But all

mechanical systems have at least a tiny bit of damping (let’s not worry about superfluids and such), so we’ll
ignore the γ = 0 case.
28 CHAPTER 1. OSCILLATIONS

Mathematically, the point is that the first two terms on the lefthand side of the F = ma
equation in Eq. (81) are negligible:

ẍ + γ ẋ + ω02 x = F cos ωt. (112)

This follows from the fact that since x(t) = A cos(ωt + φ), the coefficients of each of the
sinusoidal terms on the lefthand side are proportional to ω 2 , γω, and ω02 , respectively. And
since we’re assuming both ω ¿ ω0 and γω ¿ ω02 , the first two terms are negligible compared
with the third. The acceleration and velocity of the mass are negligible. The position is all
that matters.
The φ ≈ 0 result in Eq. (110) can be seen as follows. We saw above that the driving
force cancels the spring force. Another way of saying this is that the driving force is 180◦
out of phase with the −kx = k(−x) spring force. This means that the driving force is in
phase with the position x. Intuitively, the larger the force you apply, the larger the spring
force and hence the larger the position x. The position just follows the force.

Fast driving (ω À ω0 )
If ω is very large compared with ω0 , then we can again simplify the expressions for φ and
A in Eqs. (87) and (88). Assuming that γ isn’t excessively large (which now means that
γ ¿ ω), we find
F Fd
φ ≈ −π, and A≈ 2 = . (113)
ω mω 2
Therefore, the position takes the form,
Fd Fd
x(t) = A cos(ωt + φ) = 2
cos(ωt − π) = − cos ωt. (114)
mω mω 2
Note that the mass times the acceleration is then mẍ = Fd cos ωt, which is the driving force.
In other words, the driving force is essentially solely responsible for the acceleration. This
makes sense: Since there are ω’s in the denominator of x(t), and since ω is assumed to be
large, we see that x(t) is very small. The mass hardly moves, so the spring and damping
forces play no role in this high-frequency motion. The driving force provides essentially all
of the force and therefore causes the acceleration.
Mathematically, the point is that the second two terms on the lefthand side of the
F = ma equation in Eq. (81) are negligible:

ẍ + γ ẋ + ω02 x = F cos ωt. (115)

As we noted after Eq. (112), the coefficients of each of the sinusoidal terms on the lefthand
side are proportional to ω 2 , γω, and ω02 , respectively. And since we’re assuming both ω0 ¿ ω
and γ ¿ ω, the second two terms are negligible compared with the first. The velocity and
position of the mass are negligible. The acceleration is all that matters.
The φ ≈ −π result in Eq. (113) can be seen as follows. Since the driving force provides
essentially all of the force, it is therefore in phase with the acceleration. But the acceleration
is always out of phase with x(t) (at least for sinusoidal motion). So the driving force is out
of phase with x(t). Hence the φ ≈ −π result and the minus sign in the expression for x(t)
in Eq. (114).

Resonance (ω = ω0 )
If ω equals ω0 , then we can again simplify the expressions for φ and A in Eqs. (87) and (88).
We don’t need to make any assumptions about γ in this case, except that it isn’t exactly
1.3. DRIVEN AND DAMPED OSCILLATIONS 29

equal to zero. We find

π F F Fd
φ=− , and A≈ = = . (116)
2 γω γω0 γmω0

Therefore, the position takes the form,

Fd Fd
x(t) = A cos(ωt + φ) = cos(ωt − π/2) = sin ωt. (117)
γmω0 γmω0

Note that the damping force is then Fdamping = −(γm)ẋ = −Fd cos ωt, which is the negative
of the driving force. In other words, the driving force essentially balances the damping force.
This makes sense: If ω = ω0 , then the system is oscillating at ω0 , so the spring and the mass
are doing just what they would be doing if the damping and driving forces weren’t present.
You can therefore consider the system to be divided into two separate systems: One is a
simple harmonic oscillator, and the other is a driving force that drags a massless object
with the same shape as the original mass (so that it recreates the damping force) back and
forth in a fluid (or whatever was providing the original damping force). None of the things
involved here (spring, mass, fluid, driver) can tell the difference between the original system
and this split system. So the effect of the driving force is to effectively cancel the damping
force, while the spring and the mass do their natural thing.
Mathematically, the point is that the first and third terms on the lefthand side of the
F = ma equation in Eq. (81) cancel each other:

ẍ + γ ẋ + ω02 x = F cos ωt. (118)

As above, the coefficients of each of the sinusoidal terms on the lefthand side are proportional
(in magnitude) to ω 2 , γω, and ω02 , respectively. Since ω = ω0 , the first and third terms
cancel (the second derivative yields a minus sign in the first term). The remaining parts
of the equation then say that the driving force balances the damping force. Note that the
amplitude must take on the special value of Fd /γmω0 for this to work.
If γ is small, then the amplitude A in Eq. (116) is large. Furthermore, for a given value
of γ, the amplitude is largest when (roughly) ω = ω0 . Hence the name “resonance.” We’ve
added the word “roughly” here because it depends on what we’re taking to be the given
quantity, and what we’re taking to be the variable. If ω is given, and if we want to find
the maximum value of the A in Eq. (88) as a function of ω0 , then we want to pick ω0 to
equal ω, because this choice makes ω02 − ω 2 equal to zero, and we can’t do any better than
that, with regard to making the denominator of A small. On the other hand, if ω0 is given,
and if we want to find the maximum value of A as a function of ω, then we need to take
the derivative
p of A with respect to ω. We did this in Eq. (90) above, and the result was
ω = ω02 − γ 2 /2. So the peaks of the curves in Fig. 20 (where A was considered to be a
function of ω) weren’t located exactly at ω = ω0 . However, we will generally be concerned
with the case of small γ (more precisely γ ¿ ω0 ), in which case the peak occurs essentially
at ω = ω0 , even when A is considered to be a function of ω.
Having noted that the amplitude is maximum when ω = ω0 , we can now see where the
φ ≈ −π/2 result in Eq. (116) comes from. If we want to make the amplitude as large as
possible, then we need to put a lot of energy into the system. Therefore, we need to do a lot
of work. So we want the driving force to act over the largest possible distance. This means
that we want the driving force to be large when the mass velocity is large. (Basically, power
is force times velocity.) In other words, we want the driving force to be in phase with the
velocity. And since x is always 90◦ behind v, x must also be 90◦ behind the force. This
agrees with the φ = −π/2 phase in x. In short, this φ = −π/2 phase implies that the force
30 CHAPTER 1. OSCILLATIONS

always points to the right when the mass is moving to the right, and always points to the
left when the mass is moving to the left. So we’re always doing positive work. For any other
phase, there are times when we’re doing negative work.
In view of Eqs. (112), (115), and (118), we see that the above three special cases are
differentiated by which one of the terms on the lefthand side of Eq. (81) survives. There
is a slight difference, though: In the first two cases, two terms disappear because they are
small. In the third case, they disappear because they are equal and opposite.

1.3.3 Power
In a driven and damped oscillator, the driving force feeds energy into the system during
some parts of the motion and takes energy out during other parts (except in the special case
of resonance where it always feeds energy in). The damping force always takes energy out,
because the damping force always points antiparallel to the velocity. For the steady-state
solution (the “particular” solution), the motion is periodic, so the energy should stay the
same on average; the amplitude isn’t changing. The average net power (work per time) from
the driving force must therefore equal the negative of the average power from the damping
force. Let’s verify this. Power is the rate at which work is done, so we have

dW dx
dW = F dx =⇒ P ≡ =F = F v. (119)
dt dt
The powers from the damping and driving forces are therefore:
power dissipated by the damping force: This equals

Pdamping = Fdamping v = (−bẋ)ẋ

¡ ¢2
= −b − ωA sin(ωt + φ)
= −b(ωA)2 sin2 (ωt + φ). (120)

Since the average value of sin2 θ over a complete cycle is 1/2 (obtained by either doing an
integral or noting that sin2 θ has the same average as cos2 θ, and these two averages add up
to 1), the average value of the power from the damping force is

1
hPdamping i = − b(ωA)2 (121)
2

power supplied by the driving force: This equals

Pdriving = Fdriving v = (Fd cos ωt)ẋ

¡ ¢
= (Fd cos ωt) − ωA sin(ωt + φ)
¡ ¢
= −Fd ωA cos ωt sin ωt cos φ + cos ωt sin φ . (122)

The results in Eqs. (120) and (122) aren’t the negatives of each other for all t, because the
energy waxes and and wanes throughout a cycle. (The one exception is on resonance with
φ = −π/2, as you can verify.) But on average they must cancel each other, as we noted
above. This is indeed the case, because in Eq. (122), the cos ωt sin ωt term averages to zero,
while the cos2 ωt term averages to 1/2. So the average value of the power from the driving
force is
1
hPdriving i = − Fd ωA sin φ. (123)
2
1.3. DRIVEN AND DAMPED OSCILLATIONS 31

Now, what is sin φ? From Fig. 17, we have sin φ = −γωA/F ≡ −γmωA/Fd . So Eq. (123)
gives
µ ¶
1 −γmωA 1 1
hPdriving i = − Fd ωA = γm(ωA)2 = b(ωA)2 (124)
2 Fd 2 2

Eqs. (121) and (124) therefore give hPdamping i + hPdriving i = 0, as desired.

What does hPdriving i look like as a function of ω? Using the expression for A in Eq. (88),
along with b ≡ γm, we have

1
hPdriving i = b(ωA)2
2
(γm)ω 2 (Fd /m)2
= · 2
2 (ω0 − ω 2 )2 + γ 2 ω 2
2
(γm)Fd γ 2 ω2
= ·
2γ 2 m2 (ω02 − ω 2 )2 + γ 2 ω 2
Fd2 γ 2 ω2 Fd2
= · 2 ≡ · f (ω). (125)
2γm (ω0 − ω 2 )2 + γ 2 ω 2 2γm

We have chosen to write the result this way because the function f (ω) is a dimensionless
function of ω. The Fd2 out front tells us that for given ω, ω0 , and γ, the average power
hPdriving i grows as the driving amplitude Fd becomes larger, which makes sense. Fig. 21
shows some plots of the dimensionless function f (ω) for a few values of γ. In other words,
it shows plots of hPdriving i in units of Fd2 /2γm ≡ Fd2 /2b.

f(ω) P(ω) in units of Fd2/2γm

1.0

0.8 γ = (0.2)ω0

0.6 γ = (0.5)ω0
0.4 γ = ω0
0.2

0.0 ω (in units of ω0)

0 1 2 3 4

Figure 21

Fig. 22 shows plots of f (ω)/γ for the same values of γ. That is, it shows plots of the
actual average power, hPdriving i, in units of Fd2 /2m. These plots are simply 1/γ times the
plots in Fig. 21.
32 CHAPTER 1. OSCILLATIONS

f(ω)/γ P(ω) in units of Fd2/2m

5

3 γ = (0.2)ω0
2
γ = (0.5)ω0
1
γ = ω0
0 ω (in units of ω0)
0 1 2 3 4

Figure 22
The curves in Fig. 22 get thinner and taller as γ gets smaller. How do the widths
depend on γ? We’ll define the “width” to be the width at half max. The maximum value
of hPdriving i (or equivalently, of f (ω)) is achieved where its derivative with respect to ω is
zero. This happens to occur right at ω = ω0 (see Problem [to be added]). The maximum
value of f (ω) then 1, as indicated in Fig. 21 So the value at half max equals 1/2. This is
obtained when the denominator of f (ω) equals 2γ 2 ω 2 , that is, when
(ω02 − ω 2 )2 = γ 2 ω 2 =⇒ ω02 − ω 2 = ±γω. (126)
There are two quadratic equations in ω here, depending on the sign. The desired width at
half max equals the difference between the positive roots, call them ω1 and ω2 , of each of
these two equations. If you want, you can use the quadratic formula to find these roots, and
then take the difference. But a cleaner way is to write down the statements that ω1 and ω2
are solutions to the equations:
ω02 − ω12 = γω1 ,
ω02 − ω22 = −γω2 , (127)
and then take the difference. The result is
ω22 − ω12 = γ(ω2 + ω1 ) =⇒ ω2 − ω1 = γ =⇒ width = γ (128)
(We have ignored the ω1 + ω2 solution to this equation, since we are dealing with positive
ω1 and ω2 .) So we have the nice result that the width at half max is exactly equal to γ.
This holds for any value of γ, even though the the plot of hPdriving i looks like a reasonably
symmetric peak only if γ is small compared with ω0 . This can be see in Fig. 23, which
shows plots of f (ω) for the reasonably small value of γ = (0.4)ω0 and the reasonably large
value of γ = 2ω0 .

f(ω) P(ω) in units of Fd2/2γm

1.0
γ = (0.4)ω0
0.8
width = (0.4)ω0
0.6
γ = 2ω0
0.4
width = 2ω0
0.2

0.0 ω (in units of ω0)

0 1 2 3 4
1.3. DRIVEN AND DAMPED OSCILLATIONS 33

Figure 23

To sum up, the maximum height of the hPdriving i curve is proportional to 1/γ (it equals
Fd2 /2γm), and the width of the curve at half max is proportional to γ (it’s just γ). So the
curves get narrower as γ decreases.
Furthermore, the curves get narrower in an absolute sense, unlike the A(ω) curves in
Fig. 20 (see the discussion at the end of the “Method 1” part of Section 1.3.1). By this we
mean that for a given value of ω (except ω0 ), say ω = (0.9)ω0 , the value of P in Fig. 22
decreases as γ decreases. Equivalently, for a given value of P , the width of the curve at this
value decreases as γ decreases. These facts¡follow from the fact ¢ that as γ → 0, the P in
Eq. (125) is proportional to the function γ/ (ω02 − ω 2 )2 + γ 2 ω 2 . For any given value of ω
(except ω0 ), this becomes very small if γ is sufficiently small.
Since the height and width of the power curve are proportional to 1/γ and γ, respectively,
we might suspect that the area under the curve is independent of γ. This is indeed the case.
The integral is very messy to calculate in closed form, but if you find the area by numerical
integration for a few values of γ, that should convince you.
Let’s now discuss intuitively why the hPdriving i curve in Fig. 22 goes to zero at ω ≈ 0
and ω ≈ ∞, and why it is large at ω = ω0 .
x,F F(t)
• ω ≈ 0: In this case, Eq. (110) gives the phase as φ ≈ 0, so the motion is in phase
with the force. The consequence of this fact is that half the time your driving force
is doing positive work, and half the time it is doing negative work. These cancel, and t
on average there is no work done. In more detail, let’s look at each quarter cycle; see x(t)
Fig. 24 (the graph is plotted with arbitrary units on the axes). As you (very slowly)
drag the mass to the right from the origin to the maximum displacement, you are +W -W +W -W
doing positive work, because your force is in the direction of the motion. But then as (+v,+F) (-v,+F) (-v,-F) (+v,-F)

you (very slowly) let the spring pull the mass back toward to the origin, you are doing Figure 24
negative work, because your force is now in the direction opposite the motion. The
same cancelation happens in the other half of the cycle.
x,F x(t) F(t)
• ω ≈ ∞: In this case, Eq. (113) gives the phase as φ ≈ −π, so the motion is out of
phase with the force. And again, this implies that half the time your driving force
is doing positive work, and half the time it is doing negative work. So again there is t
cancelation. Let’s look at each quarter cycle again; see Fig. 25. As the mass (very
quickly) moves from the origin to the maximum displacement, you are doing negative
work, because your force is in the direction opposite the motion (you are the thing -W +W -W +W
(+v,-F) (-v,-F) (-v,+F) (+v,+F)
that is slowing the mass down). But then as you (very quickly) drag the mass back
toward to the origin, you are doing positive work, because your force is now in the Figure 25
direction of the motion (you are the thing that is speeding the mass up). The same
cancelation happens in the other half of the cycle.
x,F F(t) x(t)
• ω = ω0 : In this case, Eq. (116) gives the phase as φ = −π/2, so the motion is in
“quadrature” with the force. The consequence of this fact is that your driving force is
always doing positive work. Let’s now look at each half cycle; see Fig. 26. Start with t
the moment when the mass has maximum negative displacement. For the next half
cycle until it reaches maximum positive displacement, you are doing positive work,
because both your force and the velocity point to the right. And for other half cycle +W +W +W +W
(+v,+F) (+v,+F) (-v,-F) (-v,-F)
where the mass move back to the maximum negative displacement, you are also doing
positive work, because now both your force and the velocity point to the left. In short, Figure 26
the velocity, which is obtained by taking the derivative of the position in Eq. (117),
is always in phase with the force. So you are always doing positive work, and there is
no cancelation.
34 CHAPTER 1. OSCILLATIONS

Q values
Recall the Q ≡ ω0 /γ definition in Eq. (66). Q has interpretations for both the transient
(“homogeneous”) solution and the steady-state (“particular”) solution.

• For the transient solution, we found in Eq. (67) that Q is the number of oscillations
it takes for the amplitude to decrease by a factor of e−π ≈ 4%.

For the steady-state solution, there are actually two interpretations of Q.

• The first is that it equals the ratio of the amplitude at resonance to the amplitude at
small ω. This can be seen from the expression for A in Eq. (88). For ω = ω0 we have
A = F/γω0 , while for ω ≈ 0 we have A ≈ F/ω02 . Therefore,

Aresonance F/γω0 ω0
= 2 = ≡ Q. (129)
Aω≈0 F/ω0 γ

So the larger the Q value, the larger the amplitude at resonance. The analogous
statement in terms of power is that the larger the value of Q, the larger the Fd2 /2γm =
Fd2 Q/2ω0 m value of the power at resonance.

• The second steady-state interpretation of Q comes from the fact that the widths of
both the amplitude and power curves are proportional to γ (see Eqs. (92) and (128)).
Therefore, since Q ≡ ω0 /γ, the widths of the peaks are proportional to 1/Q. So the
larger the Q value, the thinner the peaks.
Putting these two facts together, a large Q value means that the amplitude curve is
tall and thin. And likewise for the power curve.

Let’s now look at some applications of these interpretations.

Tuning forks: The transient-solution interpretation allows for an easy calculation of
Q, at least approximately. Consider a tuning fork, for example. A typical frequency is
ω = 440 s−1 (a concert A pitch). Let’s say that it takes about 5 seconds to essentially
die out (when exactly it reaches 4% of the initial amplitude is hard to tell, but we’re just
doing a rough calculation here). This corresponds to 5 · 440 ≈ 2000 oscillations. So this is
(approximately) the Q value of the tuning fork.
Radios: Both of the steady-state-solution interpretations (tall peak and thin peak) are
highly relevant in any wireless device, such at a radio, cell phone, etc. The natural frequency
of the RLC circuit in, say, a radio is “tuned” (usually by adjusting the capacitance) so that
it has a certain resonant frequency; see Problem [to be added]. If this frequency corresponds
to the frequency of the electromagnetic waves (see Chapter 8) that are emitted by a given
radio station, then a large-amplitude oscillation will be created in the radio’s circuit. This
can then be amplified and sent to the speakers, creating the sound that you hear.
The taller the power peak, the stronger the signal that is obtained. If a radio station
is very far away, then the electromagnetic wave has a very small amplitude by the time it
gets to your radio. This means that the analog of Fd in Eq. (125) is very small. So the
only chance of having a sizeable value (relative to the oscillations from the inevitable noise
of other electromagnetic waves bombarding your radio) of the Fd2 /2γm term is to have γ be
small, or equivalently Q be large. (The electrical analog of γ is the resistance of the circuit.)
However, we need two things to be true if we want to have a pleasant listening experience.
We not only need a strong signal from the station we want to listen to, we also need a weak
signal from every other station, otherwise we’ll end up with a garbled mess. The thinness
of the power curve saves the day here. If the power peak is thin enough, then a nearby
1.3. DRIVEN AND DAMPED OSCILLATIONS 35

radio-station frequency (even, say, ω = (0.99)ω0 ) will contribute negligible power to the
circuit. It’s like this second station doesn’t exist, which is exactly how we want things to
look.
Atomic clocks: Another application where the second of the steady-state-solution in-
terpretations is critical is atomic clocks. Atomic clocks involve oscillations between certain
energy levels in atoms, but there’s no need to get into the details here. Suffice it to say that
there exists a damped oscillator with a particular natural frequency, and you can drive this
oscillator. The basic goal in using an atomic clock is to measure with as much accuracy and
precision as possible the value of the natural frequency of the atomic oscillations. You can A
(Q=1.4)
do this by finding the driving frequency that produces the largest oscillation amplitude, or 3.0
ω1 ω2 ω3
equivalently that requires the largest power input. The narrower the amplitude (or power) 2.5
2.0
curve, the more confident you can be that your driving frequency ω equals the natural 1.5
frequency ω0 . This is true for the following reason. 1.0
Consider a wide amplitude curve like the first one shown in Fig. 27. It’s hard to tell, by 0.5
0.0 ω/ω0
looking at the size of the resulting amplitude, whether you’re at, say ω1 or ω2 , or ω3 (all 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
measurements have some inherent error, so you can never be sure exactly what amplitude
you’ve measured). You might define the time unit of one second under the assumption that
ω1 is the natural frequency, whereas someone else (or perhaps you on a different day) might (Q=50)
A
define a second by thinking that ω3 is the natural frequency. This yields an inconsistent 50 ω2
standard of time. Although the natural frequency of the atomic oscillations has the same 40
value everywhere, the point is that people’s opinions on what this value actually is will 30
ω1 ω3
undoubtedly vary if the amplitude curve is wide. Just because there’s a definite value out 20
there doesn’t mean that we know what it is.7 10
0 ω/ω0
If, on the other hand, we have a narrow amplitude curve like the second one shown 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
in Fig. 27, then a measurement of a large amplitude can quite easily tell you that you’re
somewhere around ω1 , versus ω2 or ω3 . Basically, the uncertainty is on the order of the Figure 27
width of the curve, so the smaller the width, the smaller the uncertainty. Atomic clocks
have very high Q values, on the order of 1017 . The largeness of this number implies a very
small amplitude width, and hence very accurate clocks.
The tall-peak property of a large Q value isn’t too important in atomic clocks. It was
important in the case of a radio, because you might want to listen to a radio station that is
far away. But with atomic clocks there isn’t an issue with the oscillator having to pick up
a weak driving signal. The driving mechanism is right next to the atoms that are housing
the oscillations.
The transient property of large Q (that a large number of oscillations will occur before
the amplitude dies out) also isn’t too important in atomic clocks. You are continuing to
drive the system, so there isn’t any danger of the oscillations dying out.

1.3.4 Further discussion of resonance

Let’s now talk a bit more about resonance. As we saw in Eq. (118), the ẍ and ω02 x terms
cancel for the steady-state solution, x(t) = A cos(ωt + φ), because ω = ω0 at resonance.
You can consider the driving force to be simply balancing the damping force, while the
spring and the mass undergo their standard simple-harmonic motion. If ω = ω0 , and if
A < Fd /γmω (which means that the system isn’t in a steady state), then the driving force
is larger than the damping force, so the motion grows. If, on the other hand, A > Fd /γmω,
7 This is the classical way of thinking about it. The (correct) quantum-mechanical description says that

there actually isn’t a definite natural frequency; the atoms themselves don’t even know what it is. All that
exists is a distribution of possible natural frequencies. But for the present purposes, it’s fine to think about
things classically.
36 CHAPTER 1. OSCILLATIONS

then the driving force is less than the damping force, so the motion shrinks. This is why
A = Fd /γmω at resonance.
As shown in Fig. 26, the force leads the motion by 90◦ at resonance, so the force is in
phase with the velocity. This leads to the largest possible energy being fed into the system,
because the work is always positive, so there is no cancelation with negative work. There
are many examples of resonance in the real world, sometimes desirable, and sometimes
undesirable. Let’s take a look at a few.

Desirable resonance
• RLC circuits: As you will find if you do Problem [to be added], you can use
Kirchhoff’s rules in an RLC circuit to derive an equation exactly analogous to the
damped/driven oscillator equation in Eq. (81). The quantities m, γ, k, and Fd in the
mechanical system are related to the quantities L, R, 1/C, and V0 in the electrical
system, respectively. Resonance allows you to pick out a certain frequency and ignore
all the others. This is how radios, cell phones, etc. work, as we discussed in the “Q
values” section above.
If you have a radio sitting on your desk, then it is being bombarded by radio waves
with all sorts of frequencies. If you want to pick out a certain frequency, then you
can “tune” your radio to that frequency by changing the radio’s natural frequency
ω0 (normally done by changing the capacitance C in the internal circuit). Assuming
that the damping in the circuit is small (this is determined by R), then from the plot
of A in Fig. 20, there will be a large oscillation in the circuit at the radio station’s
frequency, but a negligible oscillation at all the other frequencies that are bombarding
the radio.

• Musical instruments: The “pipe” of, say, a flute has various natural frequencies
(depending on which keys are pressed), and these are the ones that survive when
you blow air across the opening. We’ll talk much more about musical instruments
in Chapter 5. There is a subtlety about whether some musical instruments function
because of resonance or because of “positive feedback,” but we won’t worry about that
here.

• The ear: The hair-like nerves in the cochlea have a range of resonant frequencies
which depend on the position in the cochlea. Depending on which ones vibrate, a signal
is (somehow) sent to the brain telling it what the pitch is. It is quite remarkable how
this works.

Undesirable resonance
• Vehicle vibrations: This is particularly relevant in aircraft. Even the slightest
driving force (in particular from the engine) can create havoc if its frequency matches
up with any of the resonant frequencies of the plane. There is no way to theoretically
predict every single one of the resonant frequencies, so the car/plane/whatever has to
be tested at all frequencies by sweeping through them and looking for large amplitudes.
This is difficult, because you need the final product. You can’t do it with a prototype
in early stages of development.

• Tacoma Narrows Bridge failure: There is a famous video of this bridge oscillat-
ing wildly and then breaking apart. As with some musical instruments, this technically
shouldn’t be called “resonance.” But it’s the same basic point – a natural frequency
of the object was excited, in one way or another.
1.3. DRIVEN AND DAMPED OSCILLATIONS 37

• Millennium Bridge in London: This pedestrian bridge happened to have a lateral

resonant frequency on the order of 1 Hz. So when it started to sway (for whatever
reason), people began to walk in phase with it (which is the natural thing to do). This
had the effect of driving it more and further increasing the amplitude. Dampers where
added, which fixed the problem.
• Tall buildings: A tall building has a resonant frequency of swaying (or actually a
couple, depending on the direction; and there can be twisting, too). If the effects from
the wind or earthquakes happen to unfortuitously drive the building at this frequency,
then the sway can become noticeable. “Tuned mass dampers” (large masses connected
to damping mechanisms, in the upper floors) help alleviate this problem.
• Space station: In early 2009, a booster engine on the space station changed its
firing direction at a frequency that happened to match one of the station’s resonant
frequencies (about 0.5 Hz). The station began to sway back and forth, made noticeable
by the fact that free objects in the air were moving back and forth. Left unchecked,
a larger and larger amplitude would of course be very bad for the structure. It was
fortunately stopped in time.
Chapter 2

Normal modes
David Morin, [email protected]

In Chapter 1 we dealt with the oscillations of one mass. We saw that there were various
possible motions, depending on what was influencing the mass (spring, damping, driving
forces). In this chapter we’ll look at oscillations (generally without damping or driving)
involving more than one object. Roughly speaking, our counting of the number of masses
will proceed as: two, then three, then infinity. The infinite case is relevant to a continuous
system, because such a system contains (ignoring the atomic nature of matter) an infinite
number of infinitesimally small pieces. This is therefore the chapter in which we will make
the transition from the oscillations of one particle to the oscillations of a continuous object,
that is, to waves.
The outline of this chapter is as follows. In Section 2.1 we solve the problem of two
masses connected by springs to each other and to two walls. We will solve this in two ways
– a quick way and then a longer but more fail-safe way. We encounter the important concepts
of normal modes and normal coordinates. We then add on driving and damping forces and
apply some results from Chapter 1. In Section 2.2 we move up a step and solve the analogous
problem involving three masses. In Section 2.3 we solve the general problem involving N
masses and show that the results reduce properly to the ones we already obtained in the
N = 2 and N = 3 cases. In Section 2.4 we take the N → ∞ limit (which corresponds
to a continuous stretchable material) and derive the all-important wave equation. We then
discuss what the possible waves can look like.

k κ k
2.1 Two masses
m m
p
For a single mass on a spring, there is one natural frequency, namely k/m. (We’ll consider
Figure 1
undamped and undriven motion for now.) Let’s see what happens if we have two equal
masses and three spring arranged as shown in Fig. 1. The two outside spring constants
are the same, but we’ll allow the middle one to be different. In general, all three spring
constants could be different, but the math gets messy in that case.
Let x1 and x2 measure the displacements of the left and right masses from their respective
equilibrium positions. We can assume that all of the springs are unstretched at equilibrium,
but we don’t actually have to, because the spring force is linear (see Problem [to be added]).
The middle spring is stretched (or compressed) by x2 − x1 , so the F = ma equations on the

1
2 CHAPTER 2. NORMAL MODES

two masses are

mẍ1 = −kx1 − κ(x1 − x2 ),

mẍ2 = −kx2 − κ(x2 − x1 ). (1)

Concerning the signs of the κ terms here, they are equal and opposite, as dictated by
Newton’s third law, so they are either both right or both wrong. They are indeed both
right, as can be seen by taking the limit of, say, large x2 . The force on the left mass is then
in the positive direction, which is correct.
These two F = ma equations are “coupled,” in the sense that both x1 and x2 appear in
both equations. How do we go about solving for x1 (t) and x2 (t)? There are (at least) two
ways we can do this.

2.1.1 First method

This first method is quick, but it works only for simple systems with a sufficient amount of
symmetry. The main goal in this method is to combine the F = ma equations in well-chosen
ways so that x1 and x2 appear only in certain unique combinations. It sometimes involves a
bit of guesswork to determine what these well-chosen ways are. But in the present problem,
the simplest thing to do is add the F = ma equations in Eq. (1), and it turns out that this
is in fact one of the two useful combinations to form. The sum yields

d2 k
m(ẍ1 + ẍ2 ) = −k(x1 + x2 ) =⇒ (x1 + x2 ) = − (x1 + x2 ). (2)
dt2 m
The variables x1 and x2 appear here only in the unique combination, x1 + x2 . And further-
more, this equation is simply a harmonic-motion equation for the quantity x1 + x2 . The
solution is therefore
r
k
x1 (t) + x2 (t) = 2As cos(ωs t + φs ), where ωs ≡ (3)
m

The “s” here stands for “slow,” to be distinguished from the “fast” frequency we’ll find
below. And we’ve defined the coefficient to be 2As so that we won’t have a bunch of factors
of 1/2 in our final answer in Eq. (6) below.
No matter what complicated motion the masses are doing, the quantity x1 + x2 always
undergoes simple harmonic motion with frequency ωs . This is by no means obvious if you
look at two masses bouncing back and forth in an arbitrary manner.
The other useful combination of the F = ma equations is their difference, which conve-
niently is probably the next thing you might try. This yields

d2 k + 2κ
m(ẍ1 − ẍ2 ) = −(k + 2κ)(x1 − x2 ) =⇒ 2
(x1 − x2 ) = − (x1 − x2 ). (4)
dt m
The variables x1 and x2 now appear only in the unique combination, x1 − x2 . And again,
we have a harmonic-motion equation for the quantity x1 − x2 . So the solution is (the “f”
stands for “fast”)
r
k + 2κ
x1 (t) − x2 (t) = 2Af cos(ωf t + φf ), where ωf ≡ (5)
m

As above, no matter what complicated motion the masses are doing, the quantity x1 − x2
always undergoes simple harmonic motion with frequency ωf .
2.1. TWO MASSES 3

We can now solve for x1 (t) and x2 (t) by adding and subtracting Eqs. (3) and (5). The
result is

x1 (t) = As cos(ωs t + φs ) + Af cos(ωf t + φf ),

x2 (t) = As cos(ωs t + φs ) − Af cos(ωf t + φf ). (6)

The four constants, As , Af , φs , φf are determined by the four initial conditions, x1 (0), x2 (0),
ẋ1 (0), ẋ1 (0).
The above method will clearly work only if we’re able to guess the proper combinations of
the F = ma equations that yield equations involving unique combinations of the variables.
Adding and subtracting the equations worked fine here, but for more complicated systems
with unequal masses or with all the spring constants different, the appropriate combination
of the equations might be far from obvious. And there is no guarantee that guessing around
will get you anywhere. So before discussing the features of the solution in Eq. (6), let’s take
a look at the other more systematic and fail-safe method of solving for x1 and x2 .

2.1.2 Second method

This method is longer, but it works (in theory) for any setup. Our strategy will be to look
for simple kinds of motions where both masses move with the same frequency. We will
then build up the most general solution from these simple motions. For all we know, such
motions might not even exist, but we have nothing to lose by trying to find them. We will
find that they do in fact exist. You might want to try to guess now what they are for our
two-mass system, but it isn’t necessary to know what they look like before undertaking this
method.
Let’s guess solutions of the form x1 (t) = A1 eiωt and x2 (t) = A2 eiωt . For bookkeeping
purposes, it is convenient to write these solutions in vector form:
µ ¶ µ ¶
x1 (t) A1
= eiωt . (7)
x2 (t) A2

We’ll end up taking the real part in the end. We can alternatively guess the solution eαt
without the i, but then our α will come out to be imaginary. Either choice will get the job
done. Plugging these guesses into the F = ma equations in Eq. (1), and canceling the factor
of eiωt , yields

−mω 2 A1 = −kA1 − κ(A1 − A2 ),

−mω 2 A2 = −kA2 − κ(A2 − A1 ). (8)

In matrix form, this can be written as

µ ¶µ ¶ µ ¶
−mω 2 + k + κ −κ A1 0
= . (9)
−κ −mω 2 + k + κ A2 0

At this point, it seems like we can multiply both sides of this equation by the inverse of
the matrix. This leads to (A1 , A2 ) = (0, 0). This is obviously a solution (the masses just
sit there), but we’re looking for a nontrivial solution that actually contains some motion.
The only way to escape the preceding conclusion that A1 and A2 must both be zero is if
the inverse of the matrix doesn’t exist. Now, matrix inverses are somewhat messy things
(involving cofactors and determinants), but for the present purposes, the only fact we need to
know about them is that they involve dividing by the determinant. So if the determinant is
4 CHAPTER 2. NORMAL MODES

zero, then the inverse doesn’t exist. This is therefore what we want. Setting the determinant
equal to zero gives the quartic equation,
¯ ¯
¯ −mω 2 + k + κ −κ ¯
¯ ¯ = 0 =⇒ (−mω 2 + k + κ)2 − κ2 = 0
¯ −κ −mω 2 + k + κ ¯
−mω 2 + k + κ = ±κ
=⇒
k k + 2κ
=⇒ ω 2 = or . (10)
m m
p p
The four solutions to the quartic equation are therefore ω = ± k/m and ω = ± (k + 2κ)/m.
For the case where ω 2 = k/m, we can plug this value of ω 2 back into Eq. (9) to obtain
µ ¶µ ¶ µ ¶
1 −1 A1 0
κ = . (11)
−1 1 A2 0

Both rows of this equation yield the same result (this was the point of setting the determinant
equal to zero), namely A1 = A2 . So (A1 , A2 ) is proportional to the vector (1, 1).
For the case where ω 2 = (k + 2κ)/m, Eq. (9) gives
µ ¶µ ¶ µ ¶
−1 −1 A1 0
κ = . (12)
−1 −1 A2 0

Both rows nowpyield A1 = −A2 .pSo (A1 , A2 ) is proportional to the vector (1, −1).
With ωs ≡ k/m and ωf ≡ (k + 2κ)/m, we can write the general solution as the sum
of the four solutions we have found. In vector notation, x1 (t) and x2 (t) are given by
µ ¶ µ ¶ µ ¶ µ ¶ µ ¶
x1 (t) 1 iωs t 1 −iωs t 1 iωf t 1
= C1 e +C2 e +C3 e +C4 e−iωf t . (13)
x2 (t) 1 1 −1 −1

We now perform the usual step of invoking the fact that the positions x1 (t) and x2 (t)
must be real for all t. This yields that standard result that C1 = C2∗ ≡ (As /2)eiφs and
C3 = C4∗ ≡ (Af /2)eiφf . We have included the factors of 1/2 in these definitions so that we
won’t have a bunch of factors of 1/2 in our final answer. The imaginary parts in Eq. (13)
cancel, and we obtain
µ ¶ µ ¶ µ ¶
x1 (t) 1 1
= As cos(ωs t + φs ) + Af cos(ωf t + φf ) (14)
x2 (t) 1 −1

Therefore,

x1 (t) = As cos(ωs t + φs ) + Af cos(ωf t + φf ),

x2 (t) = As cos(ωs t + φs ) − Af cos(ωf t + φf ). (15)

This agrees with the result in Eq. (6).

As we discussed in Section 1.1.5, we could have just taken the real part of the C1 (1, 1)eiωs t
and C3 (1, −1)eiωf t solutions, instead of going through the “positions must be real” reasoning.
However, you should continue using the latter reasoning until you’re comfortable with the
short cut of taking the real part.
Remark: Note that Eq. (9) can be written in the form,
µ ¶µ ¶ µ ¶
k+κ −κ A1 A1
= mω 2 . (16)
−κ k+κ A2 A2
2.1. TWO MASSES 5

So what we did above was solve for the eigenvectors and eigenvalues of this matrix. The eigenvectors
of a matrix are the special vectors that get carried into a multiple of themselves what acted on by
the matrix. And the multiple (which is mω 2 here) is called the eigenvalue. Such vectors are indeed
special, because in general a vector gets both stretched (or shrunk) and rotated when acted on by
a matrix. Eigenvectors don’t get rotated at all. ♣

A third method of solving our coupled-oscillator problem is to solve for x2 in the first
equation in Eq. (1) and plug the result into the second. You will get a big mess of a
fourth-order differential equation, but it’s solvable by guessing x1 = Aeiωt .

2.1.3 Normal modes and normal coordinates

Normal modes
Having solved for x1 (t) and x2 (t) in various ways, let’s now look at what we’re found. If
Af = 0 in Eq. (15), then we have

x1 (t) = x2 (t) = As cos(ωs t + φs ). (17)

So both masses move in exactly the same manner. Both to the right, then both to the left,
and so on. This is shown in Fig. 2. The middle spring is never stretched, so it effectively
isn’t there. We therefore basically have two copies of a simple spring-mass
p system. This
is consistent with the fact that ωs equals the standard expression k/m, independent of
Figure 2
κ. This nice motion, where both masses move with the same frequency, is called a normal
mode. To specify what a normal mode looks like, you phave to give the frequency and also
the relative amplitudes. So this mode has frequency k/m, and the amplitudes are equal.
If, on the other hand, As = 0 in Eq. (15), then we have

x1 (t) = −x2 (t) = Af cos(ωf t + φf ). (18)

Now the masses move oppositely. Bothpoutward, then both inward, and so on. This is shown
in Fig. 3. The frequency is now ωf = (k + 2κ)/m. It makes sense that this is larger than
ωs , because the middle spring is now stretched or compressed, so itpadds to the restoring
force. This nice motion is the other normal mode. It has frequency (k + 2κ)/m, and the
Figure 3
amplitudes are equal and opposite. The task of Problem [to be added] is to deduce the
frequency ωf in a simpler way, without going through the whole process above.
Eq. (15) tells us that any arbitrary motion of the system can be thought of as a linear
combination of these two normal modes. But in the general case where both coefficients
As and Af are nonzero, it’s rather difficult to tell that the motion is actually built up from
these two simple normal-mode motions.

Normal coordinates
By adding and subtracting the expressions for x1 (t) and x1 (t) in Eq. (15), we see that for
any arbitrary motion of the system, the quantity x1 + x2 oscillates with frequency ωs , and
the quantity x1 − x2 oscillates with frequency ωf . These combinations of the coordinates
are known as the normal coordinates of the system. They are the nice combinations of the
coordinates that we found advantageous to use in the first method above.
The x1 + x2 normal coordinate is associated with the normal mode (1, 1), because they
both have frequency ωs . Equivalently, any contribution from the other mode (where x1 =
−x2 ) will vanish in the sum x1 + x2 . Basically, the sum x1 + x2 picks out the part of the
motion with frequency ωs and discards the part with frequency ωf . Similarly, the x1 − x2
normal coordinate is associated with the normal mode (1, −1), because they both have
6 CHAPTER 2. NORMAL MODES

frequency ωf . Equivalently, any contribution from the other mode (where x1 = x2 ) will
vanish in the difference x1 − x2 .
Note, however, that the association of the normal coordinate x1 + x2 with the normal
mode (1, 1) does not follow from the fact that the coefficients in x1 + x2 are both 1. Rather,
it follows from the fact that the other normal mode, namely (x1 , x2 ) ∝ (1, −1), gives no
contribution to the sum x1 +x2 . There are a few too many 1’s floating around in the present
example, so it’s hard to see which results are meaningful and which results are coincidence.
But the following example should clear things up. Let’s say we solved a problem using the
determinant method, and we found the solution to be
µ ¶ µ ¶ µ ¶
x 3 1
= B1 cos(ω1 t + φ1 ) + B2 cos(ω2 t + φ2 ). (19)
y 2 −5

Then 5x + y is the normal coordinate associated with the normal mode (3, 2), which has
frequency ω1 . This is true because there is no cos(ω2 t + φ2 ) dependence in the quantity
5x + y. And similarly, 2x − 3y is the normal coordinate associated with the normal mode
(1, −5), which has frequency ω2 , because there is no cos(ω1 t+φ1 ) dependence in the quantity
2x − 3y.

2.1.4 Beats
Let’s now apply some initial conditions to the solution in Eq. (15). We’ll take the initial
conditions to be ẋ1 (0) = ẋ2 (0) = 0, x1 (0) = 0, and x2 (0) = A. In other words, we’re pulling
the right mass to the right, and then releasing both masses from rest. It’s easier to apply
these conditions if we write the solutions for x1 (t) and x2 (t) in the form,

x1 (t) = a cos ωs t + b sin ωs t + c cos ωf t + d sin ωf t

x2 (t) = a cos ωs t + b sin ωs t − c cos ωf t − d sin ωf t. (20)

This form of the solution is obtained by using the trig sum formulas to expand the sines
and cosines in Eq. (15). The coefficients a, b, c, d are related to the constants As , Af , φs ,
φf . For example, the cosine sum formula gives a = As cos φs . If we now apply the initial
conditions to Eq. (20), the velocities ẋ1 (0) = ẋ1 (0) = 0 quickly give b = d = 0. And the
positions x1 (0) = 0 and x2 (0) = A give a = −c = A/2. So we have

A¡ ¢
x1 (t) = cos ωs t − cos ωf t ,
2
A¡ ¢
x2 (t) = cos ωs t + cos ωf t . (21)
2
For arbitrary values of ωs and ωf , this generally looks like fairly random motion, but let’s
look at a special case. If κ ¿ k, then the ωf in Eq. (5) is only slightly larger than the ωs in
Eq. (3), so something interesting happens. For frequencies that are very close to each other,
it’s a standard technique (for reasons that will become clear) to write ωs and ωf in terms of
their average and (half) difference:
ωf + ωs ωf − ωs
ωs = − ≡ Ω − ²,
2 2
ωf + ωs ωf − ωs
ωf = + ≡ Ω + ², (22)
2 2
where
ωf + ωs ωf − ωs
Ω≡ , and ²≡ . (23)
2 2
2.1. TWO MASSES 7

Using the identity cos(α ± β) = cos α cos β ∓ sin α sin β, Eq. (21) becomes
A³ ´
x1 (t) = cos(Ω − ²)t − cos(Ω + ²)t = A sin Ωt sin ²t,
2 Ω=10, ε=1
A³ ´
x2 (t) = cos(Ω − ²)t + cos(Ω + ²)t = A cos Ωt cos ²t. (24) x1=A sinΩt sinεt
2
A
If ωs is very close to ωf , then ² ¿ Ω, which means that the ²t oscillation is much slower
than that Ωt oscillation. The former therefore simply acts as an envelope for the latter.
x1 (t) and x2 (t) are shown in Fig. 4 for Ω = 10 and ² = 1. The motion sloshes back and t
1 2 3 4 5 6
forth between the masses. At the start, only the second mass is moving. But after a time
of ²t = π/2 =⇒ t = π/2², the second mass is essentially not moving and the first mass has -A
all the motion. Then after another time of π/2² it switches back, and so on. -A sinεt
This sloshing back and forth can be understood in terms of driving forces and resonance. x2=A cosΩt cosεt
At the start (and until ²t = π/2), x2 looks like cos Ωt with a slowly changing amplitude A
(assuming ² ¿ Ω). And x1 looks like sin Ωt with a slowly changing amplitude. So x2 is 90◦
ahead of x1 , because cos Ωt = sin(Ωt + π/2). This 90◦ phase difference means that the x2
t
mass basically acts like a driving force (on resonance) on the x1 mass. Equivalently, the x2 1 2 3 4 5 6
mass is always doing positive work on the x1 mass, and the x1 mass is always doing negative
work on the x2 mass. Energy is therefore transferred from x2 to x1 . -A
However, right after x2 has zero amplitude (instantaneously) at ²t = π/2, the cos ²t factor
in x2 switches sign, so x2 now looks like − cos Ωt (times a slowly-changing amplitude). And Figure 4
x1 still looks like sin Ωt. So now x2 is 90◦ behind x1 , because − cos Ωt = sin(Ωt − π/2). So
the x1 mass now acts like a driving force (on resonance) on the x2 mass. Energy is therefore
transferred from x1 back to x2 . And so on and so forth.
In the plots in Fig. 4, you can see that something goes a little haywire when the envelope
curves pass through zero at ²t = π/2, π, etc. The x1 or x2 curves skip ahead (or equivalently,
fall behind) by half of a period. If you inverted the second envelope “bubble” in the first
plot, the periodicity would then return. That is, the peaks of the fast-oscillation curve would
occur at equal intervals, even in the transition region around ²t = π.
The classic demonstration of beats consists of two identical pendulums connected by a
weak spring. The gravitational restoring force mimics the “outside” springs in the above
setup, so the same general results carry over (see Problem [to be added]). At the start, one
pendulum moves while the other is nearly stationary. But then after a while the situation
is reversed. However, if the masses of the pendulums are different, it turns out that not all
of the energy is transferred. See Problem [to be added] for the details.
When people talk about the “beat frequency,” they generally mean the frequency of
the “bubbles” in the envelope curve. If you’re listening to, say, the sound from two guitar
strings that are at nearly the same frequency, then this beat frequency is the frequency of
the waxing and waning that you hear. But note that this frequency is 2², and not ², because
two bubbles occur in each of the ²t = 2π periods of the envelope.1
k κ k
2.1.5 Driven and damped coupled oscillators m m
Consider the coupled oscillator system with two masses and three springs from Fig. 1 above, Fd cos ωt
but now with a driving force acting on one of the masses, say the left one (the x1 one); see
Fig. 5. And while we’re at it, let’s immerse the system in a fluid, so that both masses have Figure 5
a drag coefficient b (we’ll assume it’s the same for both). Then the F = ma equations are
mẍ1 = −kx1 − κ(x1 − x2 ) − bẋ1 + Fd cos ωt,
1 If you want to map the spring/mass setup onto the guitar setup, then the x in Eq. (21) represents the
1
amplitude of the sound wave at your ear, and the ωs and ωf represent the two different nearby frequencies.
The second position, x2 , doesn’t come into play (or vice versa). Only one of the plots in Fig. 4 is relevant.
8 CHAPTER 2. NORMAL MODES

mẍ2 = −kx2 − κ(x2 − x1 ) − bẋ2 . (25)

We can solve these equations by using the same adding and subtracting technique we used
in Section 2.1.1. Adding them gives

m(ẍ1 + ẍ2 ) = −k(x1 + x2 ) − b(ẋ1 + ẋ2 ) + Fd cos ωt

=⇒ z̈s + γ żs + ωs2 zs = F cos ωt, (26)

where zs ≡ x1 + x2 , γ ≡ b/m, ωs2 ≡ k/m, and F ≡ Fd /m. But this is our good ol’
driven/dampded oscillator equation, in the variable zs . We can therefore just invoke the
results from Chapter 1. The general solution is the sum of the homogeneous and particular
solutions. But the let’s just concentrate on the particular (steady state) solution here. We
can imagine that the system has been oscillating for a long time, so that the damping has
made the homogeneous solution decay to zero. For the particular solution, we can simply
copy the results from Section 1.3.1. So we have

x1 + x2 ≡ zs = As cos(ωt + φs ), (27)

where
−γω F
tan φs = , and As = p . (28)
ωs2 − ω 2 (ωs2 − ω 2 )2 + γ 2 ω 2
Similarly, subtracting the F = ma equations gives

m(ẍ1 − ẍ2 ) = −(k + 2κ)(x1 − x2 ) − b(ẋ1 − ẋ2 ) + Fd cos ωt

=⇒ z̈f + γ żf + ωf2 zf = F cos ωt, (29)

where zf ≡ x1 − x2 and ωf2 ≡ (k + 2κ)/m. Again, this is a nice driven/dampded oscillator

equation, and the particular solution is

x1 − x2 ≡ zf = Af cos(ωt + φf ), (30)

where
−γω F
tan φf = , and Af = p . (31)
ωf2− ω2 (ωf2 − ω 2 )2 + γ 2 ω 2
Adding and subtracting Eqs. (27) and (30) to solve for x1 (t) and x2 (t) gives

x1 (t) = Cs cos(ωt + φs ) + Cf cos(ωt + φf ),

x2 (t) = Cs cos(ωt + φs ) − Cf cos(ωt + φf ), (32)

where Cs ≡ As /2, and Cf ≡ Af /2.

We end up getting two resonant frequencies, which are simply the frequencies of the
normal modes, ωs and ωf . If γ is small, and if the driving frequency ω equals either ωs
or ωf , then the amplitudes of x1 and x2 are large. In the ω = ωs case, x1 and x2 are
approximately in phase with equal amplitudes (the Cs terms dominate the Cf terms). And
in the ω = ωs case, x1 and x2 are approximately out of phase with equal amplitudes (the Cf
terms dominate the Cs terms, and there is a relative minus sign). But these are the normal
modes we found in Section 2.1.3. The resonances therefore cause the system to be in the
normal modes.
In general, if there are N masses (and hence N modes), then there are N resonant
frequencies, which are the N normal-mode frequencies. So for complicated objects with
more than two pieces, there are lots of resonances.
2.2. THREE MASSES 9 k k k k

2.2 Three masses m m m

As a warmup to the general case of N masses connected by springs, let’s look at the case of Figure 6
three masses, as shown in Fig. 6. We’ll just deal with undriven and undamped motion here,
and we’ll also assume that all the spring constants are equal, lest the math get intractable.
If x1 , x2 , and x3 are the displacements of the three masses from their equilibrium positions,
then the three F = ma equations are

mẍ1 = −kx1 − k(x1 − x2 ),

mẍ2 = −k(x2 − x1 ) − k(x2 − x3 ),
mẍ3 = −k(x3 − x2 ) − kx3 . (33)

You can check that all the signs of the k(xi − xj ) terms are correct, by imagining that, say,
one of the x’s is very large. It isn’t so obvious which combinations of these equations yield
equations involving only certain unique combinations of the x’s (the normal coordinates), so
we won’t be able to use the method of Section 2.1.1. We will therefore use the determinant
method from Section 2.1.2 and guess a solution of the form
   
x1 A1
 x2  =  A2  eiωt , (34)
x3 A3

with the goal of solving for ω, and also for the amplitudes A1 , A2 , and A3 (up to an overall
factor). Plugging this guess into Eq. (33) and putting all the terms on the lefthand side,
and canceling the eiωt factor, gives
    
−ω 2 + 2ω02 −ω02 0 A1 0
 −ω02 −ω 2 + 2ω02 −ω02   A2  =  0  , (35)
2 2 2
0 −ω0 −ω + 2ω0 A3 0

where ω02 ≡ k/m. As in the earlier two-mass case, a nonzero solution for (A1 , A2 , A3 ) exists
only if the determinant of this matrix is zero. Setting it equal to zero gives
³ ´ ³ ´
(−ω 2 + 2ω02 ) (−ω 2 + 2ω02 )2 − ω04 + ω02 − ω02 (−ω 2 + 2ω02 ) = 0
=⇒ (−ω 2 + 2ω02 )(ω 4 − 4ω02 ω 2 + 2ω04 ) = 0. (36)

Although this is technically a 6th-order equation, it’s really just a cubic equation in ω 2 . But
since we know that (−ω 2 + 2ω02 ) is a factor, in the end it boils down to a quadratic equation
in ω 2 .
Remark: If you had multiplied everything out and lost the information that (−ω 2 + 2ω02 ) is a
factor, you could still easily see that ω 2 = 2ω02 must be a root, because an easy-to-see normal
mode is one where the middle mass stays fixed and the outer masses move in opposite directions.
In this case the middle mass is essentially a brick wall, so the outer masses are connected to two
springs whose other ends are fixed. The effective spring constant is then 2k, which means that the
√
frequency is 2ω0 . ♣

Using the quadratic formula, the roots to Eq. (36) are

√
ω 2 = 2ω02 , and ω 2 = (2 ± 2)ω0 . (37)

Plugging these values back into Eq. (35) to find the relations among A1 , A2 , and A3 gives
 10 CHAPTER 2. NORMAL MODES

the three normal modes:2

   
√ A1 1
ω = ± 2 ω0 =⇒  A2  ∝  0 ,
A3 −1
   
q A1 1
√ √
ω=± 2+ 2 ω0 =⇒  A2  ∝  − 2 ,
A3 1
   
q
√ A1 √1
ω = ± 2 − 2 ω0 =⇒  A2  ∝  2 . (38)
A3 1

The most general solution is obtained by taking an arbitrary linear combination of the
six solutions corresponding to the six possible values of ω (don’t forget the three negative
solutions):
     
x1 1 √ 1 √
 x2  = C1  0  ei 2ω0 t + C2  0  e−i 2ω0 t + · · · . (39)
x3 −1 −1

However, the x’s must be real, so C√

2 must be the complex conjugate of C1 . Likewise for the
two C’s√corresponding to the (1, − 2, 1) mode, and also for the two C’s corresponding to
the (1, 2, 1) mode. Following the procedure that transformed Eq. (13) into Eq. (14), we
see that the most general solution can be written as
   
x1 1 ³√ ´
 x2  = Am  0  cos 2 ω0 t + φm
x3 −1
 
1 µq ¶
√ √
+Af  − 2  cos 2 + 2 ω0 t + φf
1
 
µq ¶
√1 √
+As  2  cos 2 − 2 ω0 t + φs . (40)
1

The subscripts “m,” “f,” and “s” stand for middle, fast, and slow. The six unknowns, Am ,
Af , As , φm , φf , and φs are determined by the six initial conditions (three positions and three
medium: (-1,0,1) velocities). If Am is the only nonzero coefficient, then the motion is purely in the middle
mode. Likewise for the cases where only Af or only As is nonzero. Snapshots of these modes
are shown in Fig. 7. You should convince yourself that they qualitatively make sense. If you
want to get quantitative, the√task of Problem [to be added] is to give a force argument that
explains the presence of the 2 in the amplitudes of the fast and slow modes.

fast: (1,- 2 ,1)

2.3 N masses
2.3.1 Derivation of the general result
Let’s now consider the general case of N masses between two fixed walls. The masses are
all equal to m, and the spring constants are all equal to k. The method we’ll use below will
slow: (1, 2 ,1) 2 Only two of the equations in Eq. (35) are needed. The third equation is redundant; that was the point

of setting the determinant equal to zero.

Figure 7
2.3. N MASSES 11

actually work even if we don’t have walls at the ends, that is, even if the masses extend
infinitely in both directions. Let the displacements of the masses relative to their equilibrium
positions be x1 , x2 ,. . . , xN . If the displacements of the walls are called x0 and xN +1 , then
the boundary conditions that we’ll eventually apply are x0 = xN +1 = 0.
The force on the nth mass is

Fn = −k(xn − xn−1 ) − k(xn − xn+1 ) = kxn−1 − 2kxn + kxn+1 . (41)

So we end up with a collection of F = ma equations that look like

mẍn = kxn−1 − 2kxn + kxn+1 . (42)

These can all be collected into the matrix equation,

    
.. .. ..
 .   .  . 

2  xn−1 
  · · · k −2k k   xn−1 
d   
m 2 x = k −2k k   xn . (43)
dt     
n
 xn+1   k −2k k ···   
     xn+1 
.. .. ..
. . .
In principle, we could solve for the normal modes by guessing a solution of the form,
   
.. ..
 .   . 
 xn−1   An−1 
   
 xn  =  An  eiωt , (44)
   
 xn+1   An+1 
   
.. ..
. .
and then setting the resulting determinant equal to zero. This is what we did in the N = 2
and N = 3 cases above. However, for large N , it would be completely intractable to solve
for the ω’s by using the determinant method. So we’ll solve it in a different way, as follows.
We’ll stick with the guess in Eq. (44), but instead of the determinant method, we’ll
look at each of the F = ma equations individually. Consider the nth equation. Plugging
xn (t) = An eiωt into Eq. (42) and canceling the factor of eiωt gives
¡ ¢
−ω 2 An = ω02 An−1 − 2An + An+1
An−1 + An+1 2ω02 − ω 2
=⇒ = , (45)
An ω02
p
where ω0 = k/m, as usual. This equation must hold for all values of n from 1 to N ,
so we have N equations of this form. For a given mode with a given frequency ω, the
quantity (2ω02 − ω 2 )/ω02 on the righthand side is a constant, independent of n. So the ratio
(An−1 +An+1 )/An on the lefthand side must also be independent of n. The problem therefore
reduces to finding the general form of a string of A’s that has the ratio (An−1 + An+1 )/An
being independent of n.
If someone gives you three adjacent A’s, then this ratio is determined, so you can recur-
sively find the A’s for all other n (both larger and smaller than the three you were given).
Or equivalently, if someone gives you two adjacent A’s and also ω, so that the value of
(2ω02 − ω 2 )/ω02 is known (we’re assuming that ω0 is given), then all the other A’s can be
determined. The following claim tells us what form the A’s take. It is this claim that allows
us to avoid using the determinant method.
12 CHAPTER 2. NORMAL MODES

Claim 2.1 If ω ≤ 2ω0 , then any set of An ’s satisfying the system of N equations in Eq.
(45) can be written as
An = B cos nθ + C sin nθ, (46)
for certain values of B, C, and θ. (The fact that there are three parameters here is consistent
with the fact that three A’s, or two A’s and ω, determine the whole set.)

Proof: We’ll start by defining

An−1 + An+1
cos θ ≡ . (47)
2An
As mentioned above, the righthand side is independent of n, so θ is well defined (up to the
usual ambiguities of duplicate angles; θ + 2π, and −θ, etc. also work).3 If we’re looking at
a given normal mode with frequency ω, then in view of Eq. (45), an equivalent definition of
θ is
2ω02 − ω 2
2 cos θ ≡ . (48)
ω02
These definitions are permitted only if they yield a value of cos θ that satisfies | cos θ| ≤ 1.
This condition is equivalent to the condition that ω must satisfy −2ω0 ≤ ω ≤ 2ω0 . We’ll
just deal with positive ω here (negative ω yields the same results, because only its square
enters into the problem), but we must remember to also include the e−iωt solution in the
end (as usual). So this is where the ω ≤ 2ω0 condition in the claim comes from.4
We will find that with walls at the ends, θ (and hence ω) can take on only a certain set
of discrete values. We will calculate these below. If there are no walls, that is, if the system
extends infinitely in both directions, then θ (and hence ω) can take on a continuous set of
values.
As we mentioned above, the N equations represented in Eq. (45) tell us that if we
know two of the A’s, and if we also have a value of ω, then we can use the equations to
successively determine all the other A’s. Let’s say that we know what A0 and A1 are. (In
the case where there are walls, we know know that A0 = 0, but let’s be general and not
invoke this constraint yet.) The rest of the An ’s can be determined as follows. Define B by

A0 ≡ B cos(0 · θ) + C sin(0 · θ) =⇒ A0 ≡ B. (49)

(So B = 0 if there are walls.) Once B has been defined, define C by

A1 ≡ B cos(1 · θ) + C sin(1 · θ) =⇒ A1 ≡ B cos θ + C sin θ, (50)

For any A0 and A1 , these two equations uniquely determine B and C (θ was already deter-
mined by ω). So to sum up the definitions: ω, A0 , and A1 uniquely determine θ, B and C.
(We’ll deal with the multiplicity of the possible θ values below in the “Nyquist” subsection.)
By construction of these definitions, the proposed An = B cos nθ + C sin nθ relation holds
for n = 0 and n = 1. We will now show inductively that it holds for all n.
3 The motivation for this definition is that the fraction on the righthand side has a sort of second-derivative

feel to it. The more this fraction differs from 1, the more curvature there is in the plot of the An ’s. (If the
fraction equals 1, then each An is the average of its two neighbors, so we just have a straight line.) And since
it’s a good bet that we’re going to get some sort of sinusoidal result out of all this, it’s not an outrageous
thing to define this fraction to be a sinusoidal function of a new quantity θ. But in the end, it does come a
bit out of the blue. That’s the way it is sometimes. However, you will find it less mysterious after reading
Section 2.4, where we actually end up with a true second derivative, along with sinusoidal functions of x
(the analog of n here).
4 If ω > 2ω , then we have a so-called evanescent wave. We’ll discuss these in Chapter 6. The ω = 0 and
0
ω = 2ω0 cases are somewhat special; see Problem [to be added].
2.3. N MASSES 13

If we solve for An+1 in Eq. (47) and use the inductive hypothesis that the An = B cos nθ+
C sin nθ result holds for n − 1 and n, we have
An+1 = (2 cos θ)An − An−1
³ ´ ³ ´
= 2 cos θ B cos nθ + C sin nθ − B cos(n − 1)θ + C sin(n − 1)θ
³ ´
= B 2 cos nθ cos θ − (cos nθ cos θ + sin nθ sin θ)
³ ´
+C 2 sin nθ cos θ − (sin nθ cos θ − cos nθ sin θ)
³ ´ ³ ´
= B cos nθ cos θ − sin nθ sin θ) + C sin nθ cos θ + cos nθ sin θ)
= B cos(n + 1)θ + C sin(n + 1)θ, (51)
which is the desired expression for the case of n + 1. (Note that this works independently
for the B and C terms.) Therefore, since the An = B cos nθ + C sin nθ result holds for n = 0
and n = 1, and since the inductive step is valid, the result therefore holds for all n.
If you wanted, you could have instead solved for An−1 in Eq. (51) and demonstrated
that the inductive step works in the negative direction too. Therefore, starting with two
arbitrary masses anywhere in the line, the An = B cos nθ + C sin nθ result holds even for an
infinite number of masses extending in both directions.

This claim tells us that we have found a solution of the form,

xn (t) = An eiωt = (B cos nθ + C sin nθ)eiωt . (52)
However, with the convention that ω is positive, we must remember that an e−iωt solution
works just as well. So another solution is
xn (t) = An e−iωt = (D cos nθ + E sin nθ)e−iωt . (53)
Note that the coefficients in this solution need not be the same as those in the eiωt solution.
Since the F = ma equations in Eq. (42) are all linear, the sum of two solutions is again a
solution. So the most general solution (for a given value of ω) is the sum of the above two
solutions (each of which is itself a linear combination of two solutions).
As usual, we now invoke the fact that the positions must be real. This implies that the
above two solutions must be complex conjugates of each other. And since this must be true
for all values of n, we see that B and D must be complex conjugates, and likewise for C
and E. Let’s define B = D∗ ≡ (F/2)eiβ and C = E ∗ ≡ (G/2)eiγ . There is no reason why
B, C, D, and E (or equivalently the A’s in Eq. (44)) have to be real. The sum of the two
solutions then becomes
xn (t) = F cos nθ cos(ωt + β) + G sin nθ cos(ωt + γ) (54)
As usual, we could have just taken the real part of either of the solutions to obtain this (up
to a factor of 2, which we can absorb into the definition of the constants). We can make it
look a little more symmetrical be using the trig sum formula for the cosines. This gives the
result (we’re running out of letters, so we’ll use Ci ’s for the coefficients here),

xn (t) = C1 cos nθ cos ωt + C2 cos nθ sin ωt + C3 sin nθ cos ωt + C4 sin nθ sin ωt (55)

where θ is determined by ω via Eq. (48), which we can write in the form,
µ ¶
2ω02 − ω 2
θ ≡ cos−1 (56)
2ω02
14 CHAPTER 2. NORMAL MODES

The constants C1 , C2 , C3 , C4 in Eq. (55) are related to the constants F , G, β, γ in Eq. (54)
in the usual way (C1 = F cos β, etc.). There are yet other ways to write the solution, but
we’ll save the discussion of these for Section 2.4.
Eq. (55) is the most general form of the positions for the mode that has frequency ω.
This set of the xn (t) functions (N of them) satisfies the F = ma equations in Eq. (42) (N
of them) for any values of C1 , C2 , C3 , C4 . These four constants are determined by four
initial values, for example, x0 (0), ẋ0 (0), x1 (0), and ẋ1 (0). Of course, if n = 0 corresponds
to a fixed wall, then the first two of these are zero.
Remarks:
1. Interestingly, we have found that xn (t) varies sinusoidally with position (that is, with n), as
well as with time. However, whereas time takes on a continuous set of values, the position is
relevant only at the discrete locations of the masses. For example, if the equilibrium positions
are at the locations z = na, where a is the equilibrium spacing between the masses, then we
can rewrite xn (t) in terms of z instead of n, using n = z/a. Assuming for simplicity that we
have only, say, the C1 cos nθ cos ωt part of the solution, we have
xn (t) =⇒ xz (t) = C1 cos(zθ/a) cos ωt. (57)
For a given values of θ (which is related to ω) and a, this is a sinusoidal function of z (as
well as of t). But we must remember that it is defined only at the discrete values of z of the
form, z = na. We’ll draw some nice pictures below to demonstrate the sinusoidal behavior,
when we discuss a few specific values of N .
2. We should stress the distinction between z (or equivalently n) and x. z represents the
equilibrium positions of the masses. A given mass is associated with a unique value of z. z
doesn’t change as the mass moves. xz (t), on the other hand, measures the position of a mass
(the one whose equilibrium position is z) relative to its equilibrium position (namely z). So
the total position of a given mass is z + x. The function xz (t) has dependence on both z
and t, so we could very well write it as a function of two variables, x(z, t). We will in fact
adopt this notation in Section 2.4 when we talk about continuous systems. But in the present
case where z can take on only discrete values, we’ll stick with the xz (t) notation. But either
notation is fine.
3. Eq. (55) gives the most general solution for a given value of ω, that is, for a given mode. While
the most general motion of the masses is certainly not determined by x0 (0), ẋ0 (0), x1 (0), and
ẋ1 (0), the motion for a single mode is. Let’s see why this is true. If we apply the x0 (0) and
x1 (0) boundary conditions to Eq. (55), we obtain x0 (0) = C1 and x1 (0) = C1 cos θ + C3 sin θ.
Since we are assuming that ω (and hence θ) is given, these two equations determine C1 and
C3 . But C1 and C3 in turn determine all the other xn (0) values via Eq. (55), because the
sin ωt terms are all zero at t = 0. So for a given mode, x0 (0) and x1 (0) determine all the
other initial positions. In a similar manner, the ẋ0 (0) and ẋ1 (0) values determine C2 and
C4 , which in turn determine all the other initial velocities. Therefore, since the four values
x0 (0), ẋ0 (0), x1 (0), and ẋ1 (0) give us all the initial positions and velocities, and since the
accelerations depend on the positions (from the F = ma equations in Eq. (42)), the future
motion of all the masses is determined. ♣

2.3.2 Wall boundary conditions

Let us now see what happens when we invoke the boundary conditions due to fixed walls
at the two ends. The boundary conditions at the walls are x0 (t) = xN +1 (t) = 0, for all t.
These conditions are most easily applied to xn (t) written in the form in Eq. (54), although
the form in Eq. (55) will work fine too. At the left wall, the x0 (t) = 0 condition gives
0 = x0 (t) = F cos(0) cos(ωt + β) + G sin(0) cos(ωt + γ)
= F cos(ωt + β). (58)
2.3. N MASSES 15

If this is to be true for all t, we must have F = 0. So we’re left with just the G sin nθ cos(ωt+
γ) term in Eq. (54). Applying the xN +1 (t) = 0 condition to this then gives

0 = xN +1 (t) = G sin(N + 1)θ cos(ωt + γ). (59)

One way for this to be true for all t is to have G = 0. But then all the x’s are identically
zero, which means that we have no motion at all. The other (nontrivial) way for this to be
true is to have the sin(N + 1)θ factor be zero. This occurs when
mπ
(N + 1)θ = mπ =⇒ θ = , (60)
N +1
where m is an integer. The solution for xn (t) is therefore
µ ¶
nmπ
xn (t) = G sin cos(ωt + γ) (61)
N +1

The amplitudes of the masses are then

µ ¶
nmπ
An = G sin (62)
N +1

We’ve made a slight change in notation here. The An that we’re now using for the amplitude
is the magnitude of the An that we used in Eq. (44). That An was equal to B cos nθ+C sin nθ,
which itself is some complex number which can be written in the form, |An |eiα . The solution
for xn (t) is obtained by taking the real part of Eq. (52), which yields xn (t) = |An | cos(ωt+α).
So we’re now using An to stand for |An |, lest we get tired if writing the absolute value bars
over and over.5 And α happens to equal the γ in Eq. (61).
If we invert the definition of θ in Eq. (48) to solve for ω in terms of θ, we find that the
frequency is given by
2ω02 − ω 2
2 cos θ ≡ =⇒ ω 2 = 2ω02 (1 − cos θ)
ω02
= 4ω02 sin2 (θ/2)
µ ¶
mπ
=⇒ ω = 2ω0 sin (63)
2(N + 1)

We’ve taken the positive square root, because the convention is that ω is positive. We’ll
see below that m labels the normal mode (so the “m” stands for “mode”). If m = 0 or
m = N + 1, then Eq. (61) says that all the xn ’s are identically zero, which means that we
don’t have any motion at all. So only m values in the range 1 ≤ m ≤ N are relevant. We’ll
see below in the “Nyquist” subsection that higher values of m simply give repetitions of the
modes generated by the 1 ≤ m ≤ N values.
The most important point in the above results is that if the boundary conditions are
walls at both ends, then the θ in Eq. (60), and hence the ω in Eq. (63), can take on only a
certain set of discrete values. This is consistent with our results for the N = 2 and N = 3
cases in Sections 2.1 and 2.2, where we found that there were only two or three (respectively)
allowed values of ω, that is, only two or three normal modes. Let’s now show that for N = 2
and N = 3, the preceding equations quantitatively reproduce the results from Sections 2.1
and 2.2. You can examine higher values of N in Problem [to be added].
5 If instead of taking the real part, you did the nearly equivalent thing of adding on the complex conjugate

solution in Eq. (53), then 2|An | would be the amplitude. In this case, the An in Eq. (62) stands for 2|An |.
 16 CHAPTER 2. NORMAL MODES

The N = 2 case
If N = 2, there are two possible values of m:
• m = 1: Eqs. (62) and (63) give
³ nπ ´ ³π´
An ∝ sin , and ω = 2ω0 sin . (64)
3 6
So this mode is given by
µ ¶ µ ¶ µ ¶
A1 sin(π/3) 1
∝ ∝ , and ω = ω0 . (65)
A2 sin(2π/3) 1

These agree with the first mode we found in Section 2.1.2. The frequency is ω0 , and
the masses move in phase with each other.
• m = 2: Eqs. (62) and (63) give
³ 2nπ ´ ³π´
An ∝ sin , and ω = 2ω0 sin . (66)
3 3
So this mode is given by
µ ¶ µ ¶ µ ¶
A1 sin(2π/3) 1 √
∝ ∝ , and ω= 3 ω0 . (67)
A2 sin(4π/3) −1
√
These agree with the second mode we found in Section 2.1.2. The frequency is 3 ω0 ,
and the masses move exactly out of phase with each other.
To recap, the various parameters are: N (the number of masses), m (the mode number),
and n (the label of each of the N masses). n runs from 1 to N , of course. And m effectively
also runs from 1 to N (there are N possible modes for N masses). We say “effectively”
because as we mentioned above, although m can technically take on any integer value, the
values that lie outside the 1 ≤ m ≤ N range give duplications of the modes inside this
range. See the “Nyquist” subsection below.
In applying Eqs. (62) and (63), things can get a little confusing because of all the
parameters floating around. And this is just the simple case of N = 2. Fortunately, there
(N = 2) is an extremely useful graphical way to see what’s going on. This is one situation where a
picture is indeed worth a thousand words (or equations).
If we write the argument of the sin in Eq. (62) as mπ · n/(N + 1), then we see that
for a given N , the relative amplitudes of the masses in the mth mode are obtained by
m=1 drawing a sin curve with m half oscillations, and then finding the value of this curve at
equal “1/(N + 1)” intervals along the horizontal axis. Fig. 8 shows the results for N = 2.
We’ve drawn either m = 1 or m = 2 half oscillations, and we’ve divided each horizontal
axis into N + 1 = 3 equal intervals. These curves look a lot like snapshots of beads on a
string oscillating transversely back and forth. And indeed, we will find in Chapter 4 that
the F = ma equations for transverse motion of beads on a string are exactly the same as
m=2 the equations in Eq. (42) for the longitudinal motion of the spring/mass system. But for
now, all of the displacements indicated in these pictures are in the longitudinal direction.
Figure 8 And the displacements have meaning only at the discrete locations of the masses. There
isn’t anything actually happening at the rest of the points on the curve.
2.3. N MASSES 17 m=2

We can also easily visualize what the frequencies are. If we write the argument of the
m=1
sin in Eq. (63) as π/2 · m/(N + 1) then we see that for a given N , the frequency of the 2ω0
mth mode is obtained by breaking a quarter circle (with radius 2ω0 ) into “1/(N + 1)” equal
intervals, and then finding the y values of the resulting points. Fig. 9 shows the results for
N = 2. We’ve divided the quarter circle into N + 1 = 3 equal angles of π/6, which results
in points at the angles of π/6 and π/3. It is much easier to see what’s going on by looking
at the pictures in Figs. 8 and 9 than by working with the algebraic expressions in Eqs. (62) (N = 2)
and (63).
Figure 9
The N = 3 case
If N = 3, there are three possible values of m:

• m = 1: Eqs. (62) and (63) give

³ nπ ´ ³π ´
An ∝ sin , and ω = 2ω0 sin . (68)
4 8
So this mode is given by
     
q
A1 sin(π/4) √1 √
 A2  ∝  sin(2π/4)  ∝  2  , and ω= 2− 2 ω0 , (69)
A3 sin(3π/4) −1
(N = 3)
where we have used the half-angle formula for sin(π/8) to obtain ω. (Or equivalently,
we just used the first line in Eq. (63).) These results agree with the “slow” mode we
found in Section 2.2.
m=1
• m = 2: Eqs. (62) and (63) give
³ 2nπ ´ ³ 2π ´
An ∝ sin , and ω = 2ω0 sin . (70)
4 8
So this mode is given by m=2
     
A1 sin(2π/4) 1 √
 A2  ∝  sin(4π/4)  ∝  0  , and ω= 2 ω0 . (71)
A3 sin(6π/4) −1

These agree with the “medium” mode we found in Section 2.2.

m=3
• m = 3: Eqs. (62) and (63) give
Figure 10
³ 3nπ ´ ³ 3π ´
An ∝ sin , and ω = 2ω0 sin . (72)
4 8
So this mode is given by m=3
      m=2
A1 sin(3π/4) 1 q
√ √
 A2  ∝  sin(6π/4)  ∝  − 2  , and ω= 2+ 2 ω0 . (73)
A3 sin(9π/4) 1 2ω0 m=1

These agree with the “fast” mode we found in Section 2.2..

As with the N = 2 case, it’s much easier to see what’s going on if we draw some pictures.
Fig.10shows the relative amplitudes within the three modes, and Fig.11shows the associated (N = 3)

Figure 11
 18 CHAPTER 2. NORMAL MODES

frequencies. Each horizontal axis in Fig. 10 is broken up into N + 1 = 4 equal segments,

and the quarter circle in Fig. 11 is broken up into N + 1 = 4 equal arcs.
As mentioned above, although Fig. 10 looks like transverse motion on a string, remember
that all the displacements indicated in this figure are in the longitudinal direction. For
example, in the first m = 1 mode, all three√ masses move in the same direction, but the
middle one moves farther (by a factor of 2) than the outer ones. Problem [to be added]
discusses higher values of N .

Aliasing, Nyquist frequency

Consider the N = 6 case shown in Fig. 12. Assuming that this corresponds to a normal
Figure 12 mode, which one is it? If you start with m = 1 and keep trying different values of m until
the (appropriately scaled) sin curve in Eq. (62) matches up with the six points in Fig. 12
(corresponding to n values from 1 to N ), you’ll find that m = 3 works, as shown in Fig. 13.
However, the question then arises as to whether this is the only value of m that allows
the sin curve to fit the given points. If you try some higher values of m, and if you’re
persistent, then you’ll find that m = 11 also works, as long as you throw in a minus sign
in front of the sin curve (this corresponds to the G coefficient in Eq. (61) being negative,
which is fine). This is shown in Fig. 14. And if you keep going with higher m values, you’ll
Figure 13 find that m = 17 works too, and this time there’s no need for a minus sign. This is shown
in Fig. 15. Note that the m value always equals the number of bumps (local maxima or
minima) in the sin curve.
(N = 6) It turns out that there is an infinite number of m values that work, and they fall into
two classes. If we start with particular values of N and m (6 and 3 here), then all m0 values
of the form,
m0 = 2a(N + 1) − m, (74)
also work (with a minus sign in the sin curve), where a is any integer. And all m0 values of
the form,
black: m = 3
m0 = 2a(N + 1) + m, (75)
gray: m' = 2(N+1) - m = 11
also work, where again a is any integer. You can verify these claims by using Eq. (62);
Figure 14 see Problem [to be added]. For N = 6 and m = 3, the first of these classes contains
m0 = 11, 25, 39, . . ., and the second class contains m0 = 3, 17, 31, 45, . . .. Negative values of
(N = 6) a work too, but they simply reproduce the sin curves for these m0 values, up to an overall
minus sign.
You can show using Eq. (63) that the frequencies corresponding to all of these values of
m0 (in both classes) are equal; see Problem [to be added] (a frequency of −ω yields the same
motion as a frequency of ω). So as far as the motions of the six masses go, all of these modes
yield exactly the same motions of the masses. (The other parts of the various sin curves
don’t match up, but all we care about are the locations of the masses.) It is impossible to
black: m = 3 tell which mode the masses are in. Or said more accurately, the masses aren’t really in any
gray: m' = 2(N+1) + m = 17 one particular mode. There isn’t one “correct” mode. Any of the above m or m0 values is
just as good as any other. However, by convention we label the mode with the m value in
Figure 15 the range 1 ≤ m ≤ N .
The above discussion pertains to a setup with N discrete masses in a line, with massless
springs between them. However, if we have a continuous string/mass system, or in other
words a massive spring (we’ll talk about such a system in Section 2.4), then the different
m0 values do represent physically different motions. The m = 3 and m = 17 curves in Fig.
15 are certainly different. You can think of a continuous system as a setup with N −→ ∞
masses, so all the m values in the range 1 ≤ m ≤ N =⇒ 1 ≤ m ≤ ∞ yield different modes.
In other words, each value of m yields a different mode.
2.4. N → ∞ AND THE WAVE EQUATION 19

However, if we have a continuous string, and if we only look at what is happening at

equally spaced locations along it, then there is no way to tell what mode the string is really
in (and in this case it really is in a well defined mode). If the string is in the m = 11 mode,
and if you only look at the six equally-spaced points we considered above, then you won’t
be able to tell which of the m = 3, 11, 17, 25, . . . modes is the correct one.
This ambiguity is known as aliasing, or the nyquist effect. If you look at only discrete
points in space, then you can’t tell the true spatial frequency. Or similarly, If you look at
only discrete moments in time, then you can’t tell the true temporal frequency. This effect
manifests itself in many ways in the real world. If you watch a car traveling by under a
streetlight (which emits light in quick pulses, unlike an ordinary filament lightbulb), or if
you watch a car speed by in a movie (which was filmed at a certain number of frames per
second), then the “spokes” on the tires often appear to be moving at a different angular rate
than the actual angular rate of the tire. They might even appear to be moving backwards.
This is called the “strobe” effect. There are also countless instances of aliasing in electronics.

2.4 N → ∞ and the wave equation

Let’s now consider the N → ∞ limit of our mass/spring setup. This means that we’ll now
effectively have a continuous system. This will actually make the problem easier than the
finite-N case in the previous section, and we’ll be able to use a quicker method to solve it. If
you want, you can use the strategy of taking the N → ∞ limit of the results in the previous
section. This method will work (see Problem [to be added]), but we’ll derive things from
scratch here, because the method we will use is a very important one, and it will come up
again in our study of waves.
First, a change of notation. The equilibrium position of each mass will now play a more
fundamental role and appear more regularly, so we’re going to label it with x instead of n
(or instead of the z we used in the first two remarks at the end of Section 2.3.1). So xn is the
equilibrium position of the nth mass (we’ll eventually drop the subscript n). We now need
a new letter for the displacement of the masses, because we used xn for this above. We’ll
use ξ now. So ξn is the displacement of the nth mass. The x’s are constants (they just label
the equilibrium positions, which don’t change), and the ξ’s are the things that change with
time. The actual location of the nth mass is xn + ξn , but only the ξn part will show up in
the F = ma equations, because the xn terms don’t contribute to the acceleration (because
they are constant), nor do they contribute to the force (because only the displacement from
equilibrium matters, since the spring force is linear).
Instead of the ξn notation, we’ll use ξ(xn ). And we’ll soon drop the subscript n and just
write ξ(x). All three of the ξn , ξ(xn ), ξ(x) expressions stand for the same thing, namely
the displacement from equilibrium of the mass whose equilibrium position is x (and whose
numerical label is n). ξ is a function of t too, of course, but we won’t bother writing the t
dependence yet. But eventually we’ll write the displacement as ξ(x, t).
Let ∆x ≡ xn − xn−1 be the (equal) spacing between the equilibrium positions of all the
masses. The xn values don’t change with time, so neither does ∆x. If the n = 0 mass is
located at the origin, then the other masses are located at positions xn = n∆x. In our new
notation, the F = ma equation in Eq. (42) becomes
mξ¨n = kξn−1 − 2kξn + kξn+1
¨ n) =
=⇒ mξ(x kξ(xn − ∆x) + 2kξ(xn ) + kξ(xn + ∆x)
¨
=⇒ mξ(x) = kξ(x − ∆x) + 2kξ(x) + kξ(x + ∆x). (76)
In going from the second to the third line, we are able to drop the subscript n because the
value of x uniquely determines which mass we’re looking at. If we ever care to know the
20 CHAPTER 2. NORMAL MODES

value of n, we can find it via xn = n∆x =⇒ n = x/∆x. Although the third line holds only
for x values that are integral multiples of ∆x, we will soon take the ∆x → 0 limit, in which
case the equation holds for essentially all x.
We will now gradually transform Eq. (76) into a very nice result, which is called the
wave equation. The first step actually involves going backward to the F = ma form in Eq.
(41). We have

d2 ξ(x) h³ ´ ³ ´i
m = k ξ(x + ∆x) − ξ(x) − ξ(x) − ξ(x − ∆x)
dt2 Ã ξ(x+∆x)−ξ(x) ξ(x)−ξ(x−∆x) !
2 −
m d ξ(x) ∆x ∆x
=⇒ = k∆x . (77)
∆x dt2 ∆x

We have made these judicious divisions by ∆x for the following reason. If we let ∆x → 0
(which is indeed the case if we have N → ∞ masses in the system), then we can use
the definitions of the first and second derivatives to obtain (with primes denoting spatial
derivatives)6

m d2 ξ(x) ξ 0 (x) − ξ 0 (x − ∆x)

= (k∆x)
∆x dt2 ∆x
= (k∆x)ξ 00 (x). (78)

But m/∆x is the mass density ρ. And k∆x is known as the elastic modulus, E, which
happens to have the units of force. So we obtain
d2 ξ(x)
ρ = Eξ 00 (x). (79)
dt2
Note that E ≡ k∆x is a reasonable quantity to appear here, because the spring constant
k for an infinitely small piece of spring is infinitely large (because if you cut a spring in
half, its k doubles, etc.). The ∆x in the product k∆x has the effect of yielding a finite and
informative quantity. If various people have various lengths of springs made out of a given
material, then these springs have different k values, but they all have the same E value.
Basically, if you buy a spring in a store, and if it’s cut from a large supply on a big spool,
then the spool should be labeled with the E value, because E is a property of the material
and independent of the length. k depends on the length.
Since ξ is actually a function of both x and t, let’s be explicit and write Eq. (79) as

∂ 2 ξ(x, t) ∂ 2 ξ(x, t)
ρ 2
=E (wave equation) (80)
∂t ∂x2

This is called the wave equation. This equation (or analogous equations for other systems)
will appear repeatedly throughout this book. Note that the derivatives are now written as
partial derivatives, because ξ is a function of two arguments. Up to the factors of ρ and E,
the wave equation is symmetric in x and t.
The second time derivative on the lefthand side of Eq. (80) comes from the “a” in
F = ma. The second space derivative on the righthand side comes from the fact that it
is the differences in the lengths of two springs that yields the net force, and each of these
lengths is itself the difference of the positions of two masses. So it is the difference of the
differences that we’re concerned with. In other words, the second derivative.
6 There is a slight ambiguity here. Is the (ξ(x+∆x)−ξ(x))∆x term in Eq. (77) equal to ξ 0 (x) or ξ 0 (x+∆x)?

Or perhaps ξ 0 (x + ∆x/2)? It doesn’t matter which we pick, as long as we use the same convention for the
(ξ(x) − ξ(x − ∆x))∆x term. The point is that Eq. (78) contains the first derivatives at two points (whatever
they may be) that differ by ∆x, and the difference of these yields the second derivative.
2.4. N → ∞ AND THE WAVE EQUATION 21

How do we solve the wave equation? Recall that in the finite-N case, the strategy was
to guess a solution of the form (using ξ now instead of x),
   
.. ..
 .   . 
 ξn−1   an−1 
   
 ξn  =  an  eiωt . (81)
   
 ξn+1   an+1 
   
.. ..
. .

If we relabel ξn → ξ(xn , t) → ξ(x, t), and an → a(xn ) → a(x), we can write the guess in the
more compact form,
ξ(x, t) = a(x)eiωt . (82)
This is actually an infinite number of equations (one for each x), just as Eq. (81) is an
infinite number of equations (one for each n). The a(x) function gives the amplitudes of the
masses, just as the original normal mode vector (A1 , A2 , A3 , . . .) did. If you want, you can
think of a(x) as an infinite-component vector.
Plugging this expression for ξ(x, t) into the wave equation, Eq. (80), gives

∂2 ¡ ¢ ∂2 ¡ ¢
ρ 2
a(x)eiωt = 2
Ea(x)eiωt
∂t ∂x
d2
=⇒ −ω 2 ρ a(x) = E 2 a(x)
dx
d2 ω2 ρ
=⇒ a(x) = − a(x). (83)
dx2 E
But this is our good ol’ simple-harmonic-oscillator equation, so the solution is
r
ρ
a(x) = Ae±ikx where k≡ω (84)
E

k is called the wave number. It is usually defined to be a positive number, so we’ve put in
the ± by hand. Unfortunately, we’ve already been using k as the spring constant, but there
are only so many letters! The context (and units) should make it clear which way we’re
using k. The wave number k has units of
s s
[ρ] 1 kg/m 1
[k] = [ω] = = . (85)
[E] s kg m/s2 m

So kx is dimensionless, as it should be, because it appears in the exponent in Eq. (84).

What is the physical meaning of k? If λ is the wavelength, then the kx exponent in Eq.
(84) increases by 2π whenever x increases by λ. So we have
2π
kλ = 2π =⇒ k= . (86)
λ
If k were just equal to 1/λ, then it would equal the number of wavelengths (that is, the
number of spatial oscillations) that fit into a unit length. With the 2π, it instead equals the
number of radians of spatial oscillations that fit into a unit length.
Using Eq. (84), our solution for ξ(x, t) in Eq. (82) becomes

ξ(x, t) = a(x)eiωt = Aei(±kx+ωt) . (87)

22 CHAPTER 2. NORMAL MODES

As usual, we could have done all this with an e−iωt term in Eq. (81), because only the
square of ω came into play (ω is generally assumed to be positive). So we really have the
four different solutions,
ξ(x, t) = Aei(±kx±ωt) . (88)
The most general solution is the sum of these, which gives

ξ(x, t) = A1 ei(kx+ωt) + A∗1 ei(−kx−ωt) + A2 ei(kx−ωt) + A∗2 ei(−kx+ωt) , (89)

where the complex conjugates appear because ξ must be real. There are many ways to
rewrite this expression in terms of trig functions. Depending on the situation you’re dealing
with, one form is usually easier to deal with than the others, but they’re all usable in theory.
Let’s list them out and discuss them. In the end, each form has four free parameters. We
saw above in the third remark at the end of Section 2.3.1 why four was the necessary number
in the discrete case, but we’ll talk more about this below.

• If we let A1 ≡ (B1 /2)eiφ1 and A2 ≡ (B2 /2)eiφ2 in Eq. (89), then the imaginary parts
of the exponentials cancel, and we end up with

ξ(x, t) = B1 cos(kx + ωt + φ1 ) + B2 cos(kx − ωt + φ2 ) (90)

The interpretation of these two terms is that they represent traveling waves. The
first one moves to the left, and the second one moves to the right. We’ll talk about
traveling waves below.

• If we use the trig sum formulas to expand the previous expression, we obtain

ξ(x, t) = C1 cos(kx + ωt) + C2 sin(kx + ωt) + C3 cos(kx − ωt) + C4 sin(kx − ωt)

(91)
where C1 = B1 cos φ1 , etc. This form has the same interpretation of traveling waves.
The sines and cosines are simply 90◦ out of phase.

• If we use the trig sum formulas again and expand the previous expression, we obtain

ξ(x, t) = D1 cos kx cos ωt + D2 sin kx sin ωt + D3 sin kx cos ωt + D4 cos kx sin ωt

(92)
where D1 = C1 + C3 , etc. These four terms are four different standing waves. For each
term, the masses all oscillate in phase. All the masses reach their maximum position
at the same time (the cos ωt terms at one time, and the sin ωt terms at another), and
they all pass through zero at the same time. As a function of time, the plot of each
term just grows and shrinks as a whole. The equality of Eqs. (91) and (92) implies
that any traveling wave can be written as the sum of standing waves, and vice versa.
This isn’t terribly obvious; we’ll talk about it below.

• If we collect the cos ωt terms together in the previous expression, and likewise for the
sin ωt terms, we obtain

ξ(x, t) = E1 cos(kx + β1 ) cos ωt + E2 cos(kx + β2 ) sin ωt (93)

where E1 cos β1 = D1 , etc. This form represents standing waves (the cos ωt one is 90◦
ahead of the sin ωt one in time), but they’re shifted along the x axis due to the β
phases. The spatial functions here could just as well be written in terms of sines, or
one sine and one cosine. This would simply change the phases by π/2.
2.4. N → ∞ AND THE WAVE EQUATION 23

• If we collect the cos kx terms together in Eq. (92) and likewise for the sin kx terms,
we obtain
ξ(x, t) = F1 cos(ωt + γ1 ) cos kx + F2 cos(ωt + γ2 ) sin kx (94)
where F1 cos γ1 = D1 , etc. This form represents standing waves, but they’re not 90◦
separated in time in this case, due to the γ phases. They are, however, separated by
90◦ (a quarter wavelength) in space. The time functions here could just as well be
written in terms of sines.

Remarks:
1. If there are no walls and the system extends infinitely in both directions (actually, infinite
extent in just one directionp
is sufficient), then ω can take on any value. Eq. (84) then says that
k is related to ω via k = ω ρ/E. We’ll look at the various effects of boundary conditions in
Chapter 4.
2. The fact that each of the above forms requires four parameters is probably most easily
understood by looking at the first form given in Eq. (90). The most general wave with a
given frequency ω consists of two oppositely-traveling waves, each of which is described by
two parameters (magnitude and phase). So two times two is four.
You will recall that for each of the modes in the N = 2 and N = 3 cases we discussed earlier
(and any other value of N , too), only two parameters were required: an overall factor in the
amplitudes, and a phase in time. Why only two there, but four now? The difference is due
to the fact that we had walls in the earlier cases, but no walls now. (The difference is not
due to the fact that we’re now dealing with infinite N .) The effect of the walls (actually, only
one wall is needed) is most easily seen by working with the form given in Eq. (92). Assuming
that one wall is located at x = 0, we see that the two cos kx terms can’t be present, because
the displacement must always be zero at x = 0. So D1 = D4 = 0, and we’re down to two
parameters. We’ll have much more to say about such matters in Chapter 4.
3. Remember that the above expressions for ξ(x, t), each of which contains four parameters,
represent the general solution for a given mode with frequency ω. If the system is undergoing
arbitrary motion, then it is undoubtedly in a linear combination of many different modes,
perhaps even an infinite number. So four parameters certainly don’t determine the system.
We need four times the number of modes, which might be infinite. ♣

Traveling waves
Consider one of the terms in Eq. (91), say, the cos(kx − ωt) one. Let’s draw the plot of
cos(kx − ωt), as a function of x, at two times separated by ∆t. If we arbitrarily take the
lesser time to be t = 0, the result is shown in Fig. 16. Basically, the left curve is a plot of
cos kx, and the right curve is a plot of cos(kx − φ), where φ happens to be ω∆t. It is shifted
to the right because it takes a larger value of x to obtain the same phase.

ω ∆t plots of cos(kx-ωt)
k
1.0
curve at t = 0

0.5

x
2 4 6 8

- 0.5
curve at t = ∆t
- 1.0
24 CHAPTER 2. NORMAL MODES

Figure 16

What is the horizontal shift between the curves? We can answer this by finding the
distance between the maxima, which are achieved when the argument kx − ωt equals zero
(or a multiple of 2π). If t = 0, then we have kx − ω · 0 = 0 =⇒ x = 0. And if t = ∆t, then
we have kx − ω · ∆t = 0 =⇒ x = (ω/k)∆t. So (ω/k)∆t is the horizontal shift. It takes a
time of ∆t for the wave to cover this distance, so the velocity of the wave is

(ω/k)∆t ω
v= =⇒ v= (95)
∆t k

Likewise for the sin(kx − ωt) function in Eq. (91). Similarly, the velocity of the cos(kx + ωt)
and sin(kx + ωt) curves is −ω/k.
We see that the wave cos(kx − ωt) keeps its shape and travels along at speed ω/k. Hence
the name “traveling wave.” But note that none of the masses are actually moving with this
speed. In fact, in our usual approximation of small amplitudes, the actual velocities of the
masses are very small. If we double the amplitudes, then the velocities of the masses are
doubled, but the speed of the waves is still ω/k.
As we discussed right after Eq. (92), the terms in that equation are standing waves.
They don’t travel anywhere; they just expand and contract in place. All the masses reach
their maximum position at the same time, and they all pass through zero at the same
time. This is certainly not the case with a traveling wave. Trig identities of the sort,
cos(kx − ωt) = cos kx cos ωt + sin kx sin ωt, imply that any traveling wave can be ¡written as
the sum of two standing
¢ waves. And trig identities of the sort, cos kx cos ωt = cos(kx −
ωt) + cos(kx + ωt) /2, imply that any standing wave can be written as the sum of two
opposite traveling waves. The latter of these facts is reasonably easy to visualize, but the
former is trickier. You should convince yourself that it works.

A more general solution

We’ll now present a much quicker method of finding a much more general solution (compared
with our sinusoidal solutions above) to the wave equation in Eq. (80). This is a win-win
combination. p p
From Eq. (84), we know that k = ω ρ/E. Combining this with Eq. (95) gives E/ρ =
ω/k = v =⇒ E/ρ = v 2 . In view of this relation, if we divide the wave equation in Eq. (80)
by ρ, we see that it can be written as

∂ 2 ξ(x, t) 2
2 ∂ ξ(x, t)
= v (96)
∂t2 ∂x2

Consider now the function f (x − vt), where f is an arbitrary function of its argument.
(The function f (x + vt) will work just as well.) There is no need for f to even vaguely
resemble a sinusoidal function. What happens if we plug ξ(x, t) ≡ f (x − vt) into Eq. (96)?
Does it satisfy the equation? Indeed it does, as we can see by using the chain rule. In
what follows, we’ll use the notation f 00 to denote the second derivative of f . In other words,
f 00 (x − vt) equals d2 f (z)/dz 2 evaluated at z = x − vt. (Since f is a function of only one
variable, there is no need for any partial derivatives.) Eq. (96) then becomes (using the
chain rule on the left, and also on the right in a trivial sense)

∂ 2 f (x − vt) ? ∂ 2 f (x − vt)
= v2
∂t2 ∂x2
?
⇐⇒ (−v)2 f 00 (x − vt) = v 2 · (1)2 f 00 (x − vt), (97)
2.4. N → ∞ AND THE WAVE EQUATION 25

which is indeed true.

There is a fairly easy way to see graphically why any function of the form f (x − vt)
satisfies the wave equation in Eq. (96). The function f (x −¡ vt) represents a wave moving
¢ to
the right at speed v. This is true because f (x0 − vt0 ) = f (x0 + v∆t) − v(t0 + ∆t) , which
says that if you increase t by ∆t, then you need to increase x by v∆t in order to obtain the
same value of f . This is exactly what happens if you take a curve and move it to the right
at speed v. This is basically the same reasoning that led to Eq. (95).
We now claim that any curve that moves to the right (or left) at speed v satisfies the
wave equation in Eq. (96). Consider a closeup view of the curve near a given point x0 , at
two nearby times separated by ∆t. The curve is essentially a straight line in the vicinity of
x0 , so we have the situation shown in Fig. 17.

-∆f v
f (x0-vt)
wave at t
wave at t + ∆t

v ∆t f (x0-v(t+∆t))
x0

Figure 17

The solid line shows the curve at some time t, and the dotted line shows it at time t+∆t.
The slope of the curve, which is by definition ∂f /∂x, equals the ratio of the lengths of the
legs in the right triangle shown. The length of the vertical leg equals the magnitude of the
change ∆f in the function. Since the change is negative here, the length is −∆f . But by
the definition of ∂f /∂t, the change is ∆f = (∂f /∂t)∆t. So the length of the vertical leg is
−(∂f /∂t)∆t. The length of the horizontal leg is v∆t, because the curve moves at speed v.
So the statement that ∂f /∂x equals the ratio of the lengths of the legs is

∂f −(∂f /∂t)∆t ∂f ∂f
= =⇒ = −v . (98)
∂x v∆t ∂t ∂x
If we had used the function f (x + vt), we would have obtained ∂f /∂t = v(∂f /∂x). Of
course, these results follow immediately from applying the chain rule to f (x ± vt). But it’s
nice to also see how they come about graphically.
Eq. (98) then implies the wave equation in Eq. (96), because if we take ∂/∂t of Eq. (98),
and use the fact that partial differentiation commutes (the order doesn’t matter), we obtain
µ ¶ µ ¶
∂ ∂f ∂ ∂f
= −v
∂t ∂t ∂t ∂x
2
µ ¶
∂ f ∂ ∂f
=⇒ = −v
∂t2 ∂x ∂t
µ ¶
∂ ∂f
= −v −v
∂x ∂x
2
∂ f
= v2 2 , (99)
∂x
where we have used Eq. (98) again to obtain the third line. This result agrees with Eq. (96),
as desired.
26 CHAPTER 2. NORMAL MODES

Another way of seeing why Eq. (98) implies Eq. (96) is to factor Eq. (96). Due to the
fact that partial differentiation commutes, we can rewrite Eq. (96) as
µ ¶µ ¶
∂ ∂ ∂ ∂
−v +v f = 0. (100)
∂t ∂x ∂t ∂x

We can switch the order of these “differential operators” in parentheses, so either of them
can be thought of acting on f first. Therefore, if either operator yields zero when applied
to f , then the lefthand side of the equation equals zero. In other words, if Eq. (98) is true
(with either a plus or a minus on the right side), then Eq. (96) is true.
We have therefore seen (in various ways) that any arbitrary function that takes the
form of f (x − vt) satisfies the wave equation. This seems too simple to be true. Why did
we go through the whole procedure above that involved guessing a solution of the form
ξ(x, t) = a(x)eiωt ? Well, that has always been our standard procedure, so the question we
should be asking is: Why does an arbitrary function f (x − vt) work?
Well, we gave a few reasons in Eqs. (97) and (98). But here’s another reason, one
that relates things back to our original sinusoidal solutions. f (x − vt) works because of a
combination of Fourier analysis and linearity. Fourier analysis says that any (reasonably
well-behaved) function can be written as the integral (or discrete sum, if the function is
periodic) of exponentials, or equivalently sines and cosines. That is,
Z ∞
f (z) = C(r)eirz dr. (101)
−∞

Don’t worry about the exact meaning of this; we’ll discuss it at great length in the following
chapter. But for now, you just need to know that any function f (z) can be considered to
be built up out of eirz exponential functions. The coefficient C(r) tells you how much of
the function comes from a eirz term with a particular value of r.
Let’s now pretend that we haven’t seen Eq. (97), but that we do know about Fourier
analysis. Given the result in Eq. (101), if someone gives us the function f (x − vt) out of the
blue, we can write it as Z∞
f (x − vt) = C(r)eir(x−vt) dr. (102)
−∞

But eir(x−vt) can be written as ei(kx−ωt) , where k ≡ r and ω ≡ rv. Since these values of
k and pω satisfy ω/k = v, and hence satisfy Eq. (84) (assuming that v has been chosen to
equal E/ρ), we know that all of these eir(x−vt) terms satisfy the wave equation, Eq. (80).
And since the wave equation is linear in ξ, it follows that any sum (or integral) of these
exponentials also satisfies the wave equation. Therefore, in view of Eq. (102), we see that
any arbitrary function f (x − vt) satisfies the wave equation. As stated above, both Fourier
analysis and linearity are essential in this result.
Fourier analysis plays an absolutely critical role in the study of waves. In fact, it is so
important that we’ll spend all of Chapter 3 on it. We’ll then return to our study of waves
in Chapter 4. We’ll pick up where we left off here.
Chapter 3

Fourier analysis
Copyright 2009 by David Morin, [email protected] (Version 1, November 28, 2009)
This file contains the Fourier-analysis chapter of a potential book on Waves, designed for college
sophomores.

Fourier analysis is the study of how general functions can be decomposed into trigonometric
or exponential functions with definite frequencies. There are two types of Fourier expansions:

• Fourier series: If a (reasonably well-behaved) function is periodic, then it can be

written as a discrete sum of trigonometric or exponential functions with specific fre-
quencies.

• Fourier transform: A general function that isn’t necessarily periodic (but that is still
reasonably well-behaved) can be written as a continuous integral of trigonometric or
exponential functions with a continuum of possible frequencies.

The reason why Fourier analysis is so important in physics is that many (although certainly
not all) of the differential equations that govern physical systems are linear, which implies
that the sum of two solutions is again a solution. Therefore, since Fourier analysis tells us
that any function can be written in terms of sinusoidal functions, we can limit our attention
to these functions when solving the differential equations. And then we can build up any
other function from these special ones. This is a very helpful strategy, because it is invariably
easier to deal with sinusoidal functions than general ones.
The outline of this chapter is as follows. We start off in Section 3.1 with Fourier trigono-
metric series and look at how any periodic function can be written as a discrete sum of sine
and cosine functions. Then, since anything that can be written in terms of trig functions can
also be written in terms of exponentials, we show in Section 3.2 how any periodic function
can be written as a discrete sum of exponentials. In Section 3.3, we move on to Fourier
transforms and show how an arbitrary (not necessarily periodic) function can be written
as a continuous integral of trig functions or exponentials. Some specific functions come up
often when Fourier analysis is applied to physics, so we discuss a few of these in Section 3.4.
One very common but somewhat odd function is the delta function, and this is the subject
of Section 3.5.
Section 3.6 deals with an interesting property of Fourier series near discontinuities called
the Gibbs phenomenon. This isn’t so critical for applications to physics, but it’s a very
interesting mathematical phenomenon. In Section 3.7 we discuss the conditions under which
a Fourier series actually converges to the function it is supposed to describe. Again, this
discussion is more just for mathematical interest, because the functions we deal with in

1
 2 CHAPTER 3. FOURIER ANALYSIS

physics are invariably well-enough behaved to prevent any issues with convergence. Finally,
in Section 3.8 we look at the relation between Fourier series and Fourier transforms. Using
the tools we develop in the chapter, we end up being able to derive Fourier’s theorem (which
says that any periodic function can be written as a discrete sum of sine and cosine functions)
from scratch, whereas we simply had to accept this on faith in Section 3.1.
To sum up, Sections 3.1 through 3.5 are very important for physics, while Sections 3.6
through 3.8 are more just for your amusement.

3.1 Fourier trigonometric series

Fourier’s theorem states that any (reasonably well-behaved) function can be written in terms
of trigonometric or exponential functions. We’ll eventually prove this theorem in Section
f(x) 3.8.3, but for now we’ll accept it without proof, so that we don’t get caught up in all the
details right at the start. But what we will do is derive what the coefficients of the sinusoidal
functions must be, under the assumption that any function can in fact be written in terms
x of them.
L Consider a function f (x) that is periodic on the interval 0 ≤ x ≤ L. An example is shown
in Fig. 1. Fourier’s theorem works even if f (x) isn’t continuous, although an interesting thing
Figure 1 happens at the discontinuities, which we’ll talk about in Section 3.6. Other conventions for
the interval are −L ≤ x ≤ L, or 0 ≤ x ≤ 1, or −π ≤ x ≤ π, etc. There are many different
conventions, but they all lead to the same general result in the end. If we assume 0 ≤ x ≤ L
periodicity, then Fourier’s theorem states that f (x) can be written as

∞ ·
X µ ¶ µ ¶¸
2πnx 2πnx
f (x) = a0 + an cos + bn sin (1)
n=1
L L

where the an and bn coefficients take on certain values that we will calculate below. This
expression is the Fourier trigonometric series for the function f (x). We could alternatively
not separate out the a0 term, and instead let the sum run from n = 0 to ∞, because
cos(0) = 1 and sin(0) = 0. But the normal convention is to isolate the a0 term.
With the 2π included in the arguments of the trig functions, the n = 1 terms have period
L, the n = 2 terms have period L/2, and so on. So for any integer n, an integral number
of oscillations fit into the period L. The expression in Eq. (1) therefore has a period of (at
most) L, which is a necessary requirement, of course, for it to equal the original periodic
function f (x). The period can be shorter than L if, say, only the even n’s have nonzero
coefficients (in which case the period is L/2). But it can’t be longer than L; the function
repeats at least as often as with period L.
We’re actually making two statements in Eq. (1). The first statement is that any periodic
function can be written this way. This is by no means obvious, and it is the part of the
theorem that we’re accepting here. The second statement is that the an and bn coefficients
take on particular values, assuming that the function f (x) can be written this way. It’s
reasonably straightforward to determine what these values are, in terms of f (x), and we’ll
do this below. But we’ll first need to discuss the concept of orthogonal functions.

Orthogonal functions
For given values of n and m, consider the integral,
Z L µ ¶ µ ¶
2πnx 2πmx
sin cos dx. (2)
0 L L
3.1. FOURIER TRIGONOMETRIC SERIES 3

Using the trig sum formulas, this can be written as

Z · µ ¶ µ ¶¸
1 L 2πx 2πx
sin (n + m) + sin (n − m) dx. (3)
2 0 L L
But this equals zero, because both of the terms in the integrand undergo an integral number
of complete oscillations over the interval from 0 to L, which means that the total area under
the curve is zero. (The one exception is the (n − m) term if n = m, but in that case the sine
is identically zero anyway.) Alternatively, you can simply evaluate the integrals; the results
involve cosines whose values are the same at the endpoints of integration. (But again, in
the case where n = m, the (n − m) term must be treated separately.)
The same kind of reasoning shows that the integral,
Z L µ ¶ µ ¶ Z · µ ¶ µ ¶¸
2πnx 2πmx 1 L 2πx 2πx
cos cos dx = cos (n + m) + cos (n − m) dx,
0 L L 2 0 L L
(4)
equals zero except in the special case where n = m. If n = m, the (n − m) term is identically
1, so the integral equals L/2. (Technically n = −m also yields a nonzero integral, but we’re
concerned only with positive n and m.) Likewise, the integral,
Z L µ ¶ µ ¶ Z · µ ¶ µ ¶¸
2πnx 2πmx 1 L 2πx 2πx
sin sin dx = − cos (n + m) + cos (n − m) dx,
0 L L 2 0 L L
(5)
equals zero except in the special case where n = m, in which case the integral equals L/2.
To sum up, we have demonstrated the following facts:
Z L µ ¶ µ ¶
2πnx 2πmx
sin cos dx = 0,
0 L L
Z L µ ¶ µ ¶
2πnx 2πmx L
cos cos dx = δnm ,
0 L L 2
Z L µ ¶ µ ¶
2πnx 2πmx L
sin sin dx = δnm , (6)
0 L L 2
where δnm (known as the Kronecker delta) is defined to be 1 if n = m, and zero otherwise.
In short, the integral of the product of any two of these trig functions is zero unless the two
functions are actually the same function. The fancy way of saying this is that all of these
functions are orthogonal.
We normally use the word “orthogonal” when we talk about vectors. Two vectors
are orthogonal if their inner product (also called the dot product) is zero, where the inner
product is defined to be the sum of the products of corresponding components of the vectors.
For example, in three dimensions we have

(ax , ay , az ) · (bx , by , bz ) = ax bx + ay by + az bz . (7)

For functions, if we define the inner product of the above sine and cosine functions to be the
integral of their product, then we can say that two functions are orthogonal if their inner
product is zero, just as with vectors. So our definition of the inner product of two functions,
f (x) and g(x), is Z
Inner product ≡ f (x)g(x) dx. (8)

This definition depends on the limits of integration, which can be chosen arbitrarily (we’re
taking them to be 0 and L). This definition is more than just cute terminology. The inner
4 CHAPTER 3. FOURIER ANALYSIS

product between two functions defined in this way is actually exactly the same thing as the
inner product between two vectors, for the following reason.
Let’s break up the interval 0 ≤ x ≤ L into a thousand tiny intervals and look at the
thousand values of a given function at these points. If we list out these values next to each
other, then we’ve basically formed a thousand-component vector. If we want to calculate the
inner product of two such functions, then a good approximation to the continuous integral
in Eq. (8) is the discrete sum of the thousand products of the values of the two functions
at corresponding points. (We then need to multiply by dx = L/1000, but that won’t be
important here.) But this discrete sum is exactly what you would get if you formed the
inner product of the two thousand-component vectors representing the values of the two
functions.
Breaking the interval 0 ≤ x ≤ L into a million, or a billion, etc., tiny intervals would
give an even better approximation to the integral. So in short, a function can be thought of
as an infinite-component vector, and the inner product is simply the standard inner product
of these infinite-component vectors (times the tiny interval dx).

Calculating the coefficients

Having discussed the orthogonality of functions, we can now calculate the an and bn coeffi-
cients in Eq. (1). The a0 case is special, so let’s look at that first.
RL
• Calculating a0 : Consider the integral, 0 f (x) dx. Assuming that f (x) can be
written in the form given in Eq. (1) (and we are indeed assuming that this aspect of
Fourier’s theorem is true), then all the sines and cosines undergo an integral number of
oscillations over the interval 0 ≤ x ≤ L, so they integrate to zero. The only surviving
term is the a0 one, and it yields an integral of a0 L. Therefore,
Z L Z L
1
f (x) dx = a0 L =⇒ a0 = f (x) dx (9)
0 L 0

a0 L is simply the area under the curve.

• Calculating an : Now consider the integral,
Z L µ ¶
2πmx
f (x) cos dx, (10)
0 L
and again assume that f (x) can be written in the form given in Eq. (1). Using the
first of the orthogonality relations in Eq. (6), we see that all of the sine terms in
f (x) integrate to zero when multiplied by cos(2πmx/L). Similarly, the second of the
orthogonality relations in Eq. (6) tells us that the cosine terms integrate to zero for all
values of n except in the special case where n = m, in which case the integral equals
am (L/2). Lastly, the a0 term integrates to zero, because cos(2πmx/L) undergoes an
integral number of oscillations. So we end up with
Z L µ ¶ Z µ ¶
2πmx L 2 L 2πnx
f (x) cos dx = am =⇒ an = f (x) cos dx (11)
0 L 2 L 0 L

where we have switched the m label to n, simply because it looks nicer.

• Calculating bn : Similar reasoning shows that
Z L µ ¶ Z µ ¶
2πmx L 2 L 2πnx
f (x) sin dx = bm =⇒ bn = f (x) sin dx (12)
0 L 2 L 0 L
3.1. FOURIER TRIGONOMETRIC SERIES 5

The limits of integration in the above expressions don’t actually have to be 0 and L.
They can be any two values of x that differ by L. For example, −L/4 and 3L/4 will work
just as well. Any interval of length L will yield the same an and bn coefficients, because both
f (x) and the trig functions are periodic. This can be seen in Fig. 2, where two intervals of
length L are shown. (As mentioned above, it’s fine if f (x) isn’t continuous.) These intervals
are shaded with slightly different shades of gray, and the darkest region is common to both
intervals. The two intervals yield the same integral, because for the interval on the right, we
can imagine taking the darkly shaded region and moving it a distance L to the right, which
causes the right interval to simply be the left interval shifted by L. Basically, no matter
where we put an interval of length L, each point in the period of f (x) and in the period of
the trig function gets represented exactly once.

move

f(x)

sin(2πx/L) L
L

Figure 2
f(x)

Remark: However, if we actually shift the origin (that is, define another point to be x = 0), then
x
the an and bn coefficients change. Shifting the origin doesn’t shift the function f (x), but it does shift
the sine and cosine curves horizontally, because by definition we’re using the functions sin(2πmx/L)
and cos(2πmx/L) with no phases. So it matters where we pick the origin. For example, let f (x) be x=0
the alternating step function in the first plot in Fig. 3. In the second plot, where we have shifted the
origin, the step function remains in the same position on the page, but the sine function is shifted g(x)
to the right. The integral (over any interval of length L, such as the shaded one) of the product
of f (x) and the sine curve in the first plot is not equal to the integral of the product of g(x) and
x
the shifted (by L/4) sine curve in the second plot. It is nonzero in the first plot, but zero in the
second plot (both of these facts follow from even/odd-ness). We’re using g instead of f here to
describe the given “curve” in the second plot, because g is technically a different function of x. If x=0
the horizontal shift is c (which is L/4 here), then g is related to f by g(x) = f (x + c). So the fact
that, say, the bn coefficients don’t agree is the statement that the integrals over the shaded regions
Figure 3
in the two plots in Fig. 3 don’t agree. And this in turn is the statement that
Z L ³ ´ Z L ³ ´
2πmx 2πmx
f (x) sin dx 6= f (x + c) sin dx. (13)
0
L 0
L

If f (x) has sufficient symmetry, then it is advantageous to shift the origin so that it (or technically
the new function g(x)) is an even or odd function of x. If g(x) is an even function of x, then only
the an ’s survive (the cosine terms), because the bn ’s equal the integral of an even function (namely
g(x)) times an odd function (namely sin(2πmx/L)) and therefore vanish. Similarly, if g(x) is an
odd function of x, then only the bn ’s survive (the sine terms). In particular, the odd function f (x)
in the first plot in Fig. 3 has only the bn terms, whereas the even function g(x) in the second plot
has only the an terms. But in general, a random choice of the origin will lead to a Fourier series
involving both an and bn terms. ♣
 f(x) = Ax 6 CHAPTER 3. FOURIER ANALYSIS

x
Example (Sawtooth function): Find the Fourier series for the periodic function shown
- _L _L in Fig. 4. It is defined by f (x) = Ax for −L/2 < x < L/2, and it has period L.
2 2
Figure 4 Solution: Since f (x) is an odd function of x, only the bn coefficients in Eq. (1) are nonzero.
Eq. (12) gives them as

Z L/2 ³ ´ Z L/2 ³ ´
2 2πnx 2 2πnx
bn = f (x) sin dx = Ax sin dx. (14)
L −L/2
L L −L/2
L

Integrating by parts (or just looking up the integral in a table) gives the general result,

Z ³ ´ Z
1 1
x sin(rx) dx = x − cos(rx) dx − − cos(rx) dx
r r
x 1
= − cos(rx) + 2 sin(rx). (15)
r r

Letting r ≡ 2πn/L yields

" #
³ ´ ³ ´ ¯L/2 ³ ´2 ³ ´ L/2 ¯
2A L 2πnx ¯¯ L 2πnx ¯¯
bn = −x cos ¯ + sin ¯
L 2πn L −L/2
2πn L −L/2
³ ´
AL AL
= − cos(πn) − cos(−πn) + 0
2πn 2πn
AL
= − cos(πn)
πn
AL
= (−1)n+1 . (16)
πn

Eq. (1) therefore gives the Fourier trig series for f (x) as

³ ´
AL X
∞
1 2πnx
f (x) = (−1)n+1 sin
π n L
n=1
h ³ ´ ³ ´ ³ ´ i
AL 2πx 1 4πx 1 6πx
= sin − sin + sin − ··· . (17)
π L 2 L 3 L

The larger the number of terms that are included in this series, the better the approximation
to the periodic Ax function. The partial-series plots for 1, 3, 10, and 50 terms are shown
in Fig. 5. We have arbitrarily chosen A and L to be 1. The 50-term plot gives a very good
approximation to the sawtooth function, although it appears to overshoot the value at the
discontinuity (and it also has some wiggles in it). This overshoot is an inevitable effect at a
discontinuity, known as the Gibbs phenomenon. We’ll talk about this in Section 3.6.
Interestingly, if we set x = L/4 in Eq. (17), we obtain the cool result,

³ ´
L AL 1 1 π 1 1 1
A· = 1 + 0 − + 0 + ··· =⇒ = 1 − + − + ···. (18)
4 π 3 5 4 3 5 7

Nice expressions for π like this often pop out of Fourier series.
3.2. FOURIER EXPONENTIAL SERIES 7

f(x)
0.6 (1 term) 0.6
(3 terms)
0.4 0.4
0.2 0.2
x
1.0 0.5 0.5 1.0 1.0 0.5 0.5 1.0
0.2 0.2
0.4 0.4
0.6 0.6

0.6 0.6
(10 terms) (50 terms)
0.4 0.4
0.2 0.2

1.0 0.5 0.5 1.0 1.0 0.5 0.5 1.0

0.2 0.2
0.4 0.4
0.6 0.6

Figure 5

3.2 Fourier exponential series

Any function that can be written in terms of sines and cosines can also be written in terms
of exponentials. So assuming that we can write any periodic function in the form given in
Eq. (1), we can also write it as

∞
X
f (x) = Cn ei2πnx/L (19)
n=−∞

where the Cn coefficients are given by

Z L
1
Cm = f (x)e−i2πmx/L dx (20)
L 0

as we will show below. This hold for all n, including n = 0. Eq. (19) is the Fourier
exponential series for the function f (x). The sum runs over all the integers here, whereas
it runs over only the positive integers in Eq. (1), because the negative integers there give
redundant sines and cosines.
We can obtain Eq. (20) in a manner similar to the way we obtained the an and bn
coefficients via the orthogonality relations in Eq. (6). The analogous orthogonality relation
here is
Z L ¯L
L ¯
ei2πnx/L e−i2πmx/L dx = ei2π(n−m)x/L ¯
0 i2π(n − m) 0
= 0, unless m = n, (21)

because the value of the exponential is 1 at both limits. If m = n, then the integral is simply
8 CHAPTER 3. FOURIER ANALYSIS

RL
0
1 · dx = L. So in the δnm notation, we have1
Z L
ei2πnx/L e−i2πmx/L dx = Lδnm . (22)
0

Just like with the various sine and cosine functions in the previous section, the different
exponential functions are orthogonal.2 Therefore, when we plug the f (x) from Eq. (19) into
Eq. (20), the only term that survives from the expansion of f (x) is the one where the n
value equals m. And in that case the integral is simply Cm L. So dividing the integral by L
gives Cm , as Eq. (20) states. As with a0 in the case of the trig series, Eq. (20) tells us that
C0 L is the area under the curve.

Example (Sawtooth function again): Calculate the Cn ’s for the periodic Ax function
in the previous example.

Solution: Eq. (20) gives (switching from n to m)

Z L/2
1
Cn = Axe−i2πnx/L dx. (23)
L −L/2
R
The integral of the general form xe−rx dx can be found by integration by parts (or looking
it up in a table). The result is
Z
x 1
xe−rx dx = − e−rx − 2 e−rx . (24)
r r
R
This is valid unless r = 0, in which case the integral equals x dx = x2 /2. For the specific
integral in Eq. (23) we have r = i2πn/L, so we obtain (for n 6= 0)
µ ¯ ³ ´2 ¯L/2 ¶
A xL −i2πnx/L ¯L/2 L ¯
Cn = − e ¯ + e−i2πnx/L ¯ . (25)
L i2πn −L/2 i2πn −L/2

The second of these terms yields zero because the limits produce equal terms. The first term
yields
A (L/2)L ¡ −iπn ¢
Cn = − · e + eiπn . (26)
L i2πn
The sum of the exponentials is simply 2(−1)n . So we have (getting the i out of the denomi-
nator)
iAL
Cn = (−1)n (for n 6= 0). (27)
2πn
¯L/2
If n = 0, then the integral yields C0 = (A/L)(x2 /2)¯−L/2 = 0. Basically, the area under
the curve is zero since Ax is an odd function. Putting everything together gives the Fourier
exponential series,
X iAL i2πnx/L
f (x) = (−1)n e . (28)
2πn
n6=0

This sum runs over all the integers (positive and negative) with the exception of 0.
1 The “L” in front of the integral in this equation is twice the “L/2” that appears in Eq. (6). This is no

surprise, because if we use eiθ = cos θ + i sin θ to write the integral in this equation in terms of sines and
cosines, we end up with two sin · cos terms that integrate to zero, plus a cos · cos and a sin · sin term, each
of which integrates to (L/2)δnm .
2 For complex functions, the inner product is defined to be the integral of the product of the functions,

where one of them has been complex conjugated. This complication didn’t arise with the trig functions in
Eq. (6) because they were real. But this definition is needed here, because without the complex conjugation,
the inner product of a function with itself would be zero, which wouldn’t be analogous to regular vectors.
3.2. FOURIER EXPONENTIAL SERIES 9

As a double check on this, we can write the exponential in terms of sines and cosines:
X ³ ³ ´ ³ ´´
iAL 2πnx 2πnx
f (x) = (−1)n cos + i sin . (29)
2πn L L
n6=0

Since (−1)n /n is an odd function of n, and since cos(2πnx/L) is an even function of n, the
cosine terms sum to zero. Also, since sin(2πnx/L) is an odd function of n, we can restrict
the sum to the positive integers, and then double the result. Using i2 (−1)n = (−1)n+1 , we
obtain
³ ´
AL X
∞
1 2πnx
f (x) = (−1)n+1 sin , (30)
π n L
n=1

in agreement with the result in Eq. (17).

Along the lines of this double-check we just performed, another way to calculate the
Cn ’s that avoids doing the integral in Eq. (20) is to extract them from the trig coefficients,
an and bn , if we happen to have already calculated those (which we have, in the case of the
periodic Ax function). Due to the relations,

eiz + e−iz eiz − e−iz

cos z = , and sin z = , (31)
2 2i
we can replace the trig functions in Eq. (1) with exponentials. We then quickly see that the
Cn coefficients can be obtained from the an and bn coefficients via (getting the i’s out of
the denominators)
an − ibn an + ibn
Cn = , and C−n = , (32)
2 2
and C0 = a0 . The n’s in these relations take on only positive values. For the periodic
Ax function, we know that the an ’s are zero, and that the bn ’s are given in Eq. (16) as
(−1)n+1 AL/πn. So we obtain
AL AL
Cn = −i(−1)n+1 , and C−n = i(−1)n+1 . (33)
2πn 2πn
If we define m to be −n in the second expression (so that m runs over the negative integers),
and then replace the arbitrary letter m with n, we see that both of these expressions can
be transformed into a single formula,
iAL
Cn = (−1)n , (34)
2πn
which holds for all integers (positive and negative), except 0. This result agrees with Eq.
(27).
Conversely, if we’ve already calculated the Cn ’s, and if we want to determine the an ’s
and bn ’s in the trig series without doing any integrals, then we can use e−z = cos z + i sin z
to write the exponentials in Eq. (19) in terms of sines and cosines. The result will be a trig
series in the form of Eq. (1), with an and bn given by

an = Cn + C−n , and bn = i(Cn − C−n ), (35)

and a0 = C0 . Of course, we can also obtain these relations by simply inverting the relations
in Eq. (32). We essentially used these relations in the double-check we performed in the
example above.
10 CHAPTER 3. FOURIER ANALYSIS

Real or imaginary, even or odd

If a function f (x) is real (which is generally the case in classical physics), then the nth and
(−n)th terms in Eq. (19) must be complex conjugates. Because the exponential parts are
already complex conjugates of each other, this implies that

C−n = Cn∗ (if f (x) is real. (36)

For the (real) periodic Ax function we discussed above, the Cn ’s in Eq. (27) do indeed
satisfy C−n = Cn∗ . In the opposite case where a function is purely imaginary, we must have
C−n = −Cn∗ .
Concerning even/odd-ness, a function f (x) is in general neither an even nor an odd
function of x. But if one of these special cases holds, then we can say something about the
Cn coefficients.

• If f (x) is an even function of x, then there are only cosine terms in the trig series in
Eq. (1), so the relative plus sign in the first equation in Eq. (31) implies that

Cn = C−n . (37)

If additionally f (x) is real, then the C−n = Cn∗ requirement implies that the Cn ’s are
purely real.

• If f (x) is an odd function of x, then there are only sine terms in the trig series in Eq.
(1), so the relative minus sign in the second equation in Eq. (31) implies that

Cn = −C−n . (38)

This is the case for the Cn ’s we found in Eq. (27). If additionally f (x) is real, then
the C−n = Cn∗ requirement implies that the Cn ’s are purely imaginary, which is also
the case for the Cn ’s in Eq. (27).

We’ll talk more about these real/imaginary and even/odd issues when we discuss Fourier
transforms.

3.3 Fourier transforms

The Fourier trigonometric series given by Eq. (1) and Eqs. (9,11,12), or equivalently the
Fourier exponential series given by Eqs. (19,20), work for periodic functions. But what if a
function isn’t periodic? Can we still write the function as a series involving trigonometric
or exponential terms? It turns out that indeed we can, but the series is now a continuous
integral instead of a discrete sum. Let’s see how this comes about.
The key point in deriving the continuous integral is that we can consider a nonperiodic
function to be a periodic function on the interval −L/2 ≤ x ≤ L/2, with L → ∞. This
might sound like a cheat, but it is a perfectly legitimate thing to do. Given this strategy,
we can take advantage of all the results we derived earlier for periodic functions, with the
only remaining task being to see what happens to these results in the L → ∞ limit. We’ll
work with the exponential series here, but we could just as well work with the trigonometric
series.
Our starting point will be Eqs. (19) and (20), which we’ll copy here:
∞
X Z L/2
1
f (x) = Cn ei2πnx/L , where Cn = f (x)e−i2πnx/L dx, (39)
n=−∞
L −L/2
3.3. FOURIER TRANSFORMS 11

where we have taken the limits to be −L/2 and L/2. Let’s define kn ≡ 2πn/L. The
difference between successive kn values is dkn = 2π(dn)/L = 2π/L, because dn is simply
1. Since we are interested in the L → ∞ limit, this dkn is very small, so kn is essentially a
continuous variable. In terms of kn , the expression for f (x) in Eq. (39) becomes (where we
have taken the liberty of multiplying by dn = 1)
∞
X
f (x) = Cn eikn x (dn)
n=−∞
X∞ µ ¶
L
= Cn eikn x dkn . (40)
n=−∞
2π

But since dkn is very small, this sum is essentially an integral:

Z ∞µ ¶
L
f (x) = Cn ekn x dkn
−∞ 2π
Z ∞
≡ C(kn )eikn x dkn , (41)
−∞

where C(kn ) ≡ (L/2π)Cn . We can use the expression for Cn in Eq. (39) to write C(kn ) as
(in the L → ∞ limit)
Z
L L 1 ∞
C(kn ) ≡ Cn = · f (x)e−ikn x dx
2π 2π L −∞
Z ∞
1
= f (x)e−ikn x dx. (42)
2π −∞
We might as well drop the index n, because if we specify k, there is no need to mention n.
We can always find n from k ≡ 2πn/L if we want. Eqs. (41) and (42) can then be written
as Z ∞ Z ∞
ikx 1
f (x) = C(k)e dk where C(k) = f (x)e−ikx dx (43)
−∞ 2π −∞

C(k) is known as the Fourier transform of f (x), and vice versa. We’ll talk below about how
we can write things in terms of trigonometric functions instead of exponentials. But the
term “Fourier transform” is generally taken to refer to the exponential decomposition of a
function.
Similar to the case with Fourier series, the first relation in Eq. (43) tells us that C(k)
indicates how much of the function f (x) is made up of eikx . And conversely, the second
relation in Eq. (43) tells us that f (x) indicates how much of the function C(k) is made up of
e−ikx . Of course, except in special cases (see Section 3.5), essentially none of f (x) is made
up of eikx for one particular value of k, because k is a continuous variable, so there is zero
chance of having exactly a particular value of k. But the useful thing we can say is that if dk
is small, then essentially C(k) dk of the function f (x) is made up of eikx terms in the range
from k to k + dk. The smaller dk is, the smaller the value of C(k) dk is (but see Section 3.5
for an exception to this).
Remarks:
1. These expressions in Eq. (43) are symmetric except for a minus sign in the exponent (it’s a
matter of convention as to which one has the minus √ sign), and also a 2π. If you want, you
can define things so that each
√ expression has a 1/ 2π, by simply replacing C(k) with a B(k)
function defined by B(k) ≡ 2π C(k). In any case, the product of the factors in front of the
integrals must be 1/2π.
12 CHAPTER 3. FOURIER ANALYSIS

2. With our convention that one expression has a 2π and the other doesn’t, there is an ambiguity
when we say, “f (x) and C(k) are Fourier transforms of each other.” Better terminology would
be to say that C(k) is the Fourier transform of f (x), while f (x) is the “reverse” Fourier
transform of C(k). The “reverse” just depends on where the 2π is. If x measures a position
and k measures a wavenumber, the common convention is the one in Eq. (43).
√
Making both expressions have a “1/ 2π ” in them would lead to less of a need for the word
“reverse,” although there would still be a sign convention in the exponents which would have
to be arbitrarily chosen. However, the motivation for having the first equation in Eq. (43)
stay the way it is, is that the C(k) there tells us how much of f (x) is made up of eikx , which
√
is a more natural thing to be concerned with than how much of f (x) is made up√of eikx / 2π.
The latter is analogous to expanding the Fourier series in Eq. (1) in terms of 1/ 2π times the
trig functions, which isn’t the most natural thing to do. The price to pay for the convention
in Eq. (43) is the asymmetry, but that’s the way it goes.
3. Note that both expressions in Eq. (43) involve integrals, whereas in the less symmetric Fourier-
series results in Eq. (39), one expression involves an integral and one involves a discrete sum.
The integral in the first equation in Eq. (43) is analogous to the Fourier-series sum in the first
equation in Eq. (39), but unfortunately this integral doesn’t have a nice concise name like
“Fourier series.” The name “Fourier-transform expansion” is probably the most sensible one.
At any rate, the Fourier transform itself, C(k), is analogous to the Fourier-series coefficients
Cn in Eq. (39). The transform doesn’t refer to the whole integral in the first equation in Eq.
(43). ♣

Real or imaginary, even or odd

Anything that we can do with exponentials we can also do with trigonometric functions.
So let’s see what Fourier transforms look like in terms of these. Because cosines and sines
are even and odd functions, respectively, the best way to go about this is to look at the
even/odd-ness of the function at hand. Any function f (x) can be written uniquely as the
sum of an even function and an odd function. These even and odd parts (call them fe (x)
and fo (x)) are given by
f (x) + f (−x) f (x) − f (−x)
fe (x) = and fo (x) = . (44)
2 2
You can quickly verify that these are indeed even and odd functions. And f (x) = fe (x) +
fo (x), as desired.
Remark: These functions are unique, for the following reason. Assume that there is another pair
of even and odd functions whose sum is f (x). Then they must take the form of fe (x) + g(x) and
fo (x) − g(x) for some function g(x). The first of these relations says that g(x) must be even, but
the second says that g(x) must be odd. The only function that is both even and odd is the zero
function. So the new pair of functions must actually be the same as the pair we already had. ♣

In terms of fe (x) and fo (x), the C(k) in the second equation in Eq. (43) can be written as
Z ∞
1
C(k) = f (x)e−ikx dx
2π −∞
Z ∞³ ´³ ´
1
= fe (x) + fo (x) cos(−kx) + i sin(−kx) dx. (45)
2π −∞
If we multiply out the integrand, we obtain four terms. Two of them (the fe cos and fo sin
ones) are even functions of x. And two of them (the fe sin and fo cos ones) are odd functions
of x. The odd ones integrate to zero, so we are left with
Z ∞ Z ∞
1 i
C(k) = fe (x) cos(kx) dx − fo (x) sin(kx) dx. (46)
2π −∞ 2π −∞
3.3. FOURIER TRANSFORMS 13

Up to this point we’ve been considering the even/odd-ness of f (x) as a function of x. But
let’s now consider the even/odd-ness of C(k) as a function of k. The first term in Eq. (46)
is an even function of k (because k appears only in the cosine function), and the second is
odd (because k appears only in the sine function). So we can read off the even and odd
parts of C(k):
Z ∞ Z ∞
1 i
Ce (k) = fe (x) cos(kx) dx and Co (k) = − fo (x) sin(kx) dx (47)
2π −∞ 2π −∞

Likewise, the inverse relations are quickly obtained by letting C(k) = Ce (k) + Co (k) in the
expression for f (x) in Eq. (43). The result is
Z ∞ Z ∞
f (x) = Ce (k) cos(kx) dk + i Co (k) sin(kx) dk, (48)
−∞ −∞

and so
Z ∞ Z ∞
fe (x) = Ce (k) cos(kx) dk and fo (x) = i Co (k) sin(kx) dk (49)
−∞ −∞

Similar to the case with the exponential decomposition, Ce (k) tells us how much of the
function f (x) is made up of cos(kx), and iCo (k) tells us how much of the function f (x) is
made up of sin(kx). In view of Eq. (48), the functions Ce (k) and iCo (k) can reasonably be
called the “Fourier trig transforms” of f (x).
Note that if we replace the cos(kx) in the first equations Eqs. (47) and (49) by eikx ,
the integral is unaffected, because the sine part of the exponential is an odd function (of
either x or k, as appropriate) and therefore doesn’t contribute anything to the integral.
These equations therefore tell us that Ce (k) is the Fourier transform of fe (x). And likewise
iCo (k) is the Fourier transform of fo (x). We see that the pair of functions fe (x) and Ce (k) is
completely unrelated to the pair fo (x) and Co (k), as far as Fourier transforms are concerned.
The two pairs don’t “talk” to each other.
Eq. (46) tells us that if f (x) is real (which is generally the case in classical physics),
then the real part of C(k) is even, and the imaginary part is odd. A concise way of saying
these two things is that C(k) and C(−k) are complex conjugates. That is, C(−k) = C(k)∗ .
Likewise, Eq. (48) tells us that if C(k) is real, then f (−x) = f (x)∗ . (Related statements
hold if f (x) or C(k) is imaginary.)
For the special cases of purely even/odd and real/imaginary functions, the following
facts follow from Eq. (47):
• If f (x) is even and real, then C(k) is even and real.
• If f (x) is even and imaginary, then C(k) is even and imaginary.
• If f (x) is odd and real, then C(k) is odd and imaginary.
• If f (x) is odd and imaginary, then C(k) is odd and real.
The converses are likewise all true, by Eq. (49).
Let’s now do some examples where we find the Fourier trig series and Fourier (trig)
transform of two related functions.
 f(x)
14 CHAPTER 3. FOURIER ANALYSIS
A

x
-L/2 L/2
Example (Periodic odd step function): Calculate the Fourier trig series for the periodic
odd step function shown in Fig. 6. The function takes on the values ±A.

Figure 6 Solution: The function is odd, so only the sin terms survive in Eq. (1). Eq. (12) gives the
bn coefficients as
Z L/2 ³ ´ Z L/2 ³ ´
2 2πnx 2 2πnx
bn = f (x) sin dx = 2 · A sin dx
L −L/2
L L 0
L
³ ´ ³ ´¯
4A L 2πnx ¯L/2 2A ¡ ¢
= − cos ¯ = 1 − cos(πn) . (50)
L 2πn L 0 πn
This equals 4A/πn is n is odd, and zero if n is even. So we have
X
∞ ³ ´
4A 2πnx
f (x) = sin
πn L
n=1,odd
³ ³ ´ ³ ´ ³ ´ ´
4A 2πx 1 6πx 1 10πx
= sin + sin + sin + ··· . (51)
π L 3 L 5 L
Some partial plots are shown in Fig. 7. The plot with 50 terms yields a very good approxi-
mation to the step function, although there is an overshoot at the discontinuities (and also
some fast wiggles). We’ll talk about this in Section 3.6.

f(x)
(1 term) 1.0
(3 terms)
1.0
0.5 0.5
x
- 1.0 - 0.5 0.5 1.0 -1.0 - 0.5 0.5 1.0
- 0.5 - 0.5
- 1.0 -1.0

1.0
(10 terms) 1.0
(50 terms)
0.5 0.5

-1.0 - 0.5 0.5 1.0 - 1.0 - 0.5 0.5 1.0

- 0.5 - 0.5
-1.0 -1.0

Figure 7

For the fun of it, we can plug x = L/4 into Eq. (51) to obtain
³ ´
f(x) 4A 1 1 π 1 1 1
A= 1 − + − ··· =⇒ = 1 − + − + ···, (52)
π 3 5 4 3 5 7
A
which is the same result we found in Eq. (18).
x
-L/2 L/2
Example (Non-periodic odd step function): Calculate the Fourier trig transform of
the non-periodic odd step function shown in Fig. 8. The value is ±A inside the region
−L/2 < x < L/2, and zero outside. Do this in two ways:
Figure 8
(a) Find the Fourier series for the periodic “stretched” odd step function shown in Fig. 9,
with period N L, and then take the N → ∞ limit.
3.3. FOURIER TRANSFORMS 15

f(x)
A

x
-NL/2 -L/2 L/2 NL/2

NL

Figure 9

(b) Calculate C(k) by doing the integral in Eq. (43). The result will look different from the
result in part (a), but you can demonstrate that the two are equivalent.
Solution:
(a) The function in Fig. 9 is odd, so only the sin terms survive in Eq. (1). The length of
the interval is N L, so Eq. (12) gives the bn coefficients as
Z N L/2 ³ ´
2 2πnx
bn = f (x) sin dx. (53)
NL −N L/2
NL

But f (x) is nonzero only in the interval −L/2 < x < L/2. We can consider just the
0 < x < L/2 half and then multiply by 2. So we have
Z L/2 ³ ´ ³ ´ ³ ´¯
2 2πnx 4A N L 2πnx ¯L/2
bn = 2 · A sin dx = − cos ¯
NL 0
NL N L 2πn NL 0
³ ³ ´´
2A πn
= 1 − cos . (54)
πn N
If N = 1, this agrees with the result in Eq. (50) in the previous example, as it should.
The Fourier series for the function is therefore
X
∞ ³ ´ X
∞ ³ ³ ´´ ³ ´
2πnx 2A πn 2πnx
f (x) = bn sin = 1 − cos sin . (55)
NL πn N NL
n=1 n=1

We will now find the Fourier trig transform by taking the N → ∞ limit. Define zn by
zn ≡ n/N . The changes in zn and n are related by dn = N dzn . dn is simply 1, of
course, so we can multiple the result in Eq. (55) by dn without changing anything. If
we then replace dn by N dzn , and also get rid of the n’s in favor of zn ’s, we obtain

X
∞ ³ ´ ³ ´
2A 2πzn x
f (x) = 1 − cos(πzn ) sin N dzn (56)
π(N zn ) L
n=1

In the N → ∞ limit, dzn is very small, so we can change this sum to an integral. And
we can drop the n subscript on z, because the value of z determines the associated value
of n (if we happen to care what it is). So we arrive at the desired Fourier trig transform:
Z ∞ ³ ´ ³ ´
2A 1 2πzx
f (x) = 1 − cos(πz) sin dz (for non-periodic f ). (57)
π 0
z L

If we want to put this in the standard Fourier-transform form where the integral runs
from −∞ to ∞, we just need to make the coefficient be A/π. The integrand is an even
function of z, so using a lower limit of −∞ simply doubles the integral.
The Fourier trig transform in Eq. (57) for the non-periodic odd step function should
be compared with the Fourier series for the periodic odd step function in the previous
 16 CHAPTER 3. FOURIER ANALYSIS

example. If we use the form of bn in Eq. (50) (that is, without evaluating it for even
and odd n), the series looks like

³ ´ ³ ´
2A X 1
∞
2πnx
f (x) = 1 − cos(πn) sin (for periodic f ) (58)
π n L
n=0

The only difference between this equation and Eq. (57) is that n can take on only integer
values here, whereas z is a continuous variable in Eq. (57). It is quite fascinating how the
f(x) switch to a continuous variable changes the function from a periodic one to a function
1.0 that is zero everywhere except in the region −L/2 < x < L/2.
0.5
Remark: What if we try to approximate the integral in Eq. (57) by discretizing z into small
x
-2 -1 1 2 units and calculating a discrete sum? For example, if we take the “bin” size to be ∆z = 0.2,
-0.5 then the resulting f (x) appears to have the desired value of zero outside the −L/2 < x < L/2
-1.0 region, as shown in Fig. 10 (we’ve chosen L = 1, and we’ve truncated the discrete sum at
z = 200, which is plenty large to get a good approximation). However, the zoomed-out view in
Fig. 11 shows that the function actually repeats at x = ±5. And if we zoomed out farther, we
Figure 10 would find that it repeats at ±10 and ±15, etc. In retrospect this is clear, for two reasons.
First, this discrete sum is exactly the sum in Eq. (56) (after canceling the N ’s) with dzn = 0.2,
f(x) which corresponds to N = 5. So it’s no surprise that we end up reproducing the original
“stretched” periodic function with N = 5. Second,
¡ the lowest-frequency
¢ sine function in the
1.0
discrete sum is the z = 0.2 one, which is sin 2π(0.2)x/L = sin(2πx/5L). This has a period
0.5 of 5L. And the periods of all the other sine functions are smaller than this; they take the form
x of 5L/m, where m is an integer. So the sum of all the sines in the discrete sum will certainly
- 6 - 4 -2 2 4 6 repeat after 5L (and not any sooner).
- 0.5
-1.0 This argument makes it clear that no matter how we try to approximate the Fourier-transform
integral in Eq. (57) by a discrete sum, it will inevitably eventually repeat. Even if we pick the
bin size to be a very small number like ∆z = 0.001, the function will still repeat at some point
Figure 11 (at x = ±1000L in this case). We need a truly continuous integral in order for the function to
never repeat. We’ll talk more about these approximation issues when we look at the Gaussian
function in Section 3.4.1. ♣

(b) Eq. (43) gives

Z ∞
1
C(k) = f (x)e−ikx dx
2π −∞
Z 0 Z L/2
1 1
= (−A)e−ikx dx + (A)e−ikx dx
2π −L/2
2π 0
µ ¯0 ¯L/2 ¶
1 A ¯ ¯
= · −e−ikx ¯ + e−ikx ¯
2π (−ik) −L/2 0
³ ´
iA
= − (1 − eikL/2 ) + (e−ikL/2 − 1)
2πk
³ ³ ´´
−iA kL
= 1 − cos . (59)
πk 2

So the expression for f (x) in Eq. (43) becomes

Z ∞
f (x) = C(k)eikx dk
−∞
Z ∞ ³ ³ ´´
−iA 1 kL
= 1 − cos eikx dk. (60)
π −∞
k 2

This looks different from the expression for f (x) in Eq. (57), but it had better be the
same thing, just in a different form. We can rewrite Eq. (60) as follows. Since the 1/k
factor is an odd function of k, and since cos(kL/2) is even, only the odd part of eikx
3.4. SPECIAL FUNCTIONS 17

(the sin(kx) part) survives in the integral. The integrand is then an even function of k,
so we can have the integral run from 0 to ∞ and then multiply by 2. We obtain
Z ³ ³ ´´ ¡
−iA
∞
1 kL ¢
f (x) = ·2 1 − cos i sin(kx) dk
π 0
k 2
Z ∞ ³ ³ ´´
2A 1 kL
= 1 − cos sin(kx) dk (61)
π 0
k 2

This still looks a little different from the expression for f (x) in Eq. (57), but if we define
k ≡ 2πz/L =⇒ dk = (2π/L)dz, then Eq. (61) turns into Eq. (57), as desired. So both
forms generate the same function f (x), as they must.

Remark: The “z” notation in Eq. (57) has the advantage of matching up with the “n” notation
in the Fourier series in Eq. (58). But the “k” notation in Eqs. (60) and (61) is more widely used
because it doesn’t have all the π’s floating around. The “n” notation in the Fourier series in Eq.
(58) was reasonable to use, because it made sense to count the number of oscillations that fit
into one period of the function. But when dealing with the Fourier transform of a non-periodic
function, there is no natural length scale of the function, so it doesn’t make sense to count the
number of oscillations, so in turn there is no need for the π’s. ♣

3.4 Special functions

The are a number of functions whose Fourier transforms come up often in the study of f(x)
waves. Let’s look at a few of the more common ones. 1.0
0.8
0.6
3.4.1 Gaussian 0.4
−ax2
What is the Fourier transform of the Gaussian function, f (x) = Ae , shown in Fig. 12? 0.2

(For the purposes of plotting a definite function, we have chosen A = a = 1.) If we plug x
-3 -2 -1 1 2 3
this f (x) into the second equation in Eq. (43), we obtain
Z Figure 12
∞
1 2
C(k) = Ae−ax e−ikx dx. (62)
2π −∞

In order to calculate this integral, we’ll need to complete the square in the exponent. This
gives
Z ∞
1 2 2
C(k) = Ae−a(x+ik/2a) e−k /4a dx. (63)
2π −∞
It turns out that we can just ignore the constant (as far as x goes) term, ik/2a, in the
exponent. This follows from a theorem in complex analysis involving the poles (divergences)
of a function, but we won’t get into that here. With the ik/2a removed, we now have
2 Z ∞
e−k /4a 2
C(k) = Ae−ax dx. (64)
2π −∞

R 2
Although a closed-form expression for the indefinite integral e−ax dx doesn’t exist, it is
R ∞ −ax2
fortunately possible to calculate the definite integral, −∞ e dx. We can do this by using
18 CHAPTER 3. FOURIER ANALYSIS

R∞ 2
a slick trick, as follows. Let the desired integral be labeled as I ≡ ∞
e−ax dx. Then we
have (we’ll explain the steps below)
µZ ∞ ¶ µZ ∞ ¶
2 −ax2 −ay 2
I = e dx e dy
−∞ −∞
Z ∞Z ∞
2 2
= e−ax e−ay dx dy
−∞ ∞
Z 2π Z ∞
2
= e−ar r dr dθ
0 0
Z ∞
2
= 2π re−ar dr
0
2 ¯∞
e−ar ¯¯
= −2π
2a ¯0
π
= . (65)
a
The second line follows from the fact that if you evaluate the double integral by doing the
dx integral first, and then the dy one, you’ll simply obtain the product of the two integrals
in the first line. And the third line arises from changing the Cartesian coordinates to polar
coordinates. (Note that if the limits in the original integral weren’t ±∞, then the upper
limit on r would depend on θ, and then the θ integral wouldn’t be doable.)
Taking the square root of the result in Eq. (65) gives the nice result for the definite
integral,3 r
Z ∞
−ax2 π
e dx = . (66)
−∞ a
As a check on this, if a is small, then the integral is large, which makes sense. Plugging Eq.
(66) into Eq. (64) give the desired Fourier transform,

2 r 2
e−k /4a π Ae−k /4a
C(k) = ·A =⇒ C(k) = √ (67)
2π a 2 πa

Interestingly, this is a Gaussian function of k. So the Fourier transform of √ the Gaussian

2 2
function of x, f (x) = Ae−ax , is a Gaussian function of k, C(k) = (A/2 πa )e−k /4a .
So in some sense, a Gaussian function is the “nicest” function, at least as far as Fourier
transforms are concerned. Gaussians remain Gaussians. Note the inverse dependences of
C(k) and f (x) on a. For small a, f (x) is very wide, whereas C(k) is very tall (because of
the a in the denominator) and sharply peaked (because of the a in the denominator of the
exponent).
We can double check the result for C(k) in Eq. (67) by going in the other direction.
That is, we can plug C(k) into the first equation in Eq. (43) and check that the result is the
original Guassian function, f (x). Indeed,
Z ∞
f (x) = C(k)eikx dk
−∞
Z ∞ 2
Ae−k /4a ikx
= √ e dk
−∞ 2 πa
3 As
R∞ 2
a bonus, you can efficiently calculate integrals of the form x2n e−ax dx by repeatedly differen-
−∞
tiating this expression with respect to a.
3.4. SPECIAL FUNCTIONS 19
Z ∞
A 2
/4a −ax2
= √ e−(k−2iax) e dk
2 πa −∞
−ax2 Z ∞
Ae 2
= √ e−k /4a
dk
2 πa −∞
2 r
Ae−ax π
= √
2 πa 1/4a
2
= Ae−ax , (68)

as desired. In obtaining the fourth line above, we (justifiably) ignored the constant (as far
as k goes) term, −2iax, in the exponent. In obtaining the fifth line, we used Eq. (66) with
1/4a in place of a.
R∞
Making approximations to f (x) = −∞
C(k)eikx dk
As we discussed shortly after Eq. (43), the first relation in Eq. (43) tells us that the Fourier
transform C(k) indicates how much of the function f (x) is made up of eikx . More precisely,
it tells us that C(k) dk of the function f (x) is made up of eikx terms in the range from k to
k + dk. Since f (x) is even in the present Gaussian case, we know that only the cos kx part
of eikx will be relevant, but we’ll keep writing the whole eikx anyway.
As a rough test to see if the Gaussian function C(k) in Eq. (67) does whatR ∞ it is supposed
to do (that is, produce the function f (x) when plugged into the integral −∞ C(k)eikx dk),
let’s make some approximations to this integral. There are two basic things we can do.4
• We can approximate the integral by performing a sum over discrete values of k. That e-k /4
____
2

C(k) =
is, we can break up the continuous integral into discrete “bins.” For example, if pick 0.25 2π
the “bin size” to be ∆k = 1, then this corresponds to making the approximation where 0.20
all k values in the range −0.5 < k < 0.5 have the value of C(k) at k = 0. And all k 0.15
values in the range 0.5 < k < 1.5 have the value of C(k) at k = 1. And so on. This is 0.10
shown in Fig. 13, for the values A = a = 1. 0.05
k
• The second way we can approximate the integral in Eq. (43) is to keep the integral a -6 -4 -2 2 4 6

continuous one, but have the limits of integration be k = ±` instead of ±∞. We can
Figure 13
then gradually make ` larger until we obtain the ` → ∞ limit. For example, If ` = 1,
then this means that we’re integrating only over the shaded region shown in Fig. 14.
2
e-k /4
____
Let’s see in detail what happens when we make these approximations. Consider the C(k) =
2π
first kind, where we perform a sum √ over discrete values of k. With A = a = 1, we have
0.25
2 2
f (x) = e−x and C(k) = e−k /4 /2 π. This function C(k) is very small for |k| ≥ 4, so we 0.20
0.15
can truncate the sum at k = ±4 with negligible error. With ∆k = 1, the discrete sum of
0.10
the nine terms from k = −4 to k = 4 (corresponding to the nine bins shown in Fig. 13)
0.05
yields an approximation to the f (x) in Eq. (43) that is shown in Fig. 15. This provides a k
2
remarkably good approximation to the actual Gaussian function f (x) = e−x , even though -6 -4 -2 2 4 6

our discretization of k was a fairly coarse one. If we superimposed the true Gaussian on this
Figure 14
plot, it would be hard to tell that it was actually a different curve. This is due to the high
level of smoothness of the Gaussian function. For other general functions, the agreement
invariably isn’t this good. And even for the Gaussian, if we had picked a larger bin size of f(x)
say, 2, the resulting f (x) would have been noticeably non-Gaussian. 1.0
0.8
0.6
4 Although the following discussion is formulated in terms of the Gaussian function, you can of course 0.4
0.2
apply it to any function. You are encouraged to do so for some of the other functions discussed in this
x
chapter. - 3 -2 -1 1 2 3

Sum from k = -4 to k = 4
in steps of ∆k = 1

Figure 15
 f(x)
1.0
0.8 20 CHAPTER 3. FOURIER ANALYSIS
0.6
0.4
0.2 However, all is not as well as it seems. If we zoom out and let the plot of f (x) run from,
x
-15 -10 - 5 5 10 15 say, x = −15 to x = 15, then we obtain the curve shown in Fig. 16. This is hardly a Gaussian
function. It is an (infinite) periodic string of nearly Gaussian functions. In retrospect, this
Sum from k = -4 to k = 4 should be no surprise. As we saw above in Fig. 11 and in the accompanying remark, the
in steps of ∆k = 1
sum over discrete values of k will inevitably be periodic, because there is a lowest-frequency
Figure 16 eikx component, and all other components have frequencies that are multiples of this. In
the present scenario, the k = 1 component has the lowest frequency (call is kmin ), and the
period is 2π/kmin = 2π. And indeed, the bumps in Fig. 16 occur at multiples of 2π.
f(x)
What if we make the bin size smaller? Fig. 17 shows the case where k again runs from
1.0
0.8 −4 to 4, but now with a bin size of ∆k = 0.5. So there are 17 terms in the sum, with each
0.6 term being multiplied by the ∆k = 0.5 weighting factor. The smallest (nonzero) value of k
0.4 is kmin = 0.5. We obtain basically the same shape for the bumps (they’re slightly improved,
0.2
x although there wasn’t much room for improvement over the ∆k = 1 case), but the period is
-15 -10 - 5 5 10 15 now 2π/kmin = 12π. So the bumps are spread out more. This plot is therefore better than
the one in Fig. 16, in the sense that it looks like a Gaussian for a larger range of x. If we
Sum from k = -4 to k = 4
in steps of ∆k = 0.5 kept decreasing the bin size ∆k (while still letting the bins run from k = −4 to k = 4), the
bumps would spread out farther and farther, until finally in the ∆k → 0 limit they would
Figure 17 be infinitely far apart. In other words, there would be no repetition at all, and we would
have an exact Gaussian function.
Consider now the second of the above approximation strategies, where we truncate the
(continuous) integral at k = ±`. If we let ` = 0.5, then we obtain the first plot in Fig. 18.
This doesn’t look much like a Gaussian. In particular, it dips below zero, and the value
at x = 0 isn’t anywhere near the desired value of 1. This is because we simply haven’t
integrated over enough of the C(k) curve. In the limit of very small `, the resulting f (x) is
very small, because we have included only a very small part of the entire integral.

f(x) f(x)
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
x x
-15 -10 --50.2 5 10 15 -15 -10 --50.2 5 10 15

Integral from k = -0.5 to k = 0.5 Integral from k = -1 to k = 1

f(x) f(x)
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
x x
-15 -10 --50.2 5 10 15 -15 -10 --50.2 5 10 15

Integral from k = -2 to k = 2 Integral from k = -4 to k = 4

Figure 18
We can improve things by increasing ` to 1. The result is shown in the second plot in
Fig. 18. The third plot corresponds to ` = 2, and finally the fourth plot corresponds to
` = 4. This looks very much like a Gaussian. Hardly any of the C(k) curve lies outside
k = ±4, so ` = 4 provides a very good approximation to the full ` = ∞ integral.
Compared with the first approximation method involving discrete sums, the plots gen-
erated with this second method have the disadvantage of not looking much like the desired
3.4. SPECIAL FUNCTIONS 21

Gaussian for small values of `. But they have the advantage of not being periodic (because
there is no smallest value of k, since k is a now continuous variable). So, what you see is f(x)
1.0
what you get. You don’t have to worry about the function repeating at some later point.
0.8
0.6
3.4.2 Exponential, Lorentzian 0.4
0.2
What is the Fourier transform of the exponential function, Ae−b|x| , shown in Fig. 19? (We
x
have chosen A = b = 1 in the figure.) If we plug this f (x) into Eq. (43), we obtain -4 -2 2 4
Z ∞
1 Figure 19
C(k) = Ae−b|x| e−ikx dx
2π −∞
Z 0 Z ∞
1 bx −ikx 1
= Ae e dx + Ae−bx e−ikx dx. (69)
2π −∞ 2π 0

The imaginary terms in the exponents don’t change the way these integrals are done (the
fundamental theorem of calculus still holds), so we obtain
¯0 ¯∞
Ae(b−ik)x ¯¯ Ae(−b−ik)x ¯¯
C(k) = +
2π(b − ik) ¯−∞ 2π(−b − ik) ¯0
µ ¶
A 1 1
= +
2π b − ik b + ik C(k)
Ab 0.30
= (70)
π(b2 + k 2 ) 0.25
0.20
0.15
A function of the general form c1 /(c2 + c3 k 2 ) is called a Lorentzian function of k. From the 0.10
0.05
plot of C(k) shown in Fig. 20 (with A = b = 1), we see that a Lorentzian looks somewhat like k
a Gaussian. (The main difference is that it goes to zero like 1/k 2 instead of exponentially.) -4 -2 2 4
This isn’t too much of a surprise, because the exponential function in Fig. 19 looks vaguely
similar to the Gaussian function in Fig. 12, in that they both are peaked at x = 0 and then Figure 20
taper off to zero.
As in the case of the Gaussian, the f (x) and C(k) functions here have inverse dependences
on b. For small b, f (x) is very wide, whereas C(k) is sharply peaked and very tall. C(k) is
very tall because the value at k = 0 is A/πb, which is large if b is small. And it is sharply
peaked in a relative sense because it decreases to half-max at k = b, which is small if b is
small. Furthermore, it is sharply peaked in an absolute sense because for any given value of
k, the value of C(k) is less than Ab/πk 2 , which goes to zero as b → 0.5
If you want to go in the other direction, it is possible to plug the Lorentzian C(k) into
the first equation in Eq. (43) and show that the result is in fact f (x). However, the integral f(x)
requires doing a contour integration, so we’ll just accept here that it does in fact work out. A

3.4.3 Square wave, sinc x

-a a
What is the Fourier transform of the “square wave” function shown in Fig. 21? This square
wave is defined by f (x) = A for −a ≤ x ≤ a, and f (x) = 0 otherwise. If we plug this f (x) Figure 21
into Eq. (43), we obtain
Z a
1
C(k) = Ae−ikx dx
2π −a
5 More precisely, for a given k, if you imagine decreasing b, starting with a value larger than k, then C(k)

increases to a maximum value of A/2πk at b = k, and then decreases to zero as b → 0.

 22 CHAPTER 3. FOURIER ANALYSIS

1 Ae−ikx ¯¯a
= · ¯
2π −ik −a
A e−ika − eika
= ·
C(k) 2π −ik
0.3 A sin(ka)
= (71)
0.2 πk
0.1
k This is called a sinc function of k. sinc(z) is defined generically to be sin(z)/z. A plot of
- 20 -10 10 20
C(k) is shown in Fig. 22 (with A = a = 1). It looks vaguely similar to a Gaussian and a
- 0.1
Lorentzian near k = 0, but then for larger values of k it oscillates around zero, unlike the
Figure 22 Gaussian and Lorentzian curves.
As in the Lorentzian case, it is possible to go in the other direction and plug C(k) into
the first equation in Eq. (43) and show that the result is f (x). But again it requires a
contour integration, so we’ll just accept that it works out.
The dependence of the C(k) sinc function on a isn’t as obvious as with the Gaussian
and Lorentzian functions. But using sin ² ≈ ² for small ², we see that C(k) ≈ Aa/π for k
values near zero. So if a is large (that is, f (x) is very wide), then C(k) is very tall near
k = 0. This is the same behavior that the Gaussian and Lorentzian functions exhibit.
However, the “sharply peaked” characteristic seems to be missing from C(k). Fig. 23
shows plots of C(k) for the reasonably large a values of 10 and 40. We’re still taking A
to be 1, and we’ve chosen to cut off the vertical axis at 2 and not show the peak value of
Aa/π at k = 0. The envelope of the oscillatory sine curve is simply the function A/πk, and
this is independent of a. As a increases, the oscillations become quicker, but the envelope
doesn’t fall off any faster. However, for the purposes of the Fourier transform, the important
property of C(k) is its integral, and if C(k) oscillates very quickly, then it essentially averages
out to zero. So for all practical purposes, C(k) is basically the zero function except right
near the origin. So in that sense it is sharply peaked.

C(k) C(k)
2.0 2.0
(a = 10) (a = 40)
1.5 1.5
1.0 1.0
0.5 0.5
k k
-10 -5 5 10 -10 -5 5 10
- 0.5 - 0.5
-1.0 -1.0

Figure 23

Remark: The wild oscillatory behavior of the C(k) sinc function in Fig. 23, which is absent in the
C(k) Gaussian and Lorentzian functions, is due to the discontinuity in the f (x) square wave. A
more complicated combination of eikx functions are required to generate the discontinuity, especially
if the function stays at a constant value of A for a very large interval before abruptly dropping to
zero.
So, the lower degree of “niceness” of the sinc function, compared with the Gaussian and
Lorentzian functions, can be traced to the discontinuity in the value of the f (x) square wave.
Similarly, a Lorentzian isn’t quite as nice as a Gaussian. It may be a matter of opinion, but the
Lorentzian in Fig. 20 isn’t quite as smooth as the Gaussian in Fig. 12; it’s a little less rounded. This
lower degree of niceness of the Lorentzian can be traced to the discontinuity in the first derivative
of the f (x) = Ae−b|x| exponential function. In a sense, the Gaussian function is the nicest and
3.5. THE DELTA FUNCTION 23

smoothest of all functions, and this is why its Fourier transform ends up being as nice and as
smooth as possible. In other words, it is another Gaussian function. ♣
f(x)

3.5 The delta function 1/a

3.5.1 Definition
Consider the tall and thin square-wave function shown in Fig. 24. The height and width are
related so that the area under the “curve” is 1. In the a → 0 limit, this function has an
interesting combination of properties: It is zero everywhere except at the origin, but it still
x
has area 1 because the area is independent of a. A function with these properties is called -a/2 a/2
a delta function and is denoted by δ(x). (It is sometimes called a “Dirac” delta function, to
distinguish it from the “Kronecker” delta, δnm , that came up in Sections 3.1 and 3.2.) So Figure 24
we have the following definition of a delta function:
• A delta function, δ(x), is a function that is zero everywhere except at the origin and
has area 1.
Strictly speaking, it’s unclear what sense it makes to talk about the area under a curve,
if the curve is nonzero at only one point. But there’s no need to think too hard about it,
because you can avoid the issue by simply thinking of a delta function as a square wave or
a Gaussian (or various other functions, as we’ll see below) in the sharply-peaked limit.
Consider what happens when we integrate the product of a delta function δ(x) with
some other arbitrary function f (x). We have (we’ll explain the steps below)
Z ∞ Z ∞ Z ∞
δ(x)f (x) dx = δ(x)f (0) dx = f (0) δ(x) dx = f (0). (72)
−∞ −∞ −∞

The first equality comes from the fact that the delta function is zero everywhere except at
the origin, so replacing f (x) with f (0) everywhere else has no effect on the integral; the only
relevant value of f (x) is the value at the origin. The second equality comes from pulling the
constant value, f (0), outside the integral. And the third equality comes from the fact that
the area under the delta function “curve” is 1. More generally, we have
Z ∞ Z ∞ Z ∞
δ(x − x0 )f (x) dx = δ(x − x0 )f (x0 ) dx = f (x0 ) δ(x − x0 ) dx = f (x0 ). (73)
−∞ −∞ −∞

The reasoning is the same as in Eq. (72), except that now the only value of f (x) that matters
is f (x0 ), because the delta function is zero everywhere except at x = x0 (because this is
where the argument of the delta function is zero). This result is interesting. It says that
a delta function can be used to “pick out” the value of a function at any point x = x0 , by
simply integrating the product δ(x − x0 )f (x).
A note on terminology: A delta function is technically not a function, because there is no
way for a function that is nonzero at only one point to have an area of 1. A delta function
is actually a distribution, which is something that is defined through its integral when
multiplied by another function. So a more proper definition of a delta function (we’ll still
call it that, since that’s the accepted convention) is the relation in Eq. (72) (or equivalently
Eq. (73)):
• A delta function, δ(x), is a distribution that satisfies
Z ∞
δ(x)f (x) dx = f (0). (74)
−∞
24 CHAPTER 3. FOURIER ANALYSIS

However, having said this, the most convenient way to visualize a delta function is usually
to think of it as a very tall and very thin function whose area is 1. This function could be
the thin rectangle in Fig. 24, but there are infinitely many other functions that work just
as well. For example, we can have a tall and thin triangle with base 2a and height 1/a. Or
a tall and thin Gaussian whose parameters are related in a certain way (as we’ll see below).
And so on.
As an exercise, you can verify the following properties of the delta function:

δ(x) = δ(−x)
δ(x)
δ(ax) =
|a|
δ(x − x0 )
δ(g(x)) = if g(x0 ) = 0
|g 0 (x0 )|
Z
δ 0 (x)f (x) dx = −f 0 (0). (75)

The second property follows from a change of variables in Eq. (74). The third follows
from expanding g(x) in a Taylor series around x0 and then using the second property.
The fourth follows from integration by parts. In the third property, if g(x) has multiple
zeros,
P then δ(g(x)) consists of a delta function at each zero, which means that δ(g(x)) =
δ(x − xi )/|g 0 (xi )|.

3.5.2 The Fourier transform of f (x) = 1

It turns out that the delta function plays an important role with regard to Fourier trans-
forms, in that it is the Fourier transform of one of the simplest functions of all, namely the
constant function, f (x) = 1. Consistent with the general “inverse shape” behavior of f (x)
and C(k) that we observed in the examples in Section 3.4, the function f (x) = 1 is infinitely
spread out, and the delta function is infinitely sharply peaked.
To calculate the Fourier transform of f (x) = 1, it seems like all we have to do is plug
f (x) = 1 into the second equation in Eq. (43). This gives
Z ∞
1
C(k) = (1)e−ikx dx. (76)
2π −∞

R∞
This is an interesting integral. If k = 0, then we have −∞ (1)(1) dx, which is infinite. And if
k 6= 0, then the integrand oscillates indefinitely in both the positive and negative directions,
so it’s unclear what the integral is. If we use large but finite limits, and if e−ikx undergoes
an integral number of cycles within these limits, then the integral is zero. But if there is a
“leftover” partial cycle, then the integral is nonzero. The integral therefore depends on the
limits of integration, and hence is not well defined. Therefore, our first task in calculating
the integral is to make sense of it. We can do this as follows.
Instead of having the function be identically 1 for all x, let’s have it be essentially equal
to 1 for a very large interval and then slowly taper off to zero for very large x. If it tapers off
to zero on a length scale of `, then after we calculate the integral in terms of `, we can take
the ` → ∞ limit. In this limit the function equals the original constant function, f (x) = 1.
There are (infinitely) many ways to make the function taper off to zero. We’ll consider a
few of these below, and in the process we’ll generate various representations of the delta
function. We’ll take advantage of the results for the special functions we derived in Section
3.4.
3.5. THE DELTA FUNCTION 25

3.5.3 Representations of the delta function

Gaussian
A standard way to make the f (x) = 1 function taper off is to use the Gaussian function,
2
e−ax , instead of the constant function, 1. If a is very small, then this Gaussian function is
essentially equal to 1 for a large range of x values, and
√ then it eventually tapers off to zero
for large x. It goes to zero on a length√scale of ` ∼ 1/ a. This is true because the function
−1/100
is very close to 1 for, say, x = 1/(10 a), √ where it equals e −25 ≈ 1 − 1/100. And it is
essentially equal to zero for, say, x = 5/ a, where it equals e ≈ 0. In the end, we will
take the a → 0 limit, which will make the Gaussian function be essentially equal to 1 for
all x. The point is that if we jumped right to the a = 0 case, there would be no way to get
a handle on the integral. But if we work with a finite value of a, we can actually calculate
the integral, and then we can take the a → 0 limit.
2
We calculated the Fourier transform of the Gaussian function f (x) = Ae−ax in Eq.
(67). With A = 1 here, the result is the Gaussian function of k,
2
e−k /4a
C(k) = √ . (77)
2 πa

In the a → 0 limit, the a in the exponent

√ makes this Gaussian fall off to zero very quickly
(it goes to zero on a length scale of 2 a). And the a in the denominator makes the value
at k = 0 very large. The Gaussian is therefore sharply peaked around k = 0. So the Fourier
transform of a very wide Gaussian is a sharply peaked one, as shown in Fig. 25, for a = 0.02.
This makes sense, because if f (x) is very wide, then it is made up of only very slowly varying
exponential/sinusoidal functions. In other words, C(k) is dominated by k values near zero.

(a = 0.02)
2.0 2.0 2
e-k /4a
C(k) = _____
2
f(x) = e-ax
1.5 1.5
2 πa

1.0 1.0

0.5 0.5

x k
-5 0 5 -5 0 5

Figure 25
R∞ 2 p
What is the area under the C(k) curve? Using the −∞
e−ax dx = π/a result from
Eq. (66), we have
Z ∞ Z ∞ 2
e−k /4a
C(k) dk = √ dk
−∞ −∞ 2 πa
r
1 π
= √ ·
2 πa 1/4a
= 1. (78)
2
So the area under the Fourier transform of the Gaussian function f (x) = e−ax equals 1.
This is a nice result. And it is independent of a. Even if a is very small, so that the Fourier
transform C(k) is very sharply peaked, the area is still 1. We have therefore found the
26 CHAPTER 3. FOURIER ANALYSIS

following representation of the delta function:

2
e−k /4a
δ(k) = lim √ (79)
a→0 2 πa

Remarks:
1. We should say that the above process of finding the Fourier transform of the constant function
f (x) = 1 was by no means necessary for obtaining this representation of the delta function.
Even if we just pulled Eq. (79) out of a hat, it would still be a representation, because it
has the two required properties of being infinitely sharply peaked and having area 1. But we
wanted to motivate it by considering it to be the Fourier transform of the constant function,
f (x) = 1.
2. It is no coincidence that the integral in Eq. (78) equals 1. It is a consequence of the fact that
the value of the f (x) Gaussian function equals 1 at x = 0. To see why, let’s plug the C(k)
from Eq. (77) into the first equation in Eq. (43). And let’s look at what this equation says
when x takes on the particular value of x = 0. We obtain
Z ∞ 2 Z ∞ 2
e−k /4a ik(0) e−k /4a
f (0) = √ e dk =⇒ 1 = √ dk, (80)
−∞ 2 πa −∞ 2 πa

which is just what Eq. (78) says. This is a general result: if C(k) is the Fourier transform of
f (x), then the area under the C(k) curve equals f (0). ♣

Exponential, Lorentzian
Another way to make the f (x) = 1 function taper off is to use the exponential function,
e−b|x| . If b is very small, then this exponential function is essentially equal to 1 for a large
range of x values, and then it eventually tapers off to zero for large x. It goes to zero on a
length scale of 1/b.
We calculated the Fourier transform of the exponential function Ae−b|x| in Eq. (70).
With A = 1 here, the result is the Lorentzian function of k,

b
C(k) = . (81)
π(b2 + k 2 )

In the b → 0 limit, this function is very tall, because the value at k = 0 is 1/πb. And it is
very narrow, because as we mentioned in Section 3.4.2, for any given value of k, the value
of C(k) is less than b/πk 2 , which goes to zero as b → 0. The Fourier transform of a wide
exponential is therefore a sharply peaked Lorentzian, as shown in Fig. 26.

(b = 0.1)
3.5 3.5
_______
b
3.0 f(x) = e-b|x| 3.0 C(k) =
2.5 2.5
π(b2+k2)
2.0 2.0
1.5 1.5
1.0 1.0
0.5 0.5
x k
-4 -2 0 2 4 -4 -2 0 2 4

Figure 26
3.5. THE DELTA FUNCTION 27

What is the area under the Lorentzian curve? From the above remark containing Eq.
(80), we know that
R ∞ the area must equal 1, but let’s check this by explicitly evaluating the
integral. Using −∞ b dk/(b2 + k 2 ) = tan−1 (k/b), we have
Z ∞
b dk 1³ ´ 1 ³ π −π ´
2 2
= tan−1 (∞) − tan−1 (∞) = − = 1. (82)
−∞ π(b + k ) π π 2 2

Since the area is 1, and since the curve is sharply peaked in the b → 0 limit, the Lorentzian
function in Eq. (81) therefore gives another representation of the delta function:

b
δ(k) = lim (83)
b→0 π(b2 + k2 )

This makes sense, because the b → 0 limit of the exponential function e−b|x| is essentially the
2
same function as the a → 0 limit of the Gaussian function a−ax (they both have f (x) = 1
as the limit), and we found above in Eq. (79) that the Fourier transform of a Gaussian is a
delta function in the a → 0 limit.
We can look at things the other way, too. We can approximate the f (x) = 1 function
by a very wide Lorentzian, and then take the Fourier transform of that. The Lorentzian
function that approximates f (x) = 1 is the b → ∞ limit of f (x) = b2 /(b2 + x2 ), because
this falls off from 1 when x is of order b. Assuming (correctly) that the Fourier transform
process is invertible, we know that the Fourier transform of f (x) = b2 /(b2 + x2 ) must be
an exponential function. The only question is what the factors are. We could do a contour
integration if we wanted to produce the Fourier transform from scratch, but let’s do it the
easy way, as follows.
The known Fourier-transform relation between f (x) = e−b|x| and C(k) = b/π(b2 + k 2 ),
which we derived in Eqs. (69) and (70), is given by Eq. (43) as
Z ∞ Z ∞
b b 1
e−b|x| = 2 2
e ikx
dk, where = e−b|x| e−ikx dx. (84)
−∞ π(b + k ) π(b2 + k 2 ) 2π −∞

Interchanging the x and k letters in the first of these equations, shifting some factors around,
and changing eikx to e−ikx (which doesn’t affect things since the rest of the integrand is an
even function of x) gives
Z ∞
1 b2
(b/2)e−b|k| = e−ikx dx. (85)
2π −∞ (b2 + x2 )

But this is the statement that the Fourier transform of f (x) = b2 /(b2 + x2 ) (which is
essentially the constant function f (x) = 1 if b → ∞) is C(k) = (b/2)e−b|k| .6
This C(k) = (b/2)e−b|k| exponential function is infinitely sharply-peaked in the b → ∞
limit. And you can quickly show that it has an area of 1 (but again, this follows from the
above remark containing Eq. (80)). So we have found another representation of the delta
function:
δ(k) = lim (b/2)e−b|k| (86)
b→∞

Just like with the wide Gaussian and exponential functions, the wide Lorentzian function
is essentially equal to the f (x) = 1 function, so its Fourier transform should be roughly the
same as the above ones. In other words, it should be (and is) a delta function.
6 By “Fourier transform” here, we mean the transform in the direction that has the 1/2π in front of the

integral, since that’s what we set out to calculate in Eq. (76).

28 CHAPTER 3. FOURIER ANALYSIS

Square wave, sinc

Another way to make the f (x) = 1 function “taper off” is to use a square-wave function
that takes on the constant value of 1 out to x = ±a and then abruptly drops to zero. We
calculated the Fourier transform of the square wave in Eq. (71). With A = 1 here, the result
is the “sinc” function of k,
sin(ka)
C(k) = . (87)
πk
Using sin ² ≈ ² for small ², we see that C(k) ≈ a/π for k values near zero. So if a is large
then C(k) is very tall near k = 0. The Fourier transform of a very wide square wave (large
a) is therefore a very tall sinc function, as shown in Fig. 27. This is the same behavior we
saw with the Gaussian, exponential, and Lorentzian functions.

f(x) C(k)
(a = 40) sin(ka)
0.3 C(k) = ______
10 πk
0.2
5
0.1

x k
-40 40 -2 -1 1 2

Figure 27

However, as we saw in Fig. 23 in Section 3.4.3, the “sharply peaked” characteristic

seems to be missing. The envelope of the oscillatory curve is the function 1/πk, and this
is independent of a. As a increases, the oscillations become quicker, but the envelope
doesn’t fall off any faster. But when it comes to considering C(k) to be a delta function, the
important property is its integral. If C(k) oscillates very quickly, then it essentially averages
out to zero. So for all practical purposes, C(k) is basically the zero function except right
near the origin. So in that sense it is sharply peaked.
Furthermore, we know from the above remark containing Eq. (80) (or a contour in-
tegration would also do the trick) that the (signed) area under the C(k) curve equals 1,
independent of a. We therefore have yet another representation of the delta function:

sin(ka)
δ(k) = lim (88)
a→∞ πk

We can look at things the other way, too. We can approximate the f (x) = 1 function by
a very wide sinc function, and then take the Fourier transform of that. The sinc function
that approximates f (x) = 1 is the a → 0 limit of sin(ax)/(ax), because this equals 1 at
x = 0, and it falls off from 1 when x is of order 1/a. Assuming (correctly) that the Fourier-
transform process is invertible, you can show that the above Fourier-transform relation
between the square wave and sin(ka)/πk implies (with the 2π in the opposite place; see the
step going from Eq. (84) to Eq. (85)) that the Fourier transform of f (x) = sin(ax)/(ax) is
a square-wave function of k that has height 1/2a and goes from −a to a. This is a very tall
and thin rectangle in the a → 0 limit. And the area is 1. So we have returned full circle
to our original representation of the delta function in Fig. 24 (with the a in that figure now
2a).
Again, just like with the other wide functions we discussed above, a wide sinc function
is essentially equal to the f (x) = 1 function, so its Fourier transform should be roughly the
same as the above ones. In other words, it should be (and is) a delta function.
3.6. GIBBS PHENOMENON 29

Integral representation
There is one more representation of the delta function that we shouldn’t forget to write
down. We’ve actually been using it repeatedly in this section, so it’s nothing new. As we’ve
seen many times, the Fourier transform of the function f (x) = 1 is a delta function. So we
should write this explicitly:
Z ∞
1
δ(k) = e−ikx dx (89)
2π −∞
where we haven’t bothered to write the “1” in the integrand. Of course, this integral doesn’t
quite make sense the way it is, so it’s understood that whenever it comes up, we make the
integrand taper off to zero by our method of choice (for example, any of the above methods),
and then eventually take the limit where the tapering length scale becomes infinite. More
generally, we have Z ∞
1
δ(k − k0 ) = e−i(k−k0 )x dx. (90)
2π −∞
Similar generalizations to a k − k0 argument hold for all of the above representations too,
of course.
Eq. (90) allows us to generalize the orthogonality relation in Eq. (22), which was
Z L
ei2πnx/L e−i2πmx/L dx = Lδnm . (91)
0

In that case, the inner product was defined as the integral from 0 to L. But if we define a
different inner product that involves the integral from −∞ to ∞ (we are free to define it
however we want), then Eq. (90) gives us a new orthogonality relation,
Z ∞
eik1 x e−ik2 x dx = 2πδ(k2 − k1 ). (92)
−∞

(Alternatively, we could define the inner product to be 1/2π times the integral, in which case
the 2π wouldn’t appear on the righthand side.) So the eikx functions are still orthogonal;
the inner product is zero unless the k values are equal, in which case the inner product is
infinite. We see that whether we define the inner product as the integral from 0 to L, or
from −∞ to ∞, we end up with a delta function. But in the former case it is the standard
“Kronecker” delta, whereas in latter case it is the “Dirac” delta function.

3.6 Gibbs phenomenon f(x)

A
Let’s return to the study of Fourier series, as opposed to Fourier transforms. The “Gibbs phe-
nomenon” refers to the manner in which the Fourier series for a periodic function overshoots
the values of the function on either side of a discontinuity. This effect is best illustrated x
-L/2 L/2
(for many reasons, as we will see) by the periodic step function shown in Fig. 28. This step
function is defined by ½
−A (−L/2 < x < 0)
f (x) = (93)
A (0 < x < L/2), Figure 28
and has period L. We already found the Fourier series for this function in Eq. (51). It is
4A X 1 ³ 2πnx ´
f (x) = sin
π n L
n odd
µ ³ ¶
4A 2πx ´ 1 ³ 6πx ´ 1 ³ 10πx ´
= sin + sin + sin + ··· . (94)
π L 3 L 5 L
30 CHAPTER 3. FOURIER ANALYSIS

By “n odd” we mean the positive odd integers. If we plot a few partial sums (with A =
L = 1), we obtain the plots shown in Fig. 29 (this is just a repeat of Fig. 7). The first plot
includes only the first term in the series, while the last plot includes up to the 50th term
(the n = 99 one).

f(x)
(1 term) 1.0
(3 terms)
1.0
0.5 0.5
x
-1.0 - 0.5 0.5 1.0 -1.0 - 0.5 0.5 1.0
- 0.5 - 0.5
-1.0 -1.0

1.0
(10 terms) 1.0
(50 terms)
0.5 0.5

-1.0 - 0.5 0.5 1.0 -1.0 - 0.5 0.5 1.0

- 0.5 - 0.5
-1.0 -1.0

Figure 29

The more terms we include, the more the curve looks like a step function. However,
there are two undesirable features. There are wiggles near the discontinuity, and there is
an overshoot right next to the discontinuity. As the number of terms grows, the wiggles
get pushed closer and closer to the discontinuity, in the sense that the amplitude in a given
region decreases as the number of terms, N , in the partial series increases. So in some sense
the wiggles go away as N approaches infinity. However, the overshoot unfortunately never
goes away. Fig. 30 shows the zoomed-in pictures near the discontinuity for N = 10 and
N = 100. If we included more terms, the overshoot would get pushed closer and closer to
the discontinuity, but it would still always be there.

f(x)
1.20
(10 terms) 1.20
(100 terms)
1.15 1.15
1.10 1.10
1.05 1.05
1.00 1.00
0.95 0.95
0.90 0.90
0.85 0.85
0.80 x 0.80
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20

Figure 30

It turns out that as long as the number of terms in the partial series is large, the height
of the overshoot is always about 9% of the jump in the function, independent of the (large)
number N . This is consistent with the plots in Fig. 30, where the overshoot is about 0.18,
which is 9% of the jump from −1 to 1. Furthermore, this 9% result holds for a discontinuity
in any function, not just a step function. This general result is provable in a page or two,
3.6. GIBBS PHENOMENON 31

so let’s do that here. Our strategy will be to first prove it for the above step function, and
then show how the result for any other function follows from the specific result for the step
function.
To demonstrate the 9% the result for the step function, our first task is to find the
location of the maximum value of the overshoot. Let’s assume that we’re truncating the
series in Eq. (94) at the N th term (the n = 2N − 1 one), and then we’ll take the N → ∞
limit. To find the location of the maximum, we can take the derivative of the partial series
(which we’ll call fN (x)) and set the result equal to zero. This gives
µ ³ 2πx ´ ³ 6πx ´ ³ 2(2N − 1)πx ´¶
dfN (x) 8A
0= = cos + cos + · · · + cos . (95)
dx L L L L
The smallest solution to this equation is x = L/4N . This is a solution because it makes
the arguments of the cosines in the first and last terms in Eq. (95) be supplementary, which
means that the cosines cancel. And likewise for the second and second-to-last terms. And
so on. So all of the terms cancel in pairs. (If N is odd, there is a middle term that is zero.)
Furthermore, there is no (positive) solution that is smaller than this one, for the following
reason.
For concreteness, let’s take N = 6. If we start with a very small value of x, then the
sum of the cosine terms in Eq. (95) (there are six in the N = 6 case) equals the sum of
the horizontal components of the vectors in the first diagram in Fig. 31. (The vectors make
angles 2πnx/L with the horizontal axis, where n = 1, 3, 5, 7, 9, 11.) This sum is clearly not
zero, because the four vectors are “lopsided” in the positive direction.

small x slightly larger x x = L/4N

smallest angle = 2πx/L
largest angle = 2(2N-1)πx/L (N = 6 here)
angle spacing = 4πx/L

Figure 31

If we increase x a little, we get the six vectors in the second diagram in Fig. 31. The
vectors are still lopsided in the positive direction, so the sum again isn’t zero. But if we keep
increasing x, we finally get to the situation shown in the third diagram in Fig. 31. The vectors
are now symmetric with respect to the y axis, so the sum of the horizontal components is
zero. This is the x = L/4N case we found above, where the vectors are supplementary
in pairs (setting the sum of the first and N th angles equal to π gives x = L/4N ). So
x = L/4N is indeed the smallest positive solution to Eq. (95) and is therefore the location
of the overshoot peak in Fig. 30.
Having found the location of the maximum value of the overshoot, we can plug x = L/4N
into Eq. (94) to find this maximum value. We obtain
µ ¶ µ ³ ³ (2N − 1)π ´¶
L 4A π ´ 1 ³ 3π ´ 1
fN = sin + sin + ··· + sin
4N π 2N 3 2N 2N − 1 2N
N µ ¶
4A X 1 (2m − 1)π
= sin . (96)
π m=1 2m − 1 2N
 32 CHAPTER 3. FOURIER ANALYSIS

In the N → ∞ limit, the arguments of the sines increase essentially continuously from 0
to π. Therefore, because of the (2m − 1) factor in the denominator, this sum is essentially
the integral of sin(y)/y from 0 to π, except for some factors we need to get straight. If we
multiply Eq. (96) through by 1 in the form of (π/2N )(2N/π), we obtain
µ ¶ N µ ¶
L π 4A X 2N (2m − 1)π
fN = · sin . (97)
4N 2N π m=1 (2m − 1)π 2N
Rπ¡ ¢
Each term in this sum is weighted by a factor of 1, whereas in the integral 0 sin(y)/y dy
each term is weighted by dy. But dy is the difference between successive values of (2m −
1)π/2N , which is 2π/2N . So dy = π/N . The above sum is therefore N/π times the integral
from 0 to π. So in the N → ∞ limit we obtain
µ ¶ µ Z ¶
L π 4A N π sin y
fN = · dy
4N 2N π π 0 y
Z
2A π sin y
= dy. (98)
π 0 y

Alternatively, you can systematically convert the sum in Eq. (96) to the integral in Eq. (98)
by defining
(2m − 1)π π dm
y≡ =⇒ dy = . (99)
2N N
If you multiply Eq. (96) by dm (which doesn’t affect anything, since dm = 1) and then
change variables from m to y, you will obtain Eq. (98).
The value of the actual step function just to the right of the origin is A, so to get the
overshoot, we need to subtract off A from f (L/4N ). And then we need to divide by 2A
(which is the total height of the jump) to obtain the fractional overshoot. The result is
µ ³ ¶ Z
1 L ´ 1 π sin y 1
f −A = dy − ≈ 0.0895 ≈ 9%, (100)
2A 4N π 0 y 2

where we have calculated the integral numerically. For the simple step function, we have
f(x) therefore demonstrated the desired 9% result.
As we mentioned above, this 9% result also holds for any discontinuity in any other
function. Having obtained the result for the simple step function, the generalization to
other arbitrary functions is surprisingly simple. It proceeds as follows.
Consider an arbitrary function f (x) with period L and with a discontinuity, as shown in
the first plot in Fig. 32. Without loss of generality, we can assume that the discontinuity in
x=0 x=L question (there may very well be others) is located at the origin. And we can also assume
that the discontinuity is vertically symmetric around the origin, that is, it goes from −A
to A (or vice versa) for some value of A. This assumption is valid due to the fact that
fstep(x)
shifting the function vertically simply changes the a0 value in Eq. (1), which doesn’t affect
the nature of the overshoot.
Now consider the periodic step function, fstep (x), that jumps from −A to A and also has
period L, as shown in the second plot in Fig. 32. And finally consider the function fdiff (x)
defined to be the difference between f (x) and fstep (x), as shown in the third plot in Fig.
32. So we have
f (x) = fstep (x) + fdiff (x). (101)
fdiff(x) Since fdiff (x) = f (x) − fstep (x), the plot of fdiff (x) is obtained by simply subtracting or
adding A from f (x), depending on whether the step function is positive or negative at x,
respectively. The critical point to realize is that by construction, fdiff (x) is a continuous

continuous

Figure 32
3.6. GIBBS PHENOMENON 33

function at the discontinuity at x = 0 (and x = L, etc.). Its derivative might not be

continuous (and in fact isn’t, in the present case), but that won’t matter here. Also, fdiff (x)
may very well have discontinuities elsewhere (in fact, we introduced one at x = L/2), but
these aren’t relevant as far as the discontinuity at the origin goes.
Since fdiff (x) is continuous at x = 0, its Fourier series has no overshoot there (see the first
remark below). The Fourier series for fdiff (x) therefore contributes nothing to the overshoot
of the Fourier series of f (x) at x = 0. Hence, in view of Eq. (101), the overshoot in the
Fourier series for f (x) is exactly equal to the overshoot in the Fourier series for fstep (x).
The same 9% result therefore holds, as we wanted to show.
Remarks:
1. We claimed in the previous paragraph that since fdiff (x) is continuous at x = 0, its Fourier
series has no overshoot there. Said in another way, a kink in fdiff (x) (that is, a discontinuity
in its derivative) doesn’t cause an overshoot. The can be shown with the same strategy that
we used above, where we looked at a specific function (the step function) and then argued
why any other function gave the same result.
For the present task, we’ll consider the specific triangular function shown in Fig. 35 below,
which we’ll discuss in Section 3.7. We’ll find that the Fourier series for this triangular function
is smooth at x = 0. There is no overshoot. Now, any function fdiff (x) with a kink looks
(at least locally near the kink) like a triangular function plus possibly a linear function of
x, which has the effect of tilting the triangle so that the slopes on either side of the kink
aren’t equal and opposite. The triangular function has no overshoot, and the linear function
certainly doesn’t have one, either. So we conclude that fdiff (x) doesn’t have an overshoot, as
we wanted to show.
2. The reasoning in this section can also be used to prove a quick corollary, namely, that the
value of the Fourier series at a discontinuity equals the midpoint of the jump. This follows
because it is true for the step function (the value at x = 0 is zero, because we have only sines
in the series), and it is also true for fdiff (x) (in a trivial sense, because fdiff (x) is continuous
and has no jump). Also, any a0 constant we stripped off from the original function doesn’t
affect the result, because it increases the endpoints and midpoint of the jump by the same
amount.
3. The overshoot (which occurs at x = L/4N ) gets squeezed arbitrarily close to the discontinuity
as N becomes large. So for any given nonzero value of x near the discontinuity, the overshoot
is irrelevant as far as the value of f (x) goes. Also, at any given nonzero value of x the
amplitude of the wiggles in Fig. 30 goes to zero (we’ll show this in the following subsection).
Therefore, the Gibbs phenomenon is irrelevant for any given nonzero value of x, in the N → ∞
limit. The only possibly value of x where we might have an issue is x = 0 (and x = L, etc.).
From the previous remark, we know that the value of the series at x = 0 is the midpoint of
the jump. This may or may not equal the arbitrary value we choose to assign to f (0), but
we can live with that. ♣

The shape of the wiggles

The strategy that we used above, where we turned the sum in Eq. (94) (or rather, the partial
sum of the first N terms) into the integral in Eq. (98), actually works for any small value
of x, not just the x = L/4N location of the overshoot. We can therefore use this strategy
to calculate the value of the Fourier series at any small value of x, in the large-N limit.7 In
other words, we can find the shape (and in particular the envelope) of the wiggles in Fig.
30.
7 We need x to be small compared with L, so that the sine terms in Eq. (102) below vary essentially

continuously as m varies. If this weren’t the case, then we wouldn’t be able to approximate the sum in Eq.
(102) by an integral, which we will need to do.
34 CHAPTER 3. FOURIER ANALYSIS

The procedure follows the one above. First, write the partial sum of the first N terms
in Eq. (94) as

4A X
N
1 ³ 2(2m − 1)πx ´
fN (x) = sin . (102)
π m=1 2m − 1 L

Then define
2(2m − 1)πx 4πx dm
y≡ =⇒ dy = . (103)
L L

Then multiply Eq. (102) by dm (which doesn’t affect anything, since dm = 1) and change
variables from m to y. The result is (for large N )

Z (4N −2)πx/L
2A sin y
fN (x) ≈ dy. (104)
π 0 y

As double check on this, if we let x = L/4N then the upper limit of integration is equal to
π (in the large-N limit), so this result reduces to the one in Eq. (98).
We were concerned with the N → ∞ limit in the above derivation of the overshoot, but
now we’re concerned with R ∞just large (but not infinite) N . If we actually set N = ∞ in Eq.
(104), we obtain (2A/π) 0 (sin y)/y · dy. This had better be equal to A, because that is the
value of the step function at x. And indeed it is, because this definite integral equals π/2
(see the paragraph preceding Eq. (88); it boils down to the fact that (sin y)/y is the Fourier
transform of a square wave). But the point is that for the present purposes, we want to
keep N finite, so that we can actually see the x dependence of fN (x). (So technically the
integral approximation to the sum isn’t perfect like it was in the N = ∞ case. But for large
N it’s good enough.)
To make the integral in Eq. (104) easier to deal with, let’s define n via x ≡ n(L/4N ).
The parameter n need not be an integer, but if it is, then it labels the nth local maximum
or minimum of fN (x), due to the reasoning we used in Fig. 31. The number n gives the
number of half revolutions the vectors in Fig. 31 make as they wrap around in the plane. If
n is an integer, then the sum of the horizontal projections (the cosines in Eq. (95)) is zero,
and we therefore have a local maximum or minimum of fN (x).
With this definition of n, the upper limit of the integral in Eq. (104) is essentially equal
to nπ (assuming N is large), so we obtain
Z nπ
2A sin y
fN (n) ≈ dy (where x ≡ n(L/4N )). (105)
π 0 y

Again, n need not be an integer. What does this function of n look like? The integral has
to be computed numerically, and we obtain the first plot shown in Fig. 33 (we have chosen
A = 1). The second plot is a zoomed-in (vertically) version of the first one. Note that there
is actually no N dependence in fN (n). It is simply a function of n. Therefore, the height
of the nth bump is independent of N , assuming that N is reasonably large.8 The N (and
L) dependence comes in when we convert from n back to x.

8 If N isn’t large, then we can’t make the approximation that the upper limit of the integral in Eq. (105)

equals nπ. From Eq. (104), the limit actually equals nπ(1 − 1/2N ). But if N isn’t large, then the integral
approximation to the sum isn’t so great anyway.
3.6. GIBBS PHENOMENON 35

fN(n)
fN(n)
1.2 (zoomed in vertically)
1.20
1.0
1.15
0.8 1.10
0.6 1.05
1.00
0.4
0.95
0.2 0.90
0.0 n 0.85 n
0 5 10 15 20 5 10 15 20

Figure 33
You can verify in Fig. 33 that the local maxima and minima occur at integer values of n.
This curve has the same shape as the two curves in Fig. 30, with the only difference being
the horizontal scale. Any actual step-function Fourier series such as the ones in Fig. 30
can be obtained from Fig. 33 by squashing or stretching it horizontally by the appropriate
amount. This amount is determined by making the n = 1 maximum occur at x = L/4N ,
and then the n = 2 minimum occur at x = 2L/4N , and so on. The larger N is, the smaller
these x values are, so the more squashed the wiggles are. If we increase N by a factor of,
say, 10, then the wiggles in the Fourier-series plot get squashed by a factor of 10 horizontally
(and are unchanged vertically). We can verify this in Fig. 30. If we look at, say, the fourth
maximum, we see that in the N = 10 plot it occurs at about x = 0.175, and in the N = 100
plot it occurs at slightly less than x = 0.02 (it’s hard to see exactly, but it’s believable that
it’s about 0.0175).
sin y
_______
y
0.2
The envelope
0.1
What is the envelope of each of the curves in Fig. 30? (We’ll be concerned with just the 5 10 15 20 25 30
y
asymptotic behavior, that is, not right near the discontinuity.) To answer this, we must first - 0.1
find the envelope of the fN (n) curve in Fig. 33. Fig. 34 shows a plot of (sin y)/y. Up to a - 0.2
factor of 2A/π, the function fN (n) equals the area under the (sin y)/y curve, out to y = nπ.
Figure 34
This area oscillates around its average value (which is π/2) due to the “up” and “ down”
bumps of (sin y)/y.
A local maximum of fN (n) is achieved when (sin y)/y has just completed an “up” bump,
and a local minimum is achieved when (sin y)/y has just completed a “down” bump. The
amplitude of the oscillations of fN (n) is therefore 2A/π times half of the area of one of the
bumps at y = nπ. The area of a simple sine bump is 2, so half of the area of a bump of
(sin y)/y, in a region where y is approximately equal to nπ, is (1/2) · 2/nπ = 1/nπ. (We’re
assuming that n is reasonably large, which means that the value of 1/nπ is approximately
constant over the span of a bump.) So the amplitude of the oscillations of fN (n) is
2A 1 2A A
Amplitude = · = ≈ . (106)
π nπ nπ 2 5n
You can (roughly) verify this by sight in Fig. 33. Remember that n counts the number of
maxima and minima, not just the maxima.
To find the envelope of the actual Fourier-series plots in Fig. 30, we need to convert from
n back to x. Using x ≡ n(L/4N ) =⇒ n ≡ 4N x/L, the amplitude becomes
2A AL 1
Amplitude = 2
= 2
∝ . (107)
(4N x/L)π 2π N x Nx
So up to some constants, this is a 1/x function with a 1/N coefficient. The larger N is, the
quicker it dies off to zero.
 36 CHAPTER 3. FOURIER ANALYSIS

What if f (x) isn’t a square wave and instead has a nonzero slope emanating from the
discontinuity? The reasoning above (where we wrote f (x) as fstep (x) + fdiff (x)) tells us that
we still have the same envelope of the form, AL/2π 2 N x, with the only difference being that
the envelope is now measured with respect to a line with nonzero slope.

3.7 Convergence
Fourier’s theorem states that any sufficiently well-behaved function can be written as the
series in Eq. (1), where the an and bn coefficients are given by Eqs. (9,11,12). But what
does this statement actually mean? Does it mean only that the value of the Fourier series
at a given x equals the value of the function f (x) at this x? Or does it mean that the two
f(x) functions on either side of Eq. (1) are actually the exact same function of x? In other words,
0.3
is it the case that not only the values agree, but all the derivatives do also?
0.2
0.1 It turns out that Fourier’s theorem makes only the first of these claims – that the values
x are equal. It says nothing about the agreement of the derivatives. They might agree, or
-1.0 - 0.5 - 0.1 0.5 1.0
they might not, depending on the function. Let’s give a concrete example to illustrate this.
- 0.2
- 0.3 Consider the periodic triangular function shown in Fig. 35. For ease of notation, we have
chosen the period L to be 1. The function is defined by
Figure 35 ½
x + 1/4 (−1/2 < x < 0)
f (x) = (108)
−x + 1/4 (0 < x < 1/2).
F(x) (5 terms)
0.3
The task of Problem [to be added] is to find the Fourier series for this function. The result
0.2
0.1 is
x
-1.0 - 0.5 - 0.1 0.5 1.0 2 X cos(2πnx)
F (x) =
- 0.2 π2 n2
- 0.3 n odd
µ ¶
2 cos(6πx) cos(10πx)
Figure 36 = cos(2πx) + + + · · · . (109)
π2 9 25

f'(x) We’re using F (x) to denote the Fourier series, to distinguish it from the actual function
1.0 f (x). Due to the n2 factors in the denominators, the partial sums of F (x) converge very
0.5 quickly to the actual triangular function f (x). This is evident from Fig. 36, which gives a
x very good approximation to f (x) despite including only five terms in the partial sum.
-1.0 - 0.5 0.5 1.0 What if we calculate the derivative of the Fourier series, F 0 (x), and compare it with the
- 0.5
derivative of the actual function, f 0 (x)? The slope of f (x) is either 1 or −1, so f 0 (x) is
-1.0 simply the step function shown in Fig. 37. The derivative of F (x) is easily calculated by
Figure 37 differentiating each term in the infinite sum. The result is

4 X sin(2πnx)
F 0 (x) = −
F'(x) (10 terms) π n
n odd
1.0 µ ¶
4 sin(6πx) sin(10πx)
0.5 = − sin(2πx) + + + ··· . (110)
x
π 3 5
-1.0 - 0.5 0.5 1.0
- 0.5 This is the (negative of the) Fourier series we already calculated in Eq. (94), with A = 1
-1.0
and L = 1. The plots of the partial sums of F 0 (x) with 10 and 100 terms are shown in
Fig. 38. In the infinite-sequence limit, F 0 (x) does indeed equal the f 0 (x) step function in
Fig. 37. The Gibbs phenomenon discussed in the previous section might make you think
(100 terms)
otherwise, but the Gibbs overshoot is squeezed infinitely close to the discontinuity in the
1.0 infinite-sequence limit. So for any given value of x that isn’t at the discontinuity, F 0 (x)
0.5 approaches the correct value of ±1 in the infinite-sequence limit.
-1.0 - 0.5 0.5 1.0
- 0.5
-1.0

Figure 38
 f ''(x)
3.7. CONVERGENCE 37

-1.0 1.0
What if we go a step further and calculate the second derivative of the Fourier series, x
-0.5 0.5
F 00 (x), and compare it with the second derivative of the actual function, f 00 (x)? The slope
of f 0 (x) is zero except at the discontinuities, where it is ±∞, so f 00 (x) is shown in Fig. 39.
The arrows indicate infinite values. These infinite value are actually delta functions (see
Problem [to be added]). The derivative of F 0 (x) is again easily calculated by differentiating Figure 39
each term in the infinite sum. The result is
(10 terms)
F''(x)
X 15
F 00 (x) = −8 cos(2πnx) 10
n odd 5
³ ´ x
= −8 cos(2πx) + cos(6πx) + cos(10πx) + · · · . (111) -1.0 - 0.5 - 5 0.5 1.0
-10
-15

This can also be obtained by finding the Fourier series coefficients for f 00 (x) via Eqs. (9,11,12)
(see problem [to be added]). Plots of partial sums of F 00 (x) with 10 and 40 terms are shown (40 terms)
in Fig. 40. We now have a problem. These curves don’t look anything like the plot of f 00 (x) 15
in Fig. 39. And if you plotted higher partial sums, you would find that the envelope is the 10
5
same, with the only difference being the increase in the frequency of the oscillations. The
curves definitely don’t converge to f 00 (x), a function that is zero everywhere except at the -1.0 - 0.5 - 5 0.5 1.0
discontinuities. -10
-15
Apparently, the Fourier series F 0 (x) equals the function f 0 (x) as far as the value is
concerned, but not as far as the derivative is concerned. This statement might sound a little Figure 40
odd, in view of the fact that the second plot in Fig. 38 seems to indicate that except at
F'(x) (10 terms)
the discontinuities, F 0 (x) approaches a nice straight line with slope zero. But let’s zoom in
on a piece of this “line” and see what it looks like up close. Let’s look at the middle of a - 0.96
- 0.98
step. The region between x = 0.2 and x = 0.3 is shown in Fig. 41 for partial sums with 10, - 1.00
30, and 100 terms. The value of the function converges to -1, but slope doesn’t converge to - 1.02
zero, because although the amplitude of the oscillations decreases, the frequency increases. - 1.04 x
0.22 0.24 0.26 0.28 0.30
So the slope ends up oscillating back and forth with the same amplitude. In short, the plot
for the 100-term partial sum is simply an (essentially) scaled down version of the plot for
the 10-term partial sum. This can be seen by looking at a further zoomed-in version of the - 0.96 (30 terms)
100-term partial sum, as shown in Fig. 42 (the additional zoom factor is 10 for both axes). - 0.98
This looks just like the plot for the 10-term partial sum. The proportions are the same, so - 1.00
the slope is the same. - 1.02
- 1.04
0.22 0.24 0.26 0.28 0.30
The general condition under which the value of the Fourier series F (x) equals the value
of the original function f (x) (except at isolated discontinuities) is that f (x) be square
integrable. That is, the integral (over one period) of the square of f (x) is finite. In the
- 0.96 (100 terms)
above example, f (x) is square integrable, consistent with the fact that F (x) agrees with - 0.98
f (x). Additionally, f 0 (x) is square integrable, consistent with the fact that F 0 (x) agrees - 1.00
with f 0 (x). - 1.02
- 1.04
However, f 00 (x) is not square integrable (because it contains delta functions), consistent 0.22 0.24 0.26 0.28 0.30
with the fact that F 00 (x) does not agree with f 00 (x). A quick way to see why the square of
Figure 41
a delta function isn’t integrable (in other words, why the integral is infinite) is to consider
the delta function to be a thin box with width a and height 1/a, in the a → 0 limit. The
square of this function is a thin box with width a and height 1/a2 . The area of this box is (100 terms,
F'(x) additional zoom)
a(1/a2 ) = 1/a, which diverges in the a → 0 limit. So (δ(x))2 isn’t integrable.
R Another even
- 0.996
quicker way is to use the main property of Ra delta function, namely δ(x)f (x) dx = f (0). - 0.998
Letting f (x) be a delta function here gives δ(x)δ(x) dx = δ(0) = ∞. Hence, (δ(x))2 is not - 1.000
integrable. - 1.002
- 1.004
x
0.2460.2480.2500.2520.254

Figure 42
 38 CHAPTER 3. FOURIER ANALYSIS

3.8 Relation between transforms and series

The exponential Fourier-series relations in Section 3.2 were

∞
X Z L
1
f (x) = Cn ei2πnx/L where Cn = f (x)e−i2πnx/L dx (112)
n=−∞
L 0

(We’ll work with exponential series instead of trig series in this section.) And the Fourier-
transform relations in Section 3.3 were
Z ∞ Z ∞
1
f (x) = C(k)eikx dk where C(k) = f (x)e−ikx dx (113)
−∞ 2π −∞

Eq. (112) says that any (reasonably well-behaved) periodic function can be written as a sum
of exponentials (or sines and cosines).9 But now that we’ve learned about Fourier transforms,
what if we pretend that we don’t know anything about Fourier series and instead simply
try to calculate the Fourier transform of a periodic function using Eq. (113)? A periodic
function can certainly be viewed as just a “normal” function that we might reasonably want
to find the Fourier transform of, so something sensible should come out of this endeavor.
Whatever result we obtain from Eq. (113), it had better somehow reduce to Eq. (112). In
particular, there had better not be any frequencies (k values) in the C(k) Fourier-transform
expression for f (x) besides the discrete ones of the form 2πn/L, because these are the only
ones that appear in Eq. (112). So let’s see what happens when we take the Fourier transform
of a periodic function. Actually, let’s first work backwards, starting with Eq. (112), and
see what the transform has to be. Then we’ll go in the “forward” direction and derive the
Fourier series directly by calculating the Fourier transform.

3.8.1 From series to transform

Fourier transforms involve integrals over k, so if we work backwards from the discrete sum in
Eq. (112), our goal is to somehow write this sum as an integral. This is where our knowledge
of delta functions will help us out. We know that a delta function picks out a particular
value of the integration variable, so if we have a sum of delta functions located at different
places, then they will turn a continuous integral into a discrete sum. Therefore, assuming
that we have a function f (x) that can be written in the form of Eq. (112), we’ll propose the
following Fourier transform of f (x):
Cδ(k) ∞
C-2 X
Cδ (k) = Cn δ(k − 2πn/L) (114)
C2 n=−∞
C0 C1

x where the Cn ’s are given by Eq. (112). This Cδ (k) function is shown schematically in
0 2π
__ 4π
__ Fig. 43 (it’s technically a distribution and not a function, but we won’t worry about that).
C-1 L L The arrows are the standard way to represent delta functions, with the height being the
coefficient of the delta function, which is Cn here. The value of Cδ (k) at the delta functions
Figure 43 is infinite, of course, so remember that the height of the arrow represents the area of the
delta function (which is all that matters when doing an integral involving a delta function),
and not the (infinite) value of the function.
9 The route we followed in Sections 3.1 and 3.2 was to simply accept this as fact. However, we’ll actually

prove it from scratch in Section 3.8.3 below.

3.8. RELATION BETWEEN TRANSFORMS AND SERIES 39

To determine if the Cδ (k) function in Eq. (114) is the correct Fourier transform of f (x),
we need to plug it into the first equation in Eq. (113) and see if it yields f (x). We find
Z ∞ Z ∞à X ∞
!
?
f (x) = Cδ (k)eikx dk = Cn δ(k − 2πn/L) eikx dk. (115)
−∞ −∞ n=−∞

As k runs from −∞ to ∞, the integrand is zeroR everywhere except at the delta functions,
where k takes the form, k = 2πn/L. Using the f (k)δ(k − k0 ) dk = f (k0 ) property of the
delta function, we see that the integration across each delta function simply yields the value
of the rest of the integrand evaluated at k = 2πn/L. This value is Cn ei2πnx/L . Therefore,
since n runs from −∞ to ∞, Eq. (115) becomes
∞
X
?
f (x) = Cn ei2πnx/L . (116)
n=−∞

And from the first equation in Eq. (112), the righthand side here does indeed equal f (x),
as desired. So the C(k) in Eq. (114) does in fact make the first equation in Eq. (113) true,
and is therefore the correct Fourier transform of f (x).

3.8.2 From transform to series

Given a periodic function f (x), let’s now work in the “forward” direction and derive the
expression for Cδ (k) in Eq. (114), which we just showed correctly leads to the series in Eq.
(112). We’ll derive this Cδ (k) from scratch by starting with the general Fourier-transform
expression for C(k) in Eq. (113). We’ll pretend that we’ve never heard about the concept
of a Fourier series for a periodic function.
Our goal is to show that if f (x) is periodic with period L, then the C(k) given in Eq.
(113) reduces to the Cδ (k) given in Eq. (114), where the Cn ’s are given by Eq. (112). That
is, we want to show that the C(k) in Eq. (113) equals
∞
à Z !
X 1 L −i2πnx/L
C(k) = f (x)e dx δ(k − 2πn/L) (117)
n=−∞
L 0

To demonstrate this, let’s break up the integral for C(k) in Eq. (113) into intervals of length
L. This gives
à Z Z L Z 2L !
0
1
C(k) = ··· f (x)e−ikx dx + f (x)e−ikx dx + f (x)e−ikx dx + · · · . (118)
2π −L 0 L

The f (x)’s in each of these integrals run through the same set of values, due to the period-
icity. And the e−ikx values simply shift by successive powers of e−ikL in each integral. For
R 2L
example, if we define y by x ≡ y + L and substitute this into the L integral above, we
obtain
Z 2L Z L Z L
f (x)e−ikx dx = f (y + L)e−ik(y+L) dy = e−ikL f (y)e−iky dy, (119)
L 0 0
R 2L
where we have used the f (y + L) = f (y) periodicity. So the L integral is simply e−ikL
RL R 3L RL
times the 0 integral. Likewise, the 2L integral is e−2ikL times the 0 integral. And so
RL
on. Therefore, if we factor the 0 integral out of each integral in Eq. (118), we obtain
ÃZ !
1 L ³ ´
C(k) = f (x)e−ikx dx · · · e2ikL + eikL + 1 + e−ikL + e−2ikL · · · . (120)
2π 0
40 CHAPTER 3. FOURIER ANALYSIS

We must now evaluate the sum,

S(k) ≡ · · · e2ikL + eikL + 1 + e−ikL + e−2ikL · · · . (121)

We’ll do it quantitatively below, but first let’s be qualitative to get an idea of what’s going on.
S(k) is the sum of unit vectors in the complex plane that keep rotating around indefinitely.
So they should average out to zero. The one exception occurs when eikL = 1, that is, when
k takes the form of k = 2πn/L. In this case, all the terms in the (infinite) sum equal 1
(the unit vectors don’t rotate at all), so S(k) is infinite. This reasoning is basically correct,
except that it doesn’t get quantitative about the nature of the infinity (we’ll find below that
it’s delta function). And it’s also a little sloppy in the “averaging out to zero” part, because
the sum doesn’t converge, the way it stands. All the terms have magnitude 1, so it matters
where you start and stop the sum.
Let’s now be quantitative. Following the strategy we used many times in Section 3.5,
we can get a handle on S(k) by multiplying the terms by successive powers of a number r
that is slightly less than 1, starting with the e±ikL terms and then working outward in both
directions. We’ll then take the r → 1 limit to recover the original sum. So our modified
sum (call it Sr (k)) is

Sr (k) = · · · + r3 e3ikL + r2 e2ikL + reikL + 1 + re−ikL + r2 e−2ikL + r3 e−3ikL + · · · . (122)

Summing the two geometric series on either side of the “1” turns Sr (k) into

reikL re−ikL
Sr (k) = + 1 + . (123)
1 − reikL 1 − re−ikL
Getting a common denominator and combining these terms yields

1 − r2
Sr (k) = . (124)
1 + r2 − 2r cos(kL)

If cos(kL) 6= 1, then Sr (k) equals zero in the r → 1 limit, because the denominator is
nonzero in this limit. So as we saw qualitatively above, if k 6= 2πn/L, then S(k) = 0.
However, things are tricker if cos(kL) = 1, that is, if k = 2πn/L for some integer n. In this
case we obtain Sr (k) = (1 − r2 )/(1 − r)2 = (1 + r)/(1 − r), which goes to infinity in the
r → 1 limit. We can get a handle on this infinity as follows.
Define κ by k ≡ 2πn/L + κ. We are concerned with very small κ values, because these
values correspond to k being very close to 2πn/L. In terms of κ, Sr (k) becomes (using
cos θ ≈ 1 − θ2 /2 to obtain the second line)

1 − r2
Sr (κ) =
r2
1 + − 2r cos(κL)
1 − r2
≈
1 + r − 2r(1 − κ2 L2 /2)
2

(1 + r)(1 − r)
= . (125)
(1 − r)2 + rκ2 L2

If we now let r ≡ 1 − ² (so we’re concerned with very small ²), then the (1 + r) factor in the
numerator is essentially equal to 2. So when we take the ² → 0 limit to obtain the original
sum S(k), we get
2²
S(κ) = lim S² (κ) = lim 2 . (126)
²→0 ²→0 ² + κ2 L2
3.8. RELATION BETWEEN TRANSFORMS AND SERIES 41

But from Eq. (83), this limit equals 2π δ(κL), which from the second equation in Eq. (75)
equals (2π/L)δ(κ), which in turn equals (2π/L)δ(k − 2πn/L), from the definition of κ.10
S(k) diverges for any k of the form 2nπ/L, so we arrive at

X∞
2π
S(k) = δ(k − 2πn/L). (127)
n=−∞
L

It’s legal to just add up all the different delta functions, because each one doesn’t affect
any of the others (because they’re zero everywhere except at the divergence). Plugging this
S(k) into Eq. (120) then gives
ÃZ ! ∞
1 L X
−ikx
C(k) = f (x)e dx δ(k − 2πn/L). (128)
L 0 n=−∞

The delta functions mean that only k values of the form 2πn/L are relevant, so we can
replace the k in the exponent with 2πn/L (after bringing the integral inside the sum). This
gives
∞
à Z !
X 1 L −i2πnx/L
C(k) = f (x)e dx δ(k − 2πn/L),
n=−∞
L 0
∞
X
≡ Cn δ(k − 2πn/L), (129)
n=−∞

where Cn is defined as in Eq. (112). We have therefore arrived at the desired expression for
C(k) in Eq. (117), or equivalently Eq. (114).
And as we saw in Section 3.8.1, if we plug this Fourier transform, C(k), into the first
equation in Eq. (113) to obtain f (x), we end up with the righthand side of Eq. (116), which
is the desired Fourier series of f (x). We have therefore demonstrated how the continuous
Fourier-transform expansion of a periodic function leads to the discrete Fourier series.
Remarks:
1. The delta functions in Eq. (129) occur only at k values of the form 2πn/L, so the above
derivation explains why all the k values in the Fourier series in Eqs. (1) and (19) are multiples
of 2π/L. When we introduced Eq. (1) out of the blue, you may have wondered why no other
frequencies were needed to obtain an arbitrary periodic function with period L.
For example, in the Fourier-series formalism, it’s not so obvious why a k value of, say, π/L
isn’t needed. But in the Fourier-transform formalism, it’s clear that C(π/L) equals zero,
because the terms in the sum in Eq. (120) simply alternate between 1 and −1, and therefore
add up to zero. (As usual, this can be made rigorous by tapering off the sum.) And similar
reasoning holds for all k values not of the form 2πn/L. Therefore, since non-2πn/L values
of k don’t appear in the Fourier transform of f (x), they don’t appear in the Fourier series
either.
2. For a periodic function, we found above that every value of C(k) is either zero when k 6=
2πn/L, or infinite when k = 2πn/L. There is another fairly quick way to see this, as follows.
First, consider a k value of the form, k = 2πR/L, where R is a rational number, a/b. Then
the product f (x)e−ikx is periodic with period bL. If the integral of f (x)e−ikx over this period
is exactly zero, then the integral from −∞ to ∞ is also zero (again, with the integral tapered
off). And if the integral over this period is not zero, then the integral from −∞ to ∞ is
10 It’s no surprise that we obtained the Lozentzian representation of the delta function in Eq. (126). Our

“tapering” method involving powers of r is simply a discrete version of an exponential taper. And we know
from Section 3.5.3 that an exponential taper leads to a Lorentzian delta function.
42 CHAPTER 3. FOURIER ANALYSIS

infinite, because we’re adding up a given nonzero number, whatever it may be, an infinite
number of times. Finally, since the rational numbers are dense on the real line, this result
should hold for all values of k. So for all k, C(k) is either zero or infinite. As it turns out,
C(k) is nearly always zero. ♣

3.8.3 Derivation of Fourier’s theorem

We started this chapter by invoking Fourier’s theorem. That is, we accepted the fact that
any (reasonably well-behaved) periodic function can be written as a Fourier-series sum of
sines and cosines, or exponentials; see Eqs. (1) and (19). We then used this fact to derive
(by taking the L → ∞ limit) the Fourier-transform relations in Eq. (43). It turns out,
however, that armed with our knowledge of delta functions, we can actually derive all of
these results from scratch. And the order of the derivation is reversed; we will first derive
the Fourier-transform expression for an arbitrary (reasonably well-behaved) function, and
then we will use this result to derive the Fourier-series expression for a periodic function.
Let’s be very systematic and derive things step-by-step:
1. The first step is to use the “tapering” method of Section 3.5 to derive the integral
representation of the delta function given in Eq. (90), which we will copy here:
Z ∞
1
δ(k − k0 ) = e−i(k−k0 )x dx. (130)
2π −∞

When we derived this in Section 3.5, we formulated things in terms of Fourier trans-
forms, but this language isn’t necessary. Even
R ∞ if we had never heard of Fourier trans-
forms, we could have pulled the integral −∞ e−ikx dx out of a hat, and then used
the tapering method to show that it is a delta function of k. That is, it has area 1
and is zero everywhere except at k = 0. We need not know anything about Fourier
transforms to derive this fact.
2. Given an arbitrary (not necessarily periodic) function f (x), let’s define (out of the
blue) a function C(k) by
Z ∞
1
C(k) ≡ f (x)e−ikx dx. (131)
2π −∞

The is a purely definitional statement and has no content. We haven’t done anything
with actual substance here.
3. Given this definition of C(k), we will now use the expression for the delta function in
Eq. (130) to show that Z ∞
f (x) = C(k)eikx dk. (132)
−∞

This is a statement with actual substance. To prove it, we first need to plug the above
definition of C(k) into the righthand side to obtain
Z ∞µ Z ∞ ¶
? 1 0 −ikx0
f (x) = f (x )e dx eikx dk,
0
(133)
−∞ 2π −∞

where we have been careful to label the dummy integration variable from Eq. (131)
as x0 . Switching the order of integration and rearranging things gives
Z ∞ µZ ∞ ¶
? 1 0
f (x) = f (x0 ) e−ik(x −x) dk dx0 . (134)
2π −∞ −∞
3.8. RELATION BETWEEN TRANSFORMS AND SERIES 43

From Eq. (130), the quantity in parentheses is simply 2πδ(x0 −x), with different letters.
So we obtain Z ∞
? 1
f (x) = f (x0 )δ(x0 − x) dx0 . (135)
2π −∞
And this is indeed true, because the righthand side equals f (x), due to the properties
of the delta function. We have therefore successfully derived the Fourier-transform
relations in Eq. (43). More precisely, we wrote one of them, Eq. (131), down by
definition. And we then derived the other one, Eq. (132).
R∞
We should mention again that whenever we deal with integrals of the type −∞ e−ikx dx,
it is understood that we are technically using some sort of tapering method to make
sense of them. But we know that the result will always be a delta function, so we
don’t actually have to go through the tapering procedure each time.
4. Having derived the Fourier-transform relations in Eq. (43) for arbitrary functions, our
task is now to derive the Fourier-series relations for periodic functions, Eqs. (19) and
(20), which we will copy here:
∞
X
f (x) = Cn ei2πnx/L , (136)
n=−∞

where the Cn ’s are given by

Z L
1
Cn = f (x)e−i2πnx/L dx. (137)
L 0

(Alternatively, we could use the trig series given in Eq. (3.1).) But we already per-
formed this derivation in Section 3.8.2, so we won’t go through it again. The main
point is that the periodicity of f (x) produces a collection of delta functions of k, which
turns the continuous-integral Fourier-transform expression for f (x) in Eq. (132) into
the discrete-sum Fourier-series expression for f (x) in Eq. (136).
Chapter 4

Transverse waves on a string

David Morin, [email protected]

In the previous three chapters, we built up the foundation for our study of waves. In the
remainder of this book, we’ll investigate various types of waves, such as waves on a string,
sound waves, electromagnetic waves, water waves, quantum mechanical waves, and so on.
In Chapters 4 through 6, we’ll discuss the properties of the two basic categories of waves,
namely dispersive waves, and non-dispersive waves. The rest of the book is then largely a
series of applications of these results. Chapters 4 through 6 therefore form the heart of this
book.
A non-dispersive system has the property that all waves travel with the same speed,
independent of the wavelength and frequency. These waves are the subject of this and
the following chapter (broken up into longitudinal and transverse waves, respectively). A
dispersive system has the property that the speed of a wave does depend on the wavelength
and frequency. These waves are the subject of Chapter 6. They’re a bit harder to wrap your
brain around, the main reason being the appearance of the so-called group velocity. As we’ll
see in Chapter 6, the difference between non-dispersive and dispersive waves boils down to
the fact that for non-dispersive waves, the frequency ω and wavelength k are related by a
simple proportionality constant, whereas this is not the case for dispersive waves.
The outline of this chapter is as follows. In section 4.1 we derive the wave equation for
transverse waves on a string. This equation will take exactly the same form as the wave
equation we derived for the spring/mass system in Section 2.4, with the only difference
being the change of a few letters. In Section 4.2 we discuss the reflection and transmission
of a wave from a boundary. We will see that various things can happen, depending on
exactly what the boundary looks like. In Section 4.3 we introduce the important concept
of impedance and show how our previous results can be written in terms of it. In Section
4.4 we talk about the energy and power carried by a wave. In Section 4.5 we calculate the
form of standing waves on a string that has boundary conditions that fall into the extremes
(a fixed end or a “free” end). In Section 4.6 we introduce damping, and we see how the
amplitude of a wave decreases with distance in a scenario where one end of the string is
wiggled with a constant amplitude.

4.1 The wave equation

The most common example of a non-dispersive system is a string with transverse waves on
it. We’ll see below that we obtain essentially the same wave equation for transverse waves

1
 2 CHAPTER 4. TRANSVERSE WAVES ON A STRING

on a string as we obtained for the N → ∞ limit of longitudinal waves in the mass/spring

system in Section 2.4. Either of these waves could therefore be used for our discussion of
the properties of non-dispersive systems. However, the reason why we’ve chosen to study
transverse waves on a string in this chapter is that transverse waves are generally easier to
visualize than longitudinal ones.
Consider a string with tension T and mass density µ (per unit length). Assume that
it is infinitesimally thin and completely flexible. And assume for now that it extends in-
finitely in both directions. We’ll eventually relax this restriction. Consider small transverse
displacements of the string (we’ll be quantitative about the word “small” below). Let x be
the coordinate along the string, and let ψ be the transverse displacement. (There’s no deep
reason why we’re using ψ for the displacement instead of the ξ we used in Chapter 2.) Our
2 goal is to find the most general form of ψ(x, t).
2 dψ
dx + dψ Consider two nearby points separated by a displacement dx in the longitudinal direction
along the string, and by a displacement dψ in the transverse direction. If dψ is small (more
dx precisely if the slope dψ/dx is small; see below), then we can make the approximation that
all points in the string move only in the transverse direction. That is, there is no longitudinal
Figure 1 motion. This is true because in Fig. 1 the length of the hypotenuse equals
r ³ dψ ´2 µ ¶
p
2 2
1 ³ dψ ´2 1 ³ dψ ´
dx + dψ = dx 1 + ≈ dx 1 + = dx + dψ . (1)
dx 2 dx 2 dx

This length (which is the farthest that a given point can move to the side; it’s generally less
that this) differs from the length of the long leg in Fig. 1 by an amount dψ(dψ/dx)/2, which
is only (dψ/dx)/2 times as large as the transverse displacement dψ. Since we are assuming
that the slope dψ/dx is small, we can neglect the longitudinal motion in comparison with
the transverse motion. Hence, all points essentially move only in the transverse direction.
We can therefore consider each point to be labeled with a unique value of x. That is,
the ambiguity between the original and present longitudinal positions is irrelevant. The
string will stretch slightly, but we can always assume that the amount of mass in any given
horizontal span stays essentially constant.
We see that by the phrase “small transverse displacements” we used above, we mean
T2 that the slope of the string is small. The slope is a dimensionless quantity, so it makes
sense to label it with the word “small.” It makes no sense to say that the actual transverse
θ1 displacement is small, because this quantity has dimensions.
θ2
Our strategy for finding the wave equation for the string will be to write down the trans-
x x + dx verse F = ma equation for a little piece of string in the span from x to x + dx. The situation
T1 is shown in Fig. 2. (We’ll ignore gravity here.) Let T1 and T2 be the tensions in the string at
the ends of the small interval. Since the slope dψ/dx is small, the slope is essentially equal
Figure 2
to the θ angles in the figure. (So these angles are small, even though we’ve drawn them with
reasonable sizes for the sake of clarity.) We can use the approximation cos θ ≈ 1 − θ2 /2 to
say that the longitudinal components of the tensions are equal to the tensions themselves,
up to small corrections of order θ2 ≈ (dψ/dx)2 . So the longitudinal components are (essen-
tially) equal to T1 and T2 . Additionally, from the above reasoning concerning (essentially)
no longitudinal motion, we know that there is essentially no longitudinal acceleration of the
little piece in Fig. 2. So the longitudinal forces must cancel. We therefore conclude that
T1 = T2 . Let’s call this common tension T .
However, although the two tensions and their longitudinal components are all equal, the
same thing cannot be said about the transverse components. The transverse components
differ by a quantity that is first order in dψ/dx, and this difference can’t be neglected.
This difference is what causes the transverse acceleration of the little piece, and it can be
calculated as follows.
4.1. THE WAVE EQUATION 3

In Fig. 2, the “upward” transverse force on the little piece at its right end is T sin θ1 ,
which essentially equals T times the slope, because the angle is small. So the upward force
at the right end is T ψ 0 (x + dx). Likewise, the “downward” force at the left end is −T ψ 0 (x).
The net transverse force is therefore
³ ´ ψ 0 (x + dx) − ψ 0 (x) d2 ψ(x)
Fnet = T ψ 0 (x + dx) − ψ 0 (x) = T dx ≡ T dx , (2)
dx dx2
where we have assumed that dx is infinitesimal and used the definition of the derivative to
obtain the third equality.1 Basically, the difference in the first derivatives yields the second
derivative. For the specific situation shown in Fig. 2, d2 ψ/dx2 is negative, so the piece
is accelerating in the downward direction. Since the mass of the little piece is µ dx, the
transverse F = ma equation is
d2 ψ d2 ψ d2 ψ T d2 ψ
Fnet = ma =⇒ T dx 2
= (µ dx) 2 =⇒ 2
= . (3)
dx dt dt µ dx2
Since ψ is a function of x and t, let’s explicitly include this dependence and write ψ as ψ(x, t).
We then arrive at the desired wave equation (written correctly with partial derivatives now),

∂ 2 ψ(x, t) T ∂ 2 ψ(x, t)
= (wave equation) (4)
∂t2 µ ∂x2
This takes exactly the same form as the wave equation we found in Section 2.4 for the
N → ∞ limit of the spring/mass system. The only difference is the replacement of the
quantity E/ρ with the quantity T /µ. Therefore, all of our previous results carry over here.
In particular, the solutions take the form,
s
ω T
ψ(x, t) = Aei(±kx±ωt) where = ≡c (5)
k µ

and where c is the speed of the traveling wave (from here on, we’ll generally use c instead
of v for the speed of the wave). This form does indeed represent a traveling wave, as we
saw in Section 2.4. k and ω can take on any values, as long as they’re related by ω/k = c.
The wavelength is λ = 2π/k, and the time period of the oscillation of any given point is
τ = 2π/ω. So the expression ω/k = c can be written as
2π/τ λ
= c =⇒ = c =⇒ λν = c, (6)
2π/λ τ
where ν = 1/τ is the frequency in cycles per second (Hertz).
For a given pair of k and ω values, the most general form of ψ(x, t) can be written in
many ways, as we saw in Section 2.4. From Eqs. (3.91) and (3.92) a couple of these ways
are
ψ(x, t) = C1 cos(kx + ωt) + C2 sin(kx + ωt) + C3 cos(kx − ωt) + C4 sin(kx − ωt), (7)
and
ψ(x, t) = D1 cos kx cos ωt + D2 sin kx sin ωt + D3 sin kx cos ωt + D4 cos kx sin ωt. (8)
Of course, since the wave equation is linear, the most general solution is the sum of an
arbitrary number of the expressions in,psay, Eq. (8), for different pairs of k and ω values (as
long as each pair is related by ω/k = T /µ ≡ c). For example, a possible solution is
ψ(x, t) = A cos k1 x cos ck1 t + B cos k1 x sin ck1 t + C sin k2 x sin ck2 t + etc . . . (9)
1 There is an ambiguity about whether we should write ψ 00 (x) or ψ 00 (x+dx) here. Or perhaps ψ 00 (x+dx/2).
But this ambiguity is irrelevant in the dx → 0 limit.
4 CHAPTER 4. TRANSVERSE WAVES ON A STRING

Solutions of the form f (x − ct)

As we saw in Section 2.4, any function of the form f (x − ct) satisfies the wave equation.
There are two reasons why this functional form works. The first reason, as we showed in
Eq. (2.97), is that if you simply plug ψ(x, t) =pf (x − ct) into the wave equation in Eq. (4),
you will find that it works, provided that c = T /µ.
The second reason, as we also mentioned in Section 2.4, is due to Fourier analysis and
linearity. Let’s talk a little more about this reason. It is the combination of the following
facts:
1. By Fourier analysis, we can write any function f (z) as
Z ∞
f (z) = C(k)eikz dk, (10)
−∞

where C(k) is given by Eq. (3.43). If we let z ≡ x − ct, then this becomes
Z ∞ Z ∞
f (x − ct) = C(k)eik(x−ct) dk = C(k)ei(kx−ωt) dk, (11)
−∞ −∞

where ω ≡ ck.
2. We showed in Section 2.4 (and you can quickly verify it again) that any exponential
function of the form ei(kx−ωt) satisfies the wave equation, provided that ω = ck, which
is indeed the case here.
3. Because the wave equation is linear, any linear combination of solutions is again a
solution. Therefore, the integral in Eq. (11) (which is a continuous linear sum) satisfies
the wave equation.
In this reasoning, it is critical that ω and k are related linearly by ω = ck, pwhere c
takes on a constant value, independent of ω and k (and it must be equal to T /µ, or
whatever constant appears in the wave equation). If this relation weren’t true, then the
above reasoning would be invalid, for reasons we will shortly see. When we get to dispersive
waves in Chapter 6, we will find that ω does not equal ck. In other words, the ratio ω/k
depends on k (or equivalently, on ω). Dispersive waves therefore cannot be written in the
form of f (x − ct). It is instructive to see exactly where the above reasoning breaks down
when ω 6= ck. This breakdown can be seen in mathematically, and also physically.
RMathematically,
∞
it is still true that any function f (x − ct) can be written in the form
of −∞ C(k)eik(x−ct) dk. However, it is not true that these ei(kx−(ck)t) exponential functions
are solutions to a dispersive wave equation (we’ll see in Chapter 6 what such an equation
might look like), because ω doesn’t take the form of ck for dispersive waves. The actual
solutions to a dispersive wave equation are exponentials of the form ei(kx−ωt) , where ω is
some nonlinear function of k. That is, ω does not take the form of ck. If you want, you can
write these exponential solutions as ei(kx−ck kt) , where ck ≡ ω/k is the speed of the wave
component with wavenumber k. But the point is that a single value of c doesn’t work for
all values of k.
In short, if by the ω in Eq. (11) we mean ωk (which equals ck for nondispersive waves,
but not for dispersive waves), then for dispersive waves, the above reasoning breaks down in
the second equality in Eq. (11), because the coefficient of t in the first integral is ck (times
−i), which isn’t equal to the ω coefficient in the second integral. If on the other hand we
want to keep the ω in Eq. (11) defined as ck, then for dispersive waves, the above reasoning
breaks down in step 2. The exponential function ei(kx−ωt) with ω = ck is simply not a
solution to a dispersive wave equation.
4.1. THE WAVE EQUATION 5

The physical reason why the f (x − ct) functional form doesn’t work for dispersive waves
is the following. Since the speed ck of the Fourier wave components depends on k in a
dispersive wave, the wave components move with different speeds. The shape of the wave
at some initial time will therefore not be maintained as t increases (assuming that the wave
contains more than a single Fourier component). This implies that the wave cannot be
written as f (x − ct), because this wave does keep the same shape (because the argument
x − ct doesn’t change if t increases by ∆t and x increases by c ∆t).
The distortion of the shape of the wave can readily be seen in the case where there are Cos(x) v
just two wave components, which is much easier to visualize than the continuous infinity of
components involved in a standard Fourier integral. If the two waves have (k, ω) values of
(1, 1) and (2, 1), then since the speed is ω/k, the second wave moves with half the speed of Cos(2x) v/2
the first. Fig. 3 shows the sum of these two waves at two different times. The total wave
clearly doesn’t keep the same shape, so it therefore can’t be written in the form of f (x − ct).

Fourier transform in 2-D

Having learned about Fourier transforms in Chapter 3, we can give another derivation of the
fact that any solution to the wave equation in Eq. (4) can be written in terms of ei(kx−ωt) Cos(x-t) + Cos(2x-t)
exponentials, where ω = ck. We originally derived these solutions in Section 2.4 by guessing at t = 0
exponentials, with the reasoning that since all (well enough behaved) functions can be built
up from exponentials due to Fourier analysis, it suffices to consider exponentials. However,
you might still feel uneasy about this “guessing” strategy, so let’s be a little more systematic
with the following derivation. This derivation involves looking at the Fourier transform of
a function of two variables. In Chapter 3, we considered functions of only one variable, but Cos(x-1)
the extension to two variables in quite straightforward.
Consider a wave ψ(x, t) on a string, and take a snapshot at a given time. If f (x) describes
the wave at this instant, then from 1-D Fourier analysis we can write
Cos(2x-1)
Z ∞
f (x) = C(k)eikx dk, (12)
−∞

where C(k) is given by Eq. (3.43). If we take a snapshot at a slightly later time, we can
again write ψ(x, t) in terms of its Fourier components, but the coefficients C(k) will be
slightly different. In other words, the C(k)’s are functions of time. So let’s write them as Cos(x-t) + Cos(2x-t)
C(k, t). In general, we therefore have at t = 1
Z ∞
ψ(x, t) = C(k, t)eikx dk. (13) Figure 3
−∞

This equation says that at any instant we can decompose the snapshot of the string into its
eikx Fourier components.
We can now do the same thing with the C(k, t) function that we just did with the ψ(x, t)
function. But we’ll now consider “slices” with constant k value instead of constant t value. If
g(t) describes the function C(k, t) for a particular value of k, then from 1-D Fourier analysis
we can write Z ∞
g(t) = β(ω)eiωt dω. (14)
−∞

If we consider a slightly different value of k, we can again write C(k, t) in terms of its Fourier
components, but the coefficients β(ω) will be slightly different. That is, they are functions
of k, so let’s write them as β(k, ω). In general, we have
Z ∞
C(k, t) = β(k, ω)eiωt dω. (15)
−∞
6 CHAPTER 4. TRANSVERSE WAVES ON A STRING

This equation says that for a given value of k, we can decompose the function C(k, t) into
its eiωt Fourier components. Plugging this expression for C(k, t) into Eq. (13) gives
Z ∞ µZ ∞ ¶
ψ(x, t) = β(k, ω)e dω eikx dk
iωt
−∞ −∞
Z ∞Z ∞
= β(k, ω)ei(kx+ωt) dk dω. (16)
−∞ −∞

This is a general result for any function of two variables; it has nothing to do with the wave
equation in Eq. (4). This result basically says that we can just take the Fourier transform
in each dimension separately.
Let’s now apply this general result to the problem at hand. That is, let’s plug the above
expression for ψ(x, t) into Eq. (4) and see what it tells us. We will find that ω must be
equal to ck. The function β(k, ω) is a constant as far as the t and x derivatives in Eq. (4)
are concerned, so we obtain

∂ 2 ψ(x, t) 2
2 ∂ ψ(x, t)
0 = − c
∂t2 ∂x2
Z ∞Z ∞ µ 2 i(kx+ωt) 2 i(kx+ωt)
¶
∂ e 2∂ e
= β(k, ω) − c dk dω
−∞ −∞ ∂t2 ∂x2
Z ∞Z ∞ ³ ´
= β(k, ω)ei(kx+ωt) − ω 2 − c2 (−k 2 ) dk dω. (17)
−∞ −∞

Since the ei(kx+ωt) exponentials here are linearly independent functions, the only way that
this sum can be zero for all x and t is if the coefficient of each separate ei(kx+ωt) term is
zero. That is,
β(k, ω)(ω 2 − c2 k 2 ) = 0 (18)
for all values of k and ω.2 There are two ways for this product to be zero. First, we can
have β(k, ω) = 0 for particular values of k and ω. This certainly works, but since β(k, ω)
indicates how much of ψ(x, t) is made up of ei(kx+ωt) for these particular values of k and ω,
the β(k, ω) = 0 statement tells us that this particular ei(kx+ωt) exponential doesn’t appear
in ψ(x, t). So we don’t care about how ω and k are related.
The other way for Eq. (18) to be zero is if ω 2 − c2 k 2 = 0. That is, ω = ±ck, as we
wanted to show. We therefore see that if β(k, ω) is nonzero for particular values of k and ω
(that is, if ei(kx+ωt) appears in ψ(x, t)), then ω must be equal to ±ck, if we want the wave
equation to be satisfied.

4.2 Reflection and transmission

4.2.1 Applying the boundary conditions
Instead of an infinite uniform string, let’s now consider an infinite string with density µ1
for −∞ < x < 0 and µ2 for 0 < x < ∞. Although the density isn’t uniform, the tension
is still uniform throughout the entire string, because otherwise there would be a nonzero
horizontal acceleration somewhere.
Assume that a wave of the form ψi (x, t) = fi (x − v1 t) (the “i” here is for “incident”)
starts off far to the left and heads rightward toward x = 0. It turns out that it will be much
2 Alternatively, Eq. (17) says that β(k, ω)(ω 2 − c2 k 2 ) is the 2-D Fourier transform of zero. So it must be

zero, because it can be found from the 2-D inverse-transform relations analogous to Eq. (3.43), with a zero
appearing in the integrand.
4.2. REFLECTION AND TRANSMISSION 7

more convenient to instead write the wave as

µ ¶
x
ψi (x, t) = fi t − (19)
v1

for a redefined function fi , so we’ll use this form. Note that ψ is a function
p of two variables,
whereas f is a function of only one. From Eq. (5), the speed v1 equals T /µ1 .
What happens when the wave encounters the boundary at x = 0 between the different
densities? The most general thing that can happen is that there is some reflected wave,
µ ¶
x
ψr (x, t) = fr t + , (20)
v1

moving leftward from x = 0, and also a transmitted wave,

µ ¶
x
ψt (x, t) = ft t − , (21)
v2
p
moving rightward from x = 0 (where v2 = T /µ2 ). Note the “+” sign in the argument of
fr , since the reflected wave is moving leftward.
In terms of the above functions, the complete expressions for the waves on the left and
right side of x = 0 are, respectively,

ψL (x, t) = ψi (x, t) + ψr (x, t) = fi (t − x/v1 ) + fr (t + x/v1 ),

ψR (x, t) = ψt (x, t) = ft (t − x/v2 ). (22)

If we can find the reflected and transmitted waves in terms of the incident wave, then we will
know what the complete wave looks like everywhere. Our goal is therefore to find ψr (x, t)
and ψt (x, t) in terms of ψi (x, t). To do this, we will use the two boundary conditions at
x = 0. Using Eq. (22) to write the waves in terms of the various f functions, the two
boundary conditions are:

• The string is continuous. So we must have (for all t)

"massless"
ψL (0, t) = ψR (0, t) =⇒ fi (t) + fr (t) = ft (t) (23) atom T

• The slope is continuous. This is true for the following reason. If the slope were differ- T
ent on either side of x = 0, then there would be a net (non-infinitesimal) force in some net force
direction on the atom located at x = 0, as shown in Fig. 4. This (nearly massless)
atom would then experience an essentially infinite acceleration, which isn’t physically Figure 4
possible. (Equivalently, a nonzero force would have the effect of instantaneously read-
justing the string to a position where the slope was continuous.) Continuity of the
slope gives (for all t)

∂ψL (x, t) ¯¯ ∂ψR (x, t) ¯¯ 1 1 1

¯ = ¯ =⇒ − fi0 (t) + fr0 (t) = − ft0 (t). (24)
∂x x=0 ∂x x=0 v1 v1 v2
Integrating this and getting the v’s out of the denominators gives

v2 fi (t) − v2 fr (t) = v1 ft (t) (25)

We have set the constant of integration equal to zero because we are assuming that
the string has no displacement before the wave passes by.
8 CHAPTER 4. TRANSVERSE WAVES ON A STRING

Solving Eqs. (23) and (25) for fr (t) and ft (t) in terms of fi (t) gives

v2 − v1 2v2
fr (s) = fi (s), and ft (s) = fi (s), (26)
v2 + v1 v2 + v1
where we have written the argument as s instead of t to emphasize that these relations hold
for any arbitrary argument of the f functions. The argument need not have anything to do
with the time t. The f ’s are functions of one variable, and we’ve chosen to call that variable
s here.

4.2.2 Reflection
We derived the relations in Eq. (26) by considering how the ψ(x, t)’s relate at x = 0. But
how do we relate them at other x values? We can do this in the following way. Let’s look
at the reflected wave first. If we replace the s in Eq. (26) by t + x/v1 (which we are free to
do, because the argument of the f ’s can be whatever we want it to be), then we can write
fr as µ ¶ µ ¶
x v2 − v1 −x
fr t + = fi t − . (27)
v1 v2 + v1 v1
If we recall the definition of the f ’s in Eqs. (19-21), we can write this result in terms of the
ψ’s as
v2 − v1
ψr (x, t) = ψi (−x, t) (28)
v2 + v1
This is the desired relation between ψr and ψi , and its interpretation is the following. It
says that at a given time t, the value of ψr at position x equals (v2 − v1 )/(v2 + v1 ) times
the value of ψi at position negative x. This implies that the speed of the ψr wave equals
the speed of the ψi wave (but with opposite velocity), and it also implies that the width
of the ψr wave equals the width of the ψi wave. But the height is decreased by a factor of
(v2 − v1 )/(v2 + v1 ).
Only negative values of x are relevant here, because we are dealing with the reflected
wave which exists only to the left of x = 0. Therefore, since the expression ψi (−x, t) appears
in Eq. (28), the −x here means that only positive position coordinates are relevant for the ψi
wave. You might find this somewhat disconcerting, because the ψi function isn’t applicable
to the right of x = 0. However, we can mathematically imagine ψi still proceeding to the
right. So we have the pictures shown in Fig. 5. For simplicity, let’s say that v2 = 3v1 , which
means that the (v2 − v1 )/(v2 + v1 ) factor equals 1/2. Note that in any case, this factor lies
between −1 and 1. We’ll talk about the various possibilities below.

(v2 = 3v1) (ψt not shown)

ψi
v1 ψr
v1

x=0
actual not really
waves ψi there
ψr
4.2. REFLECTION AND TRANSMISSION 9

Figure 5

In the first picture in Fig. 5, the incident wave is moving in from the left, and the reflected
wave is moving in from the right. The reflected wave doesn’t actually exist to the right of
x = 0, of course, but it’s convenient to imagine it coming in as shown. Except for the scale
factor of (v2 − v1 )/(v2 + v1 ) in the vertical direction (only), ψr is simply the mirror image
of ψi .
In the second picture in Fig. 5, the incident wave has passed the origin and continues
moving to the right, where it doesn’t actually exist. But the reflected wave is now located
on the left side of the origin and moves to the left. This is the real piece of the wave. For
simplicity, we haven’t shown the transmitted ψt wave in these pictures (we’ll deal with it
below), but it’s technically also there.
In between the two times shown in Fig. 5, things aren’t quite as clean, because there are
x values near the origin (to the left of it) where both ψi and ψr are nonzero, and we need to
add them to obtain the complete wave, ψL , in Eq. (22). But the procedure is straightforward
in principle. The two ψi and ψr waves simply pass through each other, and the value of ψi
ψr
ψL at any point to the left of x = 0 is obtained by adding the values of ψi and ψr at that
point. Remember that only the region to the left of x = 0 is real, as far as the reflected
wave is concerned. The wave to the right of x = 0 that is generated from ψi and ψr is just
a convenient mathematical construct.
Fig. 6 shows some successive snapshots that result from an easy-to-visualize incident
square wave. The bold-line wave indicates the actual wave that exists to the left of x = 0.
We haven’t drawn the transmitted wave to the right of x = 0. You should stare at this figure
until it makes sense. This wave actually isn’t physical; its derivative isn’t continuous, so it
violates the second of the above boundary conditions (although we can imagine rounding
off the corners to eliminate this issue). Also, its derivative isn’t small, which violates our
assumption at the beginning of this section. However, we’re drawing this wave so that the
important features of reflection can be seen. Throughout this book, if we actually drew
realistic waves, the slopes would be so small that it would be nearly impossible to tell what
was going on.

4.2.3 Transmission
Let’s now look at the transmitted wave. If we replace the s by t − x/v2 in Eq. (26), we can
write ft as µ ¶ µ ¶
x 2v2 (v1 /v2 )x
ft t − = fi t − . (29)
v2 v2 + v1 v1
Using the definition of the f ’s in Eqs. (19-21), we can write this in terms of the ψ’s as
Figure 6
2v2 ³ ´
ψt (x, t) = ψi (v1 /v2 )x, t (30)
v2 + v1

This is the desired relation between ψt and ψi , and its interpretation is the following. It
says that at a given time t, the value of ψt at position x equals 2v2 /(v2 + v1 ) times the value
of ψi at position (v1 /v2 )x. This implies that the speed of the ψt wave is v2 /v1 times the
speed of the ψi wave, and it also implies that the width of the ψt wave equals v2 /v1 times
the width
¡ of the ¢ψi wave. These facts are perhaps a little more obvious if we write Eq. (30)
as ψt (v2 /v1 )x, t = 2v2 /(v2 + v1 ) · ψi (x, t).
Only positive values of x are relevant here, because we are dealing with the¡transmitted¢
wave which exists only to the right of x = 0. Therefore, since the expression ψi (v1 /v2 )x, t
appears in Eq. (30), only positive position coordinates are relevant for the ψi wave. As in
 10 CHAPTER 4. TRANSVERSE WAVES ON A STRING

the case of reflection above, although the ψi function isn’t applicable for positive x, we can
mathematically imagine ψi still proceeding to the right. The situation is shown in Fig. 7. As
above, let’s say that v2 = 3v1 , which means that the 2v2 /(v2 + v1 ) factor equals 3/2. Note
that in any case, this factor lies between 0 and 2. We’ll talk about the various possibilities
below.

(v2 = 3v1)
ψt (ψr not shown)
v2 ψi
v1

actual
x=0 waves
not really ψt
there ψi

Figure 7

In the first picture in Fig. 7, the incident wave is moving in from the left, and the
ψt
transmitted wave is also moving in from the left. The transmitted wave doesn’t actually
ψi exist to the left of x = 0, of course, but it’s convenient to imagine it coming in as shown.
With v2 = 3v1 , the transmitted wave is 3/2 as tall and 3 times as wide as the incident wave.
In the second picture in Fig. 7, the incident wave has passed the origin and continues
moving to the right, where it doesn’t actually exist. But the transmitted wave is now located
on the right side of the origin and moves to the right. This is the real piece of the wave.
For simplicity, we haven’t shown the reflected ψr wave in these pictures, but it’s technically
also there.
In between the two times shown in Fig. 7, things are easier to deal with than in the
reflected case, because we don’t need to worry about taking the sum of two waves. The
transmitted wave consists only of ψt . We don’t have to add on ψi as we did in the reflected
case. In short, ψL equals ψi + ψr , whereas ψR simply equals ψt . Equivalently, ψi and ψr
have physical meaning only to the left of x = 0, whereas ψt has physical meaning only to
the right of x = 0.
Fig. 8 shows some successive snapshots that result from the same square wave we con-
sidered in Fig. 6. The bold-line wave indicates the actual wave that exists to the right of
x = 0. We haven’t drawn the reflected wave to the left of x = 0. We’ve squashed the x
axis relative to Fig. 6, to make a larger region viewable. These snapshots are a bit boring
compared with those in Fig. 6, because there is no need to add any waves. As far as ψt
is concerned on the right side of x = 0, what you see is what you get. The entire wave
(on both sides of x = 0) is obtained by juxtaposing the bold waves in Figs. 6 and 8, after
expanding Fig. 8 in the horizontal direction to make the unit sizes the same (so that the ψi
waves have the same width).
Figure 8
4.2. REFLECTION AND TRANSMISSION 11

4.2.4 The various possible cases

For convenience, let’s define the reflection and transmission coefficients as

v2 − v1 2v2
R≡ and T ≡ (31)
v2 + v1 v2 + v1

With these definitions, we can write the reflected and transmitted waves in Eqs. (28) and
(30) as

ψr (x, t) = Rψi (−x, t),

¡ ¢
ψt (x, t) = T ψi (v1 /v2 )x, t . (32)

R and T are the amplitudes of ψr and ψt relative to ψi . Note that 1 + R = T always. This
is just the statement
p of continuity of the wave at x = 0.
Since v = T /µ, and since the tension T is uniform throughout the string, we have
√ √
v1 ∝ 1/ µ1 and v2 ∝ 1/ µ2 . So we can alternatively write R and T in the terms of the
densities on either side of x = 0:
√ √ √
µ1 − µ2 2 µ1
R≡ √ √ and T ≡√ √ (33)
µ1 + µ2 µ1 + µ2
(v2 = 0)
ψi
There are various cases to consider: v1
• Brick wall on right: µ2 = ∞ (v2 = 0) =⇒ R = −1, T = 0. Nothing is
transmitted, since T = 0. And the reflected wave has the same size as the incident
wave, but is inverted due to the R = −1 value. This is shown in Fig. 9.
ψr
The inverted nature of the wave isn’t intuitively obvious, but it’s believable for the v1
following reason. When the wave encounters the wall, the wall pulls down on the
string in Fig. 9. This downward force causes the string to overshoot the equilibrium
position and end up in the inverted orientation. Of course, you may wonder why the
downward force causes the string to overshoot the equilibrium position instead of, Figure 9
say, simply returning to the equilibrium position. But by the same token, you should
wonder why a ball that collides elastically with a wall bounces back with the same
speed, as opposed to ending up at rest.
We can deal with both of these situations by invoking conservation of energy. Energy
wouldn’t be conserved if in the former case the string ended up straight, and if in the
latter case the ball ended up at rest, because there would be zero energy in the final ψi (v2 = v1/2)
state. (The wall is “infinitely” massive, so it can’t pick up any energy.) v1

• Light string on left, heavy string on right: µ1 < µ2 < ∞ (v2 < v1 ) =⇒
−1 < R < 0, 0 < T < 1. This case is in between the previous and following cases. ψt
There is partial (inverted) reflection and partial transmission. See Fig. 10 for the ψr v2
particular case where µ2 = 4µ1 =⇒ v2 = v1 /2. The reflection and transmission v1
coefficients in this case are R = −1/3 and T = 2/3.
• Uniform string: µ2 = µ1 (v2 = v1 ) =⇒ R = 0, T = 1. Nothing is reflected. The Figure 10
string is uniform, so the “boundary” point is no different from any other point. The
wave passes right through, as shown in Fig. 11. (v2 = v1)
ψi
v1

v2 = v1
ψt

Figure 11
 ψi (v2 = 2v1)
v1 12 CHAPTER 4. TRANSVERSE WAVES ON A STRING

• Heavy string on left, light string on right: 0 < µ2 < µ1 (v2 > v1 ) =⇒ 0 <
R < 1, 1 < T < 2. This case is in between the previous and following cases. There
ψr is partial reflection and partial transmission. See Fig. 12 for the particular case where
v1 µ2 = µ1 /4 =⇒ v2 = 2v1 . The reflection and transmission coefficients in this case are
R = 1/3 and T = 4/3.
v2
ψt
• Zero-mass string on right: µ2 = 0 (v2 = ∞) =⇒ R = 1, T = 2. There
Figure 12 is complete (rightside up) reflection in this case, as shown in Fig. 13. Although the
string on the right side technically moves, it has zero mass so it can’t carry any energy.
All of the energy is therefore contained in the reflected wave (we’ll talk about energy
(v2 = ) in Section 4.4). So in this sense there is total reflection. In addition to carrying no
8

ψi
v1 energy, the movement of the right part of the string isn’t even wave motion. The whole
thing always remains in a straight horizontal line and just rises and falls (technically
it’s a wave with infinite wavelength).3
As with the brick-wall case above, the right-side-up nature of the wave isn’t intuitively
obvious, but it’s believable for the following reason. When the wave encounters the
boundary, the zero-mass string on the right side is always horizontal, so it can’t apply
v1 any transverse force on the string on the left side. Since there is nothing pulling the
ψr string down, it can’t end up on the other side of the equilibrium position as it did
in the brick-wall case. The fact that it actually ends up with the same shape is a
Figure 13 consequence of energy conservation, because the massless string on the right can’t
carry any energy.

4.3 Impedance
4.3.1 Definition of impedance
In the previous section, we allowed the density to change at x = 0, but we assumed that
the tension was uniform throughout the string. Let’s now relax this condition and allow the
T2 tension to also change at x = 0. The previous treatment is easily modified, and we will find
T1 that a new quantity, called the impedance, arises.
massless ring You may be wondering how we can arrange for the tension to change at x = 0, given that
any difference in the tension should cause the atom at x = 0 to have “infinite” acceleration.
fixed pole
But we can eliminate this issue by using the setup shown in Fig. 14. The boundary between
Figure 14 the two halves of the string is a massless ring, and this ring encircles a fixed frictionless pole.
The pole balances the difference in the longitudinal components of the two tensions, so the
net longitudinal force on the ring is zero, consistent with the fact that it is constrained to
massless
ring remain on the pole and move only in the transverse direction.
T2 The net transverse force on the massless ring must also be zero, because otherwise it
θ1
would have infinite transverse acceleration. This zero-transverse-force condition is given by
T1 sin θ1 = T2 sin θ2 , where the angles are defined in Fig. 15. In terms of the derivatives on
T1 θ2 either side of x = 0, this relation can be written as (assuming, as we always do, that the
slope of the string is small)
Figure 15
∂ψL (x, t) ¯¯ ∂ψR (x, t) ¯¯
T1 ¯ = T2 ¯ . (34)
∂x x=0 ∂x x=0

3 The right part of the string must be straight (and hence horizontal, so that it doesn’t head off to ±∞)

because if it were curved, then the nonzero second derivative would imply a nonzero force on a given piece
of the string, resulting in infinite acceleration, because the piece is massless. Alternatively, the transmitted
wave is stretched horizontally by a factor v2 /v1 = ∞ compared with the incident wave. This implies that it
is essentially horizontal.
4.3. IMPEDANCE 13

In the case of uniform tension discussed in the previous section, we had T1 = T2 , so this
equation reduced to the first equality in Eq. (24). With the tensions now distinct, the only
modification to the second equality in Eq. (24) is the extra factors of T1 and T2 . So in terms
of the f functions, Eq. (34) becomes
T1 0 T1 T2
− f (t) + fr0 (t) = − ft0 (t). (35)
v1 i v1 v2
The other boundary condition (the continuity of the string) is unchanged, so all of the
results in the previous section can be carried over, with the only modification being that
wherever we had a v1 , we now have v1 /T1 . And likewise for v2 . The quantity v/T can be
written as p
v T /µ 1 1
= =√ ≡ , (36)
T T Tµ Z
where
T p
Z≡ = Tµ (37)
v
is called the impedance. We’ll discuss Z in depth below, but for now we’ll √ simply note that
the results in the previous section are modified by replacing v1 with 1/ T1 µ1 ≡ 1/Z1 , and
likewise for v2 . The reflection and transmission coefficients in Eq. (31) therefore become
1 1 2
Z2 − Z1 Z1 − Z2 Z2 2Z1
R= 1 1 = and T = 1 1 = (38)
Z2 + Z1
Z1 + Z2 Z2 + Z1
Z1 + Z2

Note that Z grows with both T and µ.

Physical meaning of impedance

What is the meaning of the impedance? It makes our formulas look nice, but does it have
any actual physical significance? Indeed it does. Consider the transverse force that the ring
applies to the string on its left. Since there is zero net force on the ring, this force also
equals the transverse force that the right string applies to the ring, which is
¯ ¯
∂ψR (x, t) ¯¯ ∂ft (t − x/v2 ) ¯¯
Fy = T2 ¯ = T2 ¯ , (39)
∂x x=0 ∂x x=0

where we have labeled the transverse direction as the y direction. But the chain rule tells
us that the x and t partial derivatives of ft are related by
∂ft (t − x/v2 ) 1 ∂ft (t − x/v2 )
=− · . (40)
∂x v2 ∂t
Substituting this into Eq. (39) and switching back to the ψR (x, t) notation gives
¯
T2 ∂ψR (x, t) ¯¯ T2
Fy = − · ¯ = − · vy ≡ −bvy , (41)
v2 ∂t x=0 v2

where vy = ∂ψR (x, t)/∂t is the transverse velocity of the ring (at x = 0), and where b is
defined to be T2 /v2 .
This force Fy (which again, is the force that the ring applies to the string on its left) has
the interesting property that it is proportional to the (negative of the) transverse velocity.
It therefore acts exactly like a damping force. If we removed the right string and replaced
the ring with a massless plate immersed in a fluid (in other words, a piston), and if we
14 CHAPTER 4. TRANSVERSE WAVES ON A STRING

arranged things (the thickness of the fluid and the cross-sectional area of the plate) so that
the damping coefficient was b, then the left string wouldn’t have any clue that the right
string was removed and replaced with the damping mechanism. As far as the left string is
concerned, the right string acts exactly like a resistance that is being dragged against.
Said in another way, if the left string is replaced by your hand, and if you move your hand
so that the right string moves just as it was moving when the left string was there, then you
can’t tell whether you’re driving the right string or driving a piston with an appropriately-
chosen damping coefficient. This is true because by Newton’s third law (and the fact that
the ring is massless), the force that the ring applies to the string on the right is +bvy . The
plus sign signifies a driving force (that is, the force is doing positive work) instead of a
damping force (where the force does negative work).
From Eq. (41), we have Fy /vy = −T2 /v2 . 4 At different points in time, the ring has
different velocities vy and exerts different forces Fy on the left string. Butpthe ratio Fy /vy is
always equal to −T2 /v2 , which is constant, given T2 and µ2 (since v2 = T2 /µ2 ). So, since
Fy /vy = −T2 /v2 is constant in time, it makes sense to give it a name, and we call it the
impedance, Z. This is consistent with the Z ≡ T /v definition in Eq. (37). From Eq. (41),
the impedance Z is simply the damping coefficient b. Large damping therefore means large
impedance, so the “impedance” name makes colloquial sense.
√
Since the impedance Z ≡ T /v equals T µ from Eq. (37), it is a property of the string
itself (given T and µ), and not of a particular wave motion on the string. From Eq. (38)
we see that if Z1 = Z2 , then R = 0 and T = 1. In other words, there is total transmission.
In this case we say that the strings are “impedance matched.” We’ll talk much more about
this below, but for now we’ll just note that there are many ways to make the impedances
match. One way is to simply have the left and right strings be identical, that is, T1 = T2
and µ1 = µ2 . In this case we have a uniform string, so the wave just moves merrily along
and nothing is reflected. But T2 = 3T1 and µ2 = µ1 /3 also yields matching impedances, as
do an infinite number of other scenarios. All we need is for the product T µ to be the same
in the two pieces, and then the impedances match and everything is transmitted. However,
in these cases, it isn’t obvious that there is no reflected wave, as it was for the uniform
string. The reason for zero reflection is that the left string can’t tell the √difference between
an identical string on the right, or a piston with a damping coefficient of T1 µ1 , or a string
with T2 = 3T1 and µ2 = µ1 /3. They all feel exactly the same, as √ far as the left string is
concerned; they all generate a transverse force of the form, Fy = − T1 µ1 · vy . So if there
is no reflection in one case (and there certainly isn’t in the case of an identical string), then
there is no reflection in any other case. As far as reflection and√transmission go, a string is
completely characterized by one quantity: p the impedance Z ≡ T µ. Nothing else matters.
Other quantities such as T , µ, and
√ v = T /µ are relevant for various other considerations,
but only the combination Z ≡ T µ appears in the R and T coefficients.
Although the word “impedance” makes colloquial sense, there is one connotation that
might be misleading. You might think that a small impedance allows a wave to transmit
easily and reflect essentially nothing back. But this isn’t the case. Maximal transmission
occurs when the impedances match, not when Z2 is small. (If Z2 is small, say Z2 = 0, then
Eq. (38) tells us that we actually have total reflection with R = 1.) When we discuss energy
in Section 4.4, we’ll see that impedance matching yields maximal energy transfer, consistent
with the fact that no energy remains on the left side, because there is no reflected wave.

4 Remember that F and v are the transverse force and velocity, which are generally very small, given
y y
our usual assumption of a small slope of the string. But T2 and v2 are the tension and wave speed on the
right side, which are “everyday-sized” quantities. What we showed in Eq. (41) was that the two ratios,
Fy /vy and −T2 /v2 , are always equal.
4.3. IMPEDANCE 15

√
Why Fy is proportional to Tµ
We saw above that the transverse force that the left string (or technically the ring at the
boundary) applies to the right string is Fy = +bvy ≡ Zvy . So if you replace the left string
with your hand, then Fy = Zvy is the transverse force than you must apply to the right
string to give it the same motion that it had when the left string was there. The impedance
Z gives a measure of how hard√it is to wiggle the end of the string back and forth. It is
therefore reasonable that Z = T2 µ2 grows with both T2 and µ2 . In particular, if µ2 is
large, then more force should be required in order to wiggle the string in a given manner.
However, although this general dependence on µ2 seems quite intuitive, you have to be
careful, because there is a common incorrect way of thinking about things. The reason why
the force grows with µ2 is not the same as the reason why the force grows with m in the
simple case of a single point mass (with no string or anything else attached to it). In that
case, if you wiggle a point mass back and forth, then the larger the mass, the larger the
necessary force, due to F = ma.
But in the case of a string, if you grab onto the leftmost atom of the right part of the
string, then this atom is essentially massless, so your force isn’t doing any of the “F = ma”
sort of acceleration. All your force is doing is simply balancing the transverse component of
the T2 tension that the right string applies to its leftmost atom. This transverse component
is nonzero due to the (in general) nonzero slope. So as far as your force is concerned, all
that matters are the values of T2 and thepslope. And the slope is where the dependence
on µ2 comes in. If µ2 is large, then v2 = T2 /µ2 is small, which means that the wave in
the right string is squashed by a factor of v2 /v1 compared with the wave on the left string.
This then means
p that the slope of the right part is larger by a factor that is proportional
to 1/v2 = µ2 /T2 , which in turn means that the transverse force is larger. Since the
transverse force is proportional
p to
√ the product of the tension and the slope, we see that it
is proportional to T2 µ2 /T2 = T2 µ2 . To sum up: µ2 affects the impedance not because
of an F = ma effect, but rather because µ affects the wave’s speed, and hence slope, which
then affects the transverse component of the force.
A byproduct
√ of this reasoning is that the dependence of the transverse force on T2 takes
the form of T2 . This comes from the expected factor of T2 which arises from the fact
that √the transverse force is proportional to the tension. But there is an additional factor
of √
1/ T2 , because the transverse force is also proportional to the slope, which behaves like
1/ T2 from the argument in the previous paragraph.

Why Fy is proportional to vy
We saw above that if the right√ string is removed and if the ring is attached to a piston with
a damping coefficient of b = T2 µ2 , then the left string can’t tell the difference. Either way,
the force on the left string takes the form of −bvy ≡ −bẏ. If instead of a piston we attach
the ring to a transverse spring, then the force that the ring applies to the left string (which
equals the force that the spring applies to the ring, since the ring is massless) is −ky. And
if the ring is instead simply a mass that isn’t attached to anything, then the force it applies
to the left string is −mÿ (equal and opposite to the Fy = may force that the string applies
to the mass). Neither of these scenarios mimics the correct −bẏ force that the right string
actually applies.
This −bẏ force from the right string is a consequence of the “wavy-ness” of waves, for
the following reason. The transverse force Fy that the right string applies to the ring is
proportional to the slope of the wave; it equals the tension times the slope, assuming the
slope is small:
∂ψR
Fy = T2 . (42)
∂x
 v2
v2 dt tanθ
wave at t 16 CHAPTER 4. TRANSVERSE WAVES ON A STRING
θ
wave at t + dt And the transverse velocity is also proportional to the (negative of the) slope, due to the
v2 dt
properties of Fig. 16. The left tilted segment is a small piece of the wave at a given time t,
Figure 16 and the right tilted segment is the piece at a later time t + dt (the wave is moving to the
right). The top dot is the location of a given atom at time t, and the bottom dot is the
location of the same atom at time t + dt. The wave moves a distance v2 dt to the right, so
from the triangle shown, the dot moves a distance (v2 dt) tan θ downward. The dot’s velocity
is therefore −v2 tan θ, which is −v2 times the slope. That is,

∂ψR
vy = −v2 . (43)
∂x
We see that both the transverse force in Eq. (42) and the transverse velocity in Eq. (43)
are proportional to the slope, with the constants of proportionality being T2 and −v2 ,
respectively. The ratio Fy /vy is therefore independent of the slope. It equals −T2 /v2 , which
is constant in time.
Fig. 16 is the geometric explanation of the mathematical relation in Eq. (40). Written
in terms of ψR instead of ft , Eq. (40) says that

∂ψR ∂ψR
= −v2 · . (44)
∂t ∂x
That is, the transverse velocity is −v2 times the slope, which is simply the geometrically-
derived result in Eq. (43). Note that this relation holds only for a single traveling wave. If
we have a wave that consists of, say, two different traveling waves, ψ(x, t) = fa (t − x/va ) +
fb (t − x/vb ), then

∂ψ ∂fa ∂fb
= + ,
∂t ∂t ∂t
∂ψ 1 ∂fa 1 ∂fa
= − · − · . (45)
∂x va ∂t vb ∂t

Looking at the righthand sides of these equations, we see that it is not the case that ∂ψ/∂t =
−v · ∂ψ/∂x for a particular value of v. It doesn’t work for v1 or v2 or anything else. This
observation is relevant to the following question.
In the discussion leading up to Eq. (41), we considered the transverse force that the ring
applies to the string on its left, and the result was −(T2 /v2 )vy . From Newton’s third law,
the transverse force that the ring applies to the string on its right is therefore +(T2 /v2 )vy .
However, shouldn’t we be able to proceed through the above derivation with “left” and
“right” reversed, and thereby conclude that the force that the ring applies to the string on
its right is equal to +(T1 /v1 )vy ? The answer had better be a “no,” because this result isn’t
consistent with the +(T2 /v2 )vy result unless T1 /v1 = T2 /v2 , which certainly doesn’t hold
for arbitrary choices of strings. Where exactly does the derivation break down? The task
of Problem [to be added] is to find out.

4.3.2 Examples of impedance matching

In this section we’ll present (without proof) a number of examples of impedance matching.
The general point of matching impedances is to yield the maximal energy transfer from one
thing to another. As we saw above in the case of the string with different densities, if the
impedances are equal, then nothing is reflected. So 100% of the energy is transferred, and
you can’t do any better than that, due to conservation of energy.
If someone gives you two impedance-matched strings (so the product T µ is the same in
both), then all of the wave is transmitted. If the above conservation-of-energy argument slips
4.3. IMPEDANCE 17

your mind, you might (erroneously) think that you could increase the amount of transmitted
energy by, say, decreasing µ2 . This has the effect of decreasing Z2 and thus increasing the
transmission coefficient T in Eq. (38). And you might think that the larger amplitude of
the transmitted wave implies a larger energy. However, this isn’t the case, because less mass
is moving now in the right string, due to the smaller value of µ2 . There are competing
effects, and it isn’t obvious from this reasoning which effect wins. But it is obvious from the
conservation-of-energy argument. If the Z’s aren’t equal, then there is nonzero reflection,
by Eq. (38), and therefore less than 100% of the energy is transmitted. This can also be
demonstrated by explicitly calculating the energy. We’ll talk about energy in Section 4.4
below.
The two basic ways to match two impedances are to (1) simply make one of them equal
to the other, or (2) keep them as they are, but insert a large number of things (whatever
type of things the two original ones are) between them with impedances that gradually
change from one to the other. It isn’t obvious that this causes essentially all of the energy
to be transferred, but this can be shown without too much difficulty. The task of Problem
[to be added] is to demonstrate this for the case of the “Gradually changing string density”
mentioned below. Without going into too much detail, here are a number of examples of
impedance matching:

Electrical circuits: The general expression for impedance is Z = F/v. In a purely

resistive circuit, the analogs of Z, F , and v are, respectively, the resistance R, the voltage
V , and the current I. So Z = F/v becomes R = V /I, which is simply Ohm’s law. If a
source has a given impedance (resistance), and if a load has a variable impedance, then we
need to make the load’s impedance equal to the source’s impedance, if the goal is to have
the maximum power delivered to the load. In general, impedances can be complex if the
circuit isn’t purely resistive, but that’s just a fancy way of incorporating a phase difference
when V and I (or F and v in general) aren’t in phase.

Gradually changing string density: If we have two strings with densities µ1 and µ2 ,
and if we insert between them a long string (much longer than the wavelength of the wave
being used) whose density gradually changes from µ1 to µ2 , then essentially all of the wave
is transmitted. See Problem [to be added].

Megaphone: A tapered megaphone works by the same principle. If you yell into a simple
cylinder, then it turns out that the abrupt change in cross section (from a radius of r to
a radius of essentially infinity) causes reflection. The impedance of a cavity depends on
the cross section. However, in a megaphone the cross section varies gradually, so not much
sound is reflected back toward your mouth. The same effect is relevant for a horn.

Ultrasound: The gel that is put on your skin has the effect of impedance matching the
waves in the device to the waves in your body.

Ball collisions: Consider a marble that collides elastically with a bowling ball. The
marble will simply bounce off and give essentially none of its energy to the bowling ball.
All of the energy will remain in the marble. (And similarly, if a bowling ball collides with
a marble, then the bowling ball will simply plow through and keep essentially all of the Figure 17
energy.) But if a series of many balls, each with slightly increasing size, is placed between
them (see Fig. 17), then it turns out that essentially all of the marble’s energy will end up in
the bowling ball. Not obvious, but true. And conversely, if the bowling ball is the one that
is initially moving (to the left), then essentially all of its energy will end up in the marble,
which will therefore be moving very fast.
It is nebulous what impedance means for one-time evens like collisions between balls,
 18 CHAPTER 4. TRANSVERSE WAVES ON A STRING

because we defined impedance for waves. But the above string of balls is certainly similar
to a longitudinal series of masses (with increasing size) and springs. The longitudinal waves
that travel along this spring/mass system consist of many “collisions” between the masses.
In the original setup with just the string of balls and no springs, when two balls collide
they smush a little and basically act like springs. Well, sort of; they can only repel and not
attract. At any rate, if you abruptly increased the size of the masses in the spring/mass
system by a large factor, then not much of the wave would make it through. But gradually
increasing the masses would be just like gradually increasing the density µ in the “Gradually
changing string density” example above.
Lever: If you try to lift a refrigerator that is placed too far out on a lever, you’re not
going to be able to do it. If you jumped on your end, you would just bounce off like on a
springboard. You’d keep all of the energy, and none of it would be transmitted. But if you
move the refrigerator inward enough, you’ll be able to lift it. However, if you move it in too
far (let’s assume it’s a point mass), then you’re back to essentially not being able to lift it,
because you’d have to move your end of the lever, say, a mile to lift the refrigerator up by
a foot. So there is an optimal placement.
Bicycle: The gears on a bike act basically the same way as a lever. If you’re in too high a
gear, you’ll find that it’s too hard on your muscles; you can’t get going fast. And likewise,
if you’re in too low a gear, your legs will just spin wildly, and you’ll be able to go only so
fast. There is an optimal gear ratio that allows you to transfer the maximum amount of
energy from chemical potential energy (from your previous meal) to kinetic energy.
Rolling a ball up a ramp: This is basically just like a lever or a bike. If the ramp is
too shallow, then the ball doesn’t gain much potential energy. And if it’s too steep, then
you might not be able to move the ball at all.

4.4 Energy
Energy
What is the energy of a wave? Or more precisely, what is the energy density per unit
length? Consider a little piece of the string between x and x + dx. In general, this piece
has both kinetic and potential energy. The kinetic energy comes from the transverse motion
(we showed in the paragraph following Eq. (1) that the longitudinal motion is negligible),
so it equals
µ ¶2
1 1 ∂ψ
2 dψ
2 Kdx = (dm)vy2 = (µ dx) . (46)
dx + dψ
2 2 ∂t
We have used the fact that since there is essentially no longitudinal motion, the mass within
dx the span from x to x + dx is always essentially equal to µ dx.
The potential energy depends on the stretch of the string. In general, a given piece of
Figure 18 the string
√ is tilted and looks like the piece shown in Fig. 18. As we saw in Eq. (1), the Taylor
series 1 + ² ≈ 1 + ²/2 gives the length of the piece as
s µ ¶2 µ ¶2
∂ψ dx ∂ψ
dx 1 + ≈ dx + . (47)
∂x 2 ∂x

The piece is therefore stretched by an amount, d` ≈ (dx/2)(∂ψ/∂x)2 . 5 This stretch is

caused by the external tension forces at the two ends. These forces do an amount T d` of
5 Actually, I think this result is rather suspect, although it doesn’t matter in the end. See the remark

below.
4.4. ENERGY 19

work, and this work shows up as potential energy in the piece, exactly in the same way that
a normal spring acquires potential energy if you grab an end and stretch it. So the potential
energy of the piece is
µ ¶2
1 ∂ψ
Udx = T dx . (48)
2 ∂x

The total energy per unit length (call it E) is therefore

µ ¶2 µ ¶2
Kdx + Udx µ ∂ψ T ∂ψ
E(x, t) = = +
dx 2 ∂t 2 ∂x
"µ ¶2 µ ¶2 #
µ ∂ψ T ∂ψ
= +
2 ∂t µ ∂x
"µ ¶2 µ ¶2 #
µ ∂ψ 2 ∂ψ
= +v , (49)
2 ∂t ∂x
p
where we have used v = T /µ. This expression for E(x, t) is valid for an arbitrary wave.
But let’s now look at the special case of a single traveling wave, which takes the form of
ψ(x, t) = f (x ± vt). For such a wave, the energy density can be further simplified. As
we have seen many times, the partial derivatives of a single traveling wave are related by
∂ψ/∂t = ±v ∂ψ/∂x. So the two terms in the expression for E(x, t) are equal at a given point
and at a given time. We can therefore write

µ ¶2 µ ¶2
∂ψ 2 ∂ψ
E(x, t) = µ or E(x, t) = µv (for traveling waves) (50)
∂t ∂x

√ p
Or equivalently, we can use Z ≡ T µ and v = T /µ to write energy density
µ ¶2 µ ¶2
Z ∂ψ ∂ψ
E(x, t) = , or E(x, t) = Zv . (51)
v ∂t ∂x

For sinusoidal traveling waves, the energy density is shown in Fig. 19 (with arbitrary units
on the axes). The energy-density curve moves right along with the wave. wave
v
Remark : As mentioned in Footnote 5, the length of the string given in Eq. (47) and the resulting zero stretch, max stretch,
expression for the potential energy given in Eq. (48) are highly suspect. The reason is the following. zero speed max speed
In writing down Eq. (47), we made the assumption that all points on the string move in the
transverse direction; we assumed that the longitudinal motion is negligible. This is certainly the Figure 19
case (in an exact sense) if the string consists of little masses that are hypothetically constrained
to ride along rails pointing in the transverse direction. If these masses are connected by little
stretchable pieces of massless string, then Eq. (48) correctly gives the potential energy.
However, all that we were able to show in the reasoning following Eq. (1) was that the points
on the string move in the transverse direction, up to errors of order dx(∂ψ/∂x)2 . We therefore have
no right to trust the result in Eq. (48), because it is of the same order. But even if this result is
wrong and if the stretching of the string is distributed differently from Eq. (47), the total amount
of stretching is the same. Therefore, because the work done on the string, T d`, is linear in d`, the
total potential energy is independent of the particular stretching details. The (T /2)(∂ψ/∂x)2 result
therefore correctly yields the average potential energy density, even though it may be incorrect at
individual points. And since we will rarely be concerned with more than the average, we can use
Eq. (48), and everything is fine. ♣
 20 CHAPTER 4. TRANSVERSE WAVES ON A STRING
vy = dψ/dt
Power
slope = dψ/dx
What is the power transmitted across a given point on the string? Equivalently, what is the
rate of energy flow past a given point? Equivalently again, at what rate does the string to
Figure 20 the left of a point do work on the string to the right of the point? In Fig. 20, the left part
of the string pulls on the dot with a transverse force of Fy = −T ∂ψ/∂x. The power flow
across the dot (with rightward taken to be positive) is therefore
µ ¶µ ¶
dW Fy dψ ∂ψ ∂ψ ∂ψ
P (x, t) = = = Fy = F vy = −T . (52)
dt dt ∂t ∂x ∂t

This expression for P (x, t) is valid for an arbitrary wave. But as with the energy density
above, we can simplify the expression if we consider the special case of single traveling wave.
If ψ(x, t) p
= f (x ± vt),
√ then ∂ψ/∂x = ±(1/v)∂ψ/∂t. So we can write the power as (using
T /v = T / T /µ = T µ ≡ Z)
µ ¶2 µ ¶2
T ∂ψ ∂ψ
P (x, t) = ∓ =⇒ P (x, t) = ∓Z = ∓vE(x, t) (53)
v ∂t ∂t

where we have used Eq. (51). We see that the magnitude of the power is simply the wave
speed times the energy density. It is positive for a rightward traveling wave f (x − vt), and
negative for a leftward traveling wave f (x + vt) (we’re assuming that v is a positive quantity
here). This makes sense, because the energy plot in Fig. 19 just travels along with the wave
at speed v.

Momentum
A wave on a string carries energy due to the transverse motion. Does such a wave carry
momentum? Well, there is certainly nonzero momentum in the transverse direction, but it
averages out to zero because half of the string is moving one way, and half is moving the
other way.
What about the longitudinal direction? We saw above that the points on the string move
only negligibly in the longitudinal direction, so there is no momentum in that direction. Even
though a traveling wave makes it look like things are moving in the longitudinal direction,
there is in fact no such motion. Every point in the string is moving only in the transverse
direction. Even if the points did move non-negligible distances, the momentum would still
average out to zero, consistent with the fact that there is no overall longitudinal motion
of the string. The general kinematic relation p = mvCM holds, so if the CM of the string
doesn’t move, then the string has no momentum.
There are a few real-world examples that might make you think that standard traveling
waves can carry momentum. One example is where you try (successfully) to move the other
end of a rope (or hose, etc.) you’re holding, which lies straight on the ground, by flicking
the rope. This causes a wave to travel down the rope, which in turn causes the other end to
move farther away from you. Everyone has probably done this at one time or another, and
it works. However, you can bet that you moved your hand forward during the flick, and this
is what gave the rope some longitudinal momentum. You certainly must have moved your
hand forward, of course, because otherwise the far end couldn’t have gotten farther away
from you (assuming that the rope can’t stretch significantly). If you produce a wave on a
rope by moving your hand only up and down (that is, transversely), then the rope will not
have any longitudinal momentum.
Another example that might make you think that waves carry momentum is the case
of sound waves. Sound waves are longitudinal waves, and we’ll talk about these in the
4.5. STANDING WAVES 21

following chapter. But for now we’ll just note that if you’re standing in front of a very large
speaker (large enough so that you feel the sound vibrations), then it seems like the sound is
applying a net force on you. But it isn’t. As we’ll see in the next chapter, the pressure on
you alternates sign, so half the time the sound wave is pushing you away from the speaker,
and half the time it’s drawing you closer. So it averages out to zero.6
An exception to this is the case of a pulse or an explosion. In this case, something at the
source of the pulse must have actually moved forward, so there is some net momentum. If
we have only half (say, the positive half) of a cycle of a sinusoidal pressure wave, then this
part can push you away without the (missing) other half drawing you closer. But this isn’t
how normal waves (with both positive and negative parts) work.

4.5 Standing waves

4.5.1 Semi-infinite string
Fixed end
Consider a leftward-moving sinusoidal wave that is incident on a brick wall at its left end,
located at x = 0. (We’ve having the wave move leftward instead of the usual rightward for
later convenience.) The most general form of a leftward-moving sinusoidal wave is

ψi (x, t) = A cos(ωt + kx + φ), (54)

p
where ω and k satisfy ω/k = T /µ = v. A is the amplitude, and φ depends on the
arbitrary choice of the t = 0 time. Since the brick wall has “infinite” impedance, Eq. (38)
gives R = −1. Eq. (32) then gives the reflected rightward-moving wave as

ψr (x, t) = Rψi (−x, t) = −A cos(ωt − kx + φ). (55)

The total wave is therefore

ψ(x, t) = ψi (x, t) + ψr (x, t) = A cos(ωt + φ + kx) − A cos(ωt + φ − kx)

= −2A sin(ωt + φ) sin kx (56)

As a double check, this satisfies the boundary condition ψ(0, t) = 0 for all t. The sine
function of x is critical here. A cosine function wouldn’t satisfy the boundary condition at
x = 0. In contrast with this, it doesn’t matter whether we have a sine or cosine function of
t, because a phase shift in φ can turn one into the other.
For a given value of t, a snapshot of this wave is a sinusoidal function of x. The wavelength
of this function is λ = 2π/k, and the amplitude is |2A sin(ωt + φ)|. For a given value of x,
each point oscillates as a sinusoidal function of t. The period of this function is τ = 2π/ω,
and the amplitude is |2A sin kx|. Points for which kx equals nπ always have ψ(x, t) = 0, so
they never move. These points are called nodes.
Fig. 21 shows the wave at a number of different times. A wave such as the one in Eq.
(fixed left end)
(56) is called a “standing wave.” All points on the string have the same phase (or differ by
π), as far as the oscillations in time go. That is, all of the points come to rest at the same Figure 21
time (at the maximal displacement from equilibrium), and they all pass through the origin
at the same time, etc. This is not true for traveling waves. In particular, in a traveling
6 The opening scene from the first Back to the Future movie involves Marty McFly standing in front of

a humongous speaker with the power set a bit too high. He then gets blown backwards when he plays a
chord. This isn’t realistic, but ingenious movies are allowed a few poetic licenses.
 22 CHAPTER 4. TRANSVERSE WAVES ON A STRING

wave, the points with ψ = 0 are moving with maximal speed, and the points with maximum
ψ are instantaneously at rest.
If you don’t want to have to invoke the R = −1 coefficient, as we did above, another
way of deriving Eq. (56) is to apply the boundary condition at x = 0 to the most general
form of the wave given in Eq. (8). Since ψ(0, t) = 0 for all t, we can have only the sin kx
terms in Eq. (8). Therefore,
ψ(x, t) = D2 sin kx sin ωt + D3 sin kx cos ωt
= (D2 sin ωt + D3 cos ωt) sin kx
≡ B sin(ωt + φ) sin kx, (57)
where B and φ are determined by B cos φ = D2 and B sin φ = D3 .
massless ring
Free end
Consider now a leftward-moving sinusoidal wave that has its left end free (located at x = 0).
fixed pole By “free” here, we mean that a massless ring at the end of the string is looped around a
fixed frictionless pole pointing in the transverse direction; see Fig. 22. So the end is free
Figure 22 to move transversely but not longitudinally. The pole makes it possible to maintain the
tension T in the string. Equivalently, you can consider the string to be infinite, but with a
density of µ = 0 to the left of x = 0.
As above, the most general form of a leftward-moving sinusoidal wave is
ψi (x, t) = A cos(ωt + kx + φ), (58)
Since the massless ring (or equivalently the µ = 0 string) has zero impedance, Eq. (38) gives
R = +1. Eq. (32) then gives the reflected rightward-moving wave as
ψr (x, t) = Rψi (−x, t) = +A cos(ωt − kx + φ). (59)
The total wave is therefore
ψ(x, t) = ψi (x, t) + ψr (x, t) = A cos(ωt + φ + kx) + A cos(ωt + φ − kx)
= 2A cos(ωt + φ) cos kx (60)
As a double check, this satisfies the boundary condition, ∂ψ/∂x|x=0 = 0 for all t. The slope
fixed pole must always be zero at x = 0, because otherwise there would be a net transverse force on
the massless ring, and hence infinite acceleration. If we choose to construct this setup with
a µ = 0 string to the left of x = 0, then this string will simply rise and fall, always remaining
horizontal. You can assume that the other end is attached to something very far to the left
of x = 0.
Fig. 23 shows the wave at a number of different times. This wave is similar to the one
in Fig. 21 (it has the same amplitude, wavelength, and period), but it is shifted a quarter
cycle in both time and space. The time shift isn’t of too much importance, but the space
(free left end)
shift is critical. The boundary at x = 0 now corresponds to an “antinode,” that is, a point
Figure 23 with the maximum oscillation amplitude.
As in the fixed-end case, if you don’t want to have to invoke the R = 1 coefficient, you
can apply the boundary condition, ∂ψ/∂x|x=0 = 0, to the most general form of the wave
given in Eq. (8). This allows only the cos kx terms.
In both this case and the fixed-endp case, ω and k can take on a continuous set of values,
as long as they are related by ω/k = T /µ = v.7 In the finite-string cases below, we will
find that they can take on only discrete values.
7 Even
though the above standing waves don’t travel anywhere and thus don’t have a speed, it still makes
p
sense to label the quantity T /µ as v, because standing waves can be decomposed into two oppositely-
moving traveling waves, as shown in Eqs. (56) and (60).
4.5. STANDING WAVES 23

4.5.2 Finite string

We’ll now consider the three possible (extreme) cases for the boundary conditions at the
two ends of a finite string. We can have both ends fixed, or one fixed and one free, or both
free. Let the ends be located at x = 0 and x = L. In general, the boundary conditions
(for all t) are ψ = 0 at a fixed end (because the end never moves), and ∂ψ/∂x = 0 at a
free end (because the slope must be zero so that there is no transverse force on the massless
endpoint).

Two fixed ends

If both ends of the string are fixed, then the two boundary conditions are ψ(0, t) = 0 and
ψ(L, t) = 0 for all t. Eq. (56) gives the most general form of a wave (with particular ω and
k values) satisfying the first of these conditions. So we just need to demand that the second
one is also true. If we plug x = L into Eq. (56), the only way to have ψ(L, t) = 0 for all t is
to have sin kL = 0. This implies that kL must equal nπ for some integer n. So
nπ
kn = , (61)
L
where we have added the subscript to indicate that k can take on a discrete set of values
associated with the integers n. n signifies which “mode” the string is in. The wavelength
is λn = 2π/kn = 2L/n. So the possible wavelengths are all integral divisors of 2L (which
is twice the length of the string). Snapshots of the first few modes are shown below in the
first set of waves in Fig. 24. The n values technically start at n = 0, but in this case ψ is
identically equal to zero since sin(0) = 0. This is certainly a physically possible location of
the string, but it is the trivial scenario. So the n values effectively
p start at 1.
The angular frequency ω is still related to k by ω/k = T /µ = v, so we have ωn = vkn .
Remember that v depends only on T and µ, and not on n (although this won’t be true when
we get to dispersion in Chapter 6). The frequency in Hertz is then
ωn vkn v(nπ/L) nv
νn = = = = . (62)
2π 2π 2π 2L
The possible frequencies are therefore integer multiples of the “fundamental” frequency,
ν1 = v/2L. In short, the additional boundary condition at the right end constrains the
system so that only discrete values of ω and k are allowed. Physically, the wave must
undergo an integral number of half oscillations in space, in order to return to the required
value of zero at the right end. This makes it clear thatp λ (and hence k) can take on only
discrete values. And since the ratio ω/k is fixed to be T /µ, this means that ω (and ν)
can take on only discrete values too. To summarize:
2L nv
λn = and νn = (63)
n 2L
The product of these quantities is λn νn = v, as it should be.
Since the wave equation in Eq. (4) is linear, the most general motion of a string with two
fixed ends is an arbitrary linear combination of the solutions in Eq. (56), with the restriction
that k takes the form kn = nπ/L (and ω/k must equal v). So the most general expression for
ψ(x, t) is (we’ll start the sum at n = 0, even though this term doesn’t contribute anything)
∞
X nπ
ψ(x, t) = Bn sin(ωn t + φn ) sin kn x where kn = , and ωn = vkn . (64)
n=0
L

The B here equals the −2A from Eq. (56). Note that the amplitudes and phases of the
various modes can in general be different.
24 CHAPTER 4. TRANSVERSE WAVES ON A STRING

One fixed end, one free end

Now consider the case where one end is fixed and the other is free. Let’s take the fixed end
to be the one at x = 0. The case where the fixed end is located at x = L gives the same
general result; it’s just the mirror image of the result we will obtain here.
The two boundary conditions are ψ(0, t) = 0 and ∂ψ/∂x|x=L = 0 for all t. Eq. (56) again
gives the most general form of a wave satisfying the first of these conditions. So we just need
to demand that the second one is also true. From Eq. (56), the slope ∂ψ/∂x is proportional
to cos kx. The only way for this to be zero at x = L is for kL to equal (n + 1/2)π for some
integer n. So
(n + 1/2)π
kn = . (65)
L
n starts at zero here. Unlike in the two-fixed-ends case, the n = 0 value now gives a
nontrivial wave.
The wavelength is λn = 2π/kn = 2L/(n + 1/2). These wavelengths are most easily seen
in pictures, and snapshots of the first few modes are shown below in the second set of waves
in Fig. 24. The easiest way to describe the wavelengths in words is to note that the number
of oscillations that fit on the string is L/λn = n/2 + 1/4. So for the lowest mode (the n = 0
one), a quarter of a wavelength fits on the string. The higher modes are then obtained by
successively adding on half an oscillation (which ensures that x = L is always located at
an antinode, with slope zero). The frequency νn can be found via Eq. (62), and we can
summarize the results:
2L (n + 1/2)v
λn = and νn = (66)
n + 1/2 2L

Similar to Eq. (64), the most general motion of a string with one fixed end and one free end
is a linear combination of the solutions in Eq. (56):
∞
X (n + 1/2)π
ψ(x, t) = Bn sin(ωn t + φn ) sin kn x where kn = , and ωn = vkn .
n=0
L
(67)
If we instead had the left end as the free one, then Eq. (60) would be the relevant equation,
and the sin kx here would instead be a cos kx. As far as the dependence on time goes, it
doesn’t matter whether it’s a sine or cosine function of t, because a redefinition of the t = 0
point yields a phase shift that can turn sines into cosines, and vice versa. We aren’t free to
redefine the x = 0 point, because we have a physical wall there.

Two free ends

Now consider the case with two free ends. The two boundary conditions are ∂ψ/∂x|x=0 = 0
and ∂ψ/∂x|x=L = 0 for all t. Eq. (60) gives the most general form of a wave satisfying the
first of these conditions. So we just need to demand that the second one is also true. From
Eq. (60), the slope ∂ψ/∂x is proportional to sin kx. The only way for this to be zero at
x = L is for kL to equal nπ for some integer n. So
nπ
kn = , (68)
L
which is the same as in the case of two fixed ends. The wavelength is λn = 2π/kn = 2L/n.
So the possible wavelengths are all integral divisors of 2L, again the same as in the case of
two fixed ends. Snapshots of the first few modes are shown below in the third set of waves
in Fig. 24.
4.5. STANDING WAVES 25

The n values technically start at n = 0. In this case ψ has no dependence on x, so we

simply have a flat line (not necessarily at ψ = 0). The line just sits there at rest, because
the frequency is again given by Eq. (62) and is therefore zero (the kn values, and hence ωn
values, are the same as in the two-fixed-ends case, so ω0 = 0). This case isn’t as trivial as
the n = 0 case for two fixed ends; the resulting constant value of ψ might be necessary to
satisfy the initial conditions of the string. This constant value is analogous to the a0 term
in the Fourier-series expression in Eq. (3.1).
As with two fixed ends, an integral number of half oscillations must fit into L, but they
now start and end at antinodes instead of nodes. The frequency νn is the same as in the
two-fixed-ends case, so as in that case we have:

2L nv
λn = and νn = (69)
n 2L

Similar to Eqs. (64) and (67), the most general motion of a string with two free ends is a
linear combination of the solutions in Eq. (60):
∞
X nπ
ψ(x, t) = Bn cos(ωn t + φn ) cos kn x where kn = , and ωn = vkn . (70)
n=0
L

Fig. 24 summarizes the above results.

2L/λ both ends fixed one fixed, one free both ends free

4
7/2
3
5/2
2
3/2
1
1/2
0 fixed end free end

Figure 24

Power in a standing wave

We saw above in Section 4.4 that not only do traveling waves contain energy, they also
contain an energy flow along the string. That is, they transmit power. A given point on
the string does work (which may be positive or negative, depending on the direction of the
wave’s velocity) on the part of the string to its right. And it does the opposite amount of
work on the string to its left.
A reasonable question to ask now is: Is there energy flow in standing waves? There is
certainly an energy density, because in general the string both moves and stretches. But is
there any energy transfer along the string?
Intuitively, a standing wave is the superposition of two oppositely-moving traveling waves
with equal amplitudes. These traveling waves have equal and opposite energy flow, on
 26 CHAPTER 4. TRANSVERSE WAVES ON A STRING

average, so we expect the net energy flow in a standing wave to be zero, on average. (We’ll
see below where this “on average” qualification arises.) Alternatively, we can note that
there can’t be any net energy flow in either direction in a standing wave, due to the left-
right symmetry of the system. If you flip the paper over, so that right and left are reversed,
then a standing wave looks exactly the same, whereas a traveling wave doesn’t, because it’s
now moving in the opposite direction.
Mathematically, we can calculate the energy flow (that is, the power) as follows. The
expression for the power in Eq. (52) is valid for an arbitrary wave. This is the power flow
across a given point, with rightward taken to be positive. Let’s see what Eq. (52) reduces
to for a standing wave. We’ll take our standing wave to be A sin ωt sin kx. (We could have
any combination of sines and cosines here; they all give the same general result.) Eq. (52)
becomes
µ ¶µ ¶
∂ψ ∂ψ
P (x, t) = −T
∂x ∂t
¡ ¢¡ ¢
= −T kA sin ωt cos kx ωA cos ωt sin kx
¡ ¢¡ ¢
= −T A2 ωk sin kx cos kx sin ωt cos ωt . (71)

In general, this is nonzero, so there is energy flow across a given point. However, at a given
value of x, the average (over one period) of sin ωt cos ωt (which equals (1/2) sin 2ωt) is zero.
So the average power is zero, as we wanted to show.
The difference between a traveling wave and a standing wave is the following. Mathe-
matically: for a traveling wave of the form A cos(kx − ωt), the two derivatives in Eq. (71)
produce the same sin(kx − ωt) function, so we end up with the square of a function, which
is always positive. There can therefore be no cancelation. But for a standing wave, Eq. (71)
yields a bunch of different functions and no squares, and they average out to zero.
Physically: in a traveling wave, the transverse force that a given dot on the string applies
to the string on its right is always in phase (or 180◦ out of phase, depending on the direction
of the wave’s motion) with the velocity of the dot. This is due to the fact that ∂ψ/∂x is
proportional to ∂ψ/∂t for a traveling wave. So the power, which is the product of the
transverse force and the velocity, always has the same sign. There is therefore never any
cancelation between positive and negative amounts of work being done.
However, for the standing wave A sin ωt sin kx, the transverse force is proportional to
W=0 −∂ψ/∂x = −kA sin ωt cos kx, while the velocity is proportional to ∂ψ/∂t = ωA cos ωt sin kx.
W<0 For a given value of x, the sin kx and cos kx functions are constant, so the t dependence
W=0 tells us that the transverse force is 90◦ ahead of the velocity. So half the time the force is
in the same direction as the velocity, and half the time it is in the opposite direction. The
W>0 product integrates to zero, as we saw in Eq. (71).
W=0 t The situation is summarized in Fig. 25, which shows a series of nine snapshots throughout
W<0 a full cycle of a standing wave. W is the work that the dot on the string does on the string
W=0 to its right. Half the time W is positive, and half the time it is negative. The stars show
the points with maximum energy density. When the string is instantaneously at rest at
W>0
maximal curvature, the nodes have the greatest energy density, in the form of potential
W=0 t=0 energy. The nodes are stretched maximally, and in contrast there is never any stretching
at the antinodes. There is no kinetic energy anywhere in the string. A quarter cycle later,
Figure 25
when the string is straight and moving quickest, the antinodes have the greatest energy
density, in the form of kinetic energy. The antinodes are moving fastest, and in contrast
there is never any motion at the nodes. There is no potential energy anywhere in the string.
We see that the energy continually flows back and forth between the nodes and antinodes.
Energy never flows across a node (because a node never moves and therefore can do no
work), nor across an antinode (because an antinode never applies a transverse force and
4.6. ATTENUATION 27

therefore can do no work). The energy in each “half bump” (a quarter of a wavelength)
of the sinusoidal curve is therefore constant. It flows back and forth between one end (a
node/antinode) and the other end (an antinode/node). In other words, it flows back and
forth across a point such as the dot we have chosen in the figure. This is consistent with
the fact that the dot is doing work (positive or negative), except at the quarter-cycle points
where W = 0.

4.6 Attenuation
What happens if we add some damping to a transverse wave on a string? This damping
could arise, for example, by immersing the string in a fluid. As with the drag force in
the spring/mass system we discussed in Chapter 1, we’ll assume that this drag force is
proportional to the (transverse) velocity of the string. Now, we usually idealize a string as
having negligible thickness, but such a string wouldn’t experience any damping. So for the
purposes of the drag force, we’ll imagine that the string has some thickness that produces
a drag force of −(β dx)vy on a length dx of string, where β is the drag coefficient per unit
length. The longer the piece, the larger the drag force.
The transverse forces on a little piece of string are the drag force along with the force
arising from the tension and the curvature of the string, which we derived in Section 4.1.
So the transverse F = ma equation for a little piece is obtained by taking the F = ma
equation in Eq. (3) and tacking on the drag force. Since vy = ∂ψ/∂t, the desired F = ma
(or rather, ma = F ) equation is

∂2ψ ∂2ψ ∂ψ ∂2ψ ∂ψ 2

2∂ ψ
(µ dx) = T dx − (β dx) =⇒ + Γ = v (72)
∂t2 ∂x2 ∂t ∂t2 ∂t ∂x2

where Γ ≡ β/µ and v 2 = T /µ. To solve this equation, we’ll use our trusty method of
guessing an exponential solution. If we guess

ψ(x, t) = Dei(ωt−kx) (73)

and plug this into Eq. (72), we obtain, after canceling the factor of Dei(ωt−kx) ,

−ω 2 + Γ(iω) = −v 2 k 2 . (74)

This equation tells us how ω and k are related, but it doesn’t tell us what the motion looks
like. The motion can take various forms, depending on what the given boundary conditions
are. To study a concrete example, let’s look at the following system.
Consider a setup where the left end of the string is located at x = 0 (and it extends
rightward to x = ∞), and we arrange for that end to be driven up and down sinusoidally
with a constant amplitude A. In this scenario, ω must be real, because if it had a complex
value of ω = a + bi, then the eiωt factor in ψ(x, t) would involve a factor of e−bt , which
decays with time. But we’re assuming a steady-state solution with constant amplitude A at
x = 0, so there can be no decay in time. Therefore, ω must be real. The i in Eq. (74) then
implies that k must have an imaginary part. Define K and −iκ be the real and imaginary
parts of k, that is,
1p 2
k= ω − iΓω ≡ K − iκ. (75)
v
If you want to solve for K and κ in terms of ω, Γ, and v, you can square both sides of
this equation and solve a quadratic equation in K 2 or κ2 . But we won’t need the actual
values. You can show however, that if K and ω have the same sign (which they do, since
 28 CHAPTER 4. TRANSVERSE WAVES ON A STRING

we’re looking at a wave that travels rightward from x = 0), then κ is positive. Plugging
k ≡ K − iκ into Eq. (73) gives

ψ(x, t) = De−κx ei(ωt−Kx) . (76)

A similar solution exists with the opposite sign in the imaginary exponent (but the e−κx
stays the same). The sum of these two solutions gives the actual physical (real) solution,

ψ(x, t) = Ae−κx cos(ωt − Kx + φ) (77)

Ae -κx cos(ωt-Kx+φ) where the phase φ comes from the possible phase of D, which may be complex. (Equivalently,
1.0 you can just take the real part of the solution in Eq. (76).) The coefficient A has been
0.5 Ae -κx determined by the boundary condition that the amplitude equals A at x = 0. Due to the
0.0 e−κx factor, we see that ψ(x, t) decays with distance, and not time. ψ(x, t) is a rightward-
10 20 30 40 50 60
- 0.5 traveling wave with the function Ae−κx as its envelope. A snapshot in time is shown in
- 1.0 Fig. 26, where we have arbitrarily chosen A = 1, κ = 1/30, K = 1, and φ = π/3. The
snapshot corresponds to t = 0. Such a wave is called an attenuated wave, because it tapers
Figure 26 off as x grows.
Let’s consider the case of small damping. If Γ is small (more precisely, if Γ/ω is small),
then we can use a Taylor series to write the k in Eq. (75) as
r µ ¶
ω iΓ ω iΓ ω iΓ
k= 1− ≈ 1− = − ≡ K − iκ. (78)
v ω v 2ω v 2v

Therefore, κ = Γ/2v. The envelope therefore takes the form, Ae−Γx/2v . So after each
distance of 2v/Γ, the amplitude decreases by a factor 1/e. If Γ is very small, then this
distance is very large, which makes sense. Note that this distance 2v/Γ doesn’t depend
on ω. No matter how fast or slow the end of the string is wiggled, the envelope dies off
on the same distance scale of 2v/Γ (unless ω is slow enough so that we can’t work in the
approximation where Γ/ω is small). Note also that K ≈ ω/v in the Γ → 0 limit. This must
be the case, of course, because we must obtain the undamped result of k = ω/v when Γ = 0.
The opposite case of large damping (more precisely, large Γ/ω) is the subject of Problem
[to be added].
If we instead have a setup with a uniform wave (standing or traveling) on a string, and if
we then immerse the whole thing at once in a fluid, then we will have decay in time, instead
of distance. The relation in Eq. (74) will still be true, but we will now be able to say that k
must be real, because all points on the string are immersed at once, so there is no preferred
value of x, and hence no decay as a function of x. The i in Eq. (74) then implies that ω must
have an imaginary part, which leads to a time-decaying e−αt exponential factor in ψ(x, t).
Chapter 5

Longitudinal waves
David Morin, [email protected]

In Chapter 4 we discussed transverse waves, in particular transverse waves on a string. We’ll

now move on to longitudinal waves. Each point in the medium (whatever it consists of)
still oscillates back and forth around its equilibrium position, but now in the longitudinal
instead of the transverse direction. Longitudinal waves are a bit harder to visualize than
transverse waves, partly because everything is taking place along only one dimension, and
partly because of the way the forces arise, as we’ll see. Most of this chapter will be spent
on sound waves, which are the prime example of longitudinal waves.
The outline of this chapter is as follows. As a warm up, in Section 5.1 we take another
look at the longitudinal spring/mass system we originally studied in Section 2.4, where we
considered at the continuum limit (the N → ∞ limit). In Section 5.2 we study actual
sound waves. We derive the wave equation (which takes the same form as all the other wave
equations we’ve seen so far), and then look at the properties of the waves. In Section 5.3
we apply our knowledge of sound waves to musical instruments.

5.1 Springs and masses revisited

Recall that the wave equation for the continuous spring/mass system was given in Eq. (2.80)
as
∂ 2 ψ(x, t) E ∂ 2 ψ(x, t)
= , (1)
∂t2 µ ∂x2
where ψ is the longitudinal position relative to equilibrium, µ is the mass density, and E is
the elastic modulus. This wave equation is very similar to the one for transverse waves on
a string, which was given in Eq. (4.4) as
∂ 2 ψ(x, t) T ∂ 2 ψ(x, t)
2
= , (2)
∂t µ ∂x2
where ψ is the transverse position relative to equilibrium, µ is the mass density, and T is
the tension. ψ
B
These equations take exactly the same form, so all of the same results hold. However, the
fact that ψ is a longitudinal position in the former case, whereas it is a transverse position A E
x
in the latter, makes the former case a little harder to visualize. For example, if we plot ψ for C
a sinusoidal traveling wave (either transverse or longitudinal), we have the picture shown in D
Fig. 1. The interpretation of this picture depends on what kind of wave we’re talking about.
Figure 1
1
 2 CHAPTER 5. LONGITUDINAL WAVES

For a transverse wave, ψ is the transverse displacement, so Fig. 1 is what the string
ψ actually looks like from the side. The wave is therefore very easy to visualize – you just
B
need to look at the figure. It’s also fairly easy to see what the various points in Fig. 1 are
A E doing as the wave travels to the right. (Imagine that these dots are painted on the string.)
x
C Points B and D are instantaneously at rest, points A and E are moving downward, and
D point C is moving upward. To verify these facts, just draw the wave at a slightly later time.
wave at a
later time The result is shown in Fig. 2, with the new positions of the dots being represented by gray
dots. Remember that the points keep their same longitudinal position and simply move up
Figure 2 or down (or not at all). They don’t travel longitudinally along with the wave.
However, for a longitudinal wave, ψ is the longitudinal displacement, so although Fig. 1
is a perfectly valid plot of ψ, it does not indicate what the wave actually looks like. There
x is no transverse motion, so the system simply lies along a straight line. What changes is
A B C D E
the density along the line. You could therefore draw the wave by shading it as in Fig. 3,
but this is a bit harder to draw than Fig. 1. For a longitudinal wave, the statements in the
Figure 3 preceding paragraph about the motion of the various points in Fig. 1 are still true, provided
that “downward” is replaced with “leftward,” and “upward” is replaced with “rightward.
But what do things actually look like along the 1-D line? In particular, how does Fig. 3
follow from Fig. 1?
At points B and D in Fig. 3, the density of the masses equals the equilibrium density,
because nearby points all have essentially the same displacement (see Fig. 1). But at points
A and E, the density is a minimum, because points to the left of them have a negative
displacement, while points to the right have a positive displacement (again see Fig. 1). The
opposite is true for point C, so the density is maximum there. Various properties of the
wave are indicated in Fig. 4. You should stare at this figure for a while and verify all of the
stated properties. We’ll talk more about the relation among the various quantities when we
discuss Fig. 8 later on when we get to sound waves.

max +ψ, v = 0, max -a, avg ρ

ψ = 0, max +v, a = 0, max +ρ
max -ψ, v = 0, max +a, avg ρ
ψ ψ = 0, max -v, a = 0, max -ρ

rightward traveling

positive position
positive velocity
positive acceleration

Figure 4

In the Fig. 4, the relation between ψ, v, and a is the same as always, namely, a is 90◦
ahead of v, and v is 90◦ ahead of ψ. But you should think about how these relate to the
density ρ. For example, from the preceding paragraph, the (excess) ρ is proportional to the
negative of the slope (see Problem [to be added] for a rigorous derivation of this fact). But
we already know that v is proportional to the negative of the slope; see Eq. (4.44). Therefore,
the (excess) ρ is proportional to v. For a leftward traveling wave, the same statement about
5.2. SOUND WAVES 3

ρ is still true, but now v is proportional to the slope (with no negative sign). So the (excess)
ρ is proportional to −v.
We can double check that this result makes sense with the following reasoning. Since
the (excess) ρ is proportional to v, we can take the derivative of this statement to say that
∂ρ/∂x ∝ ∂v/∂x (the word “excess” is now not needed). But since a traveling wave takes
the form of ψ(x, t) = f (x − ct), the velocity v = ∂ψ/∂t also takes the functional form of
g(x − ct). Therefore, we have ∂v/∂x = −(1/c)∂v/∂t. The righthand side of this is just the
acceleration a, so the ∂ρ/∂x ∝ ∂v/∂x statement becomes

∂ρ ∂ρ
∝ −a =⇒ a∝− . (3)
∂x ∂x

Does this make sense? It says, for example, that if the density is an increasing function of
x at a given point, then the acceleration is negative there. This is indeed correct, because
a larger density means that the springs are more compressed (or less stretched), which in
turn means that they exert a larger repulsive force (or a smaller attractive force). So if the
density is an increasing function of x (that is, if ∂ρ/∂x > 0), then the springs to the right of a
given region are pushing leftward more than the springs to the left of the region are pushing
rightward. There is therefore a net negative force, which means that the acceleration a is
negative, in agreement with Eq. (3).

5.2 Sound waves

5.2.1 Notation
Sound is a longitudinal wave, in both position and pressure/density, as we’ll see. Sound can
exist in solids, liquids, and gasses, but in this chapter we’ll generally work with sound waves
in air. In air, molecules push and (effectively, relative to equilibrium) pull on each other, so
we have a sort of spring/mass system like the one we discussed above.
The main goal in this section is to derive the wave equation for sound waves in air. We’ll
find that we obtain exactly the same type of wave equation that we had in Eq. (1) for the
spring/mass system. The elastic modulus E appears there, so part of our task below will
be to find the analogous quantity for sound waves. We’ll consider only one-dimensional
waves here. That is, the waves depend only on x. Waves like this that are uniform in the
transverse y and z directions are called “plane waves.”
To emphasize the 1-D nature of the wave, let’s consider a tube of air inside a cylindrical
container, with cross-sectional area A. Fig. 5 shows a given section of air at equilibrium,
and then also at a later time. Let the ends of this section be located at x and x + ∆x at
equilibrium, and then at x + ψ(x) and x + ∆x + ψ(x + ∆x) at a later time, as shown. So
the function ψ measures the displacement from equilibrium.
If we define ∆ψ by ψ(x + ∆x) ≡ ψ(x) + ∆ψ, then ∆ψ is how much more the right
boundary of the region moves compared with the left boundary. The molecules of air are in
thermal motion, of course, so it’s not as if the molecules that form the boundary at position
x in the first picture correspond to the molecules that form the boundary at position x+ψ(x)
in the second picture. But we’ll ignore this fact and just pretend that it’s the same molecules,
for ease of discussion. It doesn’t actually matter. Note that ∆ψ ≈ (∂ψ/∂x)∆x for small
∆x, by definition of the derivative. In actual sound waves in air, ∆ψ is much less than ∆x.
In other words, ∂ψ/∂x is very small.
4 CHAPTER 5. LONGITUDINAL WAVES

(equilibrium)

x x+∆x

(later)
∆x+∆ψ

x+ψ(x) x+∆x+ψ(x+∆x)
x+∆x+ψ(x)+∆ψ

Figure 5
A note on terminology: We’re taking x to be the position of a given molecule at equilib-
rium. So even after the molecule has moved to the position x + ψ(x), it is still associated
with the same value of x. So x is analogous to the index n we used in Section 2.3 and the
beginning of Section 2.4. The movement of the particle didn’t affect its label n there, and
it doesn’t affect its label x here.
In obtaining the wave equation, we’ll need to get a handle on the pressure at the two ends
of the given section of air, and then we’ll figure out how these pressures cause the section
to move. Let the pressure in the tube at equilibrium be p0 . At sea level, the atmospheric
pressure happens to be about 14.7 lbs. per square inch. The first picture in Fig. 6 shows
the pressures at the two ends at equilibrium; it is simply p0 at both ends (and everywhere
else).

(equilibrium)

p0 p0

x x+∆x

(later)

p(x)=p0+ψp(x) p(x+∆x) = p0+ψp(x+∆x)

p0+ψp(x)+∆ψp

Figure 6
What about at a later time? Let ψp (x) be the excess pressure (above p0 ) as a function
of x. (Remember that x labels the equilibrium position of the molecules, not the present
position.) The total pressure at the left boundary of the section is then p0 +ψp (x). However,
this total pressure won’t be too important; the change, ψp (x), is what we’ll be concerned
with. At the right boundary, the total pressure is, by definition, p0 + ψp (x + ∆x). If we
define ∆ψp by ψp (x + ∆x) ≡ ψp (x) + ∆ψp , then ∆ψp is how much the pressure at the
right boundary exceeds the pressure at the left boundary. Note that ∆ψp ≈ (∂ψp /∂x)∆x
for small ∆x. In practice, ψp is much smaller than p0 . And ∆ψp is infinitesimally small,
assuming that we have picked ∆x to be infinitesimally small. The pressures at a later time
are summarized in the second picture in Fig. 6.
5.2. SOUND WAVES 5

5.2.2 The wave equation

Having introduced the necessary notation, we can now derive the wave equation for sound
waves. The derivation consists of four main steps, so let’s go through them systematically.
Our strategy will be to find the net force on a given volume of air, and then write down the
F = ma equation for that volume.

1. How the volume changes: First, we need to determine how the volume of a gas
changes when the pressure is changed. Qualitatively, if we increase the pressure on
a given volume, the volume decreases. By how much? The decrease should certainly
be proportional to the volume, because if we put two copies of a given volume next
to each other, we will obtain twice the decrease. And the decrease should also be
proportional to the pressure increase, provided that the increase is small. This is a
reasonable claim, but not terribly obvious. We’ll derive it in step 2 below. Assuming
that it is true, we can write (recalling that ψp is defined to be the increase in pressure
relative to equilibrium) the change in volume from equilibrium as ∆V ∝ −V ψp . This
correctly incorporates the two proportionality facts above. The minus sign is due to
the fact that an increase in pressure causes a decrease in volume. This equation is
valid as long as ∆V is small compared with V , as we’ll see below.
Let’s define κ to be the constant of proportionality in ∆V ∝ −V ψp . κ is known as the
compressibility. The larger κ is, the more the volume is compressed (or expanded), for
a given increase (or decrease) in pressure, ψp . In terms of κ, we have

∆V = −κV ψp . (4)

But from the first picture in Fig. 5, we see that the volume of gas is V = A∆x, where
A is the cross-sectional area. And the change in the volume between the two pictures
shown is ∆V = A∆ψ.1 Eq. (4) therefore becomes

∆V A∆ψ ∂ψ
= −κψp =⇒ = −κψp =⇒ = −κψp (5)
V A∆x ∂x

where we have taken the infinitesimal limit and changed the ∆’s to differentials (par-
tial ones, since ψ is a function of t also). The quantity ∂ψ/∂x indicates how the
displacement from equilibrium grows as a function of x. Equivalently, ∂ψ/∂x is the
“stretching fraction.” If the displacement ψ grows by, say, 1 mm over the course of
a distance of 10 cm, then the length (and hence the volume) of the region has in-
creased by 1/100, and this equals ∂ψ/∂x. Eq. (5) says that the stretching fraction is
proportional to the change in pressure, which is quite reasonable.

2. Calculating the compressibility, κ: We’ll now be rigorous about the above

statement that the decrease in volume should be proportional to the pressure increase,
provided that the increase is small. In the course of doing this, we’ll find the value of
κ.
Let’s first give a derivation that isn’t quite correct. From the ideal gas law (which we’ll
accept here; we have to start somewhere), we have pV = nRT . If the temperature
T is constant, then any changes in p and V must satisfy (p + dp)(V + dV ) = nRT .
Subtracting the original pV = nRT equation from this one, and ignoring the second-
order dp dV term (this is where the assumption of small changes comes in), we obtain
1 We can alternatively say that the volume is V = A(∆x + ∆ψ), by looking at the second picture. But

we are assuming ∆ψ ¿ ∆x, so to leading order we can ignore the A∆ψ term in the volume. However, we
can’t ignore it in the change in volume, because it’s the entire change.
6 CHAPTER 5. LONGITUDINAL WAVES

p dV + dp V = 0 =⇒ dV = −(1/p)V dp. But dp is what we’ve been calling ψp above,

so if we compare this result with Eq. (4), we see that the compressibility κ equals 1/p.
Note that what we did here was basically take the differential of the pV = C equation
(where C is a constant) to obtain p dV + dp V = 0.
However, this κ = 1/p result isn’t correct, because it is based on the assumption that T
is constant in the relation pV = nRT . But T isn’t constant in a sound wave (in either
space or time). The compressions are actually adiabatic, meaning that heat can’t flow
quickly enough to redistribute itself and even things out. Basically, the time scale
of the heat flow is large compared with the time scale of the wave oscillations. So a
region that heats up due to high pressure stays hot, until the pressure decreases. The
correct relation (which we’ll just accept here) for adiabatic processes turns out to be
pV γ = C, where γ happens to be about 7/5 for air (γ is 7/5 for a diatomic gas, and
air is 99% N2 and O2 ). Taking the differential of our new pV γ = C equation gives
µ ¶
γ−1 γ 1
p · γV dV + dp V = 0 =⇒ dV = − V dp. (6)
γp
Therefore, the correct value of the compressibility κ is
1 5
κ= ≈ . (7)
γp0 7p0
The above incorrect result wasn’t so bad; it was off by only a factor of 7/5.
3. Calculating the difference in pressure, ∆ψp : Eq. (5) involves the quantity
ψp . However, what we’re actually concerned with is not ψp but the change in ψp from
one end of the volume to the other, because in writing down the F = ma equation for
the volume, we’re concerned with the net force on it, and this involves the difference
in the pressures at the ends. So let’s see how ψp changes with x. Differentiating Eq.
(5) with respect to x gives
∂ψp 1 ∂2ψ 1 ∂2ψ
=− =⇒ ∆ψp = − ∆x, (8)
∂x κ ∂x2 κ ∂x2
where we have multiplied both sides by ∂x and switched back to the ∆ notation.
Or equivalently, we have multiplied both sides by ∆x, and then used the relation
∆ψp = (∂ψp /∂x)∆x. Note that the second derivative ∂ 2 ψ/∂x2 appears here. From
Eq. (5), a constant value of ∂ψ/∂x corresponds to a constant excess pressure ψp (x);
the tube of air just stretches uniformly if the pressure is the same everywhere. So to
obtain a varying value of ψp , we need a varying value of ∂ψ/∂x. That is, we need a
nonzero value of ∂ 2 ψ/∂x2 .
4. The F = ma equation: We can now write down the F = ma equation for a given
volume of air. The net rightward force on the tube of air in the second picture in Fig.
6 is the cross-sectional area times the difference in the pressure at the ends. So we
have, recalling the definition of ∆ψp ,
¡ ¢
Fnet = A p(x) − p(x + ∆x) = A(−∆ψp ). (9)
Using the expression for ∆ψp we found in Eq. (8), the Fnet = ma equation on a given
tube of air is (with ρ being the mass density)
∂2ψ
−A∆ψp = (ρV ) 2
µ ¶ ∂t
1 ∂2ψ ∂2ψ
=⇒ −A − 2
∆x = ρ(A∆x) 2 . (10)
κ ∂x ∂t
5.2. SOUND WAVES 7

Canceling the common factor of A∆x and using κ = 1/γp0 yields

∂2ψ γp0 ∂ 2 ψ
= · (wave equation) (11)
∂t2 ρ ∂x2

This is the desired wave equation for sound waves in air. We see that γp0 /ρ is the
coefficient that replaces the E/µ coefficient for the longitudinal spring/mass system,
or the T /µ coefficient for the transverse string.

The solutions to the wave equation are the usual exponentials,

ψ(x, t) = Aei(±kx±ωt) , (12)

p
p ω and k satisfy ω/k = γp0 /ρ. And since ω/k is the speed c of the wave, we have
where
c = γp0 /ρ. As always, the speed is simply the square root of the factor on the righthand
side. As a double check on the units, we have

γp0 (unitless) · (force/area) (kg m/s2 )/m2 m2

= = 3
= 2 , (13)
ρ mass/volume kg/m s
which has the correct units of velocity squared. What is the numerical value of the speed
of sound waves? For air at sea level, the pressure is p0 ≈ 105 kg/ms2 and the density is
ρ ≈ 1.3 kg/m3 . So we have
r s
γp0 (7/5)(105 kg/ms2 )
c= ≈ ≈ 330 m/s. (14)
ρ 1.3 kg/m3

This decreases with ρ, which makes sense because the larger the density, the more inertia
the air has, so the harder it is to accelerate it. The speed increases with p0 . This follows
from the fact that if p0 is large, then the compressibility κ is small (meaning the gas is not
easily compressed). So for a given value of ∂ 2 ψ/∂x2 , the force on the left side of Eq. (10) is
large, which implies large accelerations.
The two partial derivatives in Eq. (11) come about in the usual way. The second time
derivative comes from the “a” in F = ma, and the second space derivative comes from the
fact that it is the difference in the first derivatives that gives the net force. Eq. (5) tells us
that the first space derivative of the displacement gives a measure of the force at a given
location (just as with the spring/mass system, the first space derivative told us how much
the springs were stretched, which in turn gave the force). The difference in the force at the
two ends is therefore proportional to the second space derivative (again as it was with the
spring/mass system).
As with the other wave equations we have encountered thus far in this book, the speed
of sound waves is independent of ω and k. (This won’t be the case for the dispersion-ful
waves we discuss in the following chapter.) So all frequencies travel at the same speed. This
is fortunate, because if it weren’t true, then a music concert would sound like a complete
mess!

5.2.3 Pressure waves

Eq. (11) gives the wave equation for the displacement, ψ, from equilibrium for a molecule
whose equilibrium position is x. However, it is rather difficult to follow the motion of a single
molecule, so it would be nice to obtain the wave equation for the excess pressure ψp , because
the pressure is much easier to measure. (It is a macroscopic property of the average of a
 8 CHAPTER 5. LONGITUDINAL WAVES

large number of molecules, as opposed to the microscopic position of a particular molecule.)

In view of the relation between ψp and ψ in Eq. (5), we can generate the wave equation
for ψp by taking the ∂/∂x derivative of the wave equation in Eq. (11). Using the fact that
partial derivatives commute, we obtain
µ ¶ µ ¶ µ ¶ µ ¶
∂ ∂2ψ γp0 ∂ ∂ 2 ψ ∂ 2 ∂ψ γp0 ∂ 2 ∂ψ
= · =⇒ = · . (15)
∂x ∂t2 ρ ∂x ∂x2 ∂t2 ∂x ρ ∂x2 ∂x

But from Eq. (5) we know that ∂ψ/∂x ∝ ψp , so we obtain

∂ 2 ψp γp0 ∂ 2 ψp
= · (wave equation for pressure) (16)
ψ,ψp ψ
ψp ∂t2 ρ ∂x2

This is the same wave equation as the one for the displacement ψ in Eq. (11). So everything
x that is true about ψ is also true about ψp . The only difference is that since ψp ∝ −∂ψ/∂x
(the minus sign is important here), the phase of ψp is 90◦ behind the phase of ψ. This is
shown in Fig. 7. The pressure (and hence also the density) reaches its maximum value a
quarter cycle after the displacement does. This is consistent with the values of ψ and ρ
Figure 7
given in Fig. 4.

5.2.4 Impedance
What is the impedance of air? In other words, what is the force per velocity that a given
region applies to an adjacent region, as a wave propagates? Remember that impedance is a
property of the medium and not the wave, even though it is generally easiest to calculate it
by considering the properties of a traveling wave. (However, when we discuss dispersion-ful
systems in the next chapter, we will find that the impedance depends on the frequency of
the wave.)
The velocity of a “sheet” of molecules whose equilibrium position is x is simply v(x) =
∂ψ(x)/∂t. To find the force, consider a cross-sectional area A. We can use Eq. (5) to write
the (excess) force that the sheet exerts on the region to its right as
µ ¶
1 ∂ψ
F = Aψp = A − . (17)
κ ∂x

And since we are working with a traveling wave (no need for it to be sinusoidal), we have
the usual relationship between ∂ψ/∂x and ∂ψ/∂t, namely ∂ψ/∂x = ∓(1/c)(∂ψ/∂t) (the
minus sign is associated with a rightward traveling wave). So Eq. (17) becomes
µ ¶µ ¶
1 1 ∂ψ A ∂ψ
F =A − ∓ =± · . (18)
κ c ∂t κc ∂t

The force that the sheet feels from the region on its right is the negative of this, but the sign
isn’t important when calculating the impedance Z, because Z is defined to be the magnitude
of F/v. Using v = ∂ψ/∂t, Eq. (18) gives the impedance as

F A
Z≡ = . (19)
v κc
The impedance per unit area is the more natural thing to talk about, because Z/A is
independent of the specific cross section chosen. The force F can be made arbitrarily large
by making the area A arbitrarily large, so Z = F/v isn’t too meaningful. When people
talk about the impedance of air, they usually mean “impedance per area,” that is, force
5.2. SOUND WAVES 9

per velocity per area. However, we’ll stick with the Z = F/v definition of impedance,
in which case Eq. (19) tells us that the impedance per unit area is Z/A = 1/κc. But
√
c = 1/ κρ =⇒ κ = 1/ρc2 (this follows from writing the coefficient in Eq. (11) in terms of
p
κ). So we have Z/A = ρc. Using c = γp0 /ρ from Eq. (14), we can write this alternatively
as
Z √
= ρc = γρp0 (20)
A

5.2.5 Energy, Power

Energy
What is the energy density of a (longitudinal) sound wave? The kinetic energy density (per
unit volume) is simply (1/2)ρ(∂ψ/∂t)2 , because the speed of the molecules is ∂ψ/∂t. If
we want to consider instead the kinetic energy density per unit length along a tube with
cross-sectional area A, then this is (1/2)(Aρ)(∂ψ/∂t)2 , where Aρ ≡ µ is the mass density
per unit length. We are assuming here that ρ and µ are independent of position. This is
essentially true for actual sound waves, because ∂ψ/∂x is small, so the fractional change in
the density from its equilibrium value is small.
What about the potential energy density? The task of Problem [to be added] is to
show that the potential energy density (or rather, the excess over the equilibrium value) per
unit length is (1/2)Aγp0 (∂ψ/∂x)2 . The total energy density per unit length (kinetic plus
potential) is therefore
µ¶2 µ ¶2
1 ∂ψ 1 ∂ψ
E(x, t) = Aρ + Aγp0
2 ∂t 2 ∂x
"µ ¶2 µ ¶2 #
1 ∂ψ γp0 ∂ψ
= Aρ +
2 ∂t ρ ∂x
"µ ¶2 µ ¶2 #
1 ∂ψ 2 ∂ψ
= Aρ +c . (21)
2 ∂t ∂x

This is the same as the result we found in Eq. (4.49) for transverse waves, since Aρ is the
mass per unit length, µ. As with Eq. (4.49), the present expression for E(x, t) is valid for
an arbitrary wave. But if we consider the special case of a single traveling wave, then we
have the usual relation, ∂ψ/∂t = ±c ∂ψ/∂x. So the two terms in the expression for E(x, t)
are equal at a given point and at a given time. We can therefore write the energy density
per unit length as
µ ¶2
∂ψ
E(x, t) = Aρ (for traveling waves) (22)
∂t

The energy density per unit volume is then E/A = ρ(∂ψ/∂t)2 .

Power
Consider a cross-sectional “sheet” of molecules. At what rate does the air on the left of the
sheet do work on the sheet? (This is the same type of question that we asked in Section
4.4 for a transverse wave: At what rate does the string to the left of a dot do work on the
dot?) In a small amount of the time, the work done by the air is dW = F dψ = (pA) dψ.
10 CHAPTER 5. LONGITUDINAL WAVES

The power on the sheet whose equilibrium position is x is therefore

∂W ∂ψ
P = = (p0 + ψp )A . (23)
∂t ∂t
The Ap0 (∂ψ/∂t) part of this averages out to zero over time, so we’ll ignore it. The
Aψp (∂ψ/∂t) term, however, always has the same sign, for the following reason. From Eq.
(5), we have ψp = −(1/κ)(∂ψ/∂x). But as usual, ∂ψ/∂x = ∓(1/c)(∂ψ/∂t) (the minus sign
is associated with a rightward traveling wave). So the power is
µ ¶ µ ¶2
∂ψ 1 ∂ψ ∂ψ A ∂ψ
P = Aψp =A ± · =± . (24)
∂t κc ∂t ∂t κc ∂t
√
Using c = 1/ κρ =⇒ κ = 1/ρc2 , we have

µ ¶2
∂ψ
P = ±Aρc (25)
∂t

Since Z = Aρc from Eq. (20), we can also write the power as
µ ¶2
∂ψ
P = ±Z , (26)
∂t

which takes exactly the same form as the result in Eq. (4.53) for transverse waves.
If we compare Eqs. (22) and (25), we see that P = ±cE. This makes sense, because
as with transverse waves, the power must equal the product of the wave velocity and the
energy density, because the E curve moves right along with the wave.
If we want to write P in terms of the pressure ψp , we can do this in the following way.
Using Eq. (5), Eq. (25) becomes
µ ¶2 µ ¶2
∂ψ 3 2 3 1 A
P = ±Aρc ∓c = ±Aρc (−κψp ) = ±Aρc ψp2 = ± ψp2 (27)
∂x ρc2 ρc

We see that when written in terms of ψp , the power decreases with ρ and c. But when
written in terms of ψ (or rather ∂ψ/∂t) in Eq. (25), it grows with ρ and c. The latter
is fairly clear. For example, a larger ρ means that more matter is moving, so the energy
density is larger. But the dependence on ψp isn’t as obvious. It arises from the fact that
there are factors of ρ and c hidden in ψp . So, for example, if ρ is increased (for a given
function ψ), then ψp2 grows faster than ρ, so the righthand side of Eq. (27) still increases
with ρ. However, if ρ is increased for a given function ψp , then the power decreases, because
the displacement ψ has to decrease to keep ψp the same, and this effect wins out over the
increase in ρ, thereby decreasing P .

5.2.6 Qualitative description

Let’s take a look at what a given molecule in the air is doing at a few different times, as a
rightward-traveling wave passes by. A number of snapshots with phase differences of π/4
are shown in Fig. 8. The darker regions indicate a higher pressure (and density), and the
lighter regions indicate a lower pressure (and density). The vertical line, which represents
the equilibrium position of the molecule, is drawn for clarity.
The three plots at the top of the figure give the values of the various parameters at t = 0,
as functions of x. So these plots correspond to the first of the shaded snapshots. The plots
5.2. SOUND WAVES 11

for the eight other snapshots are obtained by simply shifting the three plots to the right, by
the same amount as the snapshots shift. The three plots at the left of the figure give the
values of the various parameters at x = 0, as functions of t. So these plots correspond to
the molecule in question (the little circle). The leftmost plot (the one for ψ(0, t)) is simply
a copy of the circles as they appear in the snapshots.2
Let’s discuss what’s happening in each of the snapshots. In doing this, a helpful thing to
remember is that for a rightward-traveling wave, the density (and hence pressure) is always
in phase with the velocity (as we discussed in Section 5.1). And both the pressure and the
velocity are 90◦ ahead of the position, as functions of time. (For a leftward-traveling wave,
the pressure is out of phase with the velocity, which is, as always, 90◦ ahead of the position.)
Again, the displacements are exaggerated in the figures. In reality, the displacements are
much smaller than the wavelength. The commentary on the snapshots is as follows.

plots for t=0, x (x=0)

as functions of x
ψ(x,0)
plots for x=0,
as functions of t v(x,0), ψp(x,0)

ψp(0,t) a(x,0)
ψ(0,t) v(0,t) a(0,t)

1. (t=0)

2.
0
3.

4.

t 5.

6.
0
7.

8.

9.

(Exaggerated displacements
of the molecule)

Figure 8
2 The ψ values are exaggerated for emphasis. In reality, they are much smaller than the wavelength of

the wave. But if we drew them to scale, all of the circles would essentially lie on the x = 0 line, and we
wouldn’t be able to tell that they were actually moving.
12 CHAPTER 5. LONGITUDINAL WAVES

1. In the first snapshot, the molecule is located at its equilibrium position and is moving
to the right with maximum speed.3 The pressure (and density) is also maximum. The
pressure is the same on both sides of the molecule, so there is zero net force, consistent
with the fact that it has maximum speed and hence zero acceleration.
2. The molecule is still moving to the right, but it is decelerating (a < 0) because there
is higher pressure (which goes hand-in-hand with higher density) on its right than on
its left.
3. It has now reached its maximum value of ψ and is instantaneously at rest. It has the
maximum negative acceleration, because the pressure gradient is largest here; the pres-
sure is changing most rapidly (as a function of x) halfway between the maximum and
minimum pressures. The difference between the forces on either side of the molecule
is therefore largest here, so the molecule experiences the largest acceleration.
4. It has now started moving leftward and is picking up speed due to the higher pressure
on the right.
5. It passes through equilibrium again, but now with the maximum negative velocity.
This ends the period of negative acceleration. Up to this time, there was always higher
pressure on the molecule’s right side. For the next half cycle, the higher pressure will
be on the left side, so there will be positive acceleration; see the “a(0, t)” plot in the
left part of the figure.
6. It is moving to the left but is slowing down due to the higher pressure on the left.
7. It has now reached its maximum negative value of ψ and is instantaneously at rest.
As in the third snapshot, the pressure gradient is largest here.
8. It has started moving rightward and is picking up speed due to the higher pressure on
the left.
9. We are back to the beginning of the cycle. The molecule is in the equilibrium position
and is moving to the right with maximum speed.

5.3 Musical instruments

Musical instruments (at least the wind ones) behave roughly like pipes of various sorts, so
let’s start our discussion of instruments by considering the simple case of a standing wave
in a pipe.
Consider first the case where the pipe is closed at one end, taken to be at x = 0. The
air molecules at the closed end can’t move into the “wall” at the end. And they can’t
move away from it either, because there would then be a vacuum at the wall which would
immediately suck the molecules back to the wall. This boundary condition at x = 0 tells us
that the wall must be a node of the ψ(x, t) wave. The standing wave for the displacement
must therefore be of the form,
ψ(x, t) = A sin kx cos(ωt + φ). (28)
What does the pressure wave look like? Since ψp = −(1/κ)(∂ψ/∂x) from Eq. (5), we have
Ak
ψp (x, t) = − cos kx cos(ωt + φ). (29)
κ
3 Or, it is moving to the left if we have a leftward-traveling wave. After reading through this commentary,

you should verify that everything works out for a leftward-traveling wave.
5.3. MUSICAL INSTRUMENTS 13

Looking at the x dependence of this function, we see that nodes of ψ correspond to antinodes
of ψp , and vice-versa.
If we instead have an open end at x = 0, then the boundary condition isn’t as obvious.
You might claim that now the air molecules move a maximum amount at the open end,
which means that instead of a node in ψ, we have an antinode. This is indeed correct, but
it isn’t terribly obvious. So let’s consider the pressure wave instead. If we have a standing
wave inside the pipe, then there is essentially no wave outside the pipe. (Well, there must
of course be some wave outside, given that there are sound waves hitting your ear.) So the
pressure outside must be (essentially) the atmospheric pressure p0 . In other words, ψp = 0
outside the pipe. And since the pressure must be continuous, the boundary condition at the
open end at x = 0 is ψp = 0. So the pressure has a node there. The pressure can therefore
be written as
ψp (x, t) = B sin kx cos(ωt + φ). (30)

The ψ(x, t) function that satisfies ψp = −(1/κ)(∂ψ/∂x) is then

Bκ
ψ(x, t) = cos kx cos(ωt + φ). (31)
k

(A nonzero constant of integration would just give a redefinition of the equilibrium position.)
So ψ does indeed have an antinode at x = 0, as we suspected.
The above results hold for any open or closed end, independent of where it is located.
It doesn’t have to be located at the arbitrarily-chosen position of x = 0, of course. So
Fig. 9 shows the lowest-frequency (longest-wavelength) modes for the three possible cases
of combinations of end types: closed/open, closed/closed, and open/open. In practice, the
pressure node is slightly outside the open end, because the air just outside the pipe vibrates
a little bit.

(Position ψ) (Pressure ψp)

ψ=0 ψp=0

Figure 9

An instrument like a flute is essentially open at both ends (with one end being the
mouthpiece). But most other instruments (reeds, brass, etc.) are open at one end and
essentially closed at the mouthpiece end. This is due to the fact that the vibrating reed
(or the vibrating lips in the mouthpiece) doesn’t move much (so ψ ≈ 0), but it is what is
driving the pressure wave (so ψp is maximum there). A clarinet therefore corresponds to the
first case (closed/open) in Fig. 9, while a flute corresponds to the second case (open/open).
In view of this, you can see why a clarinet can play about an octave lower (which means
half the frequency) than a flute, even though they have about the same length. The longest
 (Pressure ψp)
14 CHAPTER 5. LONGITUDINAL WAVES
L
wavelength for a clarinet (which is four times the length of the pipe) is twice a long as the
longest wavelength for a flute (which is two times the length of the pipe).4
λ=4L The lowest four standing waves for a clarinet are shown in Fig. 10 (described by the
pressure waves, which is customary). The wavelengths are 4L, 4L/3, 4L/5, 4L/7, etc. The
frequencies are inversely proportional to the wavelengths, because νλ = c =⇒ ν ∝ 1/λ. So
the frequencies are in the ratio of 1 : 3 : 5 : 7 : · · ·. These notes are very far apart. For an
λ=(4/3)L
open/open pipe like a flute, the wavelengths are 2L, 2L/2, 2L/3, 2L/4, etc., which means
that the frequencies are in the ratio of 1 : 2 : 3 : 4 : · · ·. These notes are also very far apart.
So if a clarinet or a flute didn’t have any keys, you wouldn’t be able to play anywhere near
all of the notes in a standard scale.
λ=(4/5)L
Keys remedy this problem in the following way. If all the keys are closed, then we simply
have a pipe. But if a given key is open, then this forces the pressure wave to have a node
at that point, because the pressure must match up with the atmospheric pressure there. So
we have essentially shortened the pipe by creating an effectively open end at the location
λ=(4/7)L
of the open key. With many keys, this allows for many different effective pipe lengths, and
Figure 10 hence many different notes. And also many different ways to play a given note. If we have
a particular standing wave in the instrument, and if we then open a key at the location of
a (pressure) node, then this doesn’t change anything, so we get the same note.
What about a trumpet, which has only three valves? It’s a bit complicated, but the
conical shape (at least near the end) has the effect of making the frequencies be closer
together (and also higher). And the mouthpiece helps too. The end result (if done properly)
is that the frequencies are in the ratio 2 : 3 : 4 : 5 : 6 : · · · (for some reason, the 1 is missing)
instead of the 1 : 3 : 5 : 7 : · · · ratios for the closed/open case in Fig. 10. This is indeed the
ratio of the frequencies of the notes (C,G,C,E,G,. . . ) that can be be played on a trumpet
without pressing down any valves. The valves then change the length of the pipe in a
straightforward manner.
The flared bell of a trumpet has the effect (compared with a cylinder of the same length)
of raising the low notes more than the high notes, because the long wavelengths (low notes)
can’t follow the bell as easily, so they’re reflected sooner than the short wavelengths.5 The
longer wavelengths therefore effectively see a shorter pipe. However, another effect of the bell
is that because the short wavelengths follow it so easily (right out to the outside atmospheric
pressure), there isn’t much reflection for these waves, so it’s harder to get a standing wave.
The high notes are therefore less well defined, and thus blend together (which is quite evident
if you’ve ever heard a trumpet player screeching away in the high register).
Everything you ever wanted to know about the physics of musical instruments can be
found on this website: http://www.phys.unsw.edu.au/music

4 The third option in Fig. 9, the closed/closed pipe, isn’t too conducive to making music. Such an

instrument couldn’t have any keys, because they provide openings to the outside world. And furthermore,
you couldn’t blow into one like in reed or brass instrument, because there would be no place for the air to
come out. And you can’t blow across an opening like in a flute, because that’s an open end.
5 The fact that the shorter wavelengths (high notes) can follow the bell is the same effect as in the

“Gradually changing string density” impedance-matching example we discussed in Section 4.3.2.

Chapter 6

Dispersion
David Morin, [email protected]

The waves we’ve looked at so far in this book have been “dispersionless” waves, that is, waves
whose speed is independent of ω and k. In all of the systems we’ve studied (longitudinal
spring/mass, transverse string, longitudinal sound), we ended up with a wave equation of
the form,
∂2ψ 2∂ ψ
2
= c , (1)
∂t2 ∂x2
where c depends on various parameters in the setup. The solutions to this equation can be
built up from exponential functions, ψ(x, t) = Aei(kx−ωt) . Plugging this function into Eq.
(1) gives
ω 2 = c2 k 2 . (2)
This is the so-called dispersion relation for the above wave equation. But as we’ll see, it
is somewhat of a trivial dispersion relation, in the sense that there is no dispersion. We’ll
explain what we mean by this below.
The velocity of the wave is ω/k = ±c, which is independent of ω and k. More precisely,
this is the phase velocity of the wave, to distinguish it from the group velocity which we’ll
define below. The qualifier “phase” is used here, because the speed of a sinusoidal wave
sin(kx − ωt) is found by seeing how fast a point with constant phase, kx − ωt, moves. So
the phase velocity is given by
d(kx − ωt) dx dx ω
kx − ωt = Constant =⇒ = 0 =⇒ k − ω = 0 =⇒ = , (3)
dt dt dt k
as desired.
As we’ve noted many times, a more general solution to the wave equation in Eq. (1) is
any function of the form f (x − ct); see Eq. (2.97). So the phase velocity could reasonably
be called the “argument velocity,” because c is the speed with which a point with constant
argument, x − ct, of the function f moves.
However, not all systems have the property that the phase velocity ω/k is constant (that
is, independent of ω and k). It’s just that we’ve been lucky so far. We’ll now look at a
so-called dispersive system, in which the phase velocity isn’t constant. We’ll see that things
get more complicated for a number of reasons. In particular, a new feature that arises is
the group velocity.
The outline of this chapter is as follows. In Section 6.1 we discuss a classic example of a
dispersive system: transverse waves in a setup consisting of a massless string with discrete

1
 2 CHAPTER 6. DISPERSION

point masses attached to it. We will find that ω/k is not constant. That is, the speed of
a wave depends on its ω (or k) value. In Section 6.2 we discuss evanescent waves. Certain
dispersive systems support sinusoidal waves only if the frequency is above or below a certain
cutoff value. We will determine what happens when these bounds are crossed. In Section
6.3 we discuss the group velocity, which is the speed with which a wave packet (basically,
a bump in the wave) moves. We will find that this speed is not the same as the phase
velocity. The fact that these two velocities are different is a consequence of the fact that in
a dispersive system, waves with different frequencies move with different speeds. The two
velocities are the same in a non-dispersive system, which is why there was never any need
m m m m m to introduce the group velocity in earlier chapters.

l l l l
6.1 Beads on a string
Figure 1
Consider a system that is made up of beads on a massless string. The beads have mass
m and are glued to the string with separation `, as shown in Fig. 1. The tension is T .
We’ll assume for now that the system extends infinitely in both directions. The goal of this
n+1 section is to determine what transverse waves on this string look like. We’ll find that they
T behave fundamentally different from the waves on the continuous string that we discussed
n in Chapter 4. However, we’ll see that in a certain limit they behave the same.
θ2
n-1 T
θ1 l ψ -ψ We’ll derive the wave equation for the beaded string by writing down the transverse
l
n+1 n F = ma equation on a given bead. Consider three adjacent beads, label by n − 1, n, and
ψn - ψn-1 n + 1, as shown in Fig. 2. For small transverse displacements, ψ, we can assume (as we did
in Chapter 4) that the beads move essentially perpendicular to the equilibrium line of the
Figure 2 string. And as in Chapter 4, the tension is essentially constant, for small displacements.
So the transverse F = ma equation on the middle mass in Fig. 2 is (using sin θ ≈ tan θ for
small angles)

mψ¨n = −T sin θ1 + T sin θ2

µ ¶ µ ¶
ψn − ψn−1 ψn+1 − ψn
= −T +T
` `
T
=⇒ ψ¨n = ω02 (ψn+1 − 2ψn + ψn−1 ), where ω02 = . (4)
m`
This has exactly the same form as the F = ma equation for the longitudinal spring/mass
system we discussed in Section 2.3; see Eq. (2.41). The only difference is that ω02 now equals
T /m` instead of k/m. We can therefore carry over all of the results from Section 2.3. You
should therefore reread that section before continuing onward here.
We will, however, make one change in notation. The results in Section 2.3 were written
in terms of n, which labeled the different beads. But for various reasons, we’ll now find it
more convenient to work in terms of the position, x, along the string (as we did in Section
2.4). The solutions in Eq. (2.55) are linear combinations of functions of the form,

ψn (t) = trig(nθ) trig(ωt), (5)

where each “trig” means either sine or cosine, and where we are now using ψ to label the
displacement (which is now transverse). θ can take on a continuous set of values, because
we’re assuming for now that the string extends infinitely in both directions, so there’s aren’t
any boundary conditions that restrict θ. ω can also take on a continuous set of values, but
it must be related to θ by Eq. (2.56):
µ ¶ µ ¶
2ω02 − ω 2 2 2 1 − cos θ θ
2 cos θ ≡ =⇒ ω = 4ω0 =⇒ ω = 2ω 0 sin . (6)
ω02 2 2
6.1. BEADS ON A STRING 3

Let’s now switch from the nθ notation in Eq. (5) to the more common kx notation. But
remember that we only care about x when it is a multiple of `, because these are the locations
of the beads. We define k by

kx ≡ nθ =⇒ k(n`) = nθ =⇒ k` = θ. (7)

We have chosen the x = 0 point on the string to correspond to the n = 0 bead. The ψn (t)
in Eq. (5) now becomes
ψ(x, t) = trig(kx) trig(ωt). (8)

In the old notation, θ gave a measure of how fast the wave oscillated as a function of n, In
the new notation, k gives a measure of how fast the wave oscillates as a function of x. k
and θ differ simply by a factor of the bead spacing, `. Plugging θ = k` into Eq. (6) gives
the relation between ω and k:
µ ¶
k`
ω(k) = 2ω0 sin (dispersion relation) (9)
2
p
where ω0 = T /m`. This is known as the dispersion relation for our beaded-string system.
It tells us how ω and k are related. It looks quite different from the ω(k) = ck dispersion
relation for a continuous string (technically ω(k) = ±ck, but we generally don’t bother with
the sign). However, we’ll see below that they agree in a certain limit.
What is the velocity of a wave with wavenumber k? (Just the phase velocity for now.
We’ll introduce the group velocity in Section 6.3.) The velocity is still ω/k (the reasoning
in Eq. (3) is still valid), so we have

ω 2ω0 sin(k`/2)
c(k) = = . (10)
k k
The main point to note here is that this velocity depends on k, unlike in the dispersionless
systems in earlier chapters. In the present system, ω isn’t proportional to k.
We can perform a double check on the velocity c(k). In the limit of very small ` (tech-
nically, in the limit of very small
p k`), we essentially have a continuous string. So Eq. (10)
had better reduce to the c = T /µ result we found in Eq. (4.5) for transverse waves on a
continuous string. Or said in another way, the velocity in Eq. (10) had better not depend
on k in this limit. And indeed, using sin ² ≈ ², we have
r s s
2ω0 sin(k`/2) 2ω0 (k`/2) T T T
c(k) = ≈ = ω0 ` = `= ≡ , (11)
k k m` m/` µ

where µ p is the mass density per unit length. So it does reduce properly to the constant
value of T /µ. Note that the condition k` ¿ 1 can be written as (2π/λ)` ¿ 1 =⇒ ` ¿ λ.
In other words, if the spacing between the beads is much shorter than the wavelength of the
wave in question, then the string acts like a continuous string. This makes sense. And it
makes sense that the condition should involve these two lengths, because they are the only
two length scales in the system.
If the ` ¿ λ condition doesn’t hold, then the value of ω/k in Eq. (10) isn’t independent
of k, so the beaded string apparently doesn’t behave like a continuous string. What does it
behave like? Well, the exact expression for ω in terms of k given in Eq. (9) yields the plot
shown in Fig. 3.
4 CHAPTER 6. DISPERSION

ω
slope = ω/k = phase velocity
2ω0

k
π/l 2π/l 3π/l 4π/l

Figure 3

There are various things to note about this figure:

1. Given a value of k and its associated value of ω, the phase velocity of the wave is ω/k.
But ω/k is the slope from the origin to the point (k, ω) in the figure, as shown. So the
phase velocity has this very simple graphical
p interpretation. As we saw in Eq. (11),
the slope starts off with a value of ω0 ` = T /µ near the origin, but then it decreases.
It then repeatedly increases and decreases as the point (k, ω) runs over the successive
bumps in the figure. As k → ∞, the slope ω/k goes to zero (we’ll talk more about
this in the third comment below).
p p
2. Note that ω0 ≡ T /m` can be written as ω0 = (1/`) T /µ, where µ = m/` is the
mass density. So if we hold T and µ constant and decrease ` (for example, if we keep
subdividing the masses, thereby making the string more and more continuous), then
ω0 grows like 1/`. So the maximum height of the bumps in Fig. 3, which is 2ω0 ,
behaves like 1/`. But the width of the bumps also behaves like 1/`. So if we decrease
` while keeping T and µ constant, the whole figure simply expands uniformly. The
linear approximation in Eq. (11) near the origin is therefore relevant for a larger range
of k values. This means that the string behaves like a continuous string for more k
values, which makes sense.
3. From Fig. 3, we see that many different values of k give the same ω value. In particular,
k1 , 2π/` − k1 , 2π/` + k1 , etc., all give the same ω. However, it turns out that only the
first half-bump of the curve (between k = 0 and k = π/`) is necessary for describing
any possible motion of the beads. The rest of the curve gives repetitions of the first
half-bump. The reason for this is that we care only about the movement of the beads,
which are located at positions of the form x = n`. We don’t care about the rest of
the string. Consider, for example, the case where the wavenumber is k2 = 2π/` − k1 .
A rightward-traveling wave with this wavenumber takes the form,
µ³ ´ ¶
2π
A cos(k2 x − ωt) = A cos − k1 (n`) − ωt
`
³ ´
= A cos 2nπ − k1 (n`) − ωt
= A cos(−k1 x − ωt). (12)

We therefore conclude that a rightward-moving wave with wavenumber 2π/` − k1 and

frequency ω gives exactly the same positions of the beads as a leftward-moving wave
with wavenumber k1 and frequency ω. (A similar statement holds with “right” and
“left” reversed.) If we had instead picked the wavenumber to be 2π/` + k1 , then
a quick sign change in Eq. (12) shows that this would yield the same positions as a
rightward-moving wave with wavenumber k1 and frequency ω. The rightward/leftward
correspondence alternates as we run through the class of equivalent k’s.
 (l=1)
ψ
6.2. EVANESCENT WAVES 5

x
It is worth emphasizing that although the waves have the same values at the positions 1 2 3 4 5 6
of the beads, the waves look quite different at other locations on the string. Fig.4 shows
the case where k1 = π/2`, and so k2 = 2π/`−k1 = 3π/2`. The two waves have common
values at positions of the form x = n` (we have arbitrarily chosen ` = 1). The k values (start)
are in the ratio of 1 to 3, so the speeds ω/k are in the ratio of 3 to 1 (because the
v 3v
ω values are the same). The k2 wave moves slower. From the previous paragraph, if
the k2 wave has speed v to the right, then the k1 wave has speed 3v to the left. If we
look at slightly later times when the waves have moved distances 3d to the left and d
to the right, we see that they still have have common values at positions of the form 1 2 3 4 5 6
x = n`. This is what Eq. (12) says in equations. The redundancy of the k values is
simply the Nyquist effect we discussed at the end of Section 2.3, so you should reread
that subsection now if you haven’t already done so. (time t)
In comment “1” above, we mentioned that as k → ∞, the phase velocity ω/k goes
to zero. It is easy to see this graphically. Fig. 5 shows waves with wavenumbers k1=π/2l k2=3π/2l
k1 = π/2`, and k2 = 6π/` − k1 = 11π/2`. The wave speed of the latter is small; velocity= -3v velocity= v
it is only 1/11 times the speed of the former. This makes sense, because the latter
wave (the very wiggly one) has to move only a small distance horizontally in order for
the dots (which always have integral values of x here) to move appreciable distances 1 2 3 4 5 6
vertically. A small movement in the wiggly wave will cause a dot to undergo, say, a
full oscillation cycle. At the location of any of the dots, the slope of the k2 wave is
always −11 times the slope of the k1 wave. So the k2 wave has to move only 1/11 as (time 2t)
far as the k1 wave, to give the same change in height of a given dot. This slope ratio
of −11 (at x values of the form n`) is evident from taking the ∂/∂x derivative of Eq. Figure 4
(12); the derivatives (the slopes) are in the ratio of the k’s.
(l=1)
4. How would the waves in Fig. 4 behave if we were instead dealing with the dispersionless ψ
system of transverse waves on a continuous string, which we discussed in Chapter 4?
(The length ` now doesn’t exist, so we’ll just consider two waves with wavenumbers
k and 3k, for some value of k.) In the dispersionless case, all waves move with the x
1 2 3 4 5 6
same speed. (We would have to be given more information to determine the direction,
because ω/k = c only up to a ± sign.) The transverse-oscillation frequency of a given
point described by the k2 wave is therefore 3 times what the frequency would be if the
point were instead described by the k1 wave, as indicated in the straight-line dispersion Figure 5
relation in Fig. 6. Physically, this fact is evident if you imagine shifting both of the
waves horizontally by, say, 1 cm on the paper. Since the k2 wavelength is 1/3 the k1 ω
wavelength, a point on the k2 curve goes through 3 times as much oscillation phase
3ck
as a point on the k2 wave. In contrast, in a dispersionful system, the speeds of the
waves don’t have to all be equal. And furthermore, for the k values associated with ck
the points on the horizontal line in Fig. 3, the speeds work out in just the right way k
so that the oscillation frequencies of the points don’t depend on which wave they’re k 3k
considered to be on.
Figure 6

6.2 Evanescent waves

6.2.1 High-frequency cutoff
The dispersion relation in Eq. (9) follows from Eq. (6), which we derived back in Section
2.3.1 (although we didn’t use the term “dispersion” there). The bulk of the derivation was
contained in Claim 2.1. But recall that this claim assumed that ω ≤ 2ω0 . This is consistent
with the fact that the largest value of ω in Fig. 3 is 2ω0 ? However, what if we grab the
end of a string and wiggle it sinusoidally with a frequency ω that is larger than 2ω0 . We’re
6 CHAPTER 6. DISPERSION

free to pick any ω we want, and the string will certainly undergo some kind of motion. But
apparently this motion, whatever it is, isn’t described by the above sinusoidal waves that
we found for the ω ≤ 2ω0 case.
If ω > 2ω0 , then the math in Section 2.3.1 that eventually led to the ω = 2ω0 sin(k`/2)
result in Eq. (9) is still perfectly valid. So if ω > 2ω0 , we conclude that sin(k`/2) must be
greater than 1. This isn’t possible if k is real, but it is possible if k is complex. So let’s plug
k ≡ K + iκ into ω = 2ω0 sin(k`/2), and see what we get. We obtain (the trig sum formula
works fine for imaginary arguments)
µ ¶
ω K` iκ`
= sin −
2ω0 2 2
µ ¶ µ ¶ µ ¶ µ ¶
K` iκ` K` iκ`
= sin cos − cos sin . (13)
2 2 2 2
By looking at the Taylor series for cosine and sine, the cos(iκ`/2) function is real because
the series has only even exponents, while the sin(iκ`/2) function is imaginary (and nonzero)
because the series has only odd exponents. But we need the righthand side of Eq. (13) to be
real, because it equals the real quantity ω/2ω0 . The only way for this to be the case is for the
cos(K`/2) coefficient of sin(iκ`/2) to be zero. Therefore, we must have K` = π, 3π, 5π, . . ..
However, along the same lines as the redundancies in the k values we discussed in the third
comment in the previous section, the 3π, 5π, . . . values for K` simply reproduce the motions
(at least at the locations of the beads) that are already described by the Ka = π value. So
we need only consider the K` = π =⇒ K = π/` value. Said in another way, if we’re ignoring
all the Nyquist redundancies, then we know that k = π/` when ω = 2ω0 (see Fig. 3). And
since the real part of k should be continuous at ω = 2ω0 (imagine increasing ω gradually
across this threshold), we conclude that K = π/` for all ω > 2ω0 . So k ≡ K + iκ becomes
π
k = + iκ. (14)
`
Plugging K = π/` into Eq. (13) yields
µ ¶ µ ¶
ω iκ` κ` ω
= (1) cos − 0 =⇒ cosh = . (15)
2ω0 2 2 2ω0
This equation determines κ. You van verify the conversion to the hyperbolic cosh function
by writing out the Taylor series for both cos(iy) and cosh(y). We’ll keep writing things in
terms of κ, but it’s understood that we can solve for it by using Eq. (15).
What does the general exponential solution, Bei(kx−ωt) , for ψ look like when k takes on
the value in Eq. (14)? (We could work in terms of trig functions, but it’s much easier to use
exponentials; we’ll take the real part in the end.) Remembering that we care only about the
position of the string at the locations of the masses, the exponential solution at positions of
the form x = n` becomes
ψ(x, t) = Bei(kx−ωt)
=⇒ ψ(n`, t) = Bei((π/`+iκ)(n`)−ωt)
= Be−κn` ei(nπ−ωt)
= Be−κn` (−1)n e−iωt
→ Ae−κn` (−1)n cos(ωt + φ) (16)

where we have taken the real part. The phase φ comes from a possible phase in B.1 If we
1 If you’re worried about the legality of going from real k values to complex ones, and if you have your

doubts that this ψ function actually does satisfy Eq. (4), you should plug it in and explicitly verify that it
works, provided that κ is given by Eq. (15). This is the task of Problem [to be added].
6.2. EVANESCENT WAVES 7

want to write this as a function of x = n`, then it equals ψ(x, t) = Ae−κx (−1)x/` cos(ωt+φ). ψ κ=0.03, ω 2ω0
But it is understood that this is valid only for x values that are integral multiples of `. 1. 0
Adjacent beads are 180◦ out of phase, due to the (−1)n factor in ψ. As a function of 0. 5
position, the wave is an alternating function that is suppressed by an exponential. A wave x
2 4 6 8 10
that dies out exponentially like this is called an evanescent wave. Two snapshots of the - 0. 5
wave at a time of maximal displacement (that is, when cos(ωt + φ) = 1) are shown in Fig. 7, - 1. 0
for the values A = 1, ` = 1. The first plot has κ = 0.03, and the second has κ = 0.3. From
Eq. (15), we then have ω ≈ (2.02)(2ω0 ) in the latter, and ω is essentially equal to 2ω0 in ψ
the former. 1. 0
κ=0.3, ω=(2.02)ω0
Since the time and position dependences in the wave appear in separate factors (and not 0. 5
in the form of a kx − ωt argument), the wave is a standing wave, not a traveling wave. As x
time goes on, each wave in Fig. 7 simply expands and contracts (and inverts), with frequency - 0. 5
2 4 6 8 10

ω. All the beads in each wave pass through equilibrium at the same time. - 1. 0

Remark: A note on terminology. In Section 4.6 we discussed attenuated waves, which are also Figure 7
waves that die out. The word “attenuation” is used in the case of actual sinusoidal waves that
decrease to zero due to an envelope curve (which is usually an exponential); see Fig. 4.26. The
word “evanescent” is used when there is no oscillatory motion; the function simply decreases to
zero. The alternating signs in Eq. (16) make things a little ambiguous. It’s semantics as to whether
Eq. (16) represents two separate exponential curves going to zero, or a very fast oscillation within
an exponential envelope. But we’ll choose to call it an evanescent wave. At any rate, the system
in Section 6.2.2 below will support waves that are unambiguously evanescent. ♣

We see that 2ω0 is the frequency above which the system doesn’t support traveling waves.
Hence the “High-frequency cutoff” name of this subsection. We can have nice traveling-wave
motion below this cutoff, but not above. Fig. 8 shows the ψ(x, t) = A cos(kx − ωt + φ) waves
that arise if we wiggle the end of the string with frequencies ω equal to (0.1)ω0 , ω0 , (1.995)ω0 ,
and 2ω0 . The corresponding values of k are determined from Eq. (9). We have arbitrarily
picked A = 1 and ` = 1. The sine waves are shown for convenience, but they aren’t really
there. We haven’t shown the actual straight-line string segments between the masses. At
the ω = 2ω0 cutoff between the traveling waves in Fig. 8 and the evanescent waves in Fig.
7, the masses form two horizontal lines that simply rise and fall.

ω=(0.1)ω0 ω=ω0
ψ
1. 0 1. 0
0. 5 0. 5
x
2 4 6 8 10 2 4 6 8 10
- 0. 5 - 0. 5
- 1. 0 - 1. 0

ω=(1.995)ω0 ω=2ω0
1. 0 1. 0
0. 5 0. 5

2 4 6 8 10 2 4 6 8 10
- 0. 5 - 0. 5
- 1. 0 - 1. 0

(sine curves not actually present)

(straight-line string segments between masses not shown)

Figure 8
8 CHAPTER 6. DISPERSION

The cutoff case of ω = 2ω0 can be considered to be either an evanescent wave or a

traveling wave. So both of these cases must reduce to the same motion when ω = ω0 . Let’s
verify this. If we consider the wave to be an evanescent wave, then with ω = 2ω0 , Eq. (15)
gives κ = 0. So there is no exponential decay, and the beads’ positions in Eq. (16) simply
alternate indefinitely to the right, in two horizontal lines. If we instead consider the wave
to be a traveling wave, then with ω = 2ω0 , Eq. (9) gives k` = π =⇒ k = π/`, which means
that the wavelength approaches 2`. The traveling wave at positions of the form x = n`
looks like
³π ´
A cos(kx − ωt + φ) = A cos · n` − ωt + φ
`
= A cos(nπ − ωt + φ)
= A cos(nπ) cos(ωt − φ)
= A(−1)n cos(ωt − φ), (17)

which agrees with Eq. (16) when κ = 0 (with a different definition of φ).
In the extreme case where ω À ω0 , Eq. (15) tells us that κ is large, which means that
the exponential factor e−κn` goes to zero very quickly. This makes sense. If you wiggle the
end of the string very quickly (as always, we’re assuming that the amplitude is small, in
particular much smaller than the bead spacing), then the bead that you’re holding onto will
move with your hand, but all the other beads will hardly move at all. This is because they
have essentially zero time to accelerate in one direction before you move your hand to the
other side and change the direction of the force.
In practice, ω doesn’t have to be much larger than 2ω0 for this lack of motion to arise.
Even if ω is only 4ω0 , then Eq. (15) gives the value of κ` as 2.6. The amplitude of the
n = 1 and n = 2 masses are then suppressed by a factors of e−κn` = e−1(2.6) ≈ 1/14, and
e−κn` = e−2(2.6) ≈ 1/200. So if you grab the n = 0 mass at the end of the string and move
it back and forth at frequency 4ω0 , you’ll end up moving the n = 1 mass a little bit, but all
the other masses will essentially not move.
p
Remark: For given values of T and µ, the relation ω0 = (1/`) T /µ (see Eq. (11)) implies that if
` is very small, you need to wiggle the string very fast to get into the ω > 2ω0 evanescent regime.
In the limit of a continuous string (` → 0), ω0 is infinite, so you can never get to the evanescent
regime. In other words, any wiggling that you do will produce a normal traveling wave. This
makes intuitive sense. It also makes dimensional-analysis sense, for the following reason. Since a
continuous string is completely defined in terms of the two parameters T (with units of kg m/s2 )
and µ (with units of kg/m), there is no way to combine these parameters to form a quantity with
the dimensions of frequency (that is, s−1 ). So for a continuous string, there is therefore no possible
frequency value that can provide the cutoff between traveling and evanescent waves. All waves
must therefore fall into one of these two categories, and it happens to be the traveling waves. If,
however, a length scale ` is introduced, then it is possible to generate a frequency (namely ω0 ),
which can provide the scale of the cutoff (which happens to be 2ω0 ). ♣

Power
If you wiggle the end of a beaded string with a frequency larger than 2ω0 , so that an
evanescent wave of the form in Eq. (16) arises, are you transmitting any net power to the
string? Since the wave dies off exponentially, there is essentially no motion far to the right.
Therefore, there is essentially no power being transmitted across a given point far to the
right. So it had better be the case that you are transmitting zero net power to the string,
because otherwise energy would be piling up indefinitely somewhere between your hand and
6.2. EVANESCENT WAVES 9 1. 0
0. 5
x
the given point far to the right. This is impossible, because there is no place for this energy 2 4 6 8 10
- 0. 5
to go.2
- 1. 0
It is easy to see directly why you transmit zero net power over a full cycle (actually over
each half cycle). Let’s start with the position shown in Fig. 9, which shows a snapshot when Figure 9
the masses all have maximal displacement. The n = 0 mass is removed, and you grab the
end of the string where that mass was. As you move your hand downward, you do negative
work on the string, because you are pulling upward (and also horizontally, but there is no
work associated with this force because your hand is moving only vertically), but your hand
is moving in the opposite direction, downward. However, after your hand passes through
equilibrium, it is still moving downward but now pulling downward too (because the string
you’re holding is now angled up to the right), so you are doing positive work on the string.
The situation is symmetric (except for signs) on each side of equilibrium, so the positive
and negative works cancel, yielding zero net work, as expected.
In short, your force is in “quadrature” with the velocity of your hand. That is, it is
90◦ out of phase with the velocity (behind it). So the product of the force and the velocity
(which is the power) cancels over each half cycle. This is exactly the same situation that
arises in a simple harmonic oscillator with a mass on a spring. You can verify that the force
and velocity are in quadrature (the force is ahead now), and there is no net work done by
the spring (consistent with the fact that the average motion of the mass does’t change over
time).
How do traveling waves (ω < 2ω0 ) differ from evanescent waves (ω > 2ω0 ), with regard
to power? For traveling waves, when you wiggle the end, your force isn’t in quadrature
with the velocity of your hand, so you end up doing net positive work. In the small-ω limit
(equivalently, the continuous-string limit), your force is exactly in phase with the velocity,
so you’re always doing positive work. However, as ω increases (with a beaded string), your
force gradually shifts from being in phase with the velocity at ω ≈ 0, to being in quadrature
with it at ω = 2ω0 , at which point no net work is being done. The task of Problem [to be
added] is to be quantitative about this.

6.2.2 Low-frequency cutoff

The above beaded string supported traveling waves below a certain cutoff frequency but not
above it. Let’s now consider the opposite situation. We’ll look at a system that supports
traveling waves above a certain cutoff frequency but not below it. Hence the “Low-frequency
cutoff” name of this subsection.
Consider a continuous string (not a beaded one) with tension T and density µ. Let the
(infinite) string be connected to a wall by an essentially continuous set of springs, initially
at their relaxed length, as shown in Fig. 10.

string
springs
wall

Figure 10

If the springs weren’t present, then the return force (for transverse waves) on a little
piece of the string with length ∆x would be T ∆x(∂ 2 ψ/∂x2 ); see Eq. (4.2). But we now also
2 We are assuming steady-state motion. At the start, when you get the string going, you are doing net

work on the string. Energy piles up at the start, because the string goes from being in equilibrium to moving
back and forth. But in steady state, the average motion of the string doesn’t change in time.
 10 CHAPTER 6. DISPERSION

have the spring force, −(σ∆x)ψ, where σ is the spring constant per unit length. (The larger
the piece, the more springs that touch it, so the larger the force.) The transverse F = ma
equation on the little piece is therefore modified from Eq. (4.3) to

∂2ψ ∂2ψ
(µ∆x) = T ∆x − (σ∆x)ψ
∂t2 ∂x2
∂2ψ ∂2ψ T σ
=⇒ = c2 2 − ωs2 ψ, where c2 ≡ and ωs2 ≡ . (18)
∂t2 ∂x µ µ

We will find that c is not the wave speed, as it was in the simple string system with no
springs. To determine the dispersion relation associated with Eq. (18), we can plug in our
ω standard exponential solution, ψ(x, t) = Aei(kx−ωt) . This tells us that ω and k are related
ω= c2k2+ωs2
by
ω 2 = c2 k 2 + ωs2 (dispersion relation) (19)
ωs This is the dispersion relation for the string-spring system. The plot of ω vs. k is shown in
ω=ck
Fig. 11. There is no (real) value of k that yields a ω smaller than ωs . However, there is an
imaginary value of k that does. If ω < ωs , then Eq. (19) gives
k
p p
Figure 11 ω 2 − ωs2 ωs2 − ω 2
k= ≡ iκ, where κ ≡ . (20)
c c
Another solution for κ is the negative of this one, but we’ll be considering below the case
where the string extends to x = +∞, so this other solution would cause ψ to diverge,
given our sign convention in the exponent of ei(kx−ωt) . Substituting k ≡ iκ into ψ(x, t) =
Aei(kx−ωt) gives

ψ(x, t) = Aei((iκ)x−ωt) = Ae−κx e−iωt → Be−κx cos(ωt + φ) (21)

where as usual we have taken the real part. We see that ψ(x, t) decays as a function of x,
and that all points on the string oscillate with the same phase as a function of t. (This is
in contrast with adjacent points having opposite phases in the above beaded-string setup.
Opposite phases wouldn’t make any sense here, because we don’t have discrete adjacent
(start) points on a continuous string.) So we have an evanescent standing wave. If we wiggle the
left end up and down sinusoidally with a frequency ω < ωs , then snapshots of the motion
take the general form shown in Fig. 12. The rate of the exponential decrease (as a function
(1/4 period later) of x) depends on ω. If ω is only slightly smaller than ωs , then the κ in Eq. (20) is small, so
the exponential curve dies out very slowly to zero. In the limit where ω ≈ 0, we’re basically
holding the string at rest in the first position in Fig. 12, and you can show from scratch by
balancing transverse forces in this static setupp that the string does indeed take the shape
(1/2 period later) of a decreasing exponential with κ ≈ ωs /c = σ/T ; see Problem [to be added]. This static
case yields the quickest spatial decay to zero.
Figure 12 It makes sense that the system doesn’t support traveling waves for ω < ωs , because even
without any tension force, the
p springs would still make the atoms in the string oscillate
with a frequency of at least σ∆x/µ∆x = ωs . Adding on a tension force will only increase
the restoring force (at least in a traveling wave, where the curvature is always toward the
x axis), and thus also the frequency. In the cutoff case where ω = ωs , the string remains
straight, and it oscillates back and forth as a whole, just as a infinite set of independent
adjacent masses on springs would oscillate.
We have been talking about evanescent waves where ω < ωs , but we can still have normal
traveling waves if ω > ωs , because the k in Eq. (20) is then real. If ω À ωs , then Eq. (20)
tells us that ω ≈ ck. In other words, the system acts like a simple string with no springs
6.2. EVANESCENT WAVES 11

attached to it. This makes sense; on the short time scale of the oscillations, the spring ω
doesn’t have time to act, so it’s effectively like it’s not there. Equivalently, the transverse ω= c2k2+ωs2
force from the springs completely dominates the spring force. If on the other hand ω is
only slightly larger than ωs , then Eq. (20) says that k is very small, which means that the
wavelength is very large. In the ω → ωs limit, we have λ → ∞. The speed of the wave is ωs
slope = ω/k
then ω/k ≈ ωs /k ≈ ∞. This can be seen graphically in Fig. 13, where the slope from the = phase velocity
origin is ω/k, which is the phase velocity (just as it was in Fig. 3). This slope can be made
k
arbitrarily large by making k be arbitrarily small. We’ll talk more about excessively large
phase velocities (in particular, larger than the speed of light) in Section 6.3.3. Figure 13
Note that the straight-line shape of the string in the ω = ωs case that we mentioned
above can be considered to be the limit of a traveling wave with an infinitely long wavelength,
and also an evanescent wave with an infinitely slow decay.

Power
As with the evanescent wave on the beaded string in Section 6.2.1, no net power can be
transmitted in the present evanescent wave, because otherwise there would be energy piling
up somewhere (because the wave dies out). But there is no place for it to pile up, because
we are assuming steady-state motion. This can be verified with the same reasoning as in
the beaded-string case; the net power you transmit to the string as you wiggle the left end
alternates sign each quarter cycle, so there is complete cancelation over a full cycle.
Let’s now consider the modified setup shown in Fig. 14. To the left of x = 0, we have
a normal string with no springs. What happens if we have a rightward-traveling wave that
comes in from the left, with a frequency ω < ωs . (Or there could even be weak springs in
the left region, as long as we have ωs,left < ω < ωs,right . This would still allow a traveling
wave in the left region.)

string
(no springs) springs
wall
x=0

Figure 14

From the same reasoning as above, the fact that the wave dies out on the right side
implies that no net power is transmitted along the string (in the steady state). However,
there certainly is power transmitted in the incoming traveling wave. Where does it go?
Apparently, there must be complete reflection at x = 0, so that all the power gets reflected
back. The spring region therefore behaves effectively like a brick wall, as far as reflection
goes. But the behavior isn’t exactly like a brick wall, because ψ isn’t constrained to be zero
at the boundary in the present case.
To figure out what the complete wave looks like, we must apply the boundary conditions
(continuity of the function and the slope) at x = 0.3 If we work with exponential solutions,
then the incoming, reflected, and transmitted waves take the form of Aei(ωt−kx) , Bei(ωt+kx) ,
and Ceiωt e−κx , respectively. The goal is to solve for B and C (which may be complex) in
terms of A, that is, to solve for the ratios B/A and C/A. For ease of computation, it is
3 As usual, the continuity of the slope follows from the fact that there can be no net force on the essentially

massless atom at the boundary. The existence of the springs in the right region doesn’t affect this fact.
12 CHAPTER 6. DISPERSION

customary to divide all of the functions by A, in which case the total waves on the two sides
of the boundary can be written as

ψL = (e−ikx + Reikx )eiωt ,

ψR = De−κx eiωt , (22)

where R ≡ B/A and D ≡ C/A. R is the complex reflection coefficient. Its magnitude
is what we normally call the “reflection coefficient,” and its phase yields the phase of the
reflected wave when we eventually take the real part to obtain the physical solution. The
task of Problem [to be added] is to apply the boundary conditions and solve for R and D.

Tunneling
What happens if we have a setup in which the region with strings is finite, as shown in Fig.
15? If a rightward-traveling wave with ω < ωs comes in from the left, some of it will make it
through, and some of it will be reflected. That is, there will be a nonzero rightward-traveling
wave in the right region, and there will be the usual combination of a traveling and standing
wave in the left region that arises from partial reflection.

Figure 15

The nonzero wave in the right region implies that power is transmitted. This is consistent
with the fact that we cannot use the reasoning that zero power follows from the fact that the
wave dies out to zero; the wave doesn’t die out to zero, because the middle region has finite
length. We can’t rule out the e+κx solution, because the x = +∞ boundary condition isn’t
relevant now. So the general (steady state) solution in the left, middle, and right regions
takes the form,

ψL = (e−ikx + Reikx )eiωt ,

ψM = (Be−κx + Ceκx )eiωt ,
ψR = T e−ikx eiωt , (23)

where R and T are the complex reflection and transmission coefficients, defined to be the
ratio of the reflected and transmitted (complex) amplitudes to the incident (complex) am-
plitude. As in Eq. (22), we have written the waves in their complex forms, with the un-
derstanding that we will take the real part to find the actual wave. The four boundary
conditions (value and slope at each of the two boundaries) allow us to solve for the four
unknowns (R, T , B, C). This is the task of Problem [to be added]; the math gets a bit
messy.
The effect where some of the wave makes it through the “forbidden” region where trav-
eling waves don’t exist is known as tunneling. The calculation in Problem [to be added] is
exactly the same as in a quantum mechanical system involving tunneling through a classi-
cally forbidden region (a region where the total energy is less than the potential energy).
In quantum mechanics, the waves are probability waves instead of transverse string waves,
so the interpretation of the waves is different. But all the math is exactly the same as in
the above string-spring system. We’ll talk much more about quantum-mechanical waves in
Chapter 11.
6.3. GROUP VELOCITY 13

6.3 Group velocity

6.3.1 Derivation
Whether a system is dispersionless (with a linear relationship between ω and k) or disper-
sionful (with a nonlinear relationship between ω and k), the phase velocity in both cases is
vp = ω/k. The phase velocity gives the speed of a single sinusoidal traveling wave. But what
if we have, say, a wave in the form of a lone bump, which (from Fourier analysis) can be
written as the sum of many (or an infinite number of) sinusoidal waves. How fast does this
bump move? If the system is dispersionless, then all of the wave components move with the
same speed vp , so the bump also moves with this speed. We discussed this effect in Section
2.4 (see Eq. (2.97)), where we noted that any function of the form f (x − ct) is a solution to a
dispersionless wave equation, that is, an equation of the form ∂ 2 ψ/∂t2 = c2 (∂ 2 ψ/∂x2 ). This
equation leads to the relation ω = ck, where c takes on a single value that is independent
of ω and k.
However, if the system is dispersionful, then ω/k depends on k (and ω), so the different
sinusoidal waves that make up the bump travel at different speeds. So it’s unclear what the
speed of the bump is, or even if the bump has a well-defined speed. It turns out that it does
in fact have a well-defined speed, and it is given by the slope of the ω(k) curve:

dω
vg = (24)
dk

This is called the group velocity, which is a sensible name considering that a bump is made
up of a group of Fourier components, as opposed to a single sinusoidal wave. Although the
components travel at different speeds, we will find below that they conspire in such a way
as to make their sum (the bump) move with speed vg = dω/dk. However, an unavoidable
consequence of the differing speeds of the components is the fact that as time goes on, the
bump will shrink in height and spread out in width, until you can hardly tell that it’s a
bump. In other words, it will disperse. Hence the name dispersion.
Since the bump consists of wave components with many different values of k, there is
an ambiguity about which value of k is the one where we should evaluate vg = dω/dk. The
general rule is that it is evaluated at the value of k that dominates the bump. That is, it is
evaluated at the peak of the Fourier transform of the bump.
We’ll now derive the result for the group velocity in Eq. (24). And because it is so
important, we derive it in three ways.

First derivation

Although we just introduced the group velocity by talking about the speed of a bump,
which consists of many Fourier components, we can actually understand what’s going on by
considering just two waves. Such a system has all the properties needed to derive the group
velocity. So consider the two waves:

ψ1 (x, t) = A cos(ω1 t − k1 x),

ψ2 (x, t) = A cos(ω2 t − k2 x). (25)

It isn’t necessary that they have equal amplitudes, but it simplifies the discussion. Let’s see
what the sum of these two waves looks like. It will be advantageous to write the ω’s and
 14 CHAPTER 6. DISPERSION

k’s in terms of their averages and differences.4 So let’s define:

ω1 + ω2 ω1 − ω2 k1 + k2 k1 − k2
ω+ ≡ , ω− ≡ , k+ ≡ , k− ≡ . (26)
2 2 2 2
Then ω1 = ω+ + ω− and ω2 = ω+ − ω− . And likewise for the k’s. So the sum of the two
waves can be written as
³ ´
ψ1 (x, t) + ψ2 (x, t) = A cos (ω+ + ω− )t − (k+ + k− )x
³ ´
+A cos (ω+ − ω− )t − (k+ − k− )x
³ ´
= A cos (ω+ t − k+ x) + (ω− t − k− x)
³ ´
+A cos (ω+ t − k+ x) − (ω− t − k− x)
¡ ¢ ¡ ¢
= 2A cos ω+ t − k+ x cos ω− t − k− x , (27)

where we have used cos(α ± β) = cos α cos β ∓ sin α sin β. We see that the sum of the two
original traveling waves can be written as the product of two other traveling waves.
ψ1+ψ2 This is a general result for any values of ω1 , ω2 , k1 , and k2 . But let’s now assume that
1.0 ω1 is very close to ω2 (more precisely, that their difference is small compared with their
0.5 sum). And likewise that k1 is very close to k2 . We then have ω− ¿ ω+ and k− ¿ k+ .
x Under these conditions, the sum¡ψ1 +ψ2 in ¢Eq. (27) is the product of the quickly-varying
¡ (in
¢
1 2 3 4 5 6
- 0.5 both space and time) wave, cos ω+ t − k+ x , and the slowly-varying wave, cos ω− t − k− x .
- 1.0 A snapshot (at the arbitrarily-chosen time of t = 0) of ψ1 + ψ2 is shown in Fig. 16. The
quickly-varying
¡ wave
¢ is the actual sum, while the slowly-varying envelope is the function
Figure 16
2A cos ω− t − k− x . We have arbitrarily picked 2A = 1 in the figure. And we have chosen
k1 = 10 and k2 = 12, which yield k+ = 11 and k− = 1. So the envelope function is cos(x),
and the wiggly function (which equals ψ1 + ψ2 ) is cos(11x) cos(x).
At t increases, the quickly- and slowly-varying waves will move horizontally. What are
the velocities of these two waves? The velocity of the quickly-wiggling wave is ω+ /k+ ,
which is essentially equal to either of ω1 /k1 and ω2 /k2 , because we are assuming ω1 ≈ ω2
and k1 ≈ k2 . So the phase velocity of the quickly-wiggling wave is essentially equal to the
phase velocity of either wave.
The velocity of the slowly-varying wave (the envelope) is
ω− ω1 − ω2
= . (28)
k− k1 − k2
(Note that this may be negative, even if the phase velocities of the original two waves are
both positive.) If we have a linear dispersion relation, ω = ck, then this speed equals
c(k1 − k2 )/(k1 − k2 ) = c. So the group velocity equals the phase velocity, and this common
velocity is constant, independent of k. But what if ω and k aren’t related linearly? Well, if
ω is given by the function ω(k), and if k1 is close to k2 , then (ω1 −ω2 )/(k1 −k2 ) is essentially
equal to the derivative, dω/dk. This is the velocity of the envelope formed by the two waves,
and it is called the the group velocity. To summarize:

ω dω
vp = and vg = (29)
k dk
In general, both of these velocities are functions of k. And in general they are not equal.
(The exception to both of these statements occurs in the case of linear dispersion.) In the
4 We did something similar to this when we talked about beats in Section 2.1.4. But things are a little

different here because the functions are now functions of both x and t, as opposed to just t.
6.3. GROUP VELOCITY 15

general case where vg 6= vp , the fast wiggles in Fig. 16 move with respect to the envelope. ω
If vp > vg , then the little wiggles pop into existence at the left end of an envelope bump slope = dω/dk = vg
(or the right end if vp < vg ). They grow and then shrink as they move through the bump, 2ω0
until finally they disappear when they reach the right end of the bump.
In the case of the beaded-string system discussed in Section 6.1, the plot of ω(k) was slope = ω/k = vp
shown in Fig. 3. So the phase and group velocities are shown graphically in Fig. 17. For k
0 < k < π/` (which is generally the part of the graph we’re concerned with), the slope of π/l
the curve at any point is less than the slope of the line from the origin to the point, so we
see that vg is always less than vp . Figure 17
In the case of the string/spring system discussed in Section 6.2.2, the plot of ω(k) was
shown in Fig. 11. So the phase and group velocities are shown graphically in Fig. 18. We ω
again see that vg is always less than vp . However, this need not be true in general, as we’ll
see in the examples in Section 6.3.2 below. For now, we’ll just note that in the particular
case where the plot of ω(k) passes through the origin, there are two basic possibilities of
what the ω(k) curve can look like, depending on whether it is concave up or down. These ωs
slope = dω/dk = vg
are shown in Fig. 19. The first case always has vg > vp , while the second always has vg < vp .
slope = ω/k = vp
In the first case, sinusoidal waves with small k (large λ) travel slower (that is, they have a
k
smaller vp ) than waves with large k (small λ). The opposite is true in the second case.
Figure 18
Second derivation
Consider two waves with different values of ω and k, as shown in the first pair of waves in ω
(vg > vp)
Fig. 20. These two waves constructively interfere at the dots, so there will be a bump there.
When and where does the next bump occur? If we can answer these questions, then we can
find the effective velocity of the bump. vg
vp

λ1
v1 k

v2
ω
vg
λ2

vp (vg < vp)

Figure 20

If v1 = v2 (that is, if ω1 /k1 = ω2 /k2 ), then both waves travel at the same speed, so the k
bump simply travels along with the waves, at their common speed. But if v1 6= v2 , then the
dots will become unaligned. If we assume that v1 > v2 (the v1 < v2 case is similar), then Figure 19
at some later time the next two peaks will line up, as shown in the second pair of waves in
Fig. 20. These peaks are marked with x’s. There will then be a bump at this new location.
(If v1 < v2 , the next alignment will occur to the left of the initial one.)
When do these next peaks line up? The initial distance between the x’s is λ2 − λ1 , and
the top wave must close this gap at a relative speed of v1 − v2 , so t = (λ2 − λ1 )/(v1 − v2 ).
Equivalently, just set λ1 +v1 t = λ2 +v2 t, because these two quantities represent the positions
of the two x’s, relative to the initial dots. Having found the time t, the position of the next
alignment is given by x = λ1 + v1 t (and also λ2 + v2 t). The velocity at which the bump
effectively travels is therefore
µ ¶
x λ1 + v 1 t λ1 v1 − v2 λ1 v1 − λ1 v2 λ2 v 1 − λ1 v 1
= = + v 1 = λ1 + v1 = +
t t t λ2 − λ1 λ2 − λ1 λ2 − λ1
 16 CHAPTER 6. DISPERSION

2π ω1 2π ω2 ω1 −ω2
λ2 v 1 − λ1 v 2 k2 k1 − k1 k2 k1 k2 ω1 − ω2
= = 2π 2π = k1 −k2
= ≡ vg . (30)
λ2 − λ1 k2 − k1 k1 k2
k1 − k2

This is the same speed we obtained in Eq. (28). This is no surprise, because we basically
did the same calculation here. In the previous derivation, we assumed that the waves were
nearly identical, whereas we didn’t assume that here. This assumption isn’t needed for the
vg = (ω1 − ω2 )/(k1 − k2 ) result. Whenever and wherever a bump in the present derivation
exists, it touches the top of the envelope curve (if we had drawn it). So what we effectively
did in this derivation is find the speed of the envelope curve. But this is exactly what we
did in the previous derivation.
In between the alignments of the dot and the x in Fig. 20, the bump disappears, then
appears in the negative direction, then disappears again before reappearing at the x. This
is consistent with the fact that the wiggly wave in Fig. 16 doesn’t always (in fact, rarely
does) touch the midpoint (the highest point) of the envelope bump. But on average, the
bump effectively moves with velocity vg = (ω1 − ω2 )/(k1 − k2 ).
Note that if k1 is very close to k2 , and if ω1 is not very close to ω2 , then vg = (ω1 −
ω2 )/(k1 − k2 ) is large. It is easy to see intuitively why this is true. We may equivalently
describe this scenario by saying that λ1 is very close to λ2 , and that v1 is not very close to
v2 (because v = ωk). The nearly equal wavelengths imply that in Fig. 20 the two x’s are
very close together. This means that it takes essentially no time for them to align (because
the velocities aren’t close to each other). The location of the alignment therefore jumps
ahead by a distance of one wavelength in essentially no time, which means that the effective
speed is large (at least as large as the λ1 /t term in Eq. (30)).
What if we have a large number of waves with roughly the same values of k (and hence
ω), with a peak of each wave lining up, as shown by the dots in Fig. 21? Since the plot of
ω(k) at any point is locally approximately a straight line, the quotient (ω1 − ω2 )/(k1 − k2 ),
which is essentially equal to the derivative dω/dk, is the same for all nearby points, as shown
in Fig. 22. This means that the next bumps (the x’s in Fig. 21) will all line up at the same
time and at the same place, because the location of all of the alignments is given by x = vg t,
Figure 21 by Eq. (30). In other words, the group velocity vg is well defined. The various waves all
travel with different phase velocities vp = ω/k, but this is irrelevant as far as the group
ω velocity goes, because vg depends on the differences in ω and k through Eq. (30), and not
on the actual values of ω and k.

Third derivation
By definition, vg is the velocity at which a bump in a wave travels. From Fourier analysis,
we know that in general a wave consists of components with many different frequencies. If
k these components are to “work together” to form a bump at a certain location, then the
phases ωi t − ki t + φi of the different components (or at least many of them) must be equal
Figure 22
at the bump, if they are to add constructively to form the bump.
Assume that we have a bump at a particular value of x and t. We are free to pick the
origins of x and t to be wherever and whenever we want, so let’s pick the bump to be located
at x = 0 and t = 0. Since the phases ωi t − ki t + φi are all equal (or at least many of them)
in a region around some particular k value, we conclude that the φi are all equal, because
x = t = 0. In other words, φ is independent of k.5
At what other values of x and t, besides (x, t) = (0, 0), is there a bump? That is, at
what other values of x and t are the phases still all equal? Well, we want ωt − kx + φ to
be independent of k near some particular k value, because then the phases of the waves for
5 You can check that the following derivation still works in the case of general initial coordinates (x , t );
0 0
see Problem [to be added]. But it’s less messy if we choose (0, 0).
6.3. GROUP VELOCITY 17

all the different k values will be equal, which means that the waves will add constructively
and therefore produce another bump. (We have dropped the index i on ω and k, and it is
understood that ω is a function ω(k) of k.) Demanding that the phase be independent of k
gives
d(ωt − kx + φ) dω x dω
0= =⇒ 0 = t − x =⇒ = , (31)
dk dk t dk
where we have used dφ/dk = 0. So we have a bump at any values of x and t satisfying this
relation. In other words, the speed of the bump is

dω
vg = , (32)
dk
in agreement with the result from the previous derivations.
Note that the phase velocity (of single traveling wave) is obtained by demanding that
the phase ωt − kx + φ of the wave be independent of time:

d(ωt − kx + φ) dx dx ω
0= =⇒ 0 = ω − k =⇒ vp ≡ = . (33)
dt dt dt k
But the group velocity (of a group of traveling waves) is obtained by demanding that the
phase ωt − kx + φ of all the different waves be independent of the wavenumber k:

d(ωt − kx + φ) dω x dω
0= =⇒ 0 = t − x =⇒ vg ≡ = . (34)
dk dk t dk
Just because the quantity dω/dk exists, there’s no guarantee that there actually will be
a noticeable bump traveling at the group velocity vg . It’s quite possible (and highly likely
if things are random) that there is no constructive interference anywhere. But what we
showed above was that if there is a bump at a given time and location, then it travels with
velocity vg = dω/dk, evaluated at the k value that dominates the bump. This value can be
found by calculating the Fourier transform of the bump.

6.3.2 Examples
Beaded string
We discussed the beaded string in
p Section 6.1. Eq. (9) gives the dispersion relation as
ω(k) = 2ω0 sin(k`/2), where ω0 ≡ T /m`. Therefore,

ω 2ω0 sin(k`/2) dω
vp = = , and vg = = ω0 ` cos(k`/2). (35)
k k dk
For small k (more precisely, for k` ¿ 1), we can use sin ² ≈ ² and cos ² ≈ 1, to quickly show
that
r s s
2ω0 (k`/2) T T T
vp ≈ = ω0 ` = `= = ,
k m` m/` µ
s
T
vg ≈ ω0 `(1) = . (36)
µ

As expected, these both agree with the (equal) phase and group velocities for a continuous
string, because k` ¿ 1 implies ` ¿ λ, which means that the string is essentially continuous
on a length scale of the wavelength.
18 CHAPTER 6. DISPERSION

String/spring system
We discussed the string/spring system in Section 6.2.2. Eq. (19) gives the dispersion relation
as ω 2 = c2 k 2 + ωs2 , where c2 ≡ T /µ and ωs2 ≡ σ/µ. Therefore,
p
ω c2 k 2 + ωs2 dω c2 k
vp = = , and vg = =p . (37)
k k dk c2 k 2 + ωs2

If ωs ≈ 0, then these reduce to vp ≈ c and vg ≈ c, as expected. If ωs is large (more precisely,

if ωs À ck), then vp is large and vg is small (more precisely, vp À c and vg ¿ c). These
facts are consistent with the slopes in Fig. 18.

Stiff string
When dealing with uniform strings, we generally assume that they are perfectly flexible.
That is, we assume that they don’t bounce back when they are bent. But if we have a “stiff
string” that offers resistance when bent, it can be shown that the wave equation picks up
an extra term and now takes the form,
· 2 µ 4 ¶¸
∂2ψ 2 ∂ ψ ∂ ψ
=c −α ,
∂t2 ∂x2 ∂x4

where α depends on various things (the cross-sectional area, Young’s modulus, etc.).6 Plug-
ging in ψ(x, t) = Aei(ωt−kx) yields the dispersion relation,
p
ω 2 = c2 k 2 + αc2 k 4 =⇒ ω = ck 1 + αk 2 . (38)

This yields

ω p dω c(1 + 2αk 2 )
vp = = c 1 + αk 2 , and vg = = √ . (39)
k dk 1 + αk 2
The dispersion relation in Eq. (38) has implications in piano tuning, because although
the strings in a piano are reasonably flexible, they aren’t perfectly so. They are slightly stiff,
with a small value of α. If they were perfectly flexible (α = 0), then the linear dispersion
relation, ω = ck, would imply that the standing-wave frequencies are simply proportional
to the mode number, n, because the wavenumbers take the usual form of k = nπ/L. So the
“first harmonic” mode (n = 2) would have twice the frequency of the fundamental mode
(n = 1). In other words, it would be an octave higher.
However, for a stiff string (α 6= 0), Eq. (38) tells us that the frequency of the first
harmonic is larger than twice the frequency of the fundamental. (The k values still take the
form of k = nπ/L. This is a consequence of the boundary conditions and is independent of
the dispersion relation.)
Consider two notes that are an “octave” apart on the piano (the reason for the quotes
will soon be clear). These notes are in tune with each other if the first harmonic of the
lower string equals the fundamental of the higher string. Your ear then won’t hear any beats
between these two modes when the strings are played simultaneously, so things will sound
nice.7 A piano is therefore tuned to make the first harmonic of the lower string equal to
6 In a nutshell, the fourth derivative comes from the facts that (1) the resistance to bending (the so-

called “bending moment”) is proportional to the curvature, which is the second derivative of ψ, and (2)
the resulting net transverse force can be shown to be proportional to the second derivative of the bending
moment.
7 Your ear only cares about beats between nearby frequencies. The relation between the two fundamentals

is irrelevant because they are so far apart. Beats don’t result from widely-different frequencies.
6.3. GROUP VELOCITY 19

the fundamental of the higher string. But since the dispersion relation tells us that the first
harmonic (of any string) has more than twice the frequency of the fundamental, we conclude
that the spacing between the fundamentals of the two strings is larger than an octave. But
this is fine because it’s what your ear wants to hear. The (equal) relation between the first
harmonic of the lower string and the fundamental of the higher string is what’s important.
The relation between the two fundamentals doesn’t matter.8

Power law

If a dispersion relation takes the form of a power law, ω = Ak r , then

ω dω
vp = = Ak r−1 , and vg = = rAk r−1 . (40)
k dk

We see that vg = rvp for any value of k. If r = 1, then we have a dispersionless system. If
r > 1, then the dispersion curve is concave up, so it looks like the first plot we showed in
Fig. 19, with vg > vp . Sinusoidal waves with small k travel slower than waves with large k.
If r < 1, then we have the second plot in Fig. 19, and these statements are reversed.

Quantum mechanics

In nonrelativistic quantum mechanics, particles are replaced by probability waves. The wave
equation (known as the Schrodinger equation) for a free particle moving in one dimension
happens to be
∂ψ h̄2 ∂ 2 ψ
ih̄ =− , (41)
∂t 2m ∂x2
where h̄ = 1.05 · 16−34 J · s is Planck’s constant. Plugging in ψ(x, t) = Aei(ωt−kx) yields the
dispersion relation,
h̄k 2
ω= . (42)
2m
We’ll give an introduction to quantum mechanics in Chapter 12, but for now we’ll just note
that the motivation for the dispersion relation (and hence the wave equation) comes from
the substitutions of E = h̄ω and p = h̄k into the standard classical relation, E = p2 /2m.
We’ll discuss the origins of these forms of E and p in Chapter 12.
The dispersion relation gives

ω h̄k dω h̄k
vp = = , and vg = = . (43)
k 2m dk m

Classically, the velocity of a particle is given by v = p/m. So if p = h̄k, then we see that
vg , and not vp , corresponds to the classical velocity of a particle. This is consistent with
the fact that a particle can be thought of as a localized bump in the probability wave,
and this bump moves with the group velocity vg . A single sinusoidal wave moving with
velocity vp doesn’t correspond to a localized particle, because the wave (which represents
the probability) extends over all space. So we shouldn’t expect vp to correspond to the
standard classical velocity of p/m.

8 A nice article on piano tuning is: Physics Today, December 2009, pp 46-49. It’s based on a letter from

Richard Feynman to his piano tuner. See in particular the “How to tune a piano” box on page 48.
20 CHAPTER 6. DISPERSION

Water waves
We’ll discuss water waves in detail in Chapter 11, but we’ll invoke some results here so that
we can see what a few phase and group velocities look like. There are three common types
of waves:
• Small ripples: If the wavelength is short enough so that the effects of surface tension
dominate the effects of gravity, then the dispersion relation takes the form, ω =
p
σk 3 /ρ, where σ is the surface tension and ρ is the mass density. The surface tension
dominates if the wavelength be small compared with about 2 cm. The dispersion
relation then gives
s s
ω σk dω 3 σk
vp = = , and vg = = . (44)
k ρ dk 2 ρ

So vg = 3vp /2. The smaller the wavelength (the larger the k), then the larger the vp .
Very small ripples travel fast.
• Long wavelengths in deep water: If the wavelength is large compared with 2 cm, then
the effects of gravity dominate. If we further assume that the wavelength is small
compared
√ with the depth of the water, then the dispersion relation takes the form,
ω = gk. This gives
r r
ω g dω 1 g
vp = = , and vg = = . (45)
k k dk 2 k

So vg = vp /2. The larger the wavelength (the smaller the k), then the larger the vp .
Long waves travel fast.
• Long wavelengths, compared with depth: If the wavelength is large compared
√ with the
depth of the water, then the dispersion relation takes the form, ω = gH k, where H
is the depth. This is a dispersionless system, with
p
vp = vg = gH. (46)

So all waves travel with the same speed (provided that the wavelength is large com-
pared with H). This has dramatic consequences with regard to tsunamis.
Consider a huge wave that is created in the ocean, most commonly by an earthquake.
If the wave has, say, an amplitude of 100 ft and a wavelength of half a mile (which is
indeed a huge wave), what will happen as it travels across the ocean? The depth of
the ocean is on the order of a few miles, so we’re in the regime of “long wavelengths in
deep water.” From above, this is a dispersive system. Different wavelengths therefore
travel with different speeds, and the wave disperses. It grows shallower and wider,
until there is hardly anything left. When it reaches the other side of the ocean, it will
be barely distinguishable from no wave at all. The fast Fourier components of the
initial bump (the ones with long wavelengths) will arrive much sooner than the slower
components, so the energy of the wave will be diluted over a long period of time.
However, consider instead a wave with an amplitude of only 5 ft, but with a wavelength
of 10 miles. (Assuming roughly the same shape, this wave has the same volume as
the one above.) What will happen to this wave as it travels across the ocean? We’re
now in the “long wavelengths, compared with depth” regime. This is a nondispersive
system, so all of the different Fourier components of the initial “bump” travel with the
same speed. The wave therefore keeps the same shape as it travels across the ocean.
6.3. GROUP VELOCITY 21

Now, a 5 ft wave might not seem severe, but when the wave reaches shallower water
near the shore, its energy gets concentrated into a smaller region, so the amplitude
grows. If the boundary between the ocean and land were a hypothetical vertical
wall extending miles downward, then the waves would simply reflect off the wall and
travel back out to sea. But in reality the boundary is sloped. In short, the very
long wavelength allows the wave to travel intact all the way across the ocean, and the
sloped shore causes the amplitude to grow once the wave arrives.
What is the speed of a tsunami wave in deep √ water? The average depth of the Pacific
Ocean is about 4000 m, so we have v = gH ≈ 200 m/s ≈ 450 mph, which is quite
fast. It takes only a little over a minute for all of the 10-mile wave to hit the shore. So
the energy is deposited in a short amount of time. It isn’t diluted over a large time as
it was with the half-mile wave above. Note that in contrast with the dramatic effects
at the shore, the wave is quite unremarkable far out to sea. It rises to a height of
5 ft over the course of many miles, so the slope at any point is extremely small. It is
impossible to spot such a wave visually, but fortunately deep-sea pressure sensors on
the ocean floor can measure changes in the water level with extreme precision.

6.3.3 Faster than c?

Group velocity
In the first derivation of the group velocity in Section 6.3.1, we found the velocity of the
envelope to be vg = (ω2 − ω1 )/(k2 − k1 ). If ω1 6= ω2 , and if the k values are close together,
then vg is large. In fact, vg can be made arbitrarily large by making k2 be arbitrarily close
to k1 . This means that it is possible for vg to exceed the speed of light. Is this bad? Does it
mean that we can send a signal faster than the speed of light (commonly denoted by “c”),
which would violate the theory of relativity? Answers: No, No.
In order for this scenario to be possible, the two individual waves we used in the derivation
of Eq. (28) already needed to be in existence over a very wide range of x values. So the
envelope in Fig. 16 is going to travel with velocity vg , independent of what anyone does.
Two people therefore can’t use this effect to send information. To communicate something,
you need to change the wave, and it can be demonstrated that the leading edge of this
change can never travel faster than c.
Demonstrating this fact requires invoking some facts about relativity. It certainly can’t
be demonstrated without invoking anything about relativity, because there is nothing at all
special about the speed of light if you haven’t invoked the postulates of relativity. One line
of reasoning is to say that if the leading edge travels faster than c, then there exists a frame
in which causality is violated. This in turn violates innumerable laws of physics (F = ma
type laws, for example). The fact of the matter is that a line of atoms can’t talk to each
other faster than c, independent of whether they’re part of a wave.
At any rate, the point of the present discussion isn’t so much to say what’s right about
the relativistic reasoning (because we haven’t introduced relativity in this book), but rather
to say what’s wrong about the reasoning that vg > c implies a contradiction with relativity.
As we stated above, the error is that no information is contained in a wave that is formed
from the superposition of two waves that already existed over a wide range of x values. If
you had set up these waves, you must have set them up a while ago.
Another way to have vg > c is shown in Fig. 23. Because the bump creeps forward, it
might be possible for vg to exceed c even if the speed of the leading edge doesn’t. But that’s
fine. The leading edge is what is telling someone that something happened, and this speed
never exceeds c.
22 CHAPTER 6. DISPERSION

Figure 23

Phase velocity
The phase velocity can also exceed c. For the string/spring system in Section 6.2.2, we
derived the dispersion relation, ω 2 = v 2 k 2 + ωs2 , where v 2 ≡ T /µ and ωs2 ≡ σ/µ. (We’re
using v here, to save c for the speed of light.) The phase velocity, vp = ω/k, is the slope of
the line from the origin to a point on the ω(k) curve. By making k as small as we want, we
can make the slope be as large as we want, as we noted in Fig. 13. Is this bad? No. Again,
we need to make a change in the wave if we want to convey information. And any signal
can travel only as fast as the leading edge of the change.
It is quite easy to create a system whose phase velocity vp is arbitrarily large. Just put a
bunch of people in a long line, and have them stand up and sit down at prearranged times.
If the person a zillion meters away stands up 1 second after the person at the front of the
line stands up, then the phase velocity is vp = 1 zillion m/s. But no information is contained
in this “wave” because the actions of the people were already decided.
This is sort of like scissors. If you have a huge pair of scissors held at a very small angle,
and if you close them, then it seems like the intersection point can travel faster than c. You
might argue that this doesn’t cause a conflict with relativity, because there is no actual
object in this system that is traveling faster than c (the intersection point isn’t an actual
object). However, although it is correct that there isn’t a conflict, this reasoning isn’t valid.
Information need not be carried by an actual object.
The correct reasoning is that the intersection point will travel faster than c only if you
prearrange for the blades to move at a given instant very far away. If you were to simply
apply forces at the handles, then the parts of the blades very far away wouldn’t know right
away that they should start moving. So the blades would bend, even if they were made out
of the most rigid material possible.
Said in another way, when we guess a solution of the form ei(ωt−kx) in our various
wave equations, it is assumed that this is the solution for all space and time. These waves
always were there, and they always will be there, so they don’t convey any information by
themselves. We have to make a change in them to send a signal. And the leading edge of
the change can travel no faster than c.
Chapter 7

2D waves and other topics

David Morin, [email protected]

This chapter is fairly short. In Section 7.1 we derive the wave equation for two-dimensional
waves, and we discuss the patterns that arise with vibrating membranes and plates. In
Section 7.2 we discuss the Doppler effect, which is relevant when the source of the wave
and/or the observer are/is moving through the medium in which the wave is traveling.

7.1 2D waves on a membrane

We studied transverse waves on a one-dimensional string in Chapter 4. Let’s now look
at transverse waves on a two-dimensional membrane, for example a soap film with a wire
boundary. Let the equilibrium position of the membrane be the x-y plane. So z is the
transverse direction (we’ll use z here instead of our customary ψ). Consider a little rect-
angle in the x-y plane with sides ∆x and ∆y. During the wave motion, the patch of the
membrane corresponding to this rectangle will be displaced in the z direction. But since
we are assuming (as always) that the transverse displacement is small (more precisely, that
the slope of the membrane is small), this patch is still approximately a rectangle. But it is
slightly curved, and it is this curvature that causes there to be a net transverse force, just
as was the case for the 1-D string. The only difference, as we’ll shortly see, is that we have
“double” the effect because the membrane is two dimensional.
As with the 1-D string, the smallness of the slope implies that all points in the membrane
move essentially only in the transverse direction; there is no motion parallel to the x-y plane.
This implies that the mass of the slightly-tilted patch is always essentially equal to σ∆x∆y,
where σ is the mass density per unit area. z
Let the surface tension be S. The units of surface tension are force/length. If you draw y
a line segment of length d` on the membrane, then the force that the membrane on one side x
of the line exerts on the membrane on the other side is S d`. So the forces on the sides of S ∆y
the little patch are S ∆x and S ∆y. A view of the patch, looking along the y direction, is
shown in Fig. 1. This profile looks exactly like the picture we had in the 1-D case (see Fig.
4.2). So just as in that case, the difference in the slope at the two ends is what causes there
to be a net force in the z direction (at least as far as the sides at x and x + ∆x go; there is S ∆y
a similar net force from the other two sides).1 The net force in the z direction due to the
1 As you look along the y axis, the patch won’t look exactly like the 1-D curved segment shown in Fig. x x + ∆x
1, because in general there is curvature in the y direction too. But this fact won’t change our results.
Figure 1
1
2 CHAPTER 7. 2D WAVES AND OTHER TOPICS

forces shown in Fig. 1 is

¡ ¢
F = S ∆y z 0 (x + ∆x) − z 0 (x)
z 0 (x + ∆x) − z 0 (x)
= S ∆y∆x
∆x
∂2z
≈ S ∆y∆x 2 . (1)
∂x
We haven’t bothered writing the arguments of the function z(x, y, t).
We can do the same thing by looking at the profile along the x direction, and we find
that the net force from the two sides at y and y + ∆y is S ∆x∆y(∂ 2 z/∂x2 ). The total
transverse force is the sum of these two results. And since the mass of the patch is σ∆x∆y,
the transverse F = ma equation (or rather the ma = F equation) for the patch is
µ 2 ¶
∂2z ∂ z ∂2z
(σ∆x∆y) 2 = S ∆x∆y + 2
∂t ∂x2 ∂y
2
µ 2 2
¶
∂ z S ∂ z ∂ z
=⇒ 2
= 2
+ 2 (wave equation) (2)
∂t σ ∂x ∂y

This looks quite similar to our old 1-D wave equation in Eq. (4.4), except that we now
have partial derivatives with respect to two spatial coordinates. How do we solve this
equation for the function z(x, y, t)? We know that any function can be written in terms of its
Fourier components. Since we have three independent variables, the Fourier decomposition
of z(x, y, t) consists of the triple integral,
Z ∞Z ∞Z ∞
z(x, y, t) = g(kx , ky , ω)ei(kx x+ky y+ωt) dkx dky dω. (3)
−∞ −∞ −∞

This comes about in the same way that the double integral came about in Eq. (4.16).
And since the wave equation in Eq. (2) is linear, it suffices to guess solutions of the form
ei(kx x+ky y+ωt) . (Both Fourier analysis and linearity are necessary for this conclusion to
hold). Plugging this guess into Eq. (2) and canceling a minus sign gives

S¡ 2 ¢
ω2 = k + ky2 (dispersion relation) (4)
σ x

This looks basically the same as the 1-D dispersion relation for transverse waves on a string,
ω 2 = c2 k 2 , but with the simple addition of a second k 2 term. However, this seemingly minor
modification has a huge consequence: In the 1-D case, only one k value corresponded to a
given ω value. But in the 2-D case, an infinite number of kx and kypvalues correspond to a
given ω value, namely all the (kx , ky ) points on a circle of radius ω σ/S.
Let’s now look at some boundary conditions. Things get very complicated with arbitrarily-
shaped boundaries, so let’s consider the case of a rectangular boundary. We can imagine
having a soap film stretched across a rectangular wire boundary. Let the sides be parallel
to the coordinate axes and have lengths Lx and Ly , and let one corner be located at the
origin. The boundary condition for the membrane is that z = 0 on the boundary, because
the membrane must be in contact with the wire. Let’s switch from exponential solutions to
trig solutions, which work much better here. We can write the trig solutions in many ways,
but we’ll choose the basis where z(x, y, t) takes the form,

z(x, y, t) = A trig(kx x) trig(ky y) trig(ωt), (5)

7.1. 2D WAVES ON A MEMBRANE 3

where “trig” means either sine or cosine. Similar to the 1-D case, the x = 0 and y = 0
boundaries tell us that we can’t have any cosine functions of x and y. So the solution must
take the form,
z(x, y, t) = A sin(kx x) sin(ky y) cos(ωt + φ). (6)
And again similar to the 1-D case, the boundary conditions at x = Lx and y = Ly restrict
kx and ky to satisfy
nπ nπ
kx x = nπ, ky y = mπ =⇒ kx = , kx = . (7)
Lx Lx
The most general solution for z is an arbitrary sum of these basis solutions, so we have
X µ ¶ µ ¶
nπx mπy
z(x, y, t) = An,m sin sin cos(ωn,m t + φn,m ),
n,m
Lx Ly
"µ ¶ µ ¶2 #
2
2 S nπ mπ
where ωn,m = + . (8)
σ Lx Ly

Each basis solution (that is, each normal mode) in this sum is a standing wave. The
constants An,m and φn,m are determined by the initial conditions. If n or m is zero, the z is
identically zero, so n and m each effectively start at 1. Note that if we have a square with
Lx = Ly ≡ L, then pairs of integers (n, m) yield identical frequencies if n21 + m21 = n22 + m22 .
A trivial case is where we simply switch the numbers, such as (1, 3) and (3, 1). But we can
also have, for example, (1, 7), (7, 1), and (5, 5).
What do these modes look like? In the case of a transverse wave on a 1-D string, it was
easy to draw a snapshot on a piece of paper. But it’s harder to do that in the present case,
because the wave takes up three dimensions. We could take a photograph of an actual 3-D
wave and then put the photograph on this page, or we could draw the wave with the aid of a
computer or with fantastic artistic skills. But let’s go a little more low-tech and low-talent.
We’ll draw the membrane in a simple binary sense, indicating only whether the z value is
positive or negative. The nodes (where z is always zero) will be indicted by dotted lines. If
we pick Lx 6= Ly to be general, then the lowest few values of n and m yield the diagrams
shown in Fig. 2. n signifies the number of (equal) regions the x direction is broken up into.
And likewise for m and the y direction.

(n,m)=(1,1) (1,2) (2,1) (2,2)

(1,3) (3,1) (2,3) (3,2) (3,3)

Figure 2
4 CHAPTER 7. 2D WAVES AND OTHER TOPICS

For each of these snapshots, a little while later when cos(ωn,m t+φn,m ) = 0, the transverse
displacement z will be zero everywhere, so the entire membrane will lie in the x-y plane.
For the next half cycle after this time, all the +’s and −’s in each figure will be reversed.
They will flip flop back and forth after each half cycle. Observe that the signs are opposite
for any two regions on either side of a dotted line, consistent with the fact that z is always
zero on the dotted-line nodes.

The solution for, say, the (3, 2) mode is

µ ¶ µ ¶
3πx 2πy
z(x, y, t) = A3,2 sin sin cos(ω3,2 t + φ3,2 ), (9)
Lx Ly

where ω3,2 is given by Eq. (8). The first sine factor here is zero for x = 0, Lx /3, 2Lx /3,
and Lx . And the second sine factor is zero for y = 0, Ly /2, and Ly . These agree with the
dotted nodes in the (3, 2) picture in Fig. 2. In each direction, the dotted lines are equally
spaced.

Note that the various An,m frequencies are not simple multiples of each other, as they are
for a vibrating string with two fixed ends (see Section 4.5.2). For example, if Lx = Ly ≡ L,
then the frequencies in Eq. (8) take the form,

r
π Sp 2
n + m2 . (10)
L σ

√ √ √ √
So the first few frequencies are ω1,1 ∝ 2, ω2,1 ∝ 5, ω2,2 ∝ 8, ω3,1 ∝ 10, and so on.
Some of these are simple multiples of each other, such as ω2,2 = 2ω1,1 , but in general the
ratios are irrational. So there are lots of messy harmonics. That’s why musical instruments
are usually one-dimensional objects. The frequencies of their modes form a nice linear
progression (or are in rational multiples of you include the effects of pressing down keys or
valves).

The other soap-film boundary that is reasonably easy to deal with is a circle. In this
case, it is advantageous to write the partial derivative in terms of polar coordinates. It can
be shown that (see Problem [to be added])

∂2 ∂2 ∂2 1 ∂ 1 ∂2
+ = + + . (11)
∂x2 ∂y 2 ∂r2 r ∂r r2 ∂θ2

When this is substituted into Eq. (2), the solutions z(r, θ, t) aren’t as simple as in the
Cartesian case, but it’s still possible to get a handle on them. They involve the so-called
Bessel functions. The pictures analogous to the ones in Fig. 2 are shown in Fig. 3. These
again can be described in terms of two numbers. In the rectangular case, the nodal lines
divided each direction evenly. But here the nodal lines are equally spaced in the θ direction,
but not in the r direction.
7.2. DOPPLER EFFECT 5

(nr,nθ)=(1,0) (1,1) (1,2)

(2,0) (2,1) (2,2)

Figure 3

Chladni plates
Consider a metal plate that is caused to vibrate by, say, running a violin bow across its edge.
If this is done properly (it takes some practice), then it is possible to excite a single mode
with a particular frequency. Another method, which is more high-tech and failsafe (but
less refined), is to blast the plate with a sound wave of a given frequency. If the frequency
matches the frequency of one of the modes, then we will have a resonance effect, in the
same manner that we obtained resonances in the coupled oscillator in Section 2.1.5. If the
plate is sprinkled with some fine sand, the sand will settle at the nodal lines (or curves),
because the plate is moving at all the other non-node points, and this motion kicks the sand
off those locations. Since the sand isn’t kicked off the nodes, that’s where it settles. This is
basically the same reason why sand collects at the side of a road and not on it. Wind from
the cars pushes the sand off the road, and there’s no force pushing it back on. The sand is
on a one-way dead-end street, so to speak.
The nodal curves (which are different for the different modes) generally take on very
interesting shapes, so we get all sorts of cool figures with the sand. Ernst Chladni (1756-
1827) studied these figures in great detail. They depend on the shape of the metal plate and
the mode that the plate is in. They also depend on the boundary conditions you choose. For
example, you can hold the plate somewhere in the interior, or on the edge. And furthermore
you can choose to hold it at any number of places. And furthermore still, there are different
ways to hold it; you can have a clamp or a hinge. Or you can even support the plate with
a string. These all give different boundary conditions. We won’t get into the details, but
note that if you grab the plate at a given point with a clamp or a hinge, you create a node
there. (See Problem [to be added] if you do want to get into the details.) [Pictures will be
added.]

7.2 Doppler effect

7.2.1 Derivation
When we talk about the frequency of a wave, we normally mean the frequency as measured
in the frame in which the air (or whatever medium is relevant) is at rest. And we also
normally assume that the source is at rest in the medium. But what if the source or the
6 CHAPTER 7. 2D WAVES AND OTHER TOPICS

observer is/are moving with respect to the air? (We’ll work in terms of sound waves here.)
What frequency does the observer hear then? We’ll find that it is modified, and this effect
is known as the Doppler effect. In everyday experience, the Doppler effect is most widely
observed with sound waves. However, it is relevant to any wave, and in particular there are
important applications with electromagnetic waves (light).
Let’s look at the two basic cases of a moving source and a moving observer. In both of
these cases, we’ll do all of the calculations in the frame of the ground, or more precisely, the
frame in which the air is at rest (on average; the molecules are of course oscillating back
and forth longitudinally).

Moving source
Assume that you are standing at rest on a windless day, and a car with a sound source (say,
a siren) on it is heading straight toward you with speed vs (“s” for source). The source
emits sound, that is, pressure waves. Let the frequency (in the source’s frame) be f cycles
per second (Hertz). Let’s look at two successive maxima of the pressure (actually any two
points whose phases differ by 2π would suffice). In the time between the instants when the
source is producing these maximum pressures, the source will travel a distance vs t, where
t = 1/f . Also during this time t, the first of the pressure maxima will travel a distance ct,
where c is the wave speed. When the second pressure maximum is produced, it is therefore
a distance of only d = ct − vs t behind the first maximum, instead of the ct distance if the
source were at rest. The wavelength is therefore smaller. The situation is summarized in
Fig. 4.

source
(start)

phase φ

vs
(time t) c

vst phase φ-2π observer

(at rest)
ct

Figure 4

The movement of the source doesn’t affect the wave speed, because the speed is a function
of only the quantities γ, p0 , and ρ (see Eq. (5.14)); the derivation in Section 5.2 assumed
nothing about the movement of the source. So the time between the arrivals at your ear
of the two successive pressure maxima is T = d/c = (c − vs )t/c. The frequency that you
observe is therefore (the subscript “ms” is for moving source)

1 c 1 c
fms = = · = f (12)
T c − vs t c − vs

This result is valid for vs < c. We’ll talk about the vs ≥ c case in Section 7.2.3.
If vs = 0, then Eq. (12) yields fms = f , of course. And if vs → c, then fms approaches
infinity. This makes sense, because the pressure maxima are separated by essentially zero
distance in this case (the wavelength is very small), so they pile up and a large number hit
7.2. DOPPLER EFFECT 7

your ear in a given time interval.2 Eq. (12) is also valid for negative vs , and we see that
fms → 0 as vs → −∞. This also makes sense, because the pressure maxima are very far
apart.

Moving observer
Let’s now have the source be stationary but the observer (you) be moving directly toward
the source with speed vo (“o” for observer). Consider two successive meetings between you
and the pressure maxima. As shown in Fig. 5, the distance between successive maxima (in
the ground frame) is simply ct, where t = 1/f .

ct
c c

(start) vo

phase φ
phase φ-2π

(time t)

source
(at rest)

Figure 5

This gap is closed at a rate of c + vo , because you are heading to the left with speed
vo and the pressure maxima are heading to the right with speed c. So the time between
successive meetings is T = ct/(c + vo ). The frequency that you observe is therefore (the
subscript “mo” is for moving observer)

1 c + vo 1 c + vo
fmo = = · = f (13)
T c t c

This result is valid for vo > −c (where negative velocities correspond to moving to the
right). If vo < −c, then you are moving to the right faster than the pressure maxima, so
they can never catch up to you as you recede away. In the cutoff case where vo = −c, we
have fmo = 0. This makes sense, because you are receding away from the pressure maxima
as fast as they are moving.
If vo = 0, then Eq. (13) yields fmo = f , of course. And if vo = c, then fmo equals 2f ,
so it doesn’t diverge as in the moving-source case. The pressure maxima are the “normal”
distance of ct apart, but the gap is being closed at twice the normal rate. If vo → ∞, then
Eq. (13) yields fmo → ∞, which makes sense. You encounter a large number of pressure
maxima in a given time simply because you are moving so fast.

2 However, if v is too close to c, then the wavelength becomes short enough to make it roughly the same
s
size as the amplitude of the displacement wave. Our assumption of small slope (which we used in Section
5.2) then breaks down, and we can’t trust any of these results. Nonlinear effects become important, but we
won’t get into those here.
8 CHAPTER 7. 2D WAVES AND OTHER TOPICS

Remarks:
1. If the observed frequency is less than f , then we say that the sound (or whatever wave) is
redshifted. If it is greater than f , then we say it is blueshifted. This terminology comes from
the fact that red light is at the low-frequency (long wavelength) end of the visible spectrum,
and blue light is at the high-frequency (short wavelength) end. This terminology is carried
over to other kinds of waves, even though there is of course nothing red or blue about, say,
sound waves. Well, unless someone is yelling “red” or “blue,” I suppose.
2. The results in Eqs. (12) and (13) don’t reduce to the same thing when vs = vo , even though
these two cases yield the same relative speed between the source and observer. This is because
the situation is not symmetrical in vs and vo ; there is a preferred frame, namely the frame in
which the air is at rest. The speed with respect to this frame matters, and not just the relative
speed between the observer and the source. It makes sense that things aren’t symmetrical,
because a speed of v = c intuitively should make fms equal to infinity, but not fmo .
3. What if both you and the source are moving toward each other with speeds vs and vo ?
Imagine a hypothetical stationary observer located somewhere between you and the source.
This observer will hear the frequency fms given in Eq. (12). For all you know, you are listening
to this stationary observer emit a sound with frequency fms , instead of the original moving
source with frequency f . (The stationary observer can emit a wave exactly in phase with the
one he hears. Or equivalently, he can duck and have the wave go right past him.) So the
frequency you hear is obtained by letting the f in Eq. (13) equal fms . The frequency when
both the source and observer are moving is therefore
³ ´³ ´
c + vo c c + vo
Fmso = = f (14)
c c − vs c − vs

4. For small vs , we can use 1/(1 − ²) ≈ 1 + ² to write the result in Eq. (12) as
³ ´
1 vs
fms = f ≈ 1+ f. (15)
1 − vs /c c
And the result in Eq. (13) can be written (exactly) as
³ ´
vo
fmo = 1 + f. (16)
c
So the two results take approximately the same form for small speeds.
5. If the source isn’t moving in a line directly toward or away from you (or vice versa), then
things are a little more complicated, but not too bad (see Problem [to be added]). The
frequency changes continuously from f c/(c − vs ) at t = −∞ to f c/(c + vs ) at t = +∞ (the
same formula with vs → −vs ). So it slides from one value to the other. You’ve undoubtedly
heard a siren doing this. If the source instead hypothetically moved in a line right through
you, then it would abruptly drop from the higher to the lower of these frequencies. So, in
the words of John Dobson, “The reason the siren slides is because it doesn’t hit you.”
6. Consider a wall (or a car, or whatever) moving with speed v toward a stationary sound
source. If the source emits a frequency f , and if the wall reflects the sound back toward the
source, what reflected frequency is observed by someone standing next to the source? The
reflection is a two-step process. First the wall acts like an observer, so from Eq. (13) it receives
a frequency of f (c + v)/c. But then it acts like a source and emits whatever frequency it
receives (imagine balls bouncing off a wall). So from Eq. (12) the observer hears a frequency
of c/(c − v) · f (c + v)/c = f (c + v)/(c − v). The task of Problem [to be added] is to find the
observed frequency if the observer (and source) is additionally moving with speed u toward
the wall (which is still moving with speed v). ♣

Some examples and applications of the Doppler effect are:

Sirens on ambulances, police cars, etc: As the sirens move past you, the pitch goes
from high to low. However, with fire trucks, most of the change in pitch of the siren is due
7.2. DOPPLER EFFECT 9

to a different effect. These sirens generally start at a given pitch and then gradually get
lower, independent of the truck’s movement. This is because most fire trucks use siren disks
to generate the sound. A siren disk is a disk with holes in it that is spun quickly in front of
a fast jet of air. The result is high pressure or low pressure, depending on whether the air
goes through a hole, or gets blocked by the disk. If the frequency with which the holes pass
in front of the jet is in the audible range, then a sound is heard. (The wave won’t be an
exact sine wave, but that doesn’t matter.) The spinning disk, however, usually moves due
to an initial kick and not a sustained motor, so it gradually slows down. Hence the gradual
decrease in pitch.
Dopper radar: Light waves reflect off a moving object, and the change in frequency is
observed (see Remark 6 above). This has applications in speed guns, weather, and medicine.
Along the same lines is Doppler sonar, in particular underwater with submarines.
Astronomy: Applications include the speed of stars and galaxies, the expansion of the
universe, and the determination of binary star systems. The spectral lines of atomic tran-
sitions are shifted due to the motion of the star or galaxy. These applications rely on the
(quite reasonable) assumption that the frequencies associated with atomic transitions are
independent of their location in the universe. That is, a hydrogen atom in a distant galaxy
is identical to a hydrogen atom here on earth. Hard to prove, of course, but a reasonable
thing to assume.
Temperature determination of stars, plasma: This makes use of the fact that not
only do spectral lines shift, they also broaden due to the large range of velocities of the
atoms in a star (the larger the temperature, the larger the range).

7.2.2 Relativity
The difference in the results in Eqs. (12) and (13) presents an issue in the context of relativity.
If a source is moving toward you with speed v and emits a certain frequency of light, then
the frequency you observe must be the same as it would be if instead you were moving
toward the source with speed v. This is true because one of the postulates of relativity is
that there is no preferred reference frame. All that matters is the relative speed.
It is critical that we’re talking about a light wave here, because light requires no medium
to propagate in. (Gravity waves would work too, since they can propagate in vacuum.) If
we were talking about a sound wave, then the air would define a preferred reference frame,
thereby allowing the two frequencies to be different, as is the case in Eqs. (12) and (13).
So which of the above results is correct for light waves? Well, actually they’re both wrong.
We derived them using nonrelativistic physics, so they work fine for everyday speeds. But
they are both invalid for relativistic speeds. Let’s see how we can correct each of them.
Let’s label the vs and vo in the above results as v.
In the “moving
p source” setup, the frequency of the source in your frame is now f /γ
(where γ = 1/ 1 − v 2 /c2 ), because the source’s clock runs slow in your frame, due to time
dilation. f /γ is the frequency in your frame with which the phase of the light wave passes
through a given value, say zero, as it leaves the source. But as in the nonrelativistic case,
this isn’t the frequency that you observe, due to the fact that the “wavefronts” (locations
of equal phase) end up closer together. This part of the calculation proceeds just as above,
so the only difference is that the emission frequency f is changed to f /γ. From Eq. (12),
you therefore observe a frequency of (f /γ)c/(c − v).
In the “moving observer” setup, the frequency we calculated in the nonrelativistic case
was the frequency as measured in the source’s frame. But your clock runs slow in the
source’s frame, due to time dilation. The frequency that you observe is therefore larger by
 10 CHAPTER 7. 2D WAVES AND OTHER TOPICS

a factor of γ. (It is larger because more wavefronts will hit your eye during the time that
one second elapses on your clock, because your clock is running slow.) From Eq. (13), you
therefore observe a frequency of γ · f (c + v)/c.
As we argued above, these two results must be equal. And indeed they are, because

f c c+v c2 1
= γf ⇐⇒ = γ2 ⇐⇒ = γ2, (17)
γ c−v c c2 − v 2 1 − v 2 /c2

which is true, as we wanted to show.

7.2.3 Shock waves

source earlier Let’s return to the world of nonrelativistic physics. In the “moving source” setup above, we
source now noted that the result isn’t valid if vs > c. So what happens in this case? Since the source is
moving faster than the sound (or whatever) wave, the source gets to the observer before the
previously-emitted wavefronts get there. If we draw a number of wavefronts (places with,
say, maximal pressure) that were emitted at various times, we obtain the picture shown in
Fig. 6.
The cone is a “shock wave” where the phases of waves emitted at different times are equal.
Figure 6 This causes constructive interference. (In the case where vs < c, the different wavefronts
never interact with each other, so there is never any constructive interference; see Fig. 8
below.) The amplitude of the wave on the surface of the cone is therefore very large. So
someone standing off to the side will hear a loud “sonic boom” when the surface of the cone
ct passes by.
θ
vt We can find the half angle of the cone in the following way. Fig. 7 shows the circular
wavefront that was emitted at time t ago, along with the original and present locations of
the source. The source travels a distance vt in this time, and the sound travels a distance
ct. So the half angle of the cone satisfies
Figure 7
c
sin θ = . (18)
v

The total angle is therefore 2θ = 2 sin−1 (c/v). This result is valid only for v ≥ c. The larger
v is, the narrower the cone. If v → ∞ then θ → 0. And if v = c then 2θ = 180◦ , so the
“cone” is very wide, to the point of being just a straight line.
A summary of the various cases of the relative size of v and c is shown in Fig. 8. In the
v = c case, it is intuitively clear that the waves pile up at the location of the moving source,
because the waves are never able to gain any ground on the source. In the v > c case,
the cone actually arises from this same effect (although to a lesser extent) for the following
reason. If v > c, there is a particular moment in time when the distance between the source
and the observer is decreasing at speed c. (This follows from continuity; the rate of decrease
is v at infinity, and zero at closest approach.) The transverse component of the source’s
velocity isn’t important for the present purposes, so at this moment the source is effectively
moving directly toward the observer with speed c. The reasoning in the v = c case then
applies. The task of Problem [to be added] is to be quantitative about this.
7.2. DOPPLER EFFECT 11

(v<c) (v=c) (v>c)

Figure 8

Shock waves exist whenever the speed of the source exceeds the speed of wave (whatever
it may be) in the medium through which the source is moving. Examples include (1) planes
exceeding the speed of sound, (2) boats exceeding the speed of water waves; however, this
subject is more complicated due to the dispersive nature of water waves – we’ll talk about
this in Chapter 11, (3) charged particles moving through a material faster than the speed
of light in that material (which equals c/n, where n is the index of refraction); this is called
“Cherenkov radiation,” and (4) the crack of a whip.
This last example is particularly interesting, because the thing that makes it possible
for the tip of a whip to travel faster than the speed of sound is impedance matching; see
the “Gradually changing string density” example in Section 4.3.2. Due to this impedance
matching, a significant amount of the initial energy that you give to the whip ends up in
the tip. And since the tip is very light, it must therefore be moving very fast. If the linear
mass density of the whip changed abruptly, then not much of the initial energy would be
transmitted across the boundary. The snap of a wet towel is also the same effect; see The
Physics Teacher, pp. 376-377 (1993).
Chapter 8

Electromagnetic waves
David Morin, [email protected]

The waves we’ve dealt with so far in this book have been fairly easy to visualize. Waves
involving springs/masses, strings, and air molecules are things we can apply our intuition to.
But we’ll now switch gears and talk about electromagnetic waves. These are harder to get
a handle on, for a number of reasons. First, the things that are oscillating are electric and
magnetic fields, which are much harder to see (which is an ironic statement, considering that
we see with light, which is an electromagnetic wave). Second, the fields can have components
in various directions, and there can be relative phases between these components (this will
be important when we discuss polarization). And third, unlike all the other waves we’ve
dealt with, electromagnetic waves don’t need a medium to propagate in. They work just
fine in vacuum. In the late 1800’s, it was generally assumed that electromagnetic waves
required a medium, and this hypothesized medium was called the “ether.” However, no one
was ever able to observe the ether. And for good reason, because it doesn’t exist.
This chapter is a bit long. The outline is as follows. In Section 8.1 we talk about waves in
an extended LC circuit, which is basically what a coaxial cable is. We find that the system
supports waves, and that these waves travel at the speed of light. This section serves as
motivation for the fact that light is an electromagnetic wave. In Section 8.2 we show how
the wave equation for electromagnetic waves follows from Maxwell’s equations. Maxwell’s
equations govern all of electricity and magnetism, so it is no surprise that they yield the
wave equation. In Section 8.3 we see how Maxwell’s equations constrain the form of the
waves. There is more information contained in Maxwell’s equations than there is in the
wave equation. In Section 8.4 we talk about the energy contained in an electromagnetic
wave, and in particular the energy flow which is described by the Poynting vetor. In Section
8.5 we talk about the momentum of an electromagnetic wave. We saw in Section 4.4 that
the waves we’ve discussed so far carry energy but not momentum. Electromagnetic waves
carry both.1 In Section 8.6 we discuss polarization, which deals with the relative phases
of the different components of the electric (and magnetic) field. In Section 8.7 we show
how an electromagnetic wave can be produced by an oscillating (and hence accelerating)
charge. Finally, in Section 8.8 we discuss the reflection and transmission that occurs when
an electromagnetic wave encounters the boundary between two different regions, such as air
1 Technically, all waves carry momentum, but this momentum is suppressed by a factor of v/c, where v is

the speed of the wave and c is the speed of light. This follows from the relativity fact that energy is equivalent
to mass. So a flow of energy implies a flow of mass, which in turn implies nonzero momentum. However,
the factor of v/c causes the momentum to be negligible unless we’re dealing with relativistic speeds.

1
2 CHAPTER 8. ELECTROMAGNETIC WAVES

and glass. We deal with both normal and non-normal angles of incidence. The latter is a
bit more involved due to the effects of polarization.

8.1 Cable waves

Before getting into Maxwell’s equations and the wave equation for light, let’s do a warmup
example and study the electromagnetic waves that propagate down a coaxial cable. This
example should help convince you that light is in fact an electromagnetic wave.
To get a handle on the coaxial cable, let’s first look at the idealized circuit shown in
Fig. 1. All the inductors are L, and all the capacitors are C. There are no resistors in the
circuit. With the charges, currents, and voltages labeled as shown, we have three facts:
1. The charge on a capacitor is Q = CV =⇒ qn = CVn .
2. The voltage across an inductor is V = L(dI/dt) =⇒ Vn−1 − Vn = L(dIn /dt).
3. Conservation of charge gives In − In+1 = dqn /dt.

In-1 In In+1
Vn-2 Vn-1 Vn Vn+1

qn-2 qn-1 qn qn+1 all inductors = L

-qn-2 -qn-1 -qn -qn+1 all capacitors = C

(ground)
V=0

Figure 1
Our goal is to produce an equation, which will end up being a wave equation, for one of
the three variables, q, I, and V (the wave equations for all of them will turn out to be the
same). Let’s eliminate q and I, in favor of V . We could manipulate the above equations in
their present form in terms of discrete quantities, and then take the continuum limit (see
Problem [to be added]). But it is much simpler to first take the continuum limit and then
do the manipulation. If we let the grid size in Fig. 1 be ∆x, then by using the definition of
the derivative, the above three facts become
q = CV,
∂V ∂I
−∆x = L ,
∂x ∂t
∂I ∂q
−∆x = . (1)
∂x ∂t
Substituting q = CV from the first equation into the third, and defining the inductance and
capacitance per unit length as L0 ≡ L/∆x and C0 ≡ C/∆x, the last two equations become
∂V ∂I ∂I ∂V
− = L0 , and − = C0 . (2)
∂x ∂t ∂x ∂t
If we take ∂/∂x of the first of these equations and ∂/∂t of the second, and then equate the
results for ∂ 2 I/∂x ∂t, we obtain

1 ∂2V ∂2V ∂ 2 V (x, t) 1 ∂ 2 V (x, t)

− = −C 0 =⇒ = (3)
L0 ∂x2 ∂t2 ∂t2 L0 C0 ∂x2
8.2. THE WAVE EQUATION 3

This is the desired wave equation, and it happens to be dispersionless. We can quickly read
off the speed of the waves, which is
1
v=√ . (4)
L0 C0
If we were to subdivide the circuit in Fig. 1 into smaller and smaller cells, L and C would
depend on ∆x (and would go to zero as ∆x → 0), so it makes sense to work with the
quantities L0 and C0 . This is especially true in the case of the actual cable we’ll discuss
below, for which the choice of ∆x is arbitrary. L0 and C0 are the meaningful quantities that
are determined by the nature of the cable.
Note that since the first fact above says that q ∝ V , the exact same wave equation holds
for q. Furthermore, if we had eliminated V instead of I in Eq. (2) by taking ∂/∂t of the
first equation and ∂/∂x of the second, we would have obtained the same wave equation for
Figure 2
I, too. So V , q, and I all satisfy the same wave equation.
Let’s now look at an actual coaxial cable. Consider a conducting wire inside a conducting
cylinder, with vacuum in the region between them, as shown in Fig. 2. Assume that the wire
is somehow constrained to be in the middle of the cylinder. (In reality, the inbetween region
is filled with an insulator which keeps the wire in place, but let’s keep things simple here
with a vacuum.) The cable has an inductance L0 per unit length, in the same way that two
parallel wires have a mutual inductance per unit length. (The cylinder can be considered
to be made up of a large number wires parallel to its axis.) It also has a capacitance C0
per unit length, because a charge difference between the wire and the cylinder will create a
voltage difference.
It can be shown that (see Problem [to be added], although it’s perfectly fine to just
accept this)
µ0 2π²0
L0 = ln(r2 /r1 ) and C0 = , (5)
2π ln(r2 /r1 )
where r2 is the radius of the cylinder, and r1 is the radius of the wire. The two physical
constants in these equations are the permeability of free space, µ0 , and the permittivity of
free space, ²0 . Their values are (µ takes on this value by definition)
µ0 = 4π · 10−7 H/m , and ²0 ≈ 8.85 · 10−12 F/m . (6)
H and F are the Henry and Farad units of inductance and capacitance. Using Eq. (5), the
wave speed in Eq. (4) equals
1 1 1
v=√ =√ ≈p ≈ 3 · 108 m/s. (7)
L0 C0 µ ²
0 0 (4π · 10 H/m)(8.85 · 10−12 F/m)
−7

This is the speed of light! We see that the voltage (and charge, and current) wave that travels
down the cable travels at the speed of light. And because there are electric and magnetic
fields in the cable (due to the capacitance and inductance), these fields also undergo wave
motion. Since the waves of these fields travel with the same speed as the original voltage
wave, it is a good bet that electromagnetic waves have something to do with light. The
reasoning here is that there probably aren’t too many things in the world that travel with the
speed of light. So if we find something that travels with this speed, then it’s probably light
(loosely speaking, at least; it need not be in the visible range). Let’s now be rigorous and
show from scratch that all electromagnetic waves travel at the speed of light (in vacuum).

8.2 The wave equation

By “from scratch” we mean by starting with Maxwell’s equations. We have to start some-
where, and Maxwell’s equations govern all of (classical) electricity and magnetism. There
4 CHAPTER 8. ELECTROMAGNETIC WAVES

are four of these equations, although when Maxwell first wrote them down, there were 22 of
them. But they were gradually rewritten in a more compact form over the years. Maxwell’s
equations in vacuum in SI units are (in perhaps overly-general form):
Differential form Integrated
Z form
ρE QE
∇·E= E · dA =
²0 Z ²0
∇ · B = ρB B · dA = QB
Z
∂B dΦB
∇×E=− + JB E · dl = − + IB
∂t Z dt
∂E dΦE
∇ × B = µ0 ²0 + µ0 JE B · dl = µ0 ²0 + µ0 IE
∂t dt
Table 1
If you erase the µ0 ’s and ²0 ’s here (which arise from the arbitrary definitions of the
various units), then these equations are symmetric in E and B, except for a couple minus
signs. The ρ’s are the electric and (hypothetical) magnetic charge densities, and the J’s are
the current densities. The Q’s are the charges enclosed by the surfaces that define the dA
integrals, the Φ’s are the field fluxes through the loops that define the dl integrals, and the
I’s are the currents through these loops.
No one has ever found an isolated magnetic charge (a magnetic monopole), and there
are various theoretical considerations that suggest (but do not yet prove) that magnetic
monopoles can’t exist, at least in our universe. So we’ll set ρB , JB , and IB equal to zero
from here on. This will make Maxwell’s equations appear non-symmetrical, but we’ll soon
be setting the analogous electric quantities equal to zero too, since we’ll be dealing with
vacuum. So in the end, the equations for our purposes will be symmetric (except for the µ0 ,
the ²0 , and a minus sign). Maxwell’s equations with no magnetic charges (or currents) are:
Differential form Integrated
Z form Known as
ρE QE
∇·E= E · dA = Gauss’ Law
²0 Z ²0
∇·B=0 B · dA = 0 No magnetic monopoles
Z
∂B dΦB
∇×E=− E · dl = − Faraday’s Law
∂t Z dt
∂E dΦE
∇ × B = µ0 ²0 + µ0 JE B · dl = µ0 ²0 + µ0 IE Ampere’s Law
∂t dt
Table 2
The last of these, Ampere’s Law, includes the so-called “displacement current,” dΦE /dt.
Our goal is to derive the wave equation for the E and B fields in vacuum. Since there
are no charges of any kind in vacuum, we’ll set ρE and JE = 0 from here on. And we’ll only
need the differential form of the equations, which are now
∇·E = 0, (8)
∇·B = 0, (9)
∂B
∇×E = − , (10)
∂t
∂E
∇×B = µ0 ²0 . (11)
∂t
These equations are symmetric in E and B except for the factor of µ0 ²0 and a minus sign.
Let’s eliminate B in favor of E and see what we get. If we take the curl of Eq. (10) and
8.2. THE WAVE EQUATION 5

then use Eq. (11) to get ride of B, we obtain

∂B
∇ × (∇ × E) = −∇ ×
∂t
∂(∇ × B)
= −
µ∂t ¶
∂ ∂E
= − µ0 ²0
∂t ∂t
2
∂ E
= −µ0 ²0 2 . (12)
∂t
On the left side, we can use the handy “BAC-CAB” formula,
A × (B × C) = B(A · C) − C(A · B). (13)
See Problem [to be added] for a derivation of this. This formula holds even if we have
differential operators (such as ∇) instead of normal vectors, but we have to be careful to
keep the ordering of the letters the same (this is evident if you go through the calculation in
Problem [to be added]). Since both A and B are equal to ∇ in the present application, the
ordering of A and B in the B(A · C) term doesn’t matter. But the C(A · B) must correctly
be written as (A · B)C. The lefthand side of Eq. (12) then becomes
∇ × (∇ × E) = ∇(∇ · E) − (∇ · ∇)E
= 0 − ∇2 E, (14)
where the zero follows from Eq. (8). Plugging this into Eq. (12) finally gives

∂2E ∂2E 1
−∇2 E = −µ0 ²0 =⇒ = ∇2 E (wave equation) (15)
∂t2 ∂t2 µ0 ²0

Note that we didn’t need to use the second of Maxwell’s equations to derive this.
In the above derivation, we could have instead eliminated E in favor of B. The same steps
hold; the minus signs end up canceling again, as you should check, and the first equation is
now not needed. So we end up with exactly the same wave equation for B:

∂2B 1
= ∇2 B (wave equation) (16)
∂t2 µ0 ²0

The speed of the waves (both E and B) is given by the square root of the coefficient on
the righthand side of the wave equation. (This isn’t completely obvious, since we’re now
working in three dimensions instead of one, but we’ll justify this in Section 8.3.1 below.)
The speed is therefore
1
c= √ ≈ 3 · 108 m/s. (17)
µ0 ²0
This agrees with the result in Eq. (7). But we now see that we don’t need a cable to
support the propagation of electromagnetic waves. They can propagate just fine in vacuum!
This is a fundamentally new feature, because every wave we’ve studied so far in this book
(longitudinal spring/mass waves, transverse waves on a string, longitudinal sound waves,
etc.), needs a medium to propagate in/on. But not so with electromagnetic waves.
Eq. (15), and likewise Eq. (16), is a vector equation. So it is actually shorthand for three
separate equations for each of the components:
µ 2 ¶
∂ 2 Ex 1 ∂ Ex ∂ 2 Ex ∂ 2 Ex
= + + , (18)
∂t2 µ0 ²0 ∂x2 ∂y 2 ∂z 2
6 CHAPTER 8. ELECTROMAGNETIC WAVES

and likewise for Ey and Ez . Each component undergoes wave motion. As far as the wave
equation in Eq. (15) is concerned, the waves for the three components are completely inde-
pendent. Their amplitudes, frequencies, and phases need not have anything to do with each
other. However, there is more information contained in Maxwell’s equations than in the
wave equation. The latter follows from the former, but not the other way around. There
is no reason why one equation that follows from four equations (or actually just three of
them) should contain as much information as the original four. In fact, it is highly unlikely.
And as we will see in Section 8.3, Maxwell’s equations do indeed further constrain the form
of the waves. In other words, although the wave equation in Eq. (15) gives us information
about the electric-field wave, it doesn’t give us all the information.

Index of refraction
In a dielectric (equivalently, an insulator), the vacuum quantities µ0 and ²0 in Maxwell’s
equations are replaced by new values, µ and ². (We’ll give some justification of this below,
but see Sections 10.11 and 11.10 in Purcell’s book for the full treatment.) Our derivation
of the wave equation for electromagnetic waves in a dielectric proceeds in exactly the same
way as for the vacuum case above, except with µ0 → µ and ²0 → ². We therefore end up
with a wave velocity equal to
1
v=√ . (19)
µ²
The index of refraction, n, of a dielectric is defined by v ≡ c/n, where c is the speed of light
in vacuum. We therefore have
r
c c µ²
v= =⇒ n = = . (20)
n v µ0 ²0

Since it happens to be the case that that µ ≈ µ0 for most dielectrics, we have the approximate
result that r
²
n≈ (if µ ≈ µ0 ). (21)
²0
And since we must always have v ≤ c, this implies n ≥ 1 =⇒ ² ≥ ²0 .
Strictly speaking, Maxwell’s equations with µ0 and ²0 work in any medium. But the
point is that if we don’t have a vacuum, then induced charges and currents may arise.
In particular there are two types of charges. There are so-called free charges, which are
additional charges that we can plop down in a material. This is normally what we think
of when we think of charge. (The term “free” is probably not the best term, because the
charges need not be free to move. We can bolt them down if we wish.) But additionally,
there are bound charges. These are the effective charges that get produced when polar
molecules align themselves in certain ways to “shield” the bound charges.
For example, if we place a positive free charge qfree in a material, then the nearby polar
molecules will align themselves so that their negative ends form a negative layer around
the free charge. The net charge inside a Gaussian surface around the charge is therefore
less than q. Call it qnet . Maxwell’s first equation is then ∇ · E = ρnet /²0 . However, it is
generally much easier to deal with ρfree than ρnet , so let’s define ² by ρnet /ρfree ≡ ²0 /² < 1. 2
Maxwell’s first equation can then be written as
ρfree
∇·E= . (22)
²
2 The fact that the shielding is always proportional to q
free (at least in non-extreme cases) implies that
there is a unique value of ² that works for all values of qfree .
8.3. THE FORM OF THE WAVES 7

The electric field in the material around the point charge is less than what it would be
in vacuum, by a factor of ²0 /² (and ² is always greater than or equal to ²0 , because it
isn’t possible to have “anti-shielding”). In a dielectric, the fact that ² is greater than ²0 is
consistent with the fact that the index of refraction n in Eq. (21) is always greater than 1,
which in turn is consistent with the fact that v is always less than c.
A similar occurrence happens with currents. There can be free currents, which are
the normal ones we think about. But there can also be bound currents, which arise from
tiny current loops of electrons spinning around within their atoms. This is a little harder
to visualize than the case with the charges, but let’s just accept here that the fourth of
Maxwell’s equations becomes ∇ × B = µ²∂E/∂t + µJfree . But as mentioned above, µ is
generally close to µ0 for most dielectrics, so this distinction usually isn’t so important.
To sum up, we can ignore all the details about what’s going on at the atomic level by
pretending that we have a vacuum with modified µ and ² values. Although there certainly
exist bound charges and currents in the material, we can sweep them under the rug and
consider only the free charges and currents, by using the modified µ and ² values.
The above modified expressions for Maxwell’s equations are correct if we’re dealing with
a single medium. But if we have two or more mediums, the correct way to write the equations
is to multiply the first equation by ² and divide the fourth equation by µ (see Problem [to
be added] for an explanation of this). The collection of all four Maxwell’s equations is then

∇·D = ρfree ,
∇·B = 0,
∂B
∇×E = − ,
∂t
∂D
∇×H = + Jfree , (23)
∂t
where D ≡ ²E and H ≡ B/µ. D is called the electric displacement vector, and H goes by
various names, including simply the “magnetic field.” But you can avoid confusing it with
B if you use the letter H and not the name “magnetic field.”

8.3 The form of the waves

8.3.1 The wavevector k
What is the dispersion relation associated with the wave equation in Eq. (15)? That is,
what is the relation between the frequency and wavenumber? Or more precisely, what is the
dispersion relation for each component of E, for example the Ex that satisfies Eq. (18)? All
of the components are in general functions of four coordinates: the three spatial coordinates
x, y, z, and the time t. So by the same reasoning as in the two-coordinate case we discussed
at the end of Section 4.1, we know that we can Fourier-decompose the function Ex (x, y, z, t)
into exponentials of the form,

Aei(kx x+ky y+kz z−ωt) ≡ Aei(k·r−ωt) , where k ≡ (kx , ky , kz ). (24)

Likewise for Ey and Ez . And likewise for the three components of B. These are traveling
waves, although we can form combinations of them to produce standing waves.
k is known as the wavevector. As we’ll see below, the magnitude k ≡ |k| plays exactly
the same role that k played in the 1-D case. That is, k is the wavenumber. It equals 2π
times the number of wavelengths that fit into a unit length. So k = 2π/λ. We’ll also see
below that the direction of k is the direction of the propagation of the wave. In the 1-D case,
 8 CHAPTER 8. ELECTROMAGNETIC WAVES

the wave had no choice but to propagate in the ±x direction. But now it can propagate in
any direction in 3-D space.
Plugging the exponential solution in Eq. (24) into Eq. (18) gives
1 |k|2
−ω 2 = (−kx2 − ky2 − kz2 ) =⇒ ω2 = =⇒ ω = c|k| (25)
µ0 ²0 µ0 ²0
√
where c = 1/ µ0 ²0 , and where we are using the convention that ω is positive. Eq. (25)
is the desired dispersion relation. It is a trivial relation, in the sense that electromagnetic
waves in vacuum are dispersionless.
When we go through the same procedure for the other components of E and B, the
“A” coefficient in Eq. (24) can be different for the 2 · 3 = 6 different components of the
fields. And technically k and ω can be different for the six components too (as long as they
satisfy the same dispersion relation). However, although we would have solutions to the six
different waves equations, we wouldn’t have solutions to Maxwell’s equations. This is one of
the cases where the extra information contained in Maxwell’s equations is important. You
can verify (see Problem [to be added]) that if you want Maxwell’s equations to hold for all
r and t, then k and ω must be the same for all six components of E and B. If we then
collect the various “A” components into the two vectors E0 and B0 (which are constants,
independent of r and t), we can write the six components of E and B in vector form as
E = E0 ei(k·r−ωt) , and B = B0 ei(k·r−ωt) , (26)
where the k vector and the ω frequency are the same in both fields. The vectors E0 and B0
can be complex. If they do have an imaginary part, it will produce a phase in the cosine
function when we take the real part of the above exponentials. This will be important when
we discuss polarization.
From Eq. (26), we see that E (and likewise B) depends on r through the dot product
wavefronts of k · r. So E has the same value everywhere on the surface defined by k · r = C, where C is
y constant E, B some constant. This surface is a plane that is perpendicular to k. This follows from the fact
that if r1 and r2 are two points on the surface, then k · (r1 − r2 ) = C − C = 0. Therefore,
the vector r1 − r2 is perpendicular to k. Since this holds for any r1 and r2 on the surface,
the surface must be a plane perpendicular to k. If we suppress the z dependence of E and
k B for the sake of drawing a picture on a page, then for a given wavevector k, Fig. 3 shows
some “wavefronts” with common phases k · r − ωt, and hence common values of E and B.
The planes perpendicular to k in the 3-D case become lines perpendicular to k in the 2-D
x case. Every point on a given plane is equivalent, as far as E and B are concerned.
How do these wavefronts move as time goes by? Well, they must always be perpendicular
to k, so all they can do is move in the direction of k. How fast do they move? The dot
Figure 3 product k · r equals kr cos θ, where θ is the angle between k and a given position r, and
where k ≡ |k| and r = |r|. If we group the product as k(r cos θ), we see that it equals
k times the projection of r along k. If we rotate our coordinate system so that a new x0
axis points in the k direction, then the projection r cos θ simply equals the x0 value of the
position. So the phase k · r − ωt equals kx0 − ωt. We have therefore reduced the problem to
a 1-D problem (at least as far as the phase is concerned), so we can carry over all of our 1-D
results. In particular, the phase velocity (and group velocity too, since the wave equation
√
in Eq. (15) is dispersionless) is v = ω/k, which we see from Eq. (25) equals c = 1/ µ0 ²0 .
Remark: We just found that the phase velocity has magnitude
ω ω ω
v= ≡ = p , (27)
k |k| kx2 + ky2 + kz2

and it points in the k̂ direction. You might wonder if the simpler expression ω/kx has any meaning.
And likewise for y and z. It does, but it isn’t a terribly useful quantity. It is the velocity at
8.3. THE FORM OF THE WAVES 9
y k
which a point with constant phase moves in the x direction, with y and z held constant. This
follows from the fact that if we let the constant y and z values be y0 and z0 , then the phase equals
kx x + ky y0 + kz z0 − ωt = kx x − ωt + C, where C is a constant. So we effectively have a 1-D problem
for which the phase velocity is ω/kx .
But note that this velocity can be made arbitrarily large, or even infinite if kx = 0. Fig. 4 shows
a situation where k points mainly in the y direction, so kx is small. Two wavefronts are shown,
and they move upward along the direction of k. In the time during which the lower wavefront
moves to the position of the higher one, a point on the x axis with a particular constant phase x
moves from one dot to the other. This means that it is moving very fast (much faster than the
wavefronts), consistent with the fact the ω/kx is very large if kx is very small. In the limit where Figure 4
the wavefronts are horizontal (kx = 0), a point of constant phase moves infinitely fast along the x
axis. The quantities ω/kx , ω/ky , and ω/kz therefore cannot be thought of as components of the
phase velocity in Eq. (27). The component of a vector should be smaller than the vector itself,
after all.
The vector that
p does correctly break up into components is the wavevector k = (kx , ky , kz ). Its
magnitude k = kx2 + ky2 + kz2 represents how much the phase of the wave increases in each unit
distance along the k̂ direction. (In other words, it equals 2π times the number of wavelengths that
fit into a unit distance.) And kx represents how much the phase of the wave increases in each unit
distance along the x direction. This is less than k, as it should be, in view of Fig. 4. For a given
distance along the x axis, the phase advances by only a small amount, compared with along the
k vector. The phase needs the entire distance between the two dots to increase by 2π along the
x axis, whereas it needs only the distance between the wavefronts to increase by 2π along the k
vector. ♣
y
E
8.3.2 Further constraints due to Maxwell’s equations B
E
Fig. 3 tells us only that points along a given line have common values of E and B. It doesn’t
tell us what these values actually are, or if they are constrained in other ways. For all we B k
know, E and B on a particular wavefront might look like the vectors shown in Fig. 5 (we E
have ignored any possible z components). But it turns out the these vectors aren’t actually B
possible. Although they satisfy the wave equation, they don’t satisfy Maxwell’s equations.
x
So let’s now see how Maxwell’s equations further constrain the form of the waves. Later
on in Section 8.8, we’ll see that the waves are even further constrained by any boundary (impossible E, B vectors)
conditions that might exist. We’ll look at Maxwell’s equations in order and see what each
of them implies. Figure 5

• Using the expression for E in Eq. (26), the first of Maxwell’s equations, Eq. (8), gives

∂Ex ∂Ey ∂Ez

∇·E=0 =⇒ + + =0
∂x ∂y ∂z
=⇒ ikx Ex + iky Ey + ikz Ez = 0
=⇒ k·E=0 (28)

This says that E is always perpendicular to k. As we see from the second line here,
each partial derivative simply turns into a factor if ikx , etc.

• The second of Maxwell’s equations, Eq. (9), gives the analogous result for B, namely,

k·B=0 (29)

So B is also perpendicular to k.
10 CHAPTER 8. ELECTROMAGNETIC WAVES

• Again using the expression for E in Eq. (26), the third of Maxwell’s equations, Eq.
(10), gives
µ ¶
∂B ∂ ∂ ∂ ∂B
∇×E=− =⇒ , , ×E=−
∂t ∂x ∂y ∂z ∂t
¡ ¢
=⇒ ikx , iky , ikz × E = −(−iω)B
=⇒ k × E = ωB (30)

Since the cross product of two vectors is perpendicular to each of them, this result
says that B is perpendicular to E. And we already know that B is perpendicular to
k, from the second of Maxwell’s equations. But technically we didn’t need to use that
equation, because the B ⊥ k result is also contained in this k × E = ωB result. Note
that as above with the divergences, each partial derivative in the curl simply turns
into a factor if ikx , etc.
We know from the first of Maxwell’s equations that E is perpendicular to k, so the
magnitude of k × E is simply |k||E| ≡ kE. The magnitude of the k × E = ωB relation
then tells us that
ω
kE = ωB =⇒ E= B =⇒ E = cB (31)
k
Therefore, the magnitudes of E and B are related by a factor of the wave speed, c.
Eq. (31) is very useful, but its validity is limited to a single traveling wave, because
the derivation of Eq. (30) assumed a unique k vector. If we form the sum of two waves
with different k vectors, then the sum doesn’t satisfy Eq. (30) for any particular vector
k. There isn’t a unique k vector associated with the wave. Likewise for Eqs. (28) and
(29).
• The fourth of Maxwell’s equations, Eq. (11), can be written as ∇ × B = (1/c2 )∂E/∂t,
so the same procedure as above yields

ω ω E
k×B=− E =⇒ B= E =⇒ B= (32)
c2 kc2 c

This doesn’t tell us anything new, because we already know that E, B, and k are all
mutually perpendicular, and also that E = cB. In retrospect, the first and third (or
alternatively the second and fourth) of Maxwell’s equations are sufficient to derive all
of the above results, which can be summarized as

E ⊥ k, B ⊥ k, E ⊥ B, E = cB (33)

If three vectors are mutually perpendicular, there are two possibilities for how they are
oriented. With the conventions of E, B, and k that we have used in Maxwell’s equations and
in the exponential solution in Eq. (24) (where the k · r term comes in with a plus sign), the
orientation is such that E, B, and k form a “righthanded” triplet. That is, E × B points in
the same direction as k (assuming, of course, that you’re defining the cross product with the
righthand rule!). You can show that this follows from the k × E = ωB relation in Eq. (30)
by either simply drawing three vectors that satisfy Eq. (30), or by using the determinant
definition of the cross product to show that a cyclic permutation of the vectors maintains
the sign of the cross product.
A snapshot (for an arbitrary value of t) of a possible electromagnetic wave is shown in
Fig. 6. We have chosen k to point along the z axis, and we have drawn the field only for
8.3. THE FORM OF THE WAVES 11

points on the z axis. But for a given value z0 , all points of the form (x, y, z0 ), which is a
plane perpendicular to the z axis, have common values of E and B. E points in the ±x
direction, and B points in the ±y direction. As time goes by, the whole figure simply slides
along the z axis at speed c. Note that E and B reach their maximum and minimum values
at the same locations. We will find below that this isn’t the case for standing waves.

E k

B z

Figure 6

What are the mathematical expressions for the E and B fields in Fig. 6? We’ve chosen
k to point along the z axis, so we have k = kẑ, which gives k · r = kz. And since E points
in the x direction, its amplitude takes the form of E0 eiφ x̂. (The coefficient can be complex,
and we have written it as a magnitude times a phase.) This then implies that B points
in the y direction (as drawn), because it must be perpendicular to both E and k. So its
amplitude takes the form of B0 eiφ ŷ = (E0 eiφ /c)ŷ. This is the same phase φ, due to Eq.
(30) and the fact that k is real, at least for simple traveling waves. The desired expressions
for E and B are obtained by taking the real part of Eq. (26), so we arrive at
E0
E = x̂E0 cos(kz − ωt + φ), and B = ŷ cos(kz − ωt + φ), (34)
c
These two vectors are in phase with each other, consistent with Fig. 6. And E, B, and k
form a righthanded triple of vectors, as required.
Remarks:
1. When we talk about polarization in Section 8.6, we will see that E and B don’t have to point
in specific directions, as they do in Fig. 6, where E points only along x̂ and B points only
along ŷ. Fig. 6 happens to show the special case of “linear polarization.”
2. The E and B waves don’t have to be sinusoidal, of course. Because the wave equation is
linear, we can build up other solutions from sinusoidal ones. And because the wave equation
is dispersionless, we know (as we saw at the end of Section 2.4) that any function of the form
f (z − vt), or equivalently f (kz − ωt), satisfies the wave equation. But the restrictions placed
by Maxwell’s equations still hold. In particular, the E field determines the B field.
3. A static solution, where E and B are constant, can technically be thought of as a sinusoidal
solution in the limit where ω = k = 0. In vacuum, we can always add on a constant field
to E or B, and it won’t affect Maxwell’s equations (and therefore the wave equation either),
because all of the terms in Maxwell’s equations in vacuum involve derivatives (either space
or time). But we’ll ignore any such fields, because they’re boring for the purposes we’ll be
concerned with. ♣
12 CHAPTER 8. ELECTROMAGNETIC WAVES

8.3.3 Standing waves

Let’s combine two waves (with equal amplitudes) traveling in opposite directions, to form
a standing wave. If we add E1 = x̂E0 cos(kz − ωt) and E2 = x̂E0 cos(−kz − ωt), we obtain

E = x̂(2E0 ) cos kz cos ωt. (35)

This is indeed a standing wave, because all z values have the same phase with respect to
time.
There are various ways to find the associated B wave. Actually, there are (at least) two
right ways and one wrong way. The wrong way is to use the result in Eq. (30) to say that
ωB = k × E. This would yield the result that B is proportional to cos kz cos ωt, which we
will find below is incorrect. The error (as we mentioned above after Eq. (31)) is that there
isn’t a unique k vector associated with the wave in Eq. (35), because it is generated by two
waves with opposite k vectors. If we insisted on using Eq. (30), we’d be hard pressed to
decide if we wanted to use kẑ or −kẑ as the k vector.
A valid method for finding B is the following. We can find the traveling B1 and B2
waves associated with each of the traveling E1 and E2 waves, and then add them. You
can quickly show (using B = (1/ω)k × E for each traveling wave separately) that B1 =
ŷ(E0 /c) cos(kz − ωt) and B2 = −ŷ(E0 /c) cos(−kz − ωt). The sum of these waves give the
desired associated B field,
B = ŷ(2E0 /c) sin kz sin ωt. (36)
Another method is to use the third of Maxwell’s equations, Eq. (10), which says that
∇ × E = −∂B/∂t. Maxwell’s equations hold for any E and B fields. We don’t have to
worry about the uniqueness of k here. Using the E in Eq. (35), the cross product ∇ × E
can be calculated with the determinant:
¯ ¯
¯ x̂ ŷ ẑ ¯¯
¯ ∂Ex ∂Ex
∇ × E = ¯¯ ∂/∂x ∂/∂y ∂/∂z ¯¯ = ŷ − ẑ
¯ Ex ∂z ∂y
0 0 ¯
= −ŷ(2E0 )k sin kz cos ωt − 0. (37)

Eq. (30) tells us that this must equal −∂B/∂t, so we conclude that

B = ŷ(2E0 )(k/ω) sin kz sin ωt = ŷ(2E0 /c) sin kz sin ωt. (38)
in agreement with Eq. (36). We have ignored any possible additive constant in B.
Having derived the associated B field in two different ways, we can look at what we’ve
found. E and B are still perpendicular to each other, which makes since, because E is the
superposition of two vectors that point in the ±x̂ direction, and E is the superposition of
two vectors that point in the ±ŷ direction. But there is a major difference between standing
waves and traveling waves. In traveling waves, E and B run along in step with each other,
as shown above in Fig. 6. They reach their maximum and minimum values at the same
times and positions. However, in standing waves E is maximum when B is zero, and also
where B is zero (and vice versa). E and B are 90◦ out of phase with each other in both
time and space. That is, the B in Eq. (36) can be written as

2E0 ³ π´ ³ π´
B = ŷ cos kz − cos ωt − , (39)
c 2 2
which you can compare with the E in Eq. (35). A few snapshots of the E and B waves are
shown in Fig. 7.
8.4. ENERGY 13

x x x
E
E

y y y

z B z z
B

(t=0) (ωt=π/4) (ωt=π/2)

Figure 7

8.4 Energy
8.4.1 The Poynting vector
The energy density of an electromagnetic field is
²0 2 1 2
E= E + B , (40)
2 2µ0
where E ≡ |E| and B ≡ |B| are the magnitudes of the fields at a given location in space and
time. We have suppressed the (x, y, z, t) arguments of E, E, and B. This energy density can
be derived in various ways (see Problem [to be added]), but we’ll just accept it here. The
goal of this section is to calculate the rate of change of E, and to then write it in a form
that allows us to determine the energy flux (the flow of energy across a given surface). We
will find that the energy flux is given by the so-called Poynting vector.
If we write E 2 and B 2 as E · E and B · B, then the rate of change of E becomes
∂E ∂E 1 ∂B
= ²0 E · + B· . (41)
∂t ∂t µ0 ∂t
(The product rule works here for the dot product of vectors for the same reason it works
for a regular product. You can verify this by explicitly writing out the components.) The
third and fourth Maxwell’s equations turn this into
µ ¶
∂E 1 1 ¡ ¢
= ²0 E · ∇×B + B· −∇×E
∂t µ0 ²0 µ0
1³ ´
= E · (∇ × B) − B · (∇ × E) .
µ0
(42)
The righthand side of this expression conveniently has the same form as the righthand side
of the vector identity (see Problem [to be added] for the derivation),
∇ · (C × D) = D · (∇ × C) − C · (∇ × D). (43)
So we now have
∂E 1
= ∇ · (B × E). (44)
∂t µ0
Now consider a given volume V in space. Integrating Eq. (44) over this volume V yields
Z Z Z
∂E 1 ∂WV 1
= ∇ · (B × E) =⇒ = (B × E) · dA, (45)
V ∂t µ0 V ∂t µ0 A
14 CHAPTER 8. ELECTROMAGNETIC WAVES

where WV is the energy contained in the volume V (we’ve run out of forms of the letter E),
and where we have used the divergence theorem to rewrite the volume integral as a surface
integral over the area enclosing the volume. dA is defined to be the vector perpendicular
to the surface (with the positive direction defined to be outward), with a magnitude equal
to the area of a little patch.
Let’s now make a slight change in notation. dA is defined to be an outward-pointing
vector, but let’s define dAin to be the inward-pointing vector, dAin ≡ −dA. Eq. (45) can
then be written as (switching the order of E and B)
Z
∂WV 1
= (E × B) · dAin . (46)
∂t µ0 A

We can therefore interpret the vector,

1
S≡ E×B (energy flux : energy/(area · time)) (47)
µ0

as giving the flux of energy into a region. This vector S is known as the Poynting vector.
And since E × B ∝ k, the Poynting vector points in the same direction as the velocity of the
wave. Integrating S over any surface (or rather, just the component perpendicular to the
surface, due to the dot product with dAin ) gives the energy flow across the surface. This
result holds for any kind of wave – traveling, standing, or whatever. Comparing the units
on both sides of Eq. (46), we see that the Poynting vector has units of energy per area per
time. So if we multiply it (or its perpendicular component) by an area, we get the energy
per time crossing the area.
The Poynting vector falls into a wonderful class of phonetically accurate theorems/results.
Others are the Low energy theorem (named after S.Y. Low) dealing with low-energy pho-
tons, and the Schwarzschild radius of a black hole (kind of like a shield).

8.4.2 Traveling waves

Let’s look at the energy density E and the Poynting vector S for a traveling wave. A traveling
wave has B = E/c, so the energy density is (using c2 = 1/µ0 ²0 in the last step)

²0 2 1 2 ²0 1 E2
E= E + B = E2 + =⇒ E = ²0 E 2 (48)
2 2µ0 2 2µ0 c2

We have suppressed the (x, y, z, t) arguments of E and E. Note that this result holds only
for traveling waves. A standing wave, for example, doesn’t have B = E/c anywhere, so E
doesn’t take this form. We’ll discuss standing waves below.
The Poynting vector for a traveling wave is
µ ¶
1 1 E
S= E×B= E k̂, (49)
µ0 µ0 c

where we have used the facts that E ⊥ B and that their cross product points in the direction
of k. Using 1/µ0 = c2 ²0 , arrive at

S = c²0 E 2 k̂ = cE k̂. (50)

This last equality makes sense, because the energy density E moves along with the wave,
which moves at speed c. So the energy per unit area per unit time that crosses a surface
8.4. ENERGY 15

is cE (which you can verify has the correct units). Eq. (50) is true at all points (x, y, z, t)
individually, and not just in an average sense. We’ll derive formulas for the averages below.
In the case of a sinusoidal traveling wave of the form,

E = x̂E0 cos(kz − ωt) and B = ŷ(E0 /c) cos(kz − ωt), (51)

the above expressions for E and S yield

E = ²0 E02 cos2 (kz − ωt) and S = c²0 E02 cos2 (kz − ωt)k̂. (52)

Since the average value of cos2 (kz − ωt) over one period (in either space or time) is 1/2, we
see that the average values of E and |S| are
1 1
Eavg = ²0 E02 and |S|avg = c²0 E02 . (53)
2 2
|S|avg is known as the intensity of the wave. It is the average amount of energy per unit
area per unit time that passes through (or hits) a surface. For example, at the location of
the earth, the radiation from the sun has an intensity of 1360 Watts/m2 . The energy comes
from traveling waves with many different frequencies, and the total intensity is just the sum
of the intensities of the individual waves (see Problem [to be added]).

8.4.3 Standing waves

Consider the standing wave in Eqs. (35) and (36). With 2E0 defined to be A, this wave
becomes
E = x̂A cos kz cos ωt, and B = ŷ(A/c) sin kz sin ωt. (54)
The energy density in Eq. (48) for a traveling wave isn’t valid here, because it assumed
B = E/c. Using the original expression in Eq. (40), the energy density for the above
standing wave is (using 1/c2 = µ0 ²0 )
²0 2 1 2 1 ¡ ¢
E= E + B = ²0 A2 cos2 kz cos2 ωt + sin2 kz sin2 ωt . (55)
2 2µ0 2
If we take the average over a full cycle in time (a full wavelength in space would work just
as well), then the cos2 ωt and sin2 ωt factors turn into 1/2’s, so the time average of E is
µ ¶
1 1 1 ²0 A2
Eavg = ²0 A2 cos2 kz + sin2 kz = , (56)
2 2 2 4

which is independent of z. It makes sense that it doesn’t depend on z, because a traveling

wave has a uniform average energy density, and a standing wave is just the sum of two
traveling waves moving in opposite directions.
The Poynting vector for our standing wave is given by Eq. (47) as (again, the traveling-
wave result in Eq. (50) isn’t valid here):

1 A2
S= E×B= k̂ cos kz sin kz cos ωt sin ωt. (57)
µ0 µ0 c
At any given value of z, the time average of this is zero (because cos ωt sin ωt = (1/2) sin 2ωt),
so there is no net energy flow in a standing wave. This makes sense, because a standing
wave is made up of two traveling waves moving in opposite directions which therefore have
opposite energy flows (on average). Similarly, for a given value of t, the spatial average is
zero. Energy sloshes back and forth between points, but there is no net flow.
16 CHAPTER 8. ELECTROMAGNETIC WAVES

Due to the fact that a standing wave is made up of two traveling waves moving in
opposite directions, you might think that the Poynting vector (that is, the energy flow)
should be identically equal to zero, for all z and t. But it isn’t, because each of the two
Poynting vectors depends on z and t, so only at certain discrete times and places do they
anti-align and exactly cancel each other. But on average they cancel.

8.5 Momentum
Electromagnetic waves carry momentum. However, all the other waves we’ve studied (longi-
tudinal spring/mass and sound waves, transverse string waves, etc.) don’t carry momentum.
(However, see Footnote 1 above.) Therefore, it is certainly not obvious that electromagnetic
waves carry momentum, because it is quite possible for waves to carry energy without also
carrying momentum.
A quick argument that demonstrates why an electromagnetic wave (that is, light) carries
momentum is the following argument from relativity. The relativistic relation between a
particle’s energy, momentum, and mass is E 2 = p2 c2 + m2 c4 (we’ll just accept this here).
For a massless particle (m = 0), this yields E 2 = p2 c2 =⇒ E = pc. Since photons (which is
what light is made of) are massless, they have a momentum given by p = E/c. We already
know that electromagnetic waves carry energy, so this relation tells us that they must also
carry momentum. In other words, a given part of an electromagnetic wave with energy E
also has momentum p = E/c.
However, although this argument is perfectly valid, it isn’t very satisfying, because (a)
it invokes a result from relativity, and (b) it invokes the fact that electromagnetic waves
(light) can be considered to be made up of particle-like objects called photons, which is
by no means obvious. But why should the particle nature of light be necessary to derive
the fact that an electromagnetic wave carries momentum? It would be nice to derive the
p = E/c result by working only in terms of waves and using only the results that we have
developed so far in this book. Or said in a different way, it would be nice to understand
how would someone living in, say, 1900 (that is, pre-relativity) would demonstrate that an
electromagnetic waves carries momentum. We can do this in the following way.
Consider a particle with charge q that is free to move around in some material, and let
it be under the influence of a traveling electromagnetic wave. The particle will experience
forces due to the E and B fields that make up the wave. There will also be damping forces
from the material. And the particle will also lose energy due to the fact that it is accelerating
and hence radiating (see Section 8.7). But the exact nature of the effects of the damping
and radiation won’t be important for this discussion.3
Assume that the wave is traveling in the z direction, and let the E field point along the x
direction. The B field then points along the y direction, because E × B ∝ k. The complete
motion of the particle will in general be quite complicated, but for the present purposes it
suffices to consider the x component of the particle’s velocity, that is, the component that is
parallel to E. 4 Due to the oscillating electric field, the particle will (mainly) oscillate back
and forth in the x direction. However, we don’t know the phase. In general, part of the
velocity will be in phase with E, and part will be ±90◦ out of phase. The latter will turn
out not to matter for our purposes,5 so we’ll concentrate on the part of the velocity that is
3 If the particle is floating in outer space, then there is no damping, so only the radiation will extract

energy from the particle.

4 If we assume that the velocity v of the particle satisfies v ¿ c (which is generally a good approximation),

then the magnetic force, qv × B is small compared with the electric force, qE. This is true because B = E/c,
so the magnetic force is suppressed by a factor of v/c (or more, depending on the angle between v and B)
compared with the electric force. The force on the particle is therefore due mainly to the electric field.
5 We’ll be concerned with the work done by the electric field, and this part of the velocity will lead to
 (some time t)

8.5. MOMENTUM 17 E
vE
in phase with E. Let’s call it vE . We then have the pictures shown in Fig. 8. k
You can quickly verify with the righthand rule that the magnetic force qvE × B points B
forward along k in both cases. vE and B switch sign in phase with each other, so the two
signs cancel, and there is a net force forward. The particle therefore picks up some forward
momentum, and this momentum must have come from the wave. In a small time dt, the
magnitude of the momentum that the wave gives to the particle is (half period later)
qvE E dt
|dp| = |FB dt| = |qvE × B| dt = qvE B dt = . (58)
c
B
What is the energy that the wave gives to the particle? That is, what is the work that k
the wave does on the particle? (In the steady state, this work is balanced, on average, by vE
the energy that the particle loses to damping and radiation.) Only the electric field does E
work on the particle. And since the electric force is qE, the amount of work done on the
particle in time dt is
dW = FE · dx = (qE)(vE dt) = qvE E dt. (59) qv B points along
◦
(The part of the velocity that is ±90 out of phase with E will lead to zero net work, on k in both cases
average; see Problem [to be added].) Comparing this result with Eq. (58), we see that Figure 8
dW
|dp| = . (60)
c
In other words, the amount of momentum the particle gains from the wave equals 1/c times
the amount of energy it gains from the wave. This holds for any extended time interval ∆t,
because any interval can be built up from infinitesimal times dt.
Since Eq. (60) holds whenever any electromagnetic wave encounters a particle, we con-
clude that the wave actually carries this amount of momentum. Even if we didn’t have a
particle in the setup, we could imagine putting one there, in which case it would acquire
the momentum given by Eq. (60). This momentum must therefore be an intrinsic property
of the wave.
Another way of writing Eq. (60) is

1 ¯¯ dp ¯¯ 1 1 dW
¯ ¯= · , (61)
A dt c A dt
where A is the cross-sectional area of the wave under consideration. The lefthand side is
the force per area (in other words, the pressure) that the wave applies to a material. And
from Eqs. (46) and (47), the righthand side is |S|/c, where S is the Poynting vector. The
pressure from an electromagnetic wave (usually called the radiation pressure) is therefore

|S| |E × B| E2
Radiation pressure = = = . (62)
c µ0 c µ0 c2

You can show (see Problem [to be added]) that the total force from the radiation pressure
from sunlight hitting the earth is roughly 6·108 kg m/s2 (treating the earth like a flat coin and
ignoring reflection, but these won’t affect the order of magnitude). This force is negligible
compared with the attractive gravitational force, which is about 3.6 · 1022 kg m/s2 . But for
a small enough sphere, these two forces are comparable (see Problem [to be added]).
Electromagnetic waves also carry angular momentum if they are polarized (see Problem
[to be added]).

zero net work being done.

 18 CHAPTER 8. ELECTROMAGNETIC WAVES

8.6 Polarization
8.6.1 Linear polarization
Consider the traveling wave in Eq. (34) (we’ll ignore the overall phase φ here):
E0
E = x̂E0 cos(kz − ωt), and B = ŷ cos(kz − ωt). (63)
c
This wave has E always pointing in the x direction and B always pointing in the y direction.
A wave like this, where the fields always point along given directions, is called a linearly
polarized wave. The direction of the linear polarization is defined to be the axis along which
the E field points.
But what if we want to construct a wave where the fields don’t always point along given
directions? For example, what if we want the E vector to rotate around in a circle instead
of oscillating back and forth along a line?
Let’s try making such a wave by adding on an E field (with the same magnitude) that
points in the y direction. The associated B field then points in the negative x direction if
we want the orientation to be the same so that the wave still travels in the same direction
y (that is, so that the k̂ vector still points in the +ẑ direction). The total wave is now
B E E0
B1 E2 E = (x̂ + ŷ)E0 cos(kz − ωt), and B = (ŷ − x̂) cos(kz − ωt). (64)
c
B2 E1 If the two waves we added are labeled as “1” and “2” respectively, then the sum given in
x Eq. (64) is shown in Fig. 9. The wave travels in the positive z direction, which is out of the
page. The E field in this wave always points along the (positive or negative) diagonal x̂ + ŷ
Figure 9 direction, and the B field always points along the ŷ − x̂ direction. So we still have a linearly
polarized
√ wave. All we’ve done is rotate the fields by 45◦ and multiply the amplitudes by
2. Therefore, if our goal is to produce a wave that isn’t linearly polarized (that is, to
produce a wave where the directions of E and B change), we’re going to have to come up
with a more clever method than adding on fields that point in different directions.
Before proceeding further, we should note that no matter what traveling wave we have,
the B field is completely determined by the E field and the k vector via Eq. (30). For a given
k vector (we’ll generally pick k to point along ẑ), this means that the E field determines
the B field (and vice versa). So we won’t bother writing down the B field anymore. We’ll
just work with the E field.
We should also note that the x and y components of B are determined separately by the
y and x components of E, respectively. This follows from Eq. (30) and the properties of the
cross product:

k × E = ωB =⇒ (kẑ) × (Ex x̂ + Ey ŷ) = Bx x̂ + By ŷ

=⇒ kEx ŷ + kEy (−x̂) = Bx x̂ + By ŷ
=⇒ Bx = −kEy , and By = kEx . (65)

We see that the Bx and Ey pair of components is “decoupled” from the By and Ex pair.
The two pairs have nothing to do with each other. Each pair can be doing whatever it feels
like, independent of the other pair. So we basically have two independent electromagnetic
waves. This is the key to understanding polarization.

8.6.2 Circular and elliptical polarization

Our above attempt at forming a non-linearly polarized wave failed because the two E fields
that we added together had the same phase. This resulted in the two pairs of components
8.6. POLARIZATION 19

(Ex and By , and Ey and Bx ) having the same phase, which in turn resulted in a simple
(tilted) line for each of the total E and B fields.
Let us therefore try some different relative phases between the components. As men-
tioned above, from here on we’ll write down only the E field. The B field can always be
obtained from Eq. (30). Let’s add a phase of, say, π/2 to the y component of E. As above,
we’ll have the magnitudes of the components be equal, so we obtain

E = x̂E0 cos(kz − ωt) + ŷE0 cos(kz − ωt + π/2)

= x̂E0 cos(kz − ωt) − ŷE0 sin(kz − ωt). (66)

What does E look like as a function of time, for a given value of z? We might as well pick
z = 0 for simplicity, in which case we have (using the facts that cosine and sine are even
y
and odd functions, respectively)

E = x̂E0 cos(ωt) + ŷE0 sin(ωt). (67)

This is the expression for a vector with magnitude E0 that swings around in a counterclock- E
wise circle in the x-y plane, as shown in Fig. 10. (And at all times, B is perpendicular to E.) x
This is our desired circular polarization. The phase difference of π/2 between the x and y
components of E causes E to move in a circle, as opposed to simply moving back and forth
along a line. This is consistent with the fact that the phase difference implies that Ex and
Ey can’t both be zero at the same time, which is a necessary property of linear polarization,
because the vector passes through the origin after each half cycle. Figure 10
If we had chosen a phase of −π/2 instead of π/2, we would still have obtained circular
polarization, but with the circle now being traced out in a clockwise sense (assuming that
ẑ still points out of the page).
A phase of zero gives linear polarization, and a phase of ±π/2 gives circular polarization.
What about something in between? If we choose a phase of, say, π/3, then we obtain

E = x̂E0 cos(kz − ωt) + ŷE0 cos(kz − ωt + π/3) (68)

As a function of time at z = 0, this takes the form,

y
ωt=π/3
E = x̂E0 cos(ωt) + ŷE0 cos(ωt − π/3). (69) (1/2,1)
ωt=2π/3
So the y component lags the x component by π/3. Ey therefore achieves its maximum value (-1/2,1/2)
E
ωt=0
(1,1/2)
at a time given by ωt = π/3 after Ex achieves its maximum value. A plot of E is shown in
Fig. 11. A few of the points are labeled with their ωt values. This case is reasonably called x
elliptical polarization. The shape in Fig. 11 is indeed an ellipse, as you can show in Problem ωt=π ωt=5π/3
[to be added]. And as you can also show in this problem, the ellipse is always titled at ±45◦ . (-1,-1/2) (1/2,-1/2)
The ellipse is very thin if the phase is near 0 or π, and it equals a diagonal line in either of ωt=4π/3
these limits. Linear and circular polarization are special cases of elliptical polarization. If (-1/2,-1)
you want to produce a tilt angle other than ±45◦ , you need to allow for Ex and Ey to have
different amplitudes (see Problem [to be added]). Figure 11

Remark: We found above in Eq. (26) that the general solution for E is

E = E0 ei(k·r−ωt) , (70)

In looking at this, it appears that the various components of E have the same phase. So were we
actually justified in throwing in the above phases of π/2 and π/3, or anything else? Yes, because
as we mentioned right after Eq. (26), the E0 vector (and likewise the B0 vector) doesn’t have to
be real. Each component can be complex and have an arbitrary phase (although the three phases
in B0 are determined by the three phases phases in E0 by Maxwell’s equations). For example, we
20 CHAPTER 8. ELECTROMAGNETIC WAVES

can have E0,x = A and E0,y = Aeiφ . When we take the real part of the solution in Eq. (70), we
then obtain

Ex = A cos(k · r − ωt) and Ey = A cos(k · r − ωt + φ). (71)

So this is the source of the relative phase, which in turn is the source of the various types of
polarizations. ♣

Standing waves can also have different types of polarizations. Such a wave can be viewed
as the sum of polarized waves traveling in opposite directions. But there are different cases
to consider, depending on the orientation of the polarizations; see Problem [to be added].

8.6.3 Wave plates

Certain anisotropic materials (that is, materials that aren’t symmetric around a given axis)
have the property that electromagnetic waves that are linearly polarized along one axis travel
at a different speed from waves that are linearly polarized along another (perpendicular)
axis. This effect is known as birefringence, or double refraction, because it has two different
speeds and hence two different indices of refraction, nx and ny . The difference in speeds
and n values arises from the difference in permittivity values, ², in the two directions. This
difference in speed implies, as we will see, that the electric field components in the two
directions will gradually get out of phase as the wave travels through the material. If the
thickness of the material is chosen properly, we can end up with a phase difference of, say,
π/2 (or anything else) which implies circular polarization. Let’s see how this phase difference
arises.

Let the two transverse directions be x and y, and let the Ex wave travel faster than
the Ey wave. As mentioned right after Eq. (65), we can consider these components to be
two separate waves. Let’s assume that linearly polarized light traveling in the z direction
impinges on the material and that it has nonzero components in both the x and y directions.
Since this single wave is driving both the Ex and Ey waves in the material, these two
components will be in phase with each other at the front end of the material. The material
is best described as a plate (hence the title of this subsection), because the dimension
along the direction of the wave’s motion is generally small compared with the other two
dimensions.

What happens as the Ex and Ey waves propagate through the material? Since the same
external waves is driving both the Ex and Ey waves in the material, the frequencies of these
waves must be equal. However, since the speeds are different, the ω = vk relation tells us
that the k values must be different. Equivalently, the relation λν = v tells us that (since
ν is the same) the wavelength is proportional to the velocity. So a smaller speed means a
shorter wavelength. A possible scenario is shown in Fig. 12. We have assumed that the
y speed is slightly smaller than the x speed, which means that λy is slightly shorter than
λx . Equivalently, ky is slightly larger than kx . Therefore, slightly more Ey waves fit in
the material than Ex waves, as shown. What does this imply about the phase difference
between the Ex and Ey waves when they exit the material?
8.6. POLARIZATION 21

(vacuum) (waveplate) (vacuum)

c vx=ω/kx c
Ex

z (fast axis)

Ey

z (slow axis)

c vy=ω/ky c

ky>kx , λy<λx , vy<vx

Figure 12

At the instant shown, the Ex field at the far end of the material has reached its maximum
value, indicated by the dot shown. (It isn’t necessary for an integral number of wavelengths
to fit into the material, but it makes things a little easier to visualize.) But Ey hasn’t
reached its maximum quite yet. The Ey wave needs to travel a little more to the right
before the crest marked by the dot reaches the far end of the material. So the phase of the
Ey wave at the far end is slightly behind the phase of the Ex wave.6
By how much is the Ey phase behind the Ex phase at the far end? We need to find
the phase of the Ey wave that corresponds to the extra little distance between the dots
shown in Fig. 12. Let the length of the material (the thickness of the wave plate) be L.
Then the number of Ex and Ey wavelengths that fit into the material are L/λx and L/λy ,
respectively, with the latter of these numbers being slightly larger. The number of extra
wavelengths of Ey compared with Ex is therefore L/λy − L/λx . Each wavelength is worth
2π radians, so the phase of the Ey wave at the far end of the material is behind the phase
of the Ex wave by an amount (we’ll give four equivalent expressions here)
µ ¶ µ ¶
1 1 1 1 ωL
∆φ = 2πL − = L(ky − kx ) = ωL − = (ny − nx ), (72)
λy λx vy vx c

where we have used vi = c/ni . In retrospect, we could have simply written down the second
of these expressions from the definition of the wavenumber k, but we have to be careful to
get the sign right. The fact that a larger number of Ey waves fit into the material means
that the phase of the Ey wave is behind the phase of the Ex wave. Of course, if Ey is behind
by a large enough phase, then it is actually better described as being ahead. For example,
being behind by 7π/4 is equivalent to being ahead by π/4. We’ll see shortly how we can use
wave plates to do various things with polarization, including creating circularly polarized
light.
6 You might think that the E phase should be ahead, because it has more wiggles in it. But this is
y
exactly backwards. Of course, if you’re counting from the left end of the plate, Ey does sweep through more
phase than Ex . But that’s not what we’re concerned with. We’re concerned with the phase of the wave as it
passes the far end of the plate. And since the crest marked with the dot in the Ey wave hasn’t reached the
end yet, the Ey phase is behind the Ex phase. So in the end, it is correct to use the simplistic reasoning of,
“a given crest on the slower wave takes longer to reach the end, so the phase of the slower wave is behind.”
 22 CHAPTER 8. ELECTROMAGNETIC WAVES

8.6.4 Making polarized light

polymers How can we produce polarized light? Let’s look at linear polarization and then circular
polarization. One way to produce linearly polarized light is to wiggle a charged particle in
a certain way, so that the radiation is linearly polarized. We’ll talk about how this works
in Section 8.7.
Another way is to create a sheet of polymers (long molecules with repeating parts) that
are stretched out parallel to each other, as shown in Fig. 13. If you have a light source that
emits a random assortment of polarizations, then the sheet can filter out the light that is
linearly polarized along a certain direction and leave you with only the light that is linearly
polarized along the orthogonal direction. This works in the following way.
The electrons in the polymers are free to vibrate in the direction along the polymer,
polarization but not perpendicular to it. From the mechanism we discussed in Section 8.5, this leads
direction to the absorption of the energy of the electric field that points along the polymer. So this
component of the electric field shrinks to zero. Only the component perpendicular to the
Figure 13 polymer survives. So we end up with linearly polarized light in the direction perpendicular
to the polymers. You therefore can’t think of the polymers as a sort of fence which lets
through the component of the field that is parallel to the boards in the fence. It’s the
opposite of this.
Now let’s see how to make circularly polarized light. As in the case of linear polarization,
we can wiggle a charged particle in a certain way. But another way is to make use of the
wave-plate results in the previous subsection. Let the fast and slow axes of the wave plate
be the x and y axes, respectively. If we send a wave into the material that is polarized in
either the x or y directions, then nothing exciting happens. The wave simply stays polarized
y (slow) along that direction. The phase difference we found in Eq. (72) is irrelevant if only one of
the components exists.
E If we want to have anything useful come out of the phase difference caused by the
plate, we need to have both components be nonzero. So let’s assume that we have linearly
polarized light that enters the material with a polarization direction at a 45◦ angle between
the axes, as shown in Fig. 14. Given the frequency ω of the light, and given the two indices
45
x (fast) of refraction, nx and ny , let’s assume that we’ve chosen the thickness L of the plate to
yield a phase difference of ∆φ = π/2 in Eq. (72). Such a plate is called a quarter-wave
Figure 14 plate, because π/2 is a quarter of a cycle. When the wave emerges from the plate, the Ey
component is then π/2 behind the Ey component. So if we choose t = 0 to be the time
when the wave enters the plate, we have (the phase advance of φ is unimportant here)

Enter plate : E ∝ x̂ cos ωt + ŷ cos ωt,

y (slow) Leave plate : E ∝ x̂ cos(ωt + φ) + ŷ cos(ωt + φ − π/2)
= x̂ cos(ωt + φ) + ŷ sin(ωt + φ). (73)
E
This wave has the same form as the wave in Eq. (67), so it is circularly polarized light with
a counterclockwise orientation, as shown in above in Fig. 10.
What if we instead have an incoming wave polarized in the direction shown in Fig. 15?
45
x (fast) The same phase difference of π/2 arises, but the coefficient of the Ex part of the wave now
has a negative sign in it. So we have
Figure 15
Enter plate : E ∝ −x̂ cos ωt + ŷ cos ωt,
Leave plate : E ∝ −x̂ cos(ωt + φ) + ŷ cos(ωt + φ − π/2)
= −x̂ cos(ωt + φ) + ŷ sin(ωt + φ). (74)

This is circularly polarized light, but now with a clockwise orientation.

8.6. POLARIZATION 23

The thickness of a quarter-wave plate (or a half-wave plate, or anything else) depends on
the wavelength of the light, or equivalently on the various other quantities in Eq. (72). Intu-
itively, a longer wavelength means a longer distance to get ahead by a given fraction of that
wavelength. So there is no “universal” quarter-wave plate that works for all wavelengths.

8.6.5 Malus’ law y

y'
If a given electromagnetic wave encounters a linear polarizer, how much of the light makes
it through? Consider light that is linearly polarized in the x̂ direction, and assume that
we have a polarizer whose axis (call it the x̂0 axis) makes an angle of θ with the x̂ axis, as
shown in Fig. 16. (This means that the polymers are oriented at an angle θ ± π/2 with the E0 cosθ
x̂ axis.) In terms of the primed coordinate system, the amplitude of the electric field is
E
x
0 0 θ
E = E0 x̂ = E0 (x̂ cos θ + ŷ sin θ). (75) E0 sinθ
x'
The y 0 component gets absorbed by the polarizer, so we’re left with only the x0 component, polarization
which is E0 cos θ. Hence, the polarizer decreases the amplitude by a factor of cos θ. The direction
intensity (that is, the energy) is proportional to the square of the amplitude, which means
Figure 16
that it is decreased by a factor of cos2 θ. Therefore,

|Eout | = |Ein | cos θ, and Iout = Iin cos2 θ (76)

The second of these relations is known as Malus’ law. Note that if θ = 90◦ , then Iout = 0.
So two successive polarizers that are oriented at 90◦ with respect to each other block all of
the light that impinges on them, because whatever light makes it through the first polarizer
gets absorbed by the second one.
What happens if we put a third polarizer between these two at an angle of 45◦ with
respect to each? It seems that adding another polarizer can only make things “worse” as
far as the transmission of light goes, so it seems like we should still get zero light popping
out the other side. However, if a fraction f of the light makes it through the first polarizer
(f depends on what kind of light you shine in), then f cos2 45◦ makes it through the middle
polarizer. And then a fraction cos2 45◦ of this light makes it through the final polarizer. So
the total amount that makes it through all three polarizers is f cos4 45◦ = f /4. This isn’t
zero! Adding the third polarizer makes things better, not worse.
This strange occurrence is due to the fact that polarizers don’t act like filters of the sort
where, say, a certain fraction of particles make it through a screen. In that kind of filter, a
screen is always “bad” as far as letting particles through goes. The difference with actual A
polarizers is that the polarizer changes the polarization direction of whatever light makes it
through. In contrast, if a particle makes it through a screen, then it’s still the same particle. B
Another way of characterizing this difference is to note that a polarization is a vector, and
vectors can be described in different ways, depending on what set of basis vectors is chosen.
In short, in Fig. 17 the projection of A onto C is zero. But if we take the projection of A
onto B, and then take the projection of the result onto C, the result isn’t zero. C
What happens if instead of inserting one intermediate polarizer at 45◦ , we insert two
polarizers at angles 30◦ and 60◦ ? Or three at 22.5◦ , 45◦ , and 67.5◦ , etc? Does more light Figure 17
or less light make it all the way through? The task of Problem [to be added] is to find out.
You will find in this problem that something interesting happens in the case of a very large
number of polarizers. The idea behind this behavior has countless applications in physics.
 24 CHAPTER 8. ELECTROMAGNETIC WAVES

8.7 Radiation from a point charge

8.7.1 Derivation
In our above discussions of electromagnetic waves, we didn’t worry about where the waves
came from. We just studied their properties, given that they existed. Let’s now look at how
an electromagnetic wave can be created in the first place. We will find that an accelerating
(no B field) charge produces an electromagnetic wave.
Consider first a stationary charge. The field is simply radial,

E q
E = r̂ , (77)
4π²0 r2

and there is no B field. This is shown in Fig. 18. If we instead have a charge moving
with constant velocity v, then the field is also radial, but it is bunched up in the transverse
direction, as shown in Fig. 19. This can be derived with the Lorentz transformations.
However, the proof isn’t important here, and neither is half of the result. All we care about
Figure 18 is the radial nature of the field. We’ll be dealing with speeds that are generally much less
than c, in which case the bunching-up effect is negligible. We therefore again have
B q
E ≈ r̂ (for v ¿ c). (78)
4π²0 r2
E
v This is shown in Fig. 20. There is also a B field if the charge is moving. It points out of
the page in the top half of Figs. 19 and 20, and into the page in the bottom half. This also
follows from the Lorentz transformations (or simply by the righthand rule if you think of the
E moving charge as a current), but it isn’t critical for the discussion. You can check with the
righthand rule that S ∝ E × B points tangentially, which means that no power is radiated
B
outward. And you can also check that S always has a forward component in the direction of
Figure 19 the charge’s velocity. This makes sense, because the field (and hence the energy) increases
as the charge moves to the right, and S measures the flow of energy.
This radial nature of E for a moving charge seems reasonable (and even perhaps obvious),
(for v << c)
but it’s actually quite bizarre. In Fig. 21, the field at point P points radially away from the
B present position of the charge. But how can P know that the charge is where it is at this
instant? What if, for example, the charge stops shortly before the position shown? The
E
field at P would still be directed radially away from where the charge would have been if it
v had kept moving with velocity v. At least for a little while. The critical fact is that the
information that the charge has stopped can travel only at speed c, so it takes a nonzero
E amount of time to reach P . After this time, the field will point radially away from the
B stopped position, as expected.
A reasonable question to ask is then: What happens to the field during the transition
Figure 20 period when it goes from being radial from one point (the projected position if the charge
kept moving) to being radial from another point (the stopped position)? In other words,
P what is the field that comes about due to the acceleration? The answer to this question will
tell us how an electromagnetic field is created and what it looks like.
For concreteness, assume that the charge is initially traveling at speed v, and then let it
decelerate with constant acceleration −a for a time ∆t (starting at t = 0) and come to rest.
So v = a∆t, and the distance traveled during the stopping period is (1/2)a∆t2 /2. Let the
v origin of the coordinate system be located at the place where the deceleration starts.
Consider the situation at time T , where T À ∆t. For example, let’s say that the charge
Figure 21 takes ∆t = 1 s to stop, and we’re looking at the setup T = 1 hour later. The distance
(1/2)a∆t2 /2 is negligible compared with the other distances we’ll be involved with, so we’ll
ignore it. At time T , positions with r > cT have no clue that the charge has started
8.7. RADIATION FROM A POINT CHARGE 25

to decelerate, so they experience a field directed radially away from the future projected
position. Conversely, positions with r < c(T −∆t) know that the charge has stopped, so they
experience a field directed radially away from the origin (or actually a position (1/2)a∆t2 /2,
but this is negligible). So we have the situation shown in Fig. 22.

r=cT
r=c(T-∆t)
D
field due to charge
C if still moving
B

field due to
stopped charge
A θ θ
vT projected position
if no stopping

negligible distance stopped position

a(∆t)2/2 not shown

Figure 22

Since we’re assuming v ¿ c, the “exterior” field lines (the ones obtained by imagining
that the charge is still moving) are essentially not compressed in the transverse direction.
That is, they are spherically symmetric, as shown in Fig. 20. (More precisely, the compres-
sion effect is of order v 2 /c2 which is small compared with the effects of order v/c that we
will find.) Consider the segments AB and CD in Fig. 22. These segments are chosen to
make the same angle θ (which can be arbitrary) with the x axis, with AB passing through
the stopped position, and the line of CD passing through the projected position. Due to
the spherically symmetric nature of both the interior and exterior field lines, the surfaces
of revolutions of AB and CD (which are parts of cones) enclose the same amount of flux.
AB and CD must therefore be part of the same field line. This means that they are indeed
connected by the “diagonal” field line BC shown.7
If we expand the relevant part of Fig. 22, we obtain a picture that takes the general form
shown in Fig. 23. Let the radial and tangential components of the E field in the transition
region be Er and Eθ . From similar triangles in the figure, we have

Eθ vT sin θ
= . (79)
Er c∆t

Note that the righthand side of this grows with T . Again, the units of Eθ and Er aren’t
distance, so the size of the E vector in Fig. 23 is meaningless. But all the matters in the
above similar-triangle argument is that the vector points in the direction shown.

7 We’re drawing both field lines and actual distances in this figure. This technically makes no sense, of

course, because the fields don’t have units of distance. But the point is to show the directions of the fields.
We could always pick pick our unit size of the fields to be the particular value that makes the lengths on
the paper be the ones shown.
 26 CHAPTER 8. ELECTROMAGNETIC WAVES

Er
(B) Eθ

E (C)
c∆t

r=cT
vTsinθ r=c(T-∆t)

(A) θ θ
B vT
D

C Figure 23
A
We now claim that Er has the same value just outside and just inside the transition
Figure 24
region. This follows from a Gauss’s-law argument. Consider the pillbox shown in Fig. 24,
which is located at the kink in the field at point C. The long sides of the box are oriented in
the tangential direction, and also in the direction perpendicular to the page. The short sides
are chosen to be infinitesimally small, so Eθ contributes essentially nothing to the flux. The
flux is therefore due only to the radial component, so we conclude that Ertransition = Eroutside .
(And likewise, a similar pillbox at point B tells us that Ertransition = Erinside , but we won’t
need this. Ertransition varies slightly over the transition region, but the change is negligible
if ∆t is small.) We know that
q q
Eroutside = = . (80)
4π²0 r2 4π²0 (cT )2

Eq. (79) then gives

vT sin θ vT sin θ q
Eθ = Er = ·
c∆t c∆t 4π²0 (cT )2
q sin θ v
= ·
4π²0 c2 (cT ) ∆t
qa sin θ
= (using a = v/∆t and r = cT ) (81)
4π²0 rc2

So the field inside the transition region is given by

µ ¶
q 1 a sin θ
(Er , Eθ ) = , (82)
4π²0 r2 rc2

Both of the components in the parentheses have units of 1/m2 , as they should. We will
explain below why this field leads to an electromagnetic wave, but first some remarks.
Remarks:
1. The location of the Eθ field (that is, the transition region) propagates outward with speed c.
2. Eθ is proportional to a. Given this fact, you can show that the only way to obtain the right
units for Eθ using a, r, c, and θ, is to have a function of the form, af (θ)/rc2 , where f (θ) is
an arbitrary function of θ. And f (θ) happens to be sin θ (times q/4π²0 ).
3. If θ = 0 or θ = π, then Eθ = 0. In other words, there is no radiation in the forward or
backward directions. Eθ is maximum at θ = ±π/2, that is, in the transverse direction.
8.7. RADIATION FROM A POINT CHARGE 27

4. The acceleration vector associated with Fig. 22 points to the left, since the charge was de-
celerating. And we found that Eθ has a rightward component. So in general, the Eθ vector
is
q a⊥ (t0 )
Eθ (r, t) = − · , (83)
4π²0 rc2
where t0 ≡ t − r/c is the time at which the kink at point C in Fig. 22 was emmitted, and
where a⊥ is the component of a that is perpendicular to the radial direction. In other words,
it is the component with magnitude a sin θ that you “see” across your vision if you are located
at position r. This is consistent with the previous remark, because a⊥ = 0 if θ = 0 or θ = π.
Note the minus sign in Eq. (83).
5. Last, but certainly not least, we have the extremely important fact: For sufficiently large r,
Er is negligible compared with Eθ . This follows from the fact that in Eq. (82), Eθ has only
one r in the denominator, whereas Er has two. So for large r, we can ignore the “standard”
radial part of the field. We essentially have only the new “strange”
√ tangential field. By “large
r,” we mean a/rc2 À 1/r2 =⇒ r À c2 /a. Or equivalently
√ ra À c. In other words, ignoring
relativity and using the kinematic relation v = 2ad, the criterion for large r is that (in an
order-of-magnitude sense) if you accelerate something with acceleration a for a distance r,
its velocity will exceed c.
The reason why Eθ becomes so much larger than Er is because there is a T in the numerator
of Eq. (79). This T follows from the fact that in Fig. 23, the Eθ component of E grows with
time (because vT , which is the projected position of the charge if it kept moving, grows with
time), whereas Er is always proportional to the constant quantity, c∆t.

The above analysis dealt with constant acceleration. However, if the acceleration is
changing, we can simply break up time into little intervals, with the above result holding for
each interval (as long as T is large enough so that all of our approximations hold). So even
if a is changing, Eθ (r, t) is proportional to whatever −a⊥ (t0 ) equaled at time t0 ≡ t − r/c.
In particular, if the charge is wiggling sinusoidally, then Eθ (r, t) is a sinusoidal wave.
The last remark above tells us that if we’re far away from an accelerating charge, then the
only electric field we see is the tangential one; there is essentially no radial component. There
is also a magnetic field, which from Maxwell’s equations can be shown to also be tangential,
perpendicular to the page in Fig. 22; see Problem [to be added]. So we have electric and
magnetic fields that oscillate in the tangential directions while propagating with speed c in
the radial direction. But this is exactly what happens with an electromagnetic wave. We
therefore conclude that an electromagnetic wave can be generated by an accelerating charge.
Of course, we know from Section 8.3 that Maxwell’s equations in vacuum imply that the
direction of the E and B fields must be perpendicular to the propagation direction, so in
retrospect we know that this also has to be the case for whatever fields popped out of the
above analysis. The main new points of this analysis are that (1) an accelerating charge can
generate the electromagnetic wave (before doing this calculation, for all we know a nonzero,
say, third derivative of the position is needed to generate a wave), and (2) the radial field
in Eq. (82) essentially disappears due to the 1/r2 vs. 1/r behavior, leaving us with only the
tangential field.
Eq. (30) gives the magnetic field as k × E = ωB =⇒ r̂ × E = cB (using k = kr̂). So in
the top half of Fig. 22, B points into the page with magnitude B = E/c. And in the bottom
half it points out of the page. These facts are consistent with the cylindrical symmetry of
the system around the horizontal axis. If the charge is accelerating instead of decelerating
as we chose above, then the E and B fields are reversed.
 28 CHAPTER 8. ELECTROMAGNETIC WAVES
E
8.7.2 Poynting vector
r
What is the energy flow of a wave generated by a sinusoidally oscillating charge? Let the
θ position of the charge be x(t) = x0 cos ωt. The acceleration is then a(t) = −ω 2 x0 cos ωt.
x0 cosωt The resulting electric field at an arbitrary point is shown in Fig. 25. The energy flow at this
point is given by the Poynting vector, which from Eqs. (49) and (50) is
1
S= E × B = c²0 E 2 r̂. (84)
µ0
Figure 25 Since E is given by the Eθ in Eq. (82), the average value of the magnitude of S is (using
a = −ω 2 x0 cos ωt, along with the fact that the average value of cos2 ωt is 1/2)
µ ¶2
q a sin θ
Savg = c²0 ·
4π²0 rc2
q2 1 1
= · (ω 2 x0 )2 sin2 θ ·
16π 2 ²0 c3 r2 2
ω 4 x20 q 2 sin2 θ 1
= · 2 (85)
32π 2 ²0 c3 r
r dθ
If you want, you can write the 1/²0 c3 part of this as µ0 /c. Note that Savg falls off like 1/r2
and is proportional to ω 4 . Note also that it is zero if θ = 0 or θ = π.
r The Poynting vector has units of Energy/(time · area). Let’s integrate it over a whole
sphere of radius r to find the total Energy/time, that is, the total power. If we parame-
θ terize the integral by θ, then we can slice the sphere into rings as shown in Fig. 26. The
circumference of a ring is 2πr sin θ, and the width is r dθ. So the total power is
Z Z π
P = Savg = Savg (2πr sin θ)(r dθ)
sphere 0
Z π
ω 4 x20 q 2 2πr2
Figure 26 = · 2 sin3 θ dθ
32π 2 ²0 c3 r 0

ω 4 x20 q 2
= (86)
12π²0 c3
Rπ
where we have used the fact that 0 sin3 θ dθ = 4/3. You can quickly verify this by writing
the integrand as (1 − cos2 θ) sin θ. There are two important features of this result for P .
First, it is independent of r. This must be the case, because if more (or less) energy crosses
a sphere at radius r1 than at radius r2 , then energy must be piling up (or be taken from)
the region in between. But this can’t be the case, because there is no place for the energy
to go. Second, P is proportional to (ω 2 x0 )2 , which is the square of the amplitude of the
acceleration. So up to constant numbers and physical constants, we have

P ∝ a20 q 2 , (87)

were a0 is the amplitude of the acceleration.

8.7.3 Blue sky

The fact that the P in Eq. (86) is proportional to ω 4 has a very noticeable consequence
in everyday life. It is the main reason why the sky is blue. (The exponent doesn’t have
to be 4. Any reasonably large number would do the trick, as we’ll see.) In a nutshell, the
8.7. RADIATION FROM A POINT CHARGE 29 all col
ors

ω 4 factor implies that blue light (which is at the high-frequency end of visible spectrum)
scatters more easily than red light (which is at the low-frequency end of visible spectrum). e
So if random white light (composed of many different frequencies) hits the air molecule y blu
stl
shown in Fig. 27, the blue light is more likely to scatter and hit your eye, whereas the other mo
colors with smaller frequencies are more likely to pass straight through. The sky in that
direction therefore looks blue to you. More precisely, the intensity (power per area) of blue
Figure 27
light that is scattered to your eye is larger than the intensity of red light by a factor of
4 4
Pblue /Pred = ωblue /ωred . And since ωblue /ωred ≈ 1.5, we have Pblue /Pred ≈ 5. So 5 times as
much blue light hits your eye.
This also explains why sunsets are red. When the sun is near the horizon, the light must
travel a large distance through the atmosphere (essentially tangential to the earth) to reach
your eye, much larger than when the sun is high up in the sky. Since blue light scatters
more easily, very little of it makes it straight to your eye. Most of it gets scattered in various
directions (and recall that none of it gets scattered directly forward, due to the sin2 θ factor
in Eq. (85)). Red light, on the other hand, scatters less easily, so it is more likely to make
it all the way through the atmosphere in a straight line from the sun to your eye. Pollution
adds to this effect, because it adds particles to the air, which strip off even more of the blue
light by scattering. So for all the bad effects of pollution, cities sometimes have the best
sunsets. A similar situation arises with smoke. If you look at the sun through the smoke of
a forest fire, it appears as a crisp red disk (but don’t look at it for too long).
The actual scattering process is a quantum mechanical one involving photons, and it
isn’t obvious how this translates to our electromagnetic waves. But for the present pur-
poses, it suffices to think about the scattering process as one where a wave with a given
intensity encounters a region of molecules, and the molecules grab chunks of energy and
throw them off in some other direction. (The electrons in the molecules are the things that
are vibrating/accelerating and creating the radiation). The point is that with blue light,
the chunks of energy are 5 times as large as they are for red light.
There are, however, a number of issues that we’ve glossed over. The problem is rather
complicated when everything is included. In particular, one issue is that in addition to the
ω 4 factor, the P in Eq. (86) is also proportional to x20 . What if the electron’s x0 value for
red light is larger than the value for blue light? It turns out that it isn’t; the x0 ’s are all
essentially the same size. This can be shown by treating the electron in the atom as an
essentially undamped driven oscillator. The natural frequency ω0 depends on the nature
of the atom, and it turns out that it is much larger than the frequency ω of light in the
visible spectrum (we’ll just accept this fact). The driven-oscillator amplitude is given in Eq.
(1.88), and when γ ≈ 0 and ω0 À ω, it reduces to being proportional to 1/ω02 . That is, it is
independent of ω, as we wanted to show.
Other issues that complicate things are: Is there multiple scattering? (The answer is
generally no.) Why is the sky not violet, in view of the fact that ωviolet > ωblue ? How
does the eye’s sensitivity come into play? (It happens to be peaked at green.) So there are
certainly more things to consider. But the ω 4 issue we covered above can quite reasonably
be called the main issue.

8.7.4 Polarization in the sky

If you look at the daytime sky with a polarizer (polarized sunglasses do the trick) and
rotate it in a certain way (the polarizer, not the sky, although that would suffice too), you
will find that a certain region of the sky look darker. The reason for this is that the light
that makes it to your eye after getting scattered from this region is polarized. To see why,
consider an electron in the air that is located at a position such that the line from it to
you is perpendicular to the line from the sun to it (which is the k̂ direction of the sun’s
sunlight electron

90
30 CHAPTER 8. ELECTROMAGNETIC WAVES
vibration
directions
light); see Fig. 28. The radiation from the sun may cause this electron to vibrate. And then
because it is vibrating/accelerating, it radiates light which may end up in your eye.
The electric field in the sun’s light lies in the plane perpendicular to k̂. So the field is
Figure 28 some combination of the two directions indicated in Fig. 28. The field can have a component
along the line from the electron to you, and also a component perpendicular to the page,
signified by the ¯ in the figure. The sun’s light is randomly polarized, so it contains some
of each of these. The electric field causes the electron to vibrate, and from the general force
law F = qE, the electron vibrates in some combination of these two directions. However,
due to the sin2 θ factor in Eq. (85), you don’t see any of the radiation that arises from the
electron vibration along the line between it and you. Therefore, the only light that reaches
your eye is the light that was created from the vibration pointing perpendicular to the page.
Hence all of the light you see is polarized in this direction. So if your sunglasses are oriented
perpendicular to this direction, then not much light makes it through, and the sky looks
dark.
Note how the three possible directions of the resulting E field got cut down to one.
First, the electric field that you see must be perpendicular to the k̂, due to the transverse
nature of light (which is a consequence of Maxwell’s equations), and due to the fact that
the electron vibrates along the direction of E. And second, the field must be perpendicular
to the line from the electron to you, due to the “no forward scattering” fact that arises from
the sin2 θ factor in Eq. (85). Alternatively, we know that the E field that you see can’t have
a longitudinal component.
It’s easy to see that conversely if the electron is not located at a position such that the
line from it to you is perpendicular to the line from the sun to it, then you will receive some
light that came from the vertical (on the page, in Fig. 28) oscillation of the electron. But
the amount will be small if the angle is near 90◦ . So the overall result is that there is a
reasonably thick band in the sky that looks fairly dark when viewed through a polarizer. If
the sun is directly overhead, then the band is a circle at the horizon. If the sun is on the
horizon, then the band is a semicircle passing directly overhead, starting and ending at the
horizon.

8.8 Reflection and transmission

We’ll now study the reflection and transmission that arise when light propagating in one
medium encounters another medium. For example, we might have light traveling through
air and then encountering a region of glass. We’ll begin with the case of normal incidence
and then eventually get to the more complicated case of non-normal incidence. For the case
of normal incidence, we’ll start off by considering only one boundary. So if light enters a
region of glass, we’ll assume that the glass extends infinitely far in the forward direction.

8.8.1 Normal incidence, single boundary

(air) (glass) Consider light that travels to the right and encounters an air/glass boundary. As with other
waves we’ve discussed, there will be a reflected wave and a transmitted wave. Because the
Ei Et wave equation in Eq. (15) is dispersionless, all waves travel with the same speed (in each
region). So we don’t need to assume anything about the shape of the wave, although we
generally take it to be a simple sinusoidal one.
Bi Bt The incident, reflected, and transmitted waves are shown in Fig. 29 (the vertical displace-
Er ment in the figure is meaningless). We have arbitrarily defined all three electric fields to be
Br positive if they point upward on the page. A negative value of Ei , Er , or Et simply means
that the vector points downward. We aren’t assuming anything about the polarization of

Figure 29
8.8. REFLECTION AND TRANSMISSION 31

the wave. No matter what kind of wave we have, the incident electric field at a given instant
in time points in some direction, and we are taking that direction to be upward (which we
can arrange by a suitable rotation of the axes). The waves we have drawn are understood
to be the waves infinitesimally close to the boundary on either side.
Given these positive conventions for the E’s, and given the known directions of the three
k̂ vectors, the positive directions of the magnetic fields are determined by k̂ × E = ωB,
and they are shown. Note that if Er has the same orientation as Ei (the actual sign will
be determined below), then Br must have the opposite orientation as Bi , because the k̂
vector for the reflected waves is reversed. If we define positive B to be out of the page,
then the total E and B fields in the left and right regions (let’s call these regions 1 and 2,
respectively) are

E1 = Ei + Er , and E2 = Et ,
B1 = Bi − Br , and B2 = Bt . (88)

What are the boundary conditions? There are four boundary conditions in all: two for
the components of the electric and magnetic fields parallel to the surface, and two for the
components perpendicular to the surface. However, we’ll need only the parallel conditions
for now, because all of the fields are parallel to the boundary. The perpendicular ones will
come into play when we deal with non-normal incidence in Section 8.8.3 below. The two (air) (glass)
parallel conditions are:

• Let’s first find the boundary condition on E k (the superscript “k” stands for parallel).
The third of Maxwell’s equations in Eq. (23) is ∇ × E = −∂B/∂t. Consider the very
thin rectangular path shown in Fig. 30. If we integrate each side of Maxwell’s equation
over the surface S bounded by this path, we obtain
Z Z
∂B
∇×E=− . (89)
S S ∂t
R Figure 30
By Stokes’ theorem, the lefthand side equals the integral E · d` over the rectangular
path. And the right side is −∂ΦB /∂t, where ΦB is the magnetic flux through the
surface S. But if we make
R the rectangle arbitrarily thin, then the flux is essentially zero.
So Eq. (89) becomes E·d` = 0. The short sides of the rectangle contribute essentially
k k
nothing to this integral, and the contribution from the long sides is E2 ` − E1 `, where
` is the length of these sides. Since this difference equals zero, we conclude that

k k
E1 = E2 (90)

The component of the electric field that is parallel to the boundary is therefore con-
tinuous across the boundary. This makes sense intuitively, because the effect of the
dielectric material (the glass) is to at most have charge pile up on the boundary, and
this charge doesn’t affect the field parallel to the boundary. (Or if it does, in the case
where the induced charge isn’t uniform, it affects the two regions in the same way.)
• Now let’s find the boundary condition on B k . Actually, what we’ll find instead is the
boundary condition on H k , where the H field is defined by H ≡ B/µ. We’re using H
here instead of B because H is what appears in the fourth of Maxwell’s equations in
Eq. (23), ∇ × H = ∂D/∂t + Jfree . We need to use this form because we’re working
with a dielectric.
There are no free currents anywhere in our setup, so Jfree = 0. We can therefore apply
to ∇ × H = ∂D/∂t the exact same reasoning with the thin rectangle that we used
32 CHAPTER 8. ELECTROMAGNETIC WAVES

above for the electric field, except now the long sides of the rectangle are perpendicular
to the page. We immediately obtain

k k
H1 = H2 (91)

k k
The boundary condition for the B k fields is then B1 /µ1 = B2 /µ2 . However, since
the µ value for most materials is approximately equal to the vacuum µ0 value, the B k
k k
fields also approximately satisfy B1 = B2 .

We can now combine the above two boundary conditions and solve for Er and Et in
terms of Ei , and also Hr and Ht in terms of Hi . (The B’s are then given by B ≡ H/µ.) We
can write the H k boundary condition in terms of E fields by using

B E E
H≡ = ≡ , where Z ≡ vµ (92)
µ vµ Z

This expression for Z is by definition the impedance for an electromagnetic field. (We’re
√
using v for the speed of light in a general material, which equals 1/ µ² from Eq. (19). We’ll
save c for the speed of light in vacuum.) Z can alternatively be written as
r
µ µ
Z ≡ vµ = √ = . (93)
µ² ²

The two boundary conditions can now be written as

k k
E1 = E2 =⇒ Ei + Er = Et ,
k k Ei Er Et
H1 = H2 =⇒ Hi − Hr = Ht =⇒ − = . (94)
Z1 Z1 Z2
These are two equations in the two unknowns Er and Et . Solving for them in terms of Ei
gives
Z2 − Z1 2Z2
Er = Ei and Et = Ei (95)
Z2 + Z1 Z2 + Z1
These are similar (but not identical) to the reflection and transmission expressions for a
transverse wave on a string; see Eq. (4.38).
We’ll generally be concerned with just the E values, because the B values can always
be found via Maxwell’s equations. But if you want to find the H and B values, you can
write the E’s in the first boundary condition in terms of the H’s. The boundary conditions
become
k k
E1 = E2 =⇒ Z 1 Hi + Z 1 Hr = Z 2 Ht ,
k k
H1 = H2 =⇒ Hi − Hr = Ht . (96)

Solving for Hr and Ht gives

Z2 − Z1 2Z1
Hr = Hi , and Ht = Hi . (97)
Z2 + Z1 Z2 + Z1
Remember that positive Hr is defined to point into the page, whereas positive Hi points
out of the page, as indicated in Fig. 29. So the signed statement for the vectors is Hr =
Hi · (Z1 − Z2 )/(Z1 + Z2 ). If you want to find the B values, they are obtained via B = µH.
8.8. REFLECTION AND TRANSMISSION 33

But again, we’ll mainly be concerned with just the E values. Note that you can also quickly
obtain these H’s by using E = Hz in the results for the E’s in Eq. (95),
We can write the above expressions in terms of the index of refraction, which we defined
in Eq. (20). We’ll need to √ make an approximation, though.√ Since most dielectrics have
µ ≈ µ0 , Eq. (21) gives n ∝ ², and Eq. (93) gives Z ∝ 1/ ². So we have Z ∝ 1/n. The
expressions for Er and Et in Eq. (95) then become

n1 − n2 2n1
Er = Ei and Et = Ei (if µ ≈ µ0 ). (98)
n1 + n2 n1 + n2

In going from, say, air to glass, we have n1 = 1 and n2 ≈ 1.5. Since n1 < n2 , this means that
Er has the opposite sign of Ei . A reflection like this, where the wave reflects off a region of
higher index n, is called a “hard reflection.” The opposite case with a lower n is called a
“soft reflection.”
What is the power in the reflected and transmitted waves? The magnitude of the Poynt-
ing vector for a traveling wave is given by Eq. (49) as S = E 2 /vµ, where we are using v
and µ instead of c and µ0 to indicate an arbitrary dielectric. But vµ is by definition the
impedance Z, so the instantaneous power is (we’ll use “P ” instead of “S” here)

E2
P = (99)
Z

Now, it must be the case that the incident power Pi equals the sum of the reflected and
transmitted powers, Pr and Pt . (All of these P ’s are the instantaneous values at the bound-
ary.) Let’s check that this is indeed true. Using Eq. (95), the reflected and transmitted
powers are
µ ¶2
Er2 Z2 − Z1 Ei2 (Z2 − Z1 )2
Pr = = = Pi ,
Z1 Z2 + Z1 Z1 (Z2 + Z1 )2
µ ¶2
Et2 2Z2 Ei2 4Z1 Z2 Ei2 4Z1 Z2
Pt = = = = Pi . (100)
Z2 Z2 + Z1 Z2 (Z2 + Z1 )2 Z1 (Z2 + Z1 )2

We then quickly see that Pr + Pt = Pi , as desired.

The following topics will eventually be added:

8.8.2 Normal incidence, double boundary

8.8.3 Non-normal incidence
Huygens’ principle, Snell’s law
Maxwell’s equations
Chapter 9

Interference and diffraction

Copyright 2010 by David Morin, [email protected] (Version 1, June 25, 2010)

This file contains the “Interference and diffraction” chapter of a potential book on Waves, designed
for college sophomores.

In this chapter we’ll study what happens when waves from two or more sources exist at a
given point in space. In the case of two waves, the total wave at the given point is the sum
of the two waves. The waves can add constructively if they are in phase, or destructively if
they are out of phase, or something inbetween for other phases. In the general case of many
waves, we need to add them all up, which involves keeping track of what all the phases
are. The results in this chapter basically boil down to (as we’ll see) getting a handle on the
phases and adding them up properly. We won’t need to worry about various other wave
topics, such as dispersion, polarization, and so on; it pretty much all comes down to phases.
The results in this chapter apply to any kind of wave, but for convenience we’ll generally
work in terms of electromagnetic waves.

The outline of this chapter is as follows. In Section 9.1 we do the warm-up case of two
waves interfering. The setup consists of a plane wave passing through two very narrow
(much narrower than the wavelength of the wave) slits in a wall, and these two slits may be
considered to be the two sources. We will calculate the interference pattern on a screen that
is located far away. We’ll be concerned with this “far-field” limit for most of this chapter,
with the exception of Section 9.5. In Section 9.2 we solve the general case of interference
from N narrow slits. In addition to showing how the phases can be added algebraically, we
show how they can be added in an extremely informative geometric manner. In Section 9.3
we switch gears from the case of many narrow slits to the case of one wide slit. The word
“diffraction” is used to describe the interference pattern that results from a slit with non-
negligible width. We will see, however, that this still technically falls into the category of N
narrow slits, because one wide slit can be considered to be a collection of a large (infinite)
number of narrow slits. In section 9.4 we combine the results of the two previous sections
and calculate the interference pattern from N wide slits. Finally, in Section 9.5 we drop the
assumption that the screen is far away from the slit(s) and discuss “near-field” interference
and diffraction. This case is a bit more complicated, but fortunately there is still a nice
geometric way of seeing how things behave. This involves a very interesting mathematical
curve known as the Cornu spriral.

1
 2 CHAPTER 9. INTERFERENCE AND DIFFRACTION
plane wave
9.1 Two-slit interference

Consider a plane wave moving toward a wall, and assume that the wavefronts are parallel to
the wall, as shown in Fig. 1. If you want, you can consider this plane wave to be generated
by a point source that is located a very large distance to the left of the wall. Let there be
two very small slits in the wall that let the wave through. (We’ll see in Section 9.3 that
by “very small,” we mean that the height is much smaller than the wavelength.) We’re
assuming that the slits are essentially infinite in length in the direction perpendicular to the
page. So they are very wide but very squat rectangles. Fig. 2 shows a head-on view from
the far-away point source.
wall
By Huygens’ principle we can consider each slit to be the source of a cylindrically propa-
Figure 1 gating wave. It is a cylindrical (and not spherical) wave because the wave has no dependence
in the direction perpendicular to the page, due to the fact that it is generated by a line source
(view from distant source) (the slit). If we had a point source instead of a line source, then we would end up with
a standard spherically propagating wave. The reason why we’re using a line source is so
that we can ignore the coordinate perpendicular to the page. However, having said this, the
wall fact that we have a cylindrical wave instead of a spherical wave will be largely irrelevant in
this
√ chapter. The main difference is that the amplitude of a cylindrical wave falls off like
1/ r (see Section [to be added] in Chapter 7) instead of the usual 1/r for a spherical wave.
slits wall But for reasons that we will see, we can usually ignore this dependence. In the end, since
we’re ignoring the coordinate perpendicular to the page, we can consider the setup to be
a planer one (in the plane of the page) and effectively think of the line source as a point
wall infinitely
wall extends source (namely, the point on the√line that lies in the page) that happens to produce a wave
in both directions whose amplitude falls off like 1/ r (although this fact won’t be very important).

The important thing to note about our setup is that the two sources are in phase due
wall extends infinitely to the assumption that the wavefronts are parallel to the wall.1 Note that instead of this
in both directions
setup with the incoming plane wave and the slits in a wall, we could of course simply have
Figure 2 two actual sources that are in phase. But it is sometimes difficult to generate two waves
that are exactly in phase. Our setup with the slits makes this automatically be the case.

As the two waves propagate outward from the slits, they will interfere. There will be
constructive interference at places where the two waves are in phase (where the pathlengths
from the two slits differ by an integral multiple of the wavelength). And there will be
destructive interference at places where the two waves are 180◦ out of phase (where the
pathlengths from the two slits differ by an odd multiple of half of the wavelength). For
example, there is constructive interference at point A in Fig. 3 and destructive interference
A at point B.

B What is the interference pattern on a screen that is located very far to the right of the
wall? Assume that the screen is parallel to the wall. The setup is shown in Fig. 4. The
distance between the slits is d, the distance to the screen is D, the lengths of the two paths
to a given point P are r1 and r2 , and θ is the angle that the line to P makes with the
normal to the wall and screen. The distance x from P to the midpoint of the screen is then
x = D tan θ.
Figure 3

1 Problem 9.1 shows how things are modified if the wavefronts aren’t parallel to the wall. This is done in

the context of the N -slit setup in Section 9.2. The modification turns out to be a trivial one.
9.1. TWO-SLIT INTERFERENCE 3

wall essentially screen

parallel P
r1
S1 x
θ r2
d

S2

Figure 4

In finding the interference pattern on the screen, we will work in the so-called far-field
limit where the screen is very far away. (We’ll discuss the near-field case in Section 9.5.)2
The quantitative definition of the far-field limit is D À d. This assumption that D is much
larger than d leads to two important facts.

• If D À d, then we can say that the two pathlengths r1 and r2 in Fig. 4 are essentially
equal in a multiplicative sense. That is, the ratio r1 /r2 is essentially equal to 1. This
follows from the fact that the additive difference |r1 − r2 | is negligible compared with
r1 and r2 (because |r1 − r2 | can’t be any larger than d, which we are assuming is
negligible compared with D, which itself is less than r1 and r2 ). This r1 /r2 ≈ 1 fact
then tells us that the amplitudes of the two waves at point P √from the two slits are
essentially equal (because the amplitudes are proportional to 1/ r, although the exact
power dependence here isn’t important). n
cree
S1 to s away
• If D À d, then we can say that the r1 and r2 paths in Fig. 4 are essentially parallel, far
and so they make essentially the same angle (namely θ) with the normal. The parallel
θ
nature of the paths then allows us to easily calculate the additive difference between d
the pathlengths. A closeup of Fig. 4 near the slits is shown in Fig. 5. The difference
in the pathlengths is obtained by dropping the perpendicular line as shown, so we see S2
θ
that the difference r2 − r1 equals d sin θ. The phase difference between the two waves d sinθ
is then
2π d sin θ Figure 5
k(r2 − r1 ) = kd sin θ = d sin θ = 2π · . (1)
λ λ
In short, d sin θ/λ is the fraction of a cycle that the longer path is ahead of the shorter
path.

Remark: We found above that r1 is essentially equal to r2 in a multiplicative sense, but not in
an additive sense. Let’s be a little more explicit about this. Let ² be defined as the difference,
² ≡ r2 − r1 . Then r2 = r1 + ², and so r2 /r1 = 1 + ²/r1 . Since r1 > D, the second term here is less
than ²/D. As we mentioned above, this quantity is negligible because ² can’t be larger than d, and
because we’re assuming D À d. We therefore conclude that r2 /r1 ≈ 1. In other words, r1 ≈ r2 in a
multiplicative sense. This then implies that the amplitudes of the two waves are essentially equal.
However, the phase difference equals k(r2 − r1 ) = 2π(r2 − r1 )/λ = 2π²/λ. So if ² is of the
same order as the wavelength, then the phase difference isn’t negligible. So r2 is not equal to r1
2 The fancier terminology for these two cases comes from the people who did pioneering work in them:

the Fraunhofer limit for far-field, and the Fresnel limit for near-field. The correct pronunciation of “Fresnel”
appears to be fray-NELL, although many people say freh-NELL.
4 CHAPTER 9. INTERFERENCE AND DIFFRACTION

in an additive sense. To sum up, the multiplicative comparison of r2 and r1 (which is relevant for
the amplitudes) involves the comparison of ² and D, and we know that ²/D is negligible in the
far-field limit. But the additive comparison of r2 and r1 (which is relevant for the phases) involves
the comparison of ² and λ, and ² may very well be of the same order as λ. ♣

Having found the phase difference in Eq. (1), we can now find the total value of the wave
at point P . Let AP be the common amplitude of each of the two waves at P . Then up to
an overall phase that depends on when we pick the t = 0 time, the total (complex) wave at
P equals

Etot (P ) = AP ei(kr1 −ωt) + AP ei(kr2 −ωt)

= AP (eikr1 + eikr2 )e−iωt . (2)

Our goal is to find the amplitude of the total wave, because that (or rather the square of it)
yields the intensity of the total wave at point P . We can find the amplitude by factoring
out the average of the two phases in the wave, as follows.
³ ´
Etot (P ) = AP eik(r1 −r2 )/2 + e−ik(r1 −r2 )/2 eik(r1 +r2 )/2 e−iωt
µ ¶ ¡ ¢
k(r1 − r2 ) i k(r1 +r2 )/2−ωt
= 2AP cos e
2
µ ¶ ¡ ¢
kd sin θ
= 2AP cos ei k(r1 +r2 )/2−ωt , (3)
2

where we have used k(r2 − r1 ) = kd sin θ from Eq. (1). The amplitude is the coefficient of
the exponential term, so we see that the total amplitude at P is
µ ¶ µ ¶
kd sin θ kd sin θ
Atot (P ) = 2AP cos −→ Atot (θ) = 2A(θ) cos , (4)
2 2

where we have rewritten AP as A(θ), and Atot (P ) as Atot (θ), to emphasize the dependence
on θ. Note that the amplitude at θ = 0 is 2A(0) cos(0) = 2A(0). Therefore,
µ ¶
Atot (θ) A(θ) kd sin θ
= cos (5)
Atot (0) A(0) 2

The intensity is proportional to the square of the amplitude, which gives

µ ¶
Itot (θ) A(θ)2 kd sin θ
= cos2 (6)
Itot (0) A(0)2 2
√
Since the amplitude of a cylindrically propagating wave is proportional to 1/ r, we have
p s s
A(θ) 1/ r(θ) r(0) D √
= p = = = cos θ. (7)
A(0) 1/ r(0) r(θ) D/ cos θ

Therefore,
µ ¶
Itot (θ) kd sin θ
= cos θ cos2
Itot (0) 2
µ ¶
πd sin θ
= cos θ cos2 . (8)
λ
9.1. TWO-SLIT INTERFERENCE 5

This result holds for all values of θ, even ones that approach 90◦ . The only approximation
we’ve made so far is the far-field one, which allows us to say that (1) the amplitudes of the
waves from the two slits are essentially equal, and (2) the two paths are essentially parallel.
The far-field approximation has nothing to do with the angle θ.
If we want to write Itot in terms of√the distance x from the√midpoint of the screen,
instead of θ, then we can use cos θ = D/ x2 + D2 and sin θ = x/ x2 + D2 . This gives
µ ¶
Itot (x) D 2 xkd
= √ cos √
Itot (0) x2 + D2 2 x2 + D2
µ ¶
D 2 xπd
= √ cos √ . (9)
x2 + D2 λ x2 + D2
Plots of Itot (x)/Itot (0) are shown in Fig. 6, for d values of (.01)λ, (0.5)λ, 5λ, and 50λ.

I(x)/I(0) I(x)/I(0)
1.0 1.0
0.8 (d = .01 λ) 0.8 (d = 0.5 λ)
0.6 0.6
0.4 0.4
0.2 0.2
x/D x/D
-4 -2 0 2 4 -4 -2 0 2 4

I(x)/I(0) I(x)/I(0)
1.0 1.0
(d = 5λ) (d = 50λ)
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
x/D x/D
-4 -2 0 2 4 -4 -2 0 2 4

Figure 6

As you can see from the first plot, if d is much smaller than λ, the interference pattern
isn’t too exciting, because the two paths are essentially in phase with each other. The
most they can be out of phase is when θ → 90◦ (equivalently, x → ∞), in which case the
pathlength difference is simply b = (.01)λ, which is only 1% of a full phase. Since we have
d ¿ λ in the first plot, the cosine-squared term √ in Eq. (9) is essentially equal to 1, so
the curve reduces to a plot of the function D/ x2 + D2 . It decays to zero for the simple
intuitive reason that the farther
√ we get away from the slit, the smaller the amplitude is
(more precisely, A(θ) = A(0) cos θ). In this d ¿ λ case, we effectively have a single light
source from a single slit; interference from the two slits is irrelevant because the waves can
never be much out of phase. The function in the first plot is simply the intensity we would
see from a single slit.
The d = (0.5)λ plot gives the cutoff case when there is barely destructive interference at
x = ∞. (Of course, the amplitude of both waves is zero there, so the total intensity is zero
anyway.) The d = 5λ and d = 50λ plots exhibit noticeable interference. The local maxima
occur where the two pathlengths differ by an integral multiple of the wavelength. The local
minima √ occur where the two pathlengths differ by an odd multiple of half of the wavelength.
The D/ x2 + D2 √ function in Eq. (9) is the envelope of the cosine-squared function. In the
first plot, the D/ x2 + D2 function is all there is, because the cosine-squared function
 I(x)/I(0)
1.0 6 CHAPTER 9. INTERFERENCE AND DIFFRACTION
0.8
0.6
essentially never deviates from 1. But in the d = 5λ and d = 50λ cases, it actually goes
0.4
through some cycles.
0.2
x/D Fig. 7 shows a close-up
√ version of the d = 50λ case. For small x (equivalently, for small θ),
- 0.2 - 0.1 0.0 0.1 0.2 the ratio A(θ)/A(0) = cos θ is essentially equal to 1, so the envelope is essentially constant.
(d = 50λ) We therefore simply have a cosine-squared function with a nearly-constant amplitude. In
practice, we’re usually concerned only with small x and θ values, in which case Eqs. (8) and
Figure 7 (9) become
µ ¶
Itot (θ) θπd
≈ cos2 (for θ ¿ 1)
Itot (0) λ
µ ¶
Itot (x) xπd
I(x)/I(0) ≈ cos2 (for x ¿ D) (10)
Itot (0) λD
1.0
0.8 For the remainder of this chapter, we will generally work in this small-angle approximation.
0.6 So we won’t need the exact (at least exact in the far-field approximation) results in Eqs. (8)
0.4
and (9).
0.2
x
The plot of Itot (x)/Itot (0) from Eq. (10) is shown in Fig. 8. The maxima occur at integer
0 ___ 2λD
λD ____ multiples of λD/d. It makes sense that the spacing grows with λ, because the larger λ is,
d d the more tilted the paths in Fig. 5 have to be to make the difference in their lengths (which
is d sin θ) be a given multiple of λ. The approximations we’ve made in Fig. 8 are that we’ve
Figure 8 ignored the facts that as we move away from the center of the screen, (a) the amplitude
A(θ) of the two waves decreases, and (b) the peaks become spaced farther apart. You can
compare Fig. 8 with the third and fourth plots in Fig. 6.
Remember that the small-angle approximation we’ve made here is different from the
“far-field” approximation. The far-field approximation is the statement that the distances
from the two slits to a given point P on the screen are essentially equal, multiplicatively.
This holds if d ¿ D. (We’ll eventually drop this assumption in Section 9.5 when we discuss
the near-field approximation.) The small-angle approximation that leads to Eq. (10) is the
statement that the distances from the two slits to different points on the screen are all
essentially equal. This holds if x ¿ D, or equivalently θ ¿ 1. Note that the small-angle
approximation has no chance of being valid unless the far-field approximation already holds.
Remark: The small-angle approximation in Eq. (10) shoves the A(θ) dependence in Eq. (6) under
the rug. Another way to get rid of this dependance is to use a cylindrical screen instead of a flat
screen, with the axis of the cylinder coinciding with the slits. So in Fig. 4 the screen would be
represented by a semicircle in the plane of the page, with the slits located at the center. In the
far-field limit, all of the paths in different θ directions now have the same length, multiplicatively.
(The difference in pathlengths to a given point on the screen is still d sin θ.) So A(θ) = A(0) for all
θ, and the A’s cancel in Eq. (6). Note, however, that the spacing between the local maxima on the
cylindrical screen still isn’t uniform, because they occur where sin θ = λ/d. And sin θ isn’t a linear
function of θ. At any rate, the reason why we generally work in terms of a flat screen isn’t that
there is anything fundamentally better about it compared with a cylindrical screen. It’s just that
in practice it’s easier to find a flat screen. ♣

9.2 N -slit interference

9.2.1 Standard derivation
Let’s now looks at the case where we have a general number, N , of equally-spaced slits,
instead of 2. The setup is shown in Fig. 9 for the case of N = 5.
9.2. N -SLIT INTERFERENCE 7

wall screen
P
r1

d
d rN
d
d

Figure 9
Similar to the N = 2 case above, we will make the far-field assumption that the distance n
cree
to the screen is much larger than the total span of the slits, which is (N − 1)d. We can then to s away
say, as we did in the N = 2 case, that all the paths to a given point P on the screen have far
essentially the same length in a multiplicative (but not additive) sense, which implies that
the amplitudes of the waves are all essentially equal. And we can also say that all the paths
are essentially parallel. A closeup version near the slits is shown in Fig. 10. Additively, each
pathlength is d sin θ longer than the one right above it. So the lengths take the form of d θ
rn = r1 + (n − 1)d sin θ.
To find the total wave at a given point at an angle θ on the screen, we need to add up d sinθ
the N individual waves (call them En ). The procedure is the same as in the N = 2 case, d
except that now we simply have more terms in the sum. In the N = 2 case we factored out
the average of the phases (see Eq. (3)), but it will now be more convenient to factor out the
d
phase of the top wave in Fig. 9 (the r1 path). The total wave at an angle θ on the screen is
then (with A(θ) being the common amplitude of all the waves)
N
X N
X d
i(krn −ωt) θ
Etot (θ) = En = A(θ)e
n=1 n=1
N
Figure 10
X
i(kr1 −ωt) ik(n−1)d sin θ
= A(θ)e e . (11)
n=1

With z ≡ eikd sin θ , the sum here is 1 + z + z 2 + · · · + z N −1 . The sum of this geometric series
is
zN − 1 eikN d sin θ − 1
=
z−1 eikd sin θ − 1
eik(N/2)d sin θ eik(N/2)d sin θ − e−ik(N/2)d sin θ
= ·
eik(1/2)d sin θ eik(1/2)d sin θ − e−ik(1/2)d sin θ
¡1 ¢
ik((N −1)/2)d sin θ sin ¡2 N kd sin θ¢ .
= e · (12)
sin 12 kd sin θ
Substituting this into Eq. (11) yields a total wave of
¡ ¢
sin 12 N kd sin θ ³ i(kr1 −ωt) ik((N −1)/2)d sin θ ´
Etot (θ) = A(θ) ¡ ¢ e e . (13)
sin 12 kd sin θ
The amplitude is the coefficient of the exponential factors, so we have
¡ ¢
sin 12 N kd sin θ sin(N α/2)
Atot (θ) = A(θ) ¡1 ¢ ≡ A(θ) (14)
sin 2 kd sin θ sin(α/2)
8 CHAPTER 9. INTERFERENCE AND DIFFRACTION

where
2πd sin θ
α ≡ kd sin θ = . (15)
λ
Since adjacent pathlengths differ by d sin θ, the physical interpretation of α is that it is the
phase difference between adjacent paths.
What is the value of Atot (θ) at the midpoint of the screen where θ = 0 (which implies
α = 0)? At α = 0, Eq. (14) yields Atot (θ) = 0/0, which doesn’t tell us much. But we can
obtain the actual value by taking the limit of small α. Using sin ² ≈ ², we have
sin(N α/2) N α/2
Atot (0) = lim Atot (θ) = lim A(θ) = A(0) = A(0) · N. (16)
θ→0 α→0 sin(α/2) α/2
It is customary to deal not with the amplitude itself, but rather with the amplitude relative
to the amplitude at θ = 0. Combining Eqs. (14) and (16) gives
Atot (θ) A(θ) sin(N α/2)
= · . (17)
Atot (0) A(0) N sin(α/2)
Since we generally deal with small angles, we’ll ignore the variation in the A(θ) coefficient.
In other words, we’ll set A(θ) ≈ A(0). This gives

Atot (α) sin(N α/2)

≈ (for small θ) (18)
Atot (0) N sin(α/2)

The intensity at θ relative to the intensity at θ = 0 is then

µ ¶2
Itot (α) sin(N α/2)
≈ (for small θ) (19)
Itot (0) N sin(α/2)

Even for large angles, the effect of A(θ) is to simply act as an envelope function of the
oscillating sine functions. We can always bring A(θ) back in if we want to, but the more
interesting behavior of Atot (θ) is the oscillatory part. We’re generally concerned with the
locations of the maxima and minima of the oscillations and not with the actual value of
the amplitude. The A(θ) factor doesn’t affect these locations.3 We’ll draw a plot of what
Itot (α)/Itot (0) looks like, but first a remark.
Remark: Technically, we’re being inconsistent here in our small-angle approximation, because
although we set A(θ) = A(0) (which from Eq. (7) is equivalent to setting cos θ = 1), we didn’t set
sin θ = θ in the expression for α. To be consistent, we should approximate α = kd sin θ by α = kd·θ.
The reason why we haven’t made this approximation is that we want to keep the locations of the
bumps in the interference pattern correct, even for large θ. And
√ besides, the function A(θ) depends
on the nature of the screen. A flat screen has A(θ)/A(0) = cos θ, which decreases with θ, while
a cylindrical screen has A(θ)/A(0) = 1, which is constant. Other shapes yield other functions of
θ. But they’re all generally slowly-varying functions of θ, compared with the oscillations of the
sin(N α/2) function (unless you use a crazily-shaped wiggly screen, which you have no good reason
to do). The main point is that the function A(θ) isn’t an inherent property of the interference
pattern; it’s a property of the screen. On the other hand, the angular locations of the maxima and
minima of the oscillations are an inherent property of the pattern. So it makes sense to keep these
locations exact and not lose this information when making the small-angle approximation. If you
want, you can write the intensity in Eq. (19) as
µ ¶2
Itot (θ) sin(N α/2)
= F (θ) (where α ≡ kd sin θ), (20)
Itot (0) N sin(α/2)
3 Strictly speaking, A(θ) does affect the locations of the maxima in a very slight manner (because when

taking the overall derivative, the derivative of A(θ) comes into play). But A(θ) doesn’t affect the locations
of the minima, because those are where Itot (α) is zero.
9.2. N -SLIT INTERFERENCE 9
I(α)/I(0)
1.0
where F (θ) is a slowly-varying function of θ that depends on the shape of the screen. We will 0.8
generally ignore this dependence and set F (θ) = 1. ♣ 0.6
0.4
What does the Itot (α)/Itot (0) ratio in Eq. (19) look like as a function of α? The plot 0.2
for N = 4 is shown in Fig. 11. If we’re actually talking about small angles, then we have α
α = kd sin θ ≈ kd · θ. But the distance from the center of the screen is x = D tan θ ≈ D · θ.
(N = 4) _ π 3π
π __ 2π
So for small angles, we have α ∝ x. You can therefore think of Fig. 11 as showing the actual 2 2
intensity on the screen as a function of x (up to a scaling constant).
Note that although we generally assume θ to be small, α is not necessarily small, because Figure 11
α ≡ kd sin θ involves a factor of k which may be large. Said in another way, α is the phase
difference between adjacent slits, so if k is large (more precisely, if λ ¿ d), then even a small
angle θ can lead to a pathlength difference (which is d sin θ) equal to λ. This corresponds
to a phase difference of α = kd sin θ = (2π/λ)d sin θ = 2π. Consistent with this, the values
on the horizontal axis in Fig. 11 are on the order of π (that is, they are not small), and the
first side peak is located at 2π.
A number of things are evident from both Fig. 11 and Eq. (19):
1. The value of Itot (α)/Itot (0) at θ = 0 is 1, by construction.
2. Itot (α)/Itot (0) has a period of 2π in α. The sine function in the denominator picks
up a minus sign when α increases by 2π, and likewise in the numerator if N is odd.
But an overall minus sign is irrelevant because the intensity involves the squares of
the sines.
3. Itot (α) has zeroes whenever N α/2 is a multiple of π, that is, whenever N α/2 = mπ =⇒
α = 2mπ/N , which means that α is an even multiple of π/N . The one exception to
this is when α/2 is also a multiple of π, that is, when α/2 = m0 π =⇒ α = 2m0 π,
because then the denominator in Eq. (19) is also zero. (In this case, Eq. (16) tells us
that the value at θ = 0 is 1. And likewise at any integer multiple of 2π. These are the
locations of the main peaks.) In the N = 4 case in Fig. 11, you can see that the zeros
do indeed occur at
0π¶, 2π , 4π , 6π , 8π
¶ ¶, 10π , 12π , 14π , 16π
¶ ¡¡, · · · . (21)
¶4 4 4 4 ¶4 4 4 4 ¡4
And likewise for negative values. In general, the number of zeros between the main
peaks is N − 1.
4. If you take the derivative of Itot (α), you will find that the local maxima (of the small
bumps) occur when tan(N α/2) = N tan(α/2). This has to be solved numerically.
However, for large N , the solutions for α are generally very close to the odd multiples
of π/N (except for values of the form of 2π ± π/N ); see Problem [to be added]. In
other words, the local maxima are approximately right between the local minima (the
zeros) which themselves occur exactly at the even multiples of π/N , except at the
integral multiples of 2π where the main peaks are. In Fig. 11 you can see that the
small bumps do indeed occur at approximately

1π
¶¶, 3π , 5π , 7π
¶¶, 9π
¶¶, 11π , 13π , 15π
¡¡, 17π
¡¡, 19π , · · · . (22)
¶4 4 4 ¶4 ¶4 4 4 ¡4 ¡4 4
And likewise for negative values. In general, the number of little bumps between the
main peaks is N − 2.
5. The little bumps in Fig. 11 have the same height, simply because there are only two
of them. For larger values of N , the bump sizes are symmetric around α = π (or in
 I(α)/I(0)
1.0 10 CHAPTER 9. INTERFERENCE AND DIFFRACTION
0.8
0.6
general any multiple of π). They are the shortest there (because the denominator in
0.4
0.2
Eq. (19) is largest at α = π), and they grow in size as they get closer to the main
α peaks. Fig. 12 shows the interference pattern for N = 8.
(N = 8) π 2π Note that if d < λ, then α ≡ kd sin θ = (2π/λ)d sin θ < 2π sin θ ≤ 2π. So α can’t achieve
the value of 2π, which means that none of the tall side peaks in Fig. 11 exist. We have only
Figure 12 one tall peak at α = 0, and then a number of small peaks. This makes sense physically,
because the main peaks occur when the waves from all the slits are in phase. And if d < λ
there is no way for the pathlengths to differ by λ, because the difference can be at most d
(which occurs at θ = 90◦ ). In general, the upper limit on α is kd, because sin θ can’t exceed
1. So no matter what the relation between d and λ is, a plot such as the one in Fig. 12
exists out to the α = kd point (which corresponds to θ = 90◦ ), and then it stops.
In the case of N = 2, we should check that the expression for Itot (α)/Itot (0) in Eq.
(19) reduces properly to the expression in Eq. (6) (with A(θ) set equal to A(0)). Indeed, if
N = 2, then the quotient in Eq. (19) becomes sin(2·α/2)/2¡ sin(α/2). Using
¢ the double-angle
formula in the numerator turns this into cos(α/2) = cos (1/2)kd sin θ , which agrees with
Eq. (6).

9.2.2 Geometric construction

Let’s now derive the amplitude in Eq. (14) in a different way. It turns out that there is an
extremely informative geometric way of seeing how this amplitude arises. The main task in
finding the amplitude is calculating the sum in Eq. (11). With α ≡ kd sin θ, this sum is
N
X
(N=4)
eik(n−1)d sin θ = 1 + eiα + ei2α + · · · + ei(N −1)α . (23)
n=1
ei 3α ei 2α
ei α Each term in this sum is a complex number with magnitude 1. If we plot these numbers
α α as vectors in the complex plane, they make angles of 0, α, 2α, etc. with respect to the x
α 1 axis. For example, in the case of N = 4 we might have the unit vectors shown in Fig. 13.
(Remember that α depends on θ, which depends on where the point P is on the screen.
So for any point P , we have a set of N vectors in the plane. The angle α between them
Figure 13 increases as P moves farther off to the side.) The easiest way to add up these vectors is to
put them tail-to-head, as shown in Fig. 14. Each of the unit vectors is tilted at an angle α
Rei φ with respect to the one before it. The desired sum is the thick vector shown. As with any
complex number, we can write this sum as a magnitude times a phase, that is, as Reiφ .
r 1 The total amplitude Atot (θ) equals R times the A(θ) in Eq. (11), because the phase
α ei(kr1 −ωt) in Eq. (11) and the phase eiφ in the sum don’t affect the amplitude. So our goal
α is to find R, which can be done in the following way. (If you want to find the value of φ, see
αα Problem [to be added].)
1 The thick vector in Fig. 14 is the base of an isosceles triangle with vertex angle 4α, which
in general is N α. So we have
1 R = 2 · r sin(N α/2), (24)
1
where r is the length shown in the figure. But from looking at any one of the four thinner
Figure 14 isosceles triangles with vertex angle α and base 1, we have

1 = 2 · r sin(α/2). (25)

Taking the quotient of the two preceding equations eliminates the length r, and we arrive
at
sin(N α/2)
R= . (26)
sin(α/2)
9.2. N -SLIT INTERFERENCE 11

This reproduces Eq. (14), because the total amplitude Atot (θ) equals R · A(θ).
As time increases, the whole picture in Fig. 14 rotates clockwise in the plane, due to the
−ωt in the phase. There is also a phase shift due to the k1 r and φ terms in the phase, but
this simply affects the starting angle. Since all the little vectors keep their same relative
orientation, the figure keeps its same shape. That is, it rotates as a rigid “object.” The sum
(the thick vector) therefore always has the same length. This (constant) length is therefore
the amplitude, while the (changing) horizontal component is the real part that as usual
gives the actual physical wave.
The above geometric construction makes it easy to see why the main peaks and all the
various local maxima and minima appear in Fig. 12. The main peaks occur when α is a
multiple of 2π, because then all the little vectors point in the same direction (rightward at
a given instant, if the first little vector points to the right at that instant). The physical
reason for this is that α = m · 2π implies that

2πd sin θ
kd sin θ = 2mπ =⇒ = 2mπ =⇒ d sin θ = mλ. (27)
λ
This says that the difference in pathlengths from adjacent slits is a multiple of the wave-
length, which in turn says that the waves from all of the slits constructively interfere. Hence
the maximal amplitude.
A local minimum (a zero) occurs if the value of α is such that the chain of little vectors
in Fig. 14 forms a closed regular polygon (possibly wrapped around multiple times). In this
case the sum (the thick vector in Fig. 14) has no length, so the amplitude is zero. The
“polygons” for the seven zeros in the N = 8 case in Fig. 12 are shown in Fig. 15. We’ve
taken the first of the vectors to always point horizontally to the right, although this isn’t
necessary. We’ve drawn the figures slightly off from the case where the sum of the vectors
is zero, to make it easier to see what’s going on. The last three figures are mirror images of
the first three.

(α= π/4)

(α= π/2)
(α= 3π/4)
(α= π)

(α= 5π/4) (α= 3π/2) (α= 7π/4)

Figure 15
12 CHAPTER 9. INTERFERENCE AND DIFFRACTION

The local maxima occur between the local minima. In the case of large N , it’s easy to
determine the approximate locations of these maxima. For large N , the vectors form an
essentially smooth curve, and the maxima occur roughly when the amplitude is a diameter
of a circle. The first few of these occurrences are shown in Fig. 16 for the case of N = 50.
We’ve made the curve spiral slightly inward so that you can see how many times it wraps
around. But in reality (in the far-field limit), the curve just keeps tracing over itself.

(N=50)

(α=3π/50) (α=5π/50) (α=7π/50)

(α=0 for comparison)

Figure 16

The maxima don’t occur exactly at the diameters, because the circle shrinks as the
little vectors wrap around further as α increases, so there are competing effects. But it
is essentially the diameter if the wrapping number is large (because in this case the circle
hardly changes size as the amplitude line swings past the diameter, so the shrinking effect
is basically nonexistent). So we want the vectors to wrap (roughly) 3/2, 5/2, 7/2, etc. times
around the circle. Since each full circle is worth 2π radians, this implies that the total angle,
which is N α, equals 3π, 5π, 7π, etc. In other words, α is an odd multiple of π/N , excluding
π/N itself (and also excluding the other multiples adjacent to multiples of 2π). This agrees
with the result in the paragraph preceding Eq. (22). The amplitude of the main peaks that
occur when α equals zero or a multiple of 2π is also shown in Fig. 16 for comparison. In
this case the circular curve is unwrapped and forms a straight line. The little tick marks
indicate the N = 50 little vectors.

9.2.3 Diffraction gratings

A diffraction grating is a series of a large number, N , of slits with a very small spacing d
between them. If a source emits light that consists of various different wavelengths, a diffrac-
tion grating provides an extremely simple method for determining what these wavelengths
are.
As we saw above, the N -slit interference pattern consists of the main peaks, plus many
smaller peaks in between. However, we will be concerned here only with the main peaks,
because these completely dominate the smaller peaks, assuming N is large. Let’s justify
this statement rigorously and discuss a few other things, and then we’ll do an example.
From the discussion that led to Eq. (22), the smaller peaks occur when α takes on values
that are approximately the odd multiples of π/N (except for values of the form 2π ± π/N ),
that is, when α equals 3π/N , 5π/N , etc. The corresponding values of Itot (α)/Itot (0) are
obtained from Eq. (19). The numerator equals (±1)2 , and since N is large we can use a
9.2. N -SLIT INTERFERENCE 13

small-angle approximation in the denominator, which turns the denominator into N α/2.
The resulting values of Itot (α)/Itot (0) are then (2/3π)2 , (2/5π)2 , etc. Note that these are
independent of N .
The first of the side peaks isn’t negligible compared with the main peak (it’s about I(α)/I(0)
(2/3π)2 = 4.5% as tall). But by the 10th peak, the height is negligible (it’s about 0.1% as 1.0
tall). However, even though the first few side peaks aren’t negligible, they are squashed very 0.8
0.6
close to the main peaks if N is large. This follows from the fact that the spacing between
0.4
the main peaks is ∆α = 2π, whereas the side peaks are on the order of π/N away from 0.2
the main peaks. The figure for N = 20 is shown in Fig. 17. We can therefore make the α
approximation that the interference pattern is non-negligible only at (or extremely close to) -2π 0 2π
the main peaks where α is a multiple of 2π. (N = 20)
When dealing with diffraction gratings, we’re generally concerned only with the location
of the bright spots in the interference pattern, and not with the actual intensity. So any extra Figure 17
intensity from the little side peaks is largely irrelevant. And since they’re squashed so close
to the main peaks, it’s impossible to tell that they’re distinct bumps anyway. The location
of the main peaks tells us what the various wavelengths are, by using kd sin θ = 2π =⇒
(2π/λ)d sin θ = 2π =⇒ λ = d sin θ. The intensity tells us how much of each wavelength the
light is made of, but for most purposes we’re not so concerned about this.
Remarks:
1. A diffraction grating should more appropriately be called an “interference grating,” because
it is simply an example of N -slit interference. It is not an example of diffraction, which we
will define and discuss in Section 9.3.1. We’ll see there that a feature of a diffraction pattern
is that there are no tall side peaks, whereas these tall side peaks are the whole point of an
“interference grating.” However, we’ll still use the term “diffraction grating” here, since this
is the generally accepted terminology.
2. If we view the interference pattern on a screen, we know that it will look basically like Fig.
17 (we’ll assume for now that only one wavelength is involved). However, if you put your eye
right behind the grating, very close to it, what do you see? If you look straight at the light
directions of first
source, then you of course see the source. But if you look off at an angle (but still through few maxima
the grating; so your eye has to be close to it), then you will also see a bright spot there. And
you will also see bright spots at other angles. The number of spots depends on the relation
between the wavelength and the spacing. We’ll discuss a concrete case in the example below.
θ1
The angles at which you see the spots are the same as the angles of the main peaks in Fig. 17,
for the following reason. Fig. 18 shows the typical locations of the first few main peaks in the θ2
interference pattern from a standard set of slits contained in a small span in a wall. Imagine
putting additional sets of slits in the wall at locations such that a given spot on the screen
(your eye) is located at the angles of successive off-center peaks. This scenario is shown in
Fig. 19. Each set of slits also produces many other peaks, of course, but you don’t see them
because your eye is at only one location.
A diffraction grating is a continuous set of slits, but most of the slits are irrelevant. The
only slits that matter are the ones that are located at positions such that the angle to your
eye is one of the main-peak angles. In other words, we can replace the entire wall in Fig. 19 Figure 18
with a continuous set of slits, and you will still see the same thing. Only the small regions of
slits shown in the figure will produce bright spots. In short, a diffraction grating acts like a
collection of interference setups at specific locations in the grating.
3. You might be concerned that if your eye is close enough to the grating, then the far-field
approximation (and hence all of the result so far in this chapter) will be invalid. After all, the
distance D from your eye to the grating isn’t large compared with the total span of the slits
in the grating. However, the far-field approximation does indeed still hold, because from the
previous remark we’re not concerned with the total span of the slits in the grating, but rather
with the span of a small region near each of the main-peak angles. Assuming that the spacing θ2
between the lines in the grating is very small (it’s generally on the order of 10−6 m), the span
θ1
eye

Figure 19
 14 CHAPTER 9. INTERFERENCE AND DIFFRACTION

of a few hundred lines will still be very small compared with the distance from your eye to
the grating (assuming that your eyelashes don’t touch it). So the far-field approximation still
holds. That is, the distances from these few hundred slits to your eye are all essentially equal
(multiplicatively).
4. In reality, most diffraction gratings are made by etching regularly-spaced lines into the mate-
rial. The exact details of the slits/etchings are’t critical. Any periodic structure with period
d will do the trick. The actual intensities depend on the details, but the locations of the main
peaks don’t. This follows from the usual argument that if d sin θ is a multiple of λ, then there
is constructive interference from the slits (whatever they may look like).
5. Problem 9.1 shows that it doesn’t matter whether or not the incident light is normal to the
wall (which is the diffraction grating here), as long as the deviation angle is small. If we
measure all angles relative to the incident angle, then all of our previous result still hold.
This is fortunate, of course, because if you hold a diffraction grating in front of your eye, it
is highly unlikely that you will be able orient it at exactly a 90◦ angle to the line between it
and the light source. ♣

Example (Blue and red light): A diffraction grating has 5000 lines per cm. Consider
a white-light source (that is, it includes all wavelengths), and assume that it is essentially a
point source far away. Taking the wavelengths of blue and red light to be roughly 4.5·10−5 cm
and 7 · 10−5 cm, find the angles at which you have to look to the side to see the off-center
blue and red maxima. What is the total number of maxima for each color that you can
theoretically see on each side of the light source?

Solution: We basically have to do the same problem here twice, once for blue light and once
for red light. As usual, the main peaks occur where the difference in pathlengths from adjacent
slits is an integral multiple of the wavelength. So we want d sin θ = mλ. (Equivalently, we
want α = m · 2π, which reduces to d sin θ = mλ.) We therefore want sin θ = mλ/d, where
d = (1 cm)/5000 = 2 · 10−4 cm.
For blue light, this gives sin θ = m(4.5 · 10−5 cm)/(2 · 10−4 cm) = m(0.225). So we have the
following four possible pairs of m and θ values:

(m, θ) : (1, 13.0◦ ) (2, 26.7◦ ) (3, 42.5◦ ) (4, 64.2◦ ) (28)

There are only four possible angles (plus their negatives), because m = 5 gives a value of
sin θ that is larger than 1.
For red light, we have sin θ = m(7 · 10−5 cm)/(2 · 10−4 cm) = m(0.35). So we have the
following two possible pairs of m and θ values:

to far away source (m, θ) : (1, 20.5◦ ) (2, 44.4◦ ) (29)

grating There are only two possible angles (plus their negatives), because m = 3 gives a value of sin θ
that is larger than 1. The red angles are larger than the corresponding blue angles because
the red wavelength is longer, so it takes a larger angle to make adjacent pathlengths differ
θ=0 1 1 bands by a wavelength (or two wavelengths, etc.).
2
13.0 The rest of the spectrum falls between blue and red, so we obtain rainbow bands of colors.
20.5
26.7
3
2 Note, however, that the first band (from 13.0◦ to 20.5◦ , although the endpoints are fuzzy) is
42.5
44.4 the only “clean” band that doesn’t overlap with another one. The second band ends at 44.4◦ ,
4 which is after the third band starts at 42.5◦ . And the third band doesn’t even finish by the
64.2
time the angle hits 90◦ . Your viewing angle has to be less than 90◦ , of course, because you
have to be looking at least a little bit toward the grating. The angles of the various bands
eye
are shown in Fig. 20. The mirror images of these angles on the left side work too.
θ = 90
Figure 20
9.3. DIFFRACTION FROM A WIDE SLIT 15

9.3 Diffraction from a wide slit

9.3.1 Derivation
We’ll now discuss what happens when a plane wave impinges on just one wide slit with width
a, instead of a number of infinitesimally thin ones. See Fig. 21. We’ll find that if the width a
isn’t negligible compared with the wavelength λ, then something interesting happens. The
interference pattern will depend on a in a particular way, whereas it didn’t depend on the
infinitesimal width in the previous sections. We’ll keep working in the far-field limit, which
here means that D À a.

wall screen
P

a θ
D

Figure 21

By Huygen’s principle, we can consider the wide slit to consist of an infinite number of
line sources (or point sources, if we ignore the direction perpendicular to the page) next to
each other, each creating a cylindrical wave. In other words, the diffraction pattern from
one continuous wide slit is equivalent to the N → ∞ limit of the N -slit result in Eq. (19).
So we’ve already done most of the work we need to do. We’ll present three ways we can go
about taking the continuum limit. But first some terminology.
The word diffraction refers to a situation with a continuous aperture. The word inter-
ference refers to a situation involving two or more apertures whose waves interfere. On one
hand, since diffraction is simply the N → ∞ limit of interference, there is technically no need
to introduce a new term for it. But on the other hand, a specific kind of pattern arises, so it
makes sense to give it its own name. Of course, we can combine interference and diffraction
by constructing a setup with waves coming from a number of wide apertures. We’ll deal
with this in Section 9.4. A name that causes confusion between the words “interference”
and “diffraction” is the diffraction grating that we discussed above. As we mentioned in the
first remark in Section 9.2.3, this should technically be called an interference grating.

N → ∞ limit

For our first derivation of the diffraction pattern, we’ll take the N → ∞ limit of Eq. (19).
The α in Eq. (19) equals kd sin θ. But if we imagine the slit of width a to consist of N
infinitesimal slits separated by a distance d = a/N , then we have α = k(a/N ) sin θ. 4 (The
N here should perhaps be N − 1, depending on where you put the slits, but this is irrelevant
4 Of course, if we actually have infinitesimal slits separated by little pieces of wall, then the intensity will

go down. But this doesn’t matter since our goal is only to find the relative intensity Itot (θ)/Itot (0). As
we’ll see below, if the distance d = a/N is much smaller than the wavelength (which it is, in the N → ∞
limit) then we actually don’t even need to have little pieces of wall separating the slits. The slits can bump
right up against each other.
 16 CHAPTER 9. INTERFERENCE AND DIFFRACTION

in the N → ∞ limit.) Plugging this value of α into Eq. (19) gives

à ¡ ka sin θ ¢ !2
Itot (θ) sin
≈ ¡ ka2 sin θ ¢ . (30)
Itot (0) N sin 2N

In the N → ∞ limit, we can use sin ² ≈ ² in the denominator to obtain

à ¡1 ¢ !2 µ ¶2
Itot (θ) sin 2 ka sin θ Itot (θ) sin(β/2)
≈ 1 =⇒ ≈ (31)
Itot (0) 2 ka sin θ
Itot (0) β/2

where
2πa sin θ
β ≡ ka sin θ = . (32)
λ
Another convention is to define β ≡ (1/2)ka sin θ, in which case the result in Eq. (31) takes
¡ ¢2
the simpler form of (sin β)/β . The reason why we chose β ≡ ka sin θ here is because it
sinc(x) parallels the definition of α in Eq. (15). The results in Eqs. (19) and (31) are then similar,
1.0
in that they both involve factors of 2. The physical meaning of α in Eq. (15) is that it is the
0.8
0.6 phase difference between adjacent paths. The physical meaning of β is that is it the phase
0.4 difference between the paths associated with the endpoints of the wide slit of width a.
0.2
x
The function (sin x)/x is knows as the “sinc” function, sinc(x) ≡ (sin x)/x. A plot is
0.2 shown in Fig. 22. It is a sine function with a 1/x envelope. The result in Eq. (31) can
2π 4π 6π therefore be written as Itot (θ)/Itot (0) = sinc2 (β/2). A plot of this is shown in Fig. 23.
The factor of 2 in the argument makes the plot expanded by a factor of 2 in the horizontal
Figure 22 direction compared with the plot in Fig. 22. Since sin θ can’t exceed 1, β can’t exceed ka.
So the plot in Fig. 23 exists out to the β = ka point (which corresponds to θ = 90◦ ), and
sinc2(β/2) then it stops.
1.0 Note that the diffraction pattern has only one tall bump, whereas the interference pat-
0.8 terns we’ve seen generally have more than one tall bump (assuming that d > λ). This is
0.6
0.4 consistent with the discussion in the second-to-last paragraph in Section 9.2.1. We saw there
0.2 that if d < λ, then there is only one tall bump. And indeed, in the present case we have
β d = a/N , which becomes infinitesimal as N → ∞. So d is certainly smaller than λ.
0.2 2π 4π 6π
The zeros of Itot (θ) occur when β is a multiple of 2π (except β = 0). And since β ≡
Figure 23 2πa sin θ/λ, this is equivalent to a sin θ being a multiple of λ. We’ll give a physical reason
for this relation below in Section 9.3.2, but first let’s give two other derivations of Eq. (31).

Geometric derivation
We can give another derivation of the diffraction pattern by using the geometric construction
in Section 9.2.2. In the N → ∞ limit, the little vectors in Fig. 14 become infinitesimal,
r
so the crooked curve becomes a smooth curve with no kinks. If β = 0 (which corresponds
to θ = 0 and hence α = 0 in Fig. 14), then all of the infinitesimal vectors are in phase, so
β we get a straight line pointing to the right. If β is nonzero, then the vectors curl around,
rβ and we get something like the picture shown in Fig. 24. The bottom infinitesimal vector
corresponds to one end of the wide slit, and the top infinitesimal vector corresponds to the
other end. The pathlength difference between the ends is a sin θ, so the phase difference is
2r sin(β/2) ka sin θ, which is by definition β. This phase difference is the angle between the top and
bottom vectors in Fig. 24. But this angle equals the central angle subtended by the arc.
Figure 24 The central angle is therefore β, as shown.
Now, the amplitude Atot (0) is the length of the straight line in the β = 0 case. But
this is also the length of the arc in Fig. 24, which we know is rβ, where r is the radius of
the circle. And Atot (θ) is the sum of all the infinitesimal vectors, which is the straight line
9.3. DIFFRACTION FROM A WIDE SLIT 17

shown. From the isosceles triangle in the figure, this sum has length 2r sin(β/2). Therefore,
since the intensity is proportional to the square of the amplitude, we have
µ ¶2 µ ¶2
Itot (θ) 2r sin(β/2) sin(β/2)
= = , (33)
Itot (0) rβ β/2

in agreement with Eq. (31).

Continuous integral
We can also find the diffraction pattern by doing a continuous integral over all the phases
from the possible paths from different parts of the wide slit. Let the slit run from y = −a/2
to y = a/2. And let B(θ) dy be the amplitude that would be present at a location θ on the
screen if only an infinitesimal slit of width dy was open. So B(θ) is the amplitude (on the
screen) per unit length (in the slit). B(θ) dy is the analog of the A(θ) in Eq. (11). If we
measure the pathlengths relative to the midpoint of the slit, then the path that starts at
position y is shorter by y sin θ (so it is longer if y < 0). It therefore has a relative phase of
e−iky sin θ . Integrating over all the paths that emerge from the different values of y (through
imaginary slits of width dy) gives the total wave at position θ on the screen as (up to an
overall phase from the y = 0 point, and ignoring the ωt part of the phase)
Z a/2 ¡ ¢
Etot (θ) = B(θ) dy e−iky sin θ . (34)
−a/2
√
This is the continuous version of the discrete sum in Eq. (11). B(θ) falls off like 1/ r, where
r = D/ cos θ. However, as in Section 9.2.1, we’ll assume that θ is small, which mean that
we can let cos θ ≈ 1. (And even if θ isn’t small, we’re not so concerned about the exact
intensities and the overall envelope of the diffraction pattern.) So we’ll set B(θ) equal to
the constant value of B(0). We therefore have
Z a/2
B(0) ³ −ik(a/2) sin θ ´
Etot (θ) ≈ B(0) e−iky sin θ dy = e − eik(a/2) sin θ
−a/2 −ik sin θ
¡ ¢
−2i sin ka sin
2
θ
= B(0) ·
−ik sin θ
¡ ¢
sin 12 ka sin θ
= B(0)a · 1 . (35)
2 ka sin θ

There aren’t any phases here, so this itself is the amplitude Atot (θ). Taking the usual limit
at θ = 0, we obtain Atot (0) = B(0)a. Therefore, Atot (θ)/Atot (0) = sin(β/2)/(β/2), where
β ≡ ka sin θ. Since the intensity is proportional to the square of the amplitude, we again
arrive at Eq. (31).

9.3.2 Width of the diffraction pattern

From Fig. 23, we see that most of the intensity of the diffraction pattern is contained within
the main bump where |β| < 2π. Numerically, you can shown that the fraction of the total
area that lies under the main bump is about 90%. So it makes sense to say that the angular
half-width of the pattern is given by

2πa sin θ λ
β = 2π =⇒ = 2π =⇒ sin θ = (36)
λ a
 18 CHAPTER 9. INTERFERENCE AND DIFFRACTION

For small θ, this becomes

λ
θ≈ (for small θ) (37)
a
Note that this is inversely proportional to a. The narrower the slit, the wider the diffraction
pattern. There are two ways of understanding the sin θ = λ/a result, or equivalently why
β=2π
the intensity is zero when this relation holds.
• As indicated in Eq. (36), sin θ = λ/a is equivalent to β = 2π. So in the geometric
construction that led to Eq. (33), this means that the arc in Fig. 24 is actually a full
circle, as shown in Fig. 25. The sum of all the infinitesimal vectors is therefore zero,
so the amplitude and intensity are zero.
Figure 25
• If sin θ = λ/a, then the pathlength from one end of the slit is a sin θ = λ longer than
the pathlength from the other end, as shown in Fig. 26. So the waves from the two
n ends are in phase. You might think that this implies that there should be constructive
cree
to s away interference (which would be the case if we simply had two infinitesimal slits separated
far
by a). But in fact it’s exactly the opposite in the present case of a continuous wide slit.
a θ We have complete destructive interference when the whole slit is taken into account,
for the following reason.
θ Imagine dividing the slit into two halves, as shown in Fig. 27. For every path in the
a sinθ = λ upper half (the lightly shaded region), there is a path in the lower half (the darkly
shaded region) that is λ/2 longer. So the two waves are exactly out of phase. Three
Figure 26 pairs of dotted lines are shown. The waves therefore cancel in pairs throughout the
slit, and we end up with zero amplitude. This is equivalent to saying that the phases
cancel at diametrically opposite points in Fig. 25.
You can quickly show by taking a derivative that the local maxima of I(β)/I(0) =
¡ ¢2
(sin(β/2))/(β/2) occur when tan β/2 = β/2. This must be solved numerically, but you
can get a sense of what the roots are by plotting the functions y = β/2 and y = tan β/2,
and then looking at where the curves intersect. In Fig. 28 the intersections are indicated
a by the dots, and you can see that the associated β values (in addition to the β = 0 root)
are close to 3π, 5π, 7π, etc. (the larger β is, the better these approximations are). These
maxima therefore occur roughly halfway between the zeros, which themselves occur when
β equals (exactly) 2π, 4π, 6π, etc. The task of Problem [to be added] is to show how the
3π, 5π, 7π, etc. values follow from each of the above two bullet-point reasonings.
(a/2) sinθ = λ/2 The (half) angular spread of the beam, θ ≈ λ/a, is large if a is small. Physically, the
reason for this is that if a is small, then the beam needs to tilt more in order to generate
Figure 27
the path differences (and hence phase differences) that lead to the total cancelation at the
first zero (at β = 2π). Said in a different way, if a is small, then the beam can tilt quite a
y y = β/2 y = tanβ/2 bit and still have all the different paths be essentially in phase.
20 If a < λ, then even if the beam is titled at θ = π/2, there still can’t be total cancelation.
15
10 So if a < λ, then the diffraction pattern has no zeros. It simply consists of one bump that
5
β is maximum at θ = 0 and decreases as θ → π/2. It actually does approach zero in this limit
-5 (assuming we have a flat screen and not a cylindrical one) because of the B(θ) factor in Eq.
-10 π 3π 5π 7π
-15 (34). θ → π/2 corresponds to points on the screen that are very far from the slit, so the
amplitude of all the waves is essentially zero.
Figure 28
In the limit a ¿ λ, all of the waves from the different points in the slit are essentially in
phase for any angle θ. Interference effects therefore don’t come into play, so the slit behaves
essentially like a point (or rather a line) source. The only θ dependence in the diffraction
pattern comes from the B(θ) factor. If we have a cylindrical screen, then we don’t even
have this factor, so the diffraction pattern is constant. If we have a flat screen and deal
only with small angles (for which B(θ) ≈ B(0)), then the diffraction pattern is constant
9.3. DIFFRACTION FROM A WIDE SLIT 19

there, too. Since we generally deal with small angles, it is customary to say that a ¿ λ
leads to a constant diffraction pattern. We now see what we meant by “narrow slits” or
“infinitesimal slits” in Sections 9.1 and 9.2. We meant that a ¿ λ. This allowed us to
ignore any nontrivial diffraction effects from the individual slits.
If we have the other extreme where a À λ, then even the slightest tilt of the beam will
lead to a pathlength difference of λ between the paths associated with the two ends of the
slit. This corresponds to the first zero at β = 2π. So the diffraction pattern is very narrow in
an angular sense. In the far-field limit, the distances on the screen arising from the angular
spread (which take the rough form of Dθ) completely dominate the initial spread of the
beam due to the thickness a of the slit. So as long as D is very large, increasing the value
of a will decrease the size of the bright spot in the screen. If the screen were right next to
the slit, then increasing a would of course increase the size of the spot. But we’re working
in the far-field limit here, where the angular spread is all that matters.
Let’s now do two examples that illustrate various aspects of diffraction. For both of
these examples, we’ll need to use the diffraction pattern from a wide slit, but with it not
normalized to the value at θ = 0. Conveniently, this is the result we found in Eq. (35),
which we’ll write in the form,
¡ ¢ ¡ ¢
sin 12 ka sin θ sin 12 ka sin θ
Atot (θ) = B(0) · 1 =⇒ Atot (θ) ∝ 1 . (38)
2 k sin θ 2 k sin θ
The B(0) term (which is simply a measure of how bright the light is as it impinges on the
slit) will cancel out in these two examples, so all that matters is the second proportionality
relation in Eq. (38). The intensity is then
à ¡ ¢ !2
sin 12 ka sin θ
Itot (θ) ∝ 1 . (39)
2 k sin θ

Example (Four times the light?): If we let θ = 0 in Eq. (39), and if we make the usual
sin ² ≈ ² approximation, we see that Itot (0) ∝ a2 . This means that if we double a, then Itot (0)
increases by a factor of 4. Intensity equals energy per unit time per unit area, so 4 times as
much energy is now hitting a given tiny region around θ = 0. Does this makes sense? Does
it mean that if we double the width of the slit, then 4 times as much light makes it through?

Solution: The answers to the above two questions are yes and no, respectively. The answer
to the second one had better be no, because otherwise energy would be created out of nowhere.
If we double the width of the slit, then our intuition is entirely correct: twice as much light
makes it through, not 4 times as much. The reason why the 4-fold increase in Itot (0) doesn’t
imply that 4 times as much light makes it through is the following.
The critical point is that although the intensity goes up by a factor of 4 at θ = 0, the
diffraction pattern gets thinner. So the range of θ vales that have a significant intensity
decreases. It turns out that the combination of theses effects leads to just a factor of 2 in
the end. This is quite believable, and we can prove it quantitatively as follows. We’ll assume
that the bulk of the diffraction pattern is contained in the region where θ is small. For the
general case without this assumption, see Problem [to be added].
If θ is small, then we can use sin θ ≈ θ in Eq. (39) to write5
à ¡1 ¢ !2
sin 2
kaθ
Itot (θ) ∝ 1
. (40)
2
kθ
5 Note that even though we’re assuming that θ is small, we cannot assume that kaθ/2 is small and thereby

make another sin ² ≈ ² approximation in the numerator. This is because k may be large, or more precisely
λ may be much smaller than a.
 20 CHAPTER 9. INTERFERENCE AND DIFFRACTION

The larger a is, the quicker sin(kaθ/2) runs through its cycles. In particular, the first zero
I(θ) (which gives the “width” of the diffraction pattern) occurs at θ = 2π/ka. This is proportional
ratio of to 1/a, so increasing a by a general factor f shrinks the pattern by a factor f in the horizontal
heights = 4 slit width = 2a direction. And since we saw above that Itot (0) ∝ a2 , increasing a by a factor f expands the
ratio of pattern by a factor of f 2 in the vertical direction. The combination of these two effects makes
widths = 1/2 the total area under the curve (which is the total intensity) increase by a factor f 2 /f = f .
slit width = a
This is consistent with the fact that f times as much light makes it through the widened slit
of width f a, as desired. This reasoning is summarized in Fig. 29 for the case where f = 2
θ (with arbitrary units on the vertical axis).
2π
____ 2π
__
Remarks: The main point here is that intensity equals energy per unit time per unit area. So we
k(2a) ka can’t conclude anything by using only the fact that Itot increases by a factor of f 2 at the specific
Figure 29 point θ = 0. We need to integrate over all θ values on the screen (and then technically multiply
by some length in the direction perpendicular to the page to obtain an actual area, but this isn’t
important for the present discussion). From Fig. 29, the curve as a whole is most certainly not simply
scaled up by a factor f 2 .
There are two issues we glossed over in the above solution. First, in finding the area under the
intensity curve, the integral should be done over the position x along the screen, and not over θ. But
since x is given by D tan θ ≈ Dθ for small θ, the integral over x is the same (up to the constant factor
D) as the integral over θ. Second, we actually showed only that Itot (θ) increases by a factor of f 2
right at the origin. What happensRat other corresponding points isn’t as obvious. If you want to be
more rigorous about the integral Itot (θ) dθ, you can let a → f a in Eq. (40), and then make the
change of variables θ0 ≡ f θ. The integral will pick up a factor of f 2 /f = f . But having said this, you
can do things completely rigorously, with no approximations, in Problem [to be added]. ♣

Example (Increasing or decreasing intensity?): Given a slit with width a, consider the
intensity at a particular point on the screen that is a reasonable distance off to the side. (By
this we mean that the distance is large compared with the width λ/a of the central bump.) If
we make a larger, will the intensity increase or decrease at the point? By intensity here, we
I(θ) mean the average intensity in a small region, so that we take the average over a few bumps
(a = 20λ) in the diffraction pattern.
On one hand, increasing a will allow more light through the slit, so the intensity should
increase. But on the other hand, increasing a will make the diffraction pattern narrower, so
the intensity should decrease. Which effect wins?

θ Solution: It turns out that these two effects exactly cancel, for the following reason. If we
0.0 0.1 0.2 0.3 0.4
take the average over a few oscillations of the Itot (θ) function in Eq. (40), the sin2 (kaθ/2)
term averages to 1/2 (we can ignore the variation of the denominator over a few oscillations
of the sine term). So the average value of Itot (θ) in a small region near a given value of θ is
I(θ) Itot,avg (θ) ∝ 2/(k2 θ2 ). This is independent of a. So the intensity at the given point doesn’t
change as we widen the slit. In short, the envelope of the wiggles in the diffraction pattern
(a = 50λ) behaves like a 1/θ2 function, and this is independent of a. Fig. 30 shows the diffraction
patterns for a = 20λ and a = 50λ. The envelope is the same for each.

θ
0.0 0.1 0.2 0.3 0.4

Figure 30
9.3.3 Relation to the Fourier transform
¡ ¢
If the sin 12 ka sin θ / 21 ka sin θ function in Eq. (31) looks familiar to you, it’s because this
A
function is basically (up to an overall constant) the Fourier transform of the square-wave
function shown in Fig. 31. We discussed this function in Chapter 3, but let’s derive the
x transform again here since it’s quick. If we let the argument of the Fourier transform be
-a/2 a/2 k sin θ instead of the usual k (we’re free to pick it to be whatever we want; if you wish, you
Figure 31
9.3. DIFFRACTION FROM A WIDE SLIT 21

can define k 0 ≡ k sin θ and work in terms of k 0 ), then Eq. (3.43) gives
Z ∞
1
C(k sin θ) = f (x)e−i(k sin θ)x dx
2π −∞
Z a/2
1
= Ae−ikx sin θ dx
2π −a/2
A e−i(ka sin θ)/2 − ei(ka sin θ)/2
=
2π −ik sin θ
¡ ¢
A −2i sin 12 ka sin θ
=
2π −ik sin θ
¡ ¢
aA sin 12 ka sin θ
= 1 . (41)
2π 2 ka sin θ

So in view of Eq. (31), the intensity on the screen is (up to an overall constant) the square
of the Fourier transform of the slit. This might seem like a random coincidence, but there’s
actually a good reason or it: In Eq. (35) we saw that the amplitude of the diffraction pattern
was obtained by integrating up a bunch of e−iky sin θ phases. But this is exactly the same
thing we do when we compute a Fourier transform. So that’s the reason, and that’s pretty
much all there is to it.
More generally, instead of a slit we can have a wall with transmittivity T (y). T (y) gives
the fraction (compared with no wall) of the amplitude coming through the wall at position
y. For example, a normal slit has T (y) = 1 inside the slit and T (y) = 0 outside the slit. But
you can imagine having a partially opaque wall where T (y) takes on values between 0 and
1 in various regions. In terms of T (y), the total wave at an angle θ on the screen is given
by Eq. (35), but with the extra factor of T (y) in the integrand:
Z ∞
Etot (θ) = B(0) T (y)e−iky sin θ dy. (42)
−∞

Note that the integral now runs from −∞ to ∞, although there may very well be only a finite
region where T (y) is nonzero. Up to an overall constant, the result of this integral is simply
T̃ (k sin θ), where T̃ denotes the Fourier transform of T (y). So the diffraction pattern is the
(absolute value of the square of the) Fourier transform of the transmittivity function. (We’re
assuming that the region of nonzero T (y) is small compared with the distance to the screen,
so that we can use the standard far-field approximation that all the paths from the different
points in the “slit” to a given point on the screen have equal lengths (multiplicatively).)
Recall the uncertainty principle from Problem [to be added] in Chapter 3, which stated
that the thinner a function f (x) is, the broader the Fourier transform f˜(k) is, and vice
versa. The present result (that the diffraction pattern is the square of the Fourier transform
of the slit) is consistent with this. A narrow slit gives a wide diffraction pattern, and a wide
slit gives a narrow (in an angular sense) pattern.
There are two ways of defining the Fourier transform. The definition we used above is
the statement in the second equation in Eq. (3.43): The Fourier transform is the result of
multiplying each f (x) value by a phase e−ikx and then integrating. This makes it clear
why the diffraction pattern is the Fourier transform of the transmittivity function, because
the diffraction pattern is the result of attaching an extra phase of e−iky sin θ to the Huygens
wavelets coming from each point in the slit.
The other definition of the Fourier transform comes from the first equation in Eq. (3.43):
The Fourier transform gives a measure of how much the function f (x) is made up of the
function eikx . (This holds in a simpler discrete manner in the case of a Fourier series for
 22 CHAPTER 9. INTERFERENCE AND DIFFRACTION

a periodic function.) Does this interpretation of the Fourier transform have an analog in
the diffraction setup? That is, does the diffraction pattern somehow give a measure of how
n much the transmittivity function is made up of the function eiky sin θ ? Indeed it does, for
cree
to s away the following reason.
far We’ll be qualitative here, but this should suffice to illustrate the general idea. Let’s
assume that we observe a large amplitude in the diffraction pattern at an angle θ. This means
that the wavelets from the various points in the slit generally constructively interfere at the
___
λ angle θ. From Fig. 32, we see that the transmittivity function must have a large component
sinθ θ with spatial period λ/ sin θ. This means that the spatial frequency of the transmittivity
function is κ = 2π/(λ/ sin θ) = (2π/λ) sin θ = k sin θ, where k is the spatial frequency of
___
λ λ the light wave. In other words, a large amplitude at angle θ means that T (y) has a large
sinθ component with spatial frequency k sin θ. The larger the amplitude at angle θ, the larger the
component of T (y) with spatial frequency k sin θ. But this is exactly the property that the
___
λ Fourier transform of T (y) has: The larger the value of T̃ (k sin θ), the larger the component
sinθ of T (y) with spatial frequency k sin θ. So this makes it believable that the amplitude of the
diffraction pattern equals the Fourier transform of T (y), with k sin θ in place of the usual k.
___
λ The actual proof of this fact is basically the statement in Eq. (42).
sinθ
θ

Figure 32 9.4 Combined Interference and diffraction

So far we’ve dealt with either N infinitesimally thin slits, or one wide slit. We’ll now combine
these two setups and consider N wide slits. Let the slits have width a, and let the spatial
period be d (this is the distance between, say, two adjacent bottom ends). Fig. 33 shows
the case with N = 3 and d = 3a. We’ll continue to work in the far-field limit.

(close-up view) (zoomed-out view)

wall screen

light r0
y θ
d
a

Figure 33

What is the amplitude of the wave at an angle θ on a distant screen? To answer this,
we can use the reasoning in the “Continuous integral” derivation of the one-wide-slit result
in Section 9.3.1. Let r0 be the pathlength from the bottom of the bottom slit, as shown
in Fig. 33. Define y to be the distance from the bottom of the bottom slit up to a given
location in a slit. Then the relevant y values are from 0 to a for the bottom slit, then d to
d + a for the next slit, and so on.
The integral that gives the total wave from the bottom slit is simply the integral in Eq.
(34), but with the integration now running from 0 to a. (We technically need to multiply
by the phase ei(kr0 −ωt) , but this phase is tacked on uniformly to all the slits, so it doesn’t
affect the overall amplitude.) The integral that gives the total wave from the second slit
is again the same, except with the integration running from d to d + a. And so on, up to
9.4. COMBINED INTERFERENCE AND DIFFRACTION 23

limits of (N − 1)d and (N − 1)d + a for the top slit. So the total wave at angle θ from all
the slits is (as usual, we’ll approximate the B(θ) in Eq. (34) by B(0))
ÃZ Z Z !
a d+a (N −1)d+a
Etot (θ) = B(0) e−iky sin θ dy + e−iky sin θ dy + · · · + e−iky sin θ dy .
0 d (N −1)d
(43)
The second integral here is simply e−ikd sin θ times the first integral, because the y values
are just shifted by a distance d. Likewise, the third integral is e−2ikd sin θ times the first.
Letting z ≡ e−ikd sin θ , we therefore have
µZ a ¶³ ´
−iky sin θ
Etot (θ) = B(0) e dy 1 + z + z 2 + · · · + z N −1 . (44)
0

Shifting the limits of this integral by −a/2 (which only introduces a phase, which doesn’t
affect the amplitude) puts it in the form of Eq. (34). So we can simply copy the result in
Eq. (35). (Or you can just compute the integral with the 0 and a limits.) And the geometric
series is the same one we calculated in Eq. (12), so we can copy that result too. (Our z here
is the complex conjugate of the z in Eq. (12), but that will only bring in an overall minus
sign in the final result, which doesn’t affect the amplitude.) So the total amplitude at angle
θ is ¡ ¢ ¡ ¢
sin 21 ka sin θ sin 12 N kd sin θ
Atot (θ) = B(0)a · 1 · ¡ ¢ . (45)
2 ka sin θ sin 12 kd sin θ
Taking the usual limit as θ → 0, the value of the amplitude at θ = 0 is B(0)aN . The
intensity relative to θ = 0 is therefore
à ¡1 ¢ ¡ ¢ !2
Itot (θ) sin 2 ka sin θ sin 12 N kd sin θ
= 1 · ¡ ¢ (46)
Itot (0) 2 ka sin θ N sin 12 kd sin θ

This result really couldn’t have come out any nicer. It is simply the product of the
results for the two separate cases we’ve discussed. The first quotient is the one-wide-slit
diffraction result, and the second quotient is the N -thin-slit interference result. Note that
since N d > a (because d > a), the second quotient oscillates faster than the first. You can
therefore think of this result as the N -thin-slit interference result modulated by (that is,
with an envelope of) the one-wide-slit diffraction result.
In retrospect, it makes sense that we obtained the product of the two earlier results. At
a given value of θ, we can think of the setup as just N -thin-slit interference, but where the
amplitude from each slit is decreased by the one-wide-slit diffraction result. This is clear if
we rearrange Eq. (45) and write it as (we’ll switch the B(0) back to B(θ))
à ¡ ¢! ¡ ¢
sin 12 ka sin θ sin 12 N kd sin θ
Atot (θ) = B(θ)a 1 · ¡ ¢ . (47)
2 ka sin θ sin 12 kd sin θ

This is the¡ N -thin-slit

¢± result, with B(θ)a (which equals the A(θ) in Eq. (14)) replaced by
B(θ)a · sin 12 ka sin θ 12 ka sin θ. Basically, at a given θ, you can’t tell the difference between
a wide slit, and an infinitesimal slit with an appropriate amount of light coming through.
Fig. 34 shows the interference/diffraction pattern for N = 4 and for various slit widths
a, given the spatial period d. The coordinate on the horizontal axis is α ≡ kd sin θ. The
a ≈ 0 plot is exactly the same as the thin-slit plot in Fig. 11, as it should be. As the width
a increases, the envelope becomes narrower. Recall from Eq. (36) that the width of the
one-wide-slit diffraction pattern is inversely proportional to a.
24 CHAPTER 9. INTERFERENCE AND DIFFRACTION

I(α)/I(0) (a 0) (a = d/4) (a = d/2)

1.0 1.0 1.0
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
α

_ π 3π
π __ 2π
2 2

1.0 (a = 3d/4) 1.0 (a = d)

0.8 0.8
(Plots for N = 4)
0.6 0.6
0.4 0.4
0.2 0.2

Figure 34

When a finally reaches the value of d in the last plot, the four slits blend together, and
we simply have one slit with width 4a. (It doesn’t make any sense to talk about a values
that are larger than d.) And the a = d plot is indeed the plot for a single wide slit with
width 4a. The only difference between it and the envelope (which comes from a slit width
a) is that it is squashed by a factor of 4 in the horizontal direction. It turns out that in
the a = d case, the zeros of the envelope fall exactly where the main peaks would be if the
envelope weren’t there (see the a ≈ 0 case). This follows from the fact that the zeros of the
diffraction envelope occur when β ≡ ka sin θ equals 2π, while the main peaks of the N -slit
interference pattern occur when α ≡ kd sin θ equals 2π. So if a = d, these occur at the same
locations.

9.5 Near-field diffraction

9.5.1 Derivation
Everything we’ve done so far in this chapter has been concerned with the far-field approx-
imation (the so-called Fraunhofer approximation). We have assumed that the distance to
the screen is large compared with the span of the slit(s). As discussed in Section 9.1, this
assumption leads to two facts:
• The pathlengths from the various points in the slit(s) to a given point on the screen
are all essentially equal in a multiplicative sense. This implies that the√amplitudes of
all the various wavelets are equal. In other words, we can ignore the 1/ r dependence
in the individual amplitudes of the cylindrically-propagating Huygens wavelets.
• The paths from the various points in the slit(s) to a given point on the screen are all
essentially parallel. This implies that the additive difference between adjacent path-
lengths equals d sin θ (or dy sin θ in the continuous case). The pathlengths therefore
take the nice general form of r0 + nd sin θ (or r0 + y sin θ), and the phases are easy to
get a handle on.
In the first of these points, note that we’re talking about the various distances from a
particular point on the screen to all the different points in the slit(s). We are not talking
about the various distances from a particular point in the slit(s) to all the different points
9.5. NEAR-FIELD DIFFRACTION 25

on the screen. These distances certainly aren’t equal; the fact that they aren’t equal is what
brought in the factor of A(θ) in, say, Eq. (4) or Eq. (14). But this lack of equality is fine;
it simply leads to an overall envelope of the interference curve. The relevant fact in the
far-field approximation is that the various distances from a particular point on the screen
to all the different points in the slit(s) are essentially equal. This lets us associate all the
different wavelets (at a given point on the screen) with a single value of A(θ), whatever that
value may be.
We’ll now switch gears and discuss the near-field approximation (the so-called Fresnel
approximation). That is, we will not assume that the distance to the screen is large compared
with the span of the slit(s). The above two points are now invalid. To be explicit, in the
near-field case:

• We cannot say that the pathlengths from the various points in the slit(s) to a given
point on the screen
√ are all equal in a multiplicative sense. We will need to take into
account the 1/ r dependence in the amplitudes.

• We cannot say that the pathlengths take the nice form of r0 + nd sin θ (or r0 + y sin θ).
We will have to calculate the lengths explicitly as a function of the position in the
slit(s).

The bad news is that all of the previous results in this chapter are now invalid. But
the good news is that they’re close to being correct. The strategy for the near-field case
is basically the same as for the far-field case, as long as we incorporate the changes in the
above two points.
The procedure is best described by an example. We’ll look at a continuous case involv- screen
ing diffraction from a wide slit, but we could of course have a near-field setup involving
interference from N narrow slits, or a combination of interference and diffraction from N
wide slits.
y
Our wide slit will actually be an infinite slit. Our goal will be to find the intensity at P
the point P directly across from the top of a “half-wall” (see Fig. 35). Since our slit is
infinitely large, we’re automatically in the near-field case, because it is impossible for the
wall-screen distance D to be much greater than the slit width a, since a = ∞. The various
pathlengths (which are infinite in number) to the given point P from all of the possible points
in the slit (three of these paths are indicated by dotted lines in Fig. 35) certainly
p cannot D
be approximated as having the same length. These paths have lengths r(y) = D2 + y 2 ,
where y is measured from the top of the wall. If we instead had an infinite pnumber of thin Figure 35
slits extending upward with separation d, the pathlengths would be rn = D2 + (nd)2 .
Since
√ the amplitudes of the various cylindrically-propagating
p wavelets are proportional
to 1/ r, we need to tack on a factor of 1/ r(y) in front of each wavelet. More precisely,
let B0 dy be the amplitude of the wave that would hit point P due to an infinitesimal span
dy in the slitp at y = 0, if the distance D were equal to 1 (in whatever units we’re using).6
Then B0 dy/ r(y) is the amplitude of the wave that hits point P due to a span dy in the
slit at height y. The length r(y) depends on where the screen is located (which gives D),
and also on the height y.
As far as the phases go, the phase of the wavelet coming from a height y in the slit is
eikr(y) , neglecting the e−iωt phase and an overall phase associated with the y = 0 path.
Using these facts about the amplitude and phase of the wavelets, we can integrate over
the entire (infinite) slit to find the total wave at the point P directly across from the top
6 B is slightly different from the B(0) in Eq. (35), because we didn’t take into account the distance to
0
the screen there. We assumed the position was fixed. But we want to be able to move the screen in the
present setup and get a handle on how this affects things.
 26 CHAPTER 9. INTERFERENCE AND DIFFRACTION

of the wall. The integral is similar to Eq. (34). But with the modified amplitude, the more
complicated phase, and the new limits of integration, we now have
Z ∞ Z ∞ √
B dy ikr(y) B0 dy
Etot (P ) = p0 e = e ik D 2 +y 2
. (48)
0 r(y) 0 (D2 + y 2 )1/4

This integral must be computed numerically, but we can get a sense of what’s going on if
we draw a picture similar to the far-field case in Fig. 14. In that figure we had little vectors
of equal length wrapping around in a circle, with successive vectors always making the same
angle with respect to each other. In the present near-field case, these two italicized words
are modified for the following reasons.
Let’s imagine discretizing the slit into equal dy intervals. Then as y increases, the
lengths of the little vectors decrease due to the (D2 + y 2 )1/4 factor in the denominator in
Eq. (48). Also, the phase doesn’t increase at a p constant rate. For small y, the phase hardly
changes at all, because the derivative of the D2 + y 2 term in the exponent is zero at
y = 0. But forp large y, the rate of change of the phase approaches a constant, because the
derivative of D2 + y 2 equals 1 for y À D. So as y increases, the angle between successive
vectors increases and asymptotically approaches a particular value. Both of these effects
(the shortening lengths and the increasing rate of change of the phase) have the effect of
decreasing the radius of curvature of the circle that is being wrapped around. In other
words, the “circles” get tighter and tighter, and instead of a circle we end up with a spiral,
as shown in Fig. 36 (we’ve arbitrarily chosen λ = D here).
λ = D, ∆y = (0.1)D In the first spiral in Fig. 36, we have discretized the integral in Eq. (48) by doing a
discrete sum over intervals with length ∆y = (0.1)D in the slit. You can see that the little
vectors get smaller as they wrap around.7 And you can also see that the angle between them
starts off near zero and then increases. The second spiral shows the continuous limit where
∆y ≈ 0. So this corresponds to the actual integral in Eq. (48). In reality, this plot was
generated by doing a discrete sum with ∆y = (0.01)D. But the little vectors are too small
to see, so the spiral is essentially continuous. So neither of these spirals actually corresponds
to the continuous integral in Eq. (48). But the second one is a very good approximation. If
you look closely, you can see that the slope of the straight line in the first spiral is slightly
different from the slope in the second.
λ = D, ∆y 0 We haven’t drawn the axes in these plots, because the absolute size of the resulting
amplitude isn’t so important. We’re generally concerned with how large the amplitude is
Figure 36 relative to a particular case. The most reasonable case to compare all others to is the one
where there is no wall at all (so the slit extends from y = −∞ to y = ∞). We’ll talk
about this below. But if you’re curious about the rough size of the spiral, the horizontal
and vertical spans (for the case in Fig. 36 where λ = D) are around (0.5)B0 .
This spiral is known as the Cornu spiral,8 or the Euler spiral. In the present case where
the upper limit on y is infinity, the spiral keeps wrapping around indefinitely (even though
we stopped drawing it after a certain point in Fig. 36). The radius gets smaller and smaller,
and the spiral approaches a definite point. This point is the sum of the infinite number
of tiny vectors. The desired amplitude of the wave at P is the distance from the origin to
this point, as indicated by the straight line in the figure. As usual, the whole figure rotates
around in the plane with frequency ω as time progresses. The horizontal component of the
straight line is the actual value of the wave.
7 We’ve stopped drawing the vectors after a certain point, but they do spiral inward all the way to the

center of the white circle you see in the figure. If we kept drawing them, they would end up forming a black
blob where the white circle presently is.
8 Technically, this name is reserved for the simpler approximate spiral we’ll discuss in Section 9.5.3. But

we’ll still use the name here.

9.5. NEAR-FIELD DIFFRACTION 27

The shape of the spiral depends on the relative size of λ and D. If we define the
dimensionless quantity z by y ≡ zD, then Eq. (48) can be written as (using k = 2π/λ and
dy = D dz) √
Z ∞
B0 D dz 2iπ(D/λ)√1+z2
Etot (P ) = e . (49)
0 (1 + z 2 )1/4
√
For a given value of D/λ, the factor of D in the numerator simply scales the whole spiral,
so it doesn’t affect the overall shape. However, the factor of D/λ in the exponent does affect
the shape, but it turns out that the dependence is fairly weak. If we instead had spherically
propagating waves with (D2 + y 2 )1/2 instead of (D2 + y 2 )1/4 in the denominator of Eq. (48),
then there would be a noticeable dependence on D/λ, especially for large λ.

9.5.2 Changing the slit

5D
What happens if instead of extending to infinity, the slit runs from y = 0 up to a finite value
ymax ? The only change in Eq. (48) is that the upper limit is now ymax . The integrand is
exactly the same. So Eqs. (48) and (49) become
Z ymax √ Z zmax √
B0 dy ik D 2 +y 2 B0 D dz 2iπ(D/λ)√1+z2 4D
Etot (P ) = e = e . (50)
0 (D2 + y 2 )1/4 0 (1 + z 2 )1/4

In terms of Fig. 36 (we’ll again assume λ = D), we now only march along the spiral until we 5λ
get to the little vector associated with ymax , the location of which can be found numerically.
(We know r(ymax ), so we know the relative phase, so we know the angle (slope) of the spiral 3D
at the stopping point.) The amplitude is then the length of the line from the origin to the 4λ
stopping point. A few cases are shown in Fig. 37.

(λ = D in all cases) 2D 3λ

D
2λ
ymax = D/2 ymax = D ymax = 3 D ymax = 3D

Figure 37
√ p P
The y = 3D case is an interesting one because it yields a pathlength of D2 + y 2 = D
2D, which equals 2λ since we’re assuming λ = D. This pathlength √is therefore λ more than (λ = D)
the pathlength associated with y = 0. So the wavelet from y = 3D is in phase with the
wavelet from y =√0. And this is exactly what we observe in the figure; the slope of the Figure 38
spiral at the y = 3D point equals the slope at the start (both slopes equal zero). A few
other values of y that yield pathlengths that are integral multiples of λ are shown in Fig. 38,
and the corresponding points in the Cornu spiral are shown in Fig. 39 (eventually the points
blend together). The spiral also has zero slope at the top of the “circles” in the spiral. These
points correspond to pathlengths of 3λ/2, 5λ/2, 7λ/2, etc. (The λ/2 is missing here because
all the pathlengths are at least D = λ.) But the associated little vectors in the spiral now pathlength = λ, 2λ, 3λ
point to the left, because the wavelets are exactly out of phase with the wavelet from y = 0
(which we defined as pointing to the right).
(λ = D)

Figure 39
 28 CHAPTER 9. INTERFERENCE AND DIFFRACTION

A(P)/B0 D Remarks:
0.7
0.6 1. Note that the distance between the first two dots along the spiral in Fig. 39 is large, and
0.5
0.4 then it decreases as we march along the spiral.√There are two reasons for this. First, there is
0.3 a large span of y values (from zero up to y = 3D) that corresponds to the region between
0.2
0.1 ymax
____ the first two dots on the spiral. This span then gets smaller as y increases, and it eventually
0.0
0 2 4 6 8 10 D approaches the wavelength λ (which we’ve chosen to equal D). Second, the amplitudes of
√
the wavelets get smaller as y increases (because the amplitude is proportional to 1/ r), so
Figure 40 the little vectors that make up the spiral get shorter as we spiral inward.
2. From Fig. 37, we see that the largest amplitude occurs for a ymax that is somewhere around
I(P) D. It happens to occur at ymax ≈ (0.935)D. If ymax is increased above this value, then
0.5 apparently the upside of having more light coming through the slit is more than canceled out
0.4 by the downside of this extra light canceling (due to the relation of the phases) some of the
0.3
0.2
light that was already there. At any local max or min of the amplitude, the line representing
0.1 ymax the amplitude is perpendicular to the tangent to the spiral.
0.0
____ √
0 2 4 6 8 10 D 3. A plot of the amplitude, A(P ) (in units of B0 D), as a function of ymax is shown in Fig. 40.
As the spiral circles around and around, the amplitude oscillates up and down. Since the
Figure 41 circles keep getting smaller, the bumps in Fig. 40 likewise keep getting smaller. The plot
oscillates around a value that happens to be about 0.5. This is the amplitude associated with
ymax = ∞. For large ymax , the period of the oscillations is essentially λ. This follows from
the fact that as we noted in Fig. 38, if y increases by λ (which corresponds to a full circle
in the spiral), then the pathlength essentially does also, if the path is roughly parallel to the
wall. A plot of the intensity (which is proportional to the amplitude squared) is shown in
Fig. 41, with arbitrary units on the vertical axis. ♣
ymax =
8 8

What happens if we put the upper limit ymax back at infinity, but now move the top of
ymin = - the wall (the bottom of the slit) downward, so that y runs from some negative value, ymin ,
to infinity? (The point in question on the screen is still the point P directly across from
(λ = D) y = 0.) To answer this, let’s first consider the case where we move the top of the wall all
the way down to y = −∞. So we have no wall at all. We claim that the total amplitude at
Figure 42
point P is given by the length of the diagonal line in Fig. 42. This is believable, of course,
because the length of this line is twice the length of the line in Fig. 36 for the case where
the “slit” was half as large. But to be rigorous, you can think of things in the following way.
In Fig. 36 imagine starting at y = +∞ and decreasing down to y = 0. This corresponds
to starting in the middle of the spiral and “unwrapping” clockwise around it until you reach
the origin. The clockwise nature is consistent with the fact that the phase decreases as
y decreases (because the pathlength decreases), and we always take positive phase to be
ymax =
8

counterclockwise. If you then want to keep going to negative values of y, you simply have to
ymin = -2D keep adding on the little vectors. But now the phase is increasing, because the pathlength is
increasing. So the spiral wraps around counterclockwise. This is indeed what is happening
(λ = D) in Fig. 42. (The spiral for the y < 0 region has to have the same shape as the spiral for the
y > 0 region, of course, due to symmetry. The only question is how it is oriented.)
Figure 43 If we want the slit to go down to a finite value of y instead of y = −∞, then we simply
need to stop marching along the spiral at the corresponding point. For example, if the wall
goes down to y = −2D, then the amplitude is given by the diagonal line in Fig. 43.
More generally, if we want to find the amplitude (still at the point P directly across from
y = 0) due to a slit that goes from a finite ymin to a finite ymax , then we just need to find
the corresponding points on the spiral and draw the line between them. For example, if a
slit goes from y = −2D to y = 3D, then the amplitude is given by the length of the diagonal
ymax = 3D line in Fig. 44. In the event that ymin and ymax are both positive (or both negative), the
ymin = -2D diagonal line is contained within the upper right (or lower left) half of the full Cornu spiral
in Fig. 42. An example of this will come up in Section 9.5.5.
(λ = D)

Figure 44
9.5. NEAR-FIELD DIFFRACTION 29

Remarks:
1. Note that in Fig. 44 the slope of the little vector at y = −2D is nonzero. This is because
we’re still measuring all the phases relative to the phase of the wavelet at y = 0. If you want,
you can measure all the phases relative to the phase at y = −2D (or any other point). But
only the relative phases matter, so this just rotates the whole figure, leaving the length of
the diagonal line (the amplitude) unchanged. (The whole figure rotates around in the plane
anyway as time goes on, due to the ωt term in the phase, which we’ve been ignoring since
we only care about the amplitude.) By convention, it is customary to draw things as we’ve
done in Fig. 42, with a slope of zero at the middle of the complete spiral.
2. In a realistic situation, the slit location is fixed, and we’re concerned with the intensity at
various points P on the screen. But instead of varying P , you can consider the equivalent
situation where P is fixed (and defined to be across from y = 0), and where the slit is moved.
This simply involves changing the values of ymin and ymax , or equivalently the endpoints
of the diagonal line on the Cornu spiral representing the amplitude. So the above analysis
actually gives the wave at any point P on the screen, not just the point across from y = 0.
3. In the earlier far-field case of interference and diffraction, the customary thing to do was to
give the intensity relative to the intensity at θ = 0. The most natural thing to compare the
near-field amplitude to is the amplitude when there is no wall. This is the amplitude shown
in Fig. 42. The Cornu spiral (the shape of which depends on the ratio D/λ in Eq. (49))
completely determines all aspects of the diffraction pattern for any location of the slit. And
the length of the diagonal line in Fig. 42 gives the general length scale of the spiral, so it
makes sense to compare all other lengths to this one. ♣

9.5.3 The D À λ limit

When dealing with light waves, it is invariably the case that D À λ. If this relation holds,
then the Cornu spiral approaches a particular shape, and we can write down an approximate
(and simpler) expression for the integral in Eq. (49). Note that D À λ does not mean that
we’re in the far-field limit. The far-field limit involves a comparison between D and the span
of the slit(s), and it results in the approximation that all of the paths from the various points
in the slit(s) to a given point on the screen are essentially equal in length (multiplicatively),
and essentially parallel. The wavelength λ has nothing to do with this.
If D À λ, the actual size (but not the shape) of the spiral depends on λ; the smaller λ
is, the smaller the spiral is. But the size (or the shape) doesn’t depend on D. Both of these
facts will follow quickly from the approximate expression we’ll derive in Eq. (51) below. The
fixed shape of the spiral is shown in Fig. 45, and it looks basically the same as Fig. 36.
The size dependence on λ is fairly easy to see physically. Even a slight increase in y from
Figure 45
y = 0 will lead to a pathlength that increases on the order of λ, if λ is small. This means
that the phases immediately start canceling each other out. The wave has no opportunity
to build up, because the phase oscillates so rapidly as a function of y. The smaller λ is, the
quicker the phases start to cancel each other.
If D À λ, we can give an approximate expression for the wave in Eq. (49). We claim that
only small values of z (much less than 1) are relevant in Eq. (49). (These values correspond
to y being much less than D.) Let’s see what Etot (P ) reduces to under the assumption that
z ¿ 1, then we’ll justify this assumption. √
If z is small, then we can use the approximation 1 + z 2 ≈ 1+z 2 /2 in both the exponent
and the denominator of Eq. (49). We can ignore the z 2 /2 term in the denominator, because
it is small compared with 1. In the exponent we have 2πiD/λ + 2πi(D/λ)z 2 /2. The first
of these terms is constant, so it just gives an overall phase in the integral, so we can ignore
it. The second term involves a z 2 , but we can’t ignore it because it also contains a factor of
 30 CHAPTER 9. INTERFERENCE AND DIFFRACTION

D/λ, which we’re assuming is large. Eq. (49) therefore reduces to (recalling z ≡ y/D)
Z zmax √ Z ymax √ ¢
iπ(D/λ)z 2
¡ 2
Etot (P ) ≈ B0 D e dz = B0 / D eiπy /Dλ dy, (51)
0 0
p
where zmax is a number much smaller than 1, but p also much larger than λ/D. And
ymax = Dzmax . The reason for this lower bound of λ/D comes from the following reasoning
that justifies why we need topconsider only z values that are much less than 1 in Eq.√(49).
If z is much larger than λ/D (which corresponds to y being much larger than λD),
but still satisfies our assumption of z ¿ 1, then the exponent in Eq. (49) is a rapidly
changing function of z. This corresponds to being deep inside the spiral where the circles
are small. By this point in the spiral, the integral in Eq. (49) has essentially reached its
limiting value, so it doesn’t matter whether we truncate the integral at this (small) value of
z or keep going to the actual upper limit of z = ∞. So if you want, you can let the upper
bounds in Eq. (51) be infinity:
Z ∞ √ Z ∞ √ ¢
2 ¡ 2

D + nλ Etot (P ) ≈ B0 D eiπ(D/λ)z dz = B0 / D eiπy /Dλ dy (52)

0 0
y
P Pictorially, if you want to get a handle on which y values correspond to which points in
D
the spiral, note that an increase in pathlength by one wavelength λ corresponds to a full
Figure 46 circle of the spiral. The value of y that yields a pathlength of the form D + nλ is found
from the right triangle in Fig. 46. The Pythagorean theorem gives
y values yielding
√
D2 + y 2 = (D + nλ)2 =⇒ y 2 = 2nDλ + n2 λ2 =⇒ y ≈ 2nDλ, (53)
pathlength = D + nλ

n =2 where we have ignored the second-order λ2 term due to the D À λ assumption. Fig. 47shows
n =1 the first 40 of these values of y for the case where D/λ = 200, although for actual setups
n =0
P involving light, this ratio is generally much higher, thereby making the approximations even
better. This figure is analogous
√ to Fig. 38. As you can see, the y values get closer together
wall as y increases, due to the n dependence. This is consistent with the above statement that
the exponent in Eq. (49) is a rapidly changing function of z.
Figure 47 Remarks:
√
1. As we noted above, the amplitude is essentially constant for small z, since 1 + z 2 ≈ 1. So
the little vectors that make up the spiral all have essentially the same length, for small z
(y ¿ D). The size of a “circle” in the spiral is therefore completely determined by how fast
the phase is changing. Since the phase changes very quickly for y À Dλ, the circles are very
small, which means that we have essentially reached the limiting value at the center of the
circle.
2. We mentioned above that if D À λ, then the size of the spiral depends on λ but not on D.
And the shape depends
p on neither. These √facts follow from Eq. (52) if we make the change
of variables, w ≡ z D/λ (which equals y/ Dλ). This turns the integral into
Z ∞
√ 2
Etot (P ) ≈ λ B0 eiπw dw. (54)
0

There are no D’s in this expression,

√ so the size and shape don’t depend on D. But the size
does depend on λ, according to λ (which decreases as λ decreases, as we argued near the
beginning of this subsection). However, the shape doesn’t depend on λ, because λ appears
only in an overall constant.
3. An interesting fact about the Cornu spiral described by Eq. (52) and shown in Fig. 45 is that
the curvature at a given point is proportional to the arclength traversed (starting from the
9.5. NEAR-FIELD DIFFRACTION 31

lower left end) to that point. The curvature is defined to be 1/R, where R is the radius of
the circle that matches up with the curve at the given point.
This property makes the Cornu spiral very useful as a transition curve in highways and
railways. If you’re driving down a highway and you exit onto an exit ramp that is shaped
like the arc of a circle, then you’ll be in for an uncomfortable jolt. Even though it seems like
the transition should be a smooth one (assuming that the tangent to the circle matches up
with the straight road), it isn’t. When you hit the circular arc, your transverse acceleration
changes abruptly from zero to v 2 /R, where R is the radius of the circle. You therefore have
to suddenly arrange for a sideways force to act on you (perhaps by pushing on the wall of
the car) to keep you in the same position with respect to the car. Consistent with this, you
will have to suddenly twist the steering wheel to immediately put it in a rotated position.
It would be much more desirable to have the curvature change in a gradual manner, ideally
at a constant rate. This way you can gradually apply a larger sideways force, and you can
gradually turn the steering wheel. No sudden movements are required. The task of Problem
9.2 is to show that the Cornu spiral does indeed have the property that the curvature is
proportional to the arclength. ♣

9.5.4 Diffraction around an object

The Cornu spiral gives the key to explaining the diffraction of light around an object. If we
shine light on an object and look at the shadow, something interesting happens near the
boundary. Fig. 48 shows the shadow of a razor blade illuminated by laser light.9 Fig. 49
shows the result of a more idealized setup with an essentially infinite straight edge (oriented
vertically on the page).10 Figure 48

Figure 49

In a normal shadow, we would naively expect to have an abrupt change from a bright
region to a dark region. Indeed, if instead of a light wave we had particles (such as baseballs)
passing by a wall, then the boundary between the “shadow” and the region containing
baseballs would be sharp. Now, even if we realize that light is a wave and can therefore
experience interference/diffraction, we might still semi-naively expect to have the same kind
of behavior on each side of the boundary, whatever that behavior might be. However, from
Fig. 49 we see that there is something fundamentally different between the bright and dark
regions. The amplitude oscillates in the bright region, but it appears to (and indeed does)
decrease monotonically in the dark region. What causes this difference? We can answer this
by looking at the Cornu spiral.
If we scan our eye across Fig. 49, this is equivalent to changing the location of point P
in Fig. 35. Points far to the left (right) in Fig. 49 correspond to P being low (high) in Fig.
35. So as we scan our eye from left to right in Fig. 49, this corresponds to P being raised
up from a large negative value to a large positive value in Fig. 35. However, as we noted in
the second remark at the end of Section 9.5.2, raising the location of point P is equivalent
9 I’mnot sure where this picture originated.
10 Thisimage comes from the very interesting webpage, http://spiff.rit.edu/richmond/occult/bessel/bessel.html,
which discusses diffraction as applied to lunar occultation.
 32 CHAPTER 9. INTERFERENCE AND DIFFRACTION

to keeping P fixed and instead lowering the top of the wall.11 Therefore, scanning our eye
from left to right in Fig. 49 corresponds to lowering the top of the wall from a large positive
value to a large negative value. And we can effectively take these values to be ±∞.
So to determine the intensity of the diffraction pattern as a function of position, we
simply need to determine the intensity at P as we lower the wall. In turn, this means that
we need to look at the length of the appropriate line in the Cornu spiral (and then square
it to go from amplitude to intensity). The line we’re concerned with always has one end
located at the center of the upper-right spiral in Fig. 42, because in our setup the upper end
of the “slit” is always located at +∞. The other end of the line corresponds to the bottom
of the slit, and since we’re lowering this position down from +∞, this other end simply
starts at the center of the upper-right spiral and then winds its way outward in the spiral.
When the top of the wall has moved all the way down to y = 0 (that is, across from P ), the
corresponding point on the spiral is as usual the center point between the two halves of the
spiral. And when the top of the wall has moved all the way down to −∞, the corresponding
point on the spiral is the center of the lower-left spiral.
What happens to the amplitude (the length of the line) as we march through this entire
process? It start out at zero when the top of the wall is at +∞, and then it monotonically
increases as we spiral outward in the upper-right spiral. It keeps increasing as we pass
through the origin, but then it reaches its maximum possible value, shown in Fig. 50. (This
spiral has D = λ, which undoubtedly isn’t the case with light. But the shape of the D À λ
spiral isn’t much different from the D = λ one.) After this point, the length of the line
Figure 50 oscillates up and down as we spiral inward in the lower-left spiral. The size of the oscillations
gradually decreases as the circles get smaller and smaller, and the line approaches the one
shown in Fig. 42, where the ends are at the centers of the two spirals. This corresponds to
the top of the wall being at y = −∞, so there is no wall at all.
The length of the amplitude line at the origin (which corresponds to P being at the edge
of the location of the naive sharp shadow) is exactly half the length that it eventually settles
down to. Since the intensity is proportional to the square of the amplitude, this means that
the intensity at the naive edge is 1/4 of the intensity far away from the shadow. Numerically,
the maximum amplitude associated with Fig. 50 is about 1.18 times the amplitude far away,
which means that the intensity is about 1.39 times as large.
I(P) Note that although the two half-spirals in Fig. 50 have the same shape, one of them
1.4 (the lower-left spiral) produces oscillations in the amplitude, while the other doesn’t. The
1.2
1.0 symmetry is broken due to where the starting point of the line is located. It is always located
0.8
0.6 at the center of the upper-right spiral, and this is what causes the different behaviors inside
0.4 and outside the shadow in Fig. 49.
0.2
yP/D The plot of the intensity (proportional to the amplitude squared) is shown in Fig. 51,
- 4 - 2 0 2 4 6 8 10
with arbitrary units on the vertical axis. The horizontal axis gives the y coordinate of P ,
(λ = D) with y = 0 being across from the top of the wall. The left part corresponds to P being low
in Fig. 35 (or equivalently, keeping P fixed and having the wall be high). In other words, P
Figure 51
is in the left part of Fig. 49, in the shadow. The right part corresponds to P being high (or
equivalently, keeping P fixed and having the wall be low). So P is in the left part of Fig.
49, outside the shadow. Moving from left to right in Fig. 51 corresponds to moving from
left to right in Fig. 49, and also to running around the spiral in the direction we discussed
above, starting at the inside of the upper-right spiral. As we mentioned above, you can see
in Fig. 51 that the intensity at yP = 0 is 1/4 of the intensity at large yP .
In the D À λ limit (which is generally applicable to any setup involving light), the
locations of the bright lines in the diffraction pattern are √ given by essentially
√ the same
reasoning that led to Eq. (53). So we essentially have y ≈ 2nDλ. The n dependence
11 With an infinite straight edge, we do indeed have the situation in Fig. 35 with a “half wall.” The case

of the razor blade is more complicated because it has holes and corners, but the general idea is the same.
9.5. NEAR-FIELD DIFFRACTION 33

implies that the bright lines get closer together as P moves farther away from the shadow
(see Fig. 47). This is what we observe in Fig. 49. Note that the angles
√ at which p
he bright
lines occur are given by (assuming the angle is small) θ ≈ y/D = 2nDλ/D = 2nλ/D.
So although the y values increase with D, the angles decrease with D.

9.5.5 Far-field limit

The expression for the wave in Eq. (48) is an exact one. It holds for arbitrary values of D and
k (or equivalently λ), and also for arbitrary values of the limits of integration associated
with the endpoints of the slit. Therefore, Eq. (48) and all conclusions drawn from the
associated Cornu spiral hold for any setup. There is no need to actually be in the near-field
regime; the results hold just as well in the far-field limit. So technically, the title of Section
9.5 should more appropriately be called “Anything-field diffraction” instead of “Near-field
diffraction.” We should therefore be able to obtain the far-field result as a limiting case of
the “near-field” result. Let’s see how this comes about.
For concreteness, let the distance to the screen be D = 100, and let the width of the slit be
a = 5. Then D À a is a fairly good approximation, so we should be able to (approximately)
extract the far-field results from the Cornu spiral. Let’s pick the wavelength to be λ = 1.
Fig. 52 shows three possible locations of the slit. The reason for the particular bounds on
the highest of these slits will be made clear shortly.

wall screen

17.3 < y < 22.3

10 < y < 15

0<y<5
P
D = 100

Figure 52

We can geometrically find the amplitudes at point P due to these three slits in the
following way. The first spiral in Fig. 53 shows the relevant part of the spiral (the thick
part) for the 0 < y < 5 slit, along with the resulting amplitude (the straight line). The
phases from the different points in the slit are roughly equal (because all of the pathlengths
are roughly the same), so the wavelets add generally constructively (they mostly point to
the right), and we end up with a decent-sized amplitude.
34 CHAPTER 9. INTERFERENCE AND DIFFRACTION

(D = 100, λ = 1, a = 5)

slit: 0<y<5 10 < y < 15 17.3 < y < 22.3

Figure 53
The second spiral shows the situation for the 10 < y < 15 slit. The phases now differ by a
larger amount, so the relevant part of the spiral curls around more, and resulting amplitude
isn’t as large. As the slit is raised, eventually we get to a point where the relevant part of
the spiral forms a complete “circle.” (It’s not an actual circle, of course, because it doesn’t
close on itself, but it’s close.) The resulting amplitude is then very small. This corresponds
to the first zero in the diffraction pattern back in Fig. 23. The reason why the amplitude
isn’t exactly zero (as it was in the far-field result) is that D/a is only 20 here. This is fairly
large, but not large enough to make the far-field approximation a highly accurate one. But
remember that the present result is the correct one. The far-field result is an approximation.
If we choose a smaller slit width a, then the relevant part of the spiral (the thick part
in Fig. 53) is shorter. It therefore needs to march deeper into the spiral to get to the point
where it forms a complete circle (because the circles keep shrinking). Since the circles get
closer together as they shrink (eventually they blend together in the figure to form a black
blob), the small sideways shift that represents the amplitude in the third spiral in Fig. 53
is very tiny if the circle is deep in the spiral. So it’s a better approximation to say that the
amplitude there is zero. And consistent with this, the far-field approximation is a better
one, because D/a is larger now. Basically, in the far-field limit, the length of the thick
section in the spiral is much smaller than the general length scale of the spiral.
Note, however, that since the Cornu spiral never crosses itself, it is impossible to ever
get an exactly complete cancelation of the wavelets and thereby a zero amplitude. There
will always be a nonzero sideways shift between the two endpoints of the “circle.” The zeros
in the far-field limit in Fig. 23 are therefore just approximations (but good approximations
if D À a).
Returning to the above case with a = 5, let’s check that the numbers work out. In
the third spiral in Fig. 53, having a complete circle means that the wavelets from the two
ends of the slit have the same phase (because they have the same slope in the spiral). So
the pathlengths from the two ends differ by one wavelength. (This is consistent with the
reasoning in the second bullet point near the beginning of Section 9.3.2.) And indeed, since
the slit runs from y = 17.3 to y = 22.3, and since D = 100, the pathlength difference is
p p
1002 + 22.32 − 1002 + 17.32 = 0.95. (55)
This isn’t exactly equal to one wavelength (which we chose to be λ = 1), but it’s close. A
larger value of D/a would make the difference be closer to one wavelength.
Note that the angle at which the point P is off to the side from the middle of the
17.3 < y < 22.3 slit (which is located at y = 19.8) is given by tan θ = 19.8/100 =⇒ θ =
11.2◦ = 0.195 rad. In the far-field approximation where the paths are essentially parallel,
the difference in pathlengths from the ends of the slit is a sin θ = (5)(sin 11.2◦ ) = 0.97, which
is approximately one wavelength, as it should be.
9.5. NEAR-FIELD DIFFRACTION 35

What if we keep spiraling down into the spiral beyond the position shown in the third
case in Fig. 53? This corresponds to raising the slit (while still keeping the width at a = 5).
Eventually we’ll get to a point where the circles are half as big, so the relevant part of the
curve (the thick part) will wrap twice around a circle. This corresponds to the second zero
in Fig. 23. The difference in the pathlengths from the ends of the slit is now (approximately)
2λ. If we keep spiraling in, the next zero occurs when we wrap three times around a circle.
And so on.
However, we should be careful with this “and so on” statement. In the present case with
a = 5 and λ = 1, it turns out that the part of the curve corresponding to the slit can wrap
around a circle at most 5 times. (And the 5th time actually occurs only in the limit where
the slit is infinitely far up along the wall.) This follows from the fact that since λ = 1, even
if the slit is located at y = ∞, the pathlength from the far end of the slit is only a = 5 longer
than the pathlength from the near end. So the phase difference can be at most 5 cycles. In
other words, the thick part of the curve can’t wrap more than 5 times around in a circle.
Without using this physical reasoning, this limit of 5 circles isn’t obvious by just looking
at the spiral. The circles get smaller and smaller, so you might think that the wrapping
number can be arbitrarily large. However, the little vectors corresponding to a given span
dy are also getting smaller (because the amplitude is small if the slit is far away), which
means that the thick part of the curve gets shorter and shorter. From simply looking at the
curve, it isn’t obvious which effect wins.
36 CHAPTER 9. INTERFERENCE AND DIFFRACTION

9.6 Problems
9.1. Non-normal incidence *
A light wave impinges on an N -slit setup at a small angle γ with respect to the normal.
Show that for small angles, the interference pattern on a far-away screen has the same
form as in Fig. 12, except that the entire plot is shifted by an angle γ. In other words,
it’s the same interference pattern, but now centered around the direction pointing
along a ray of light (or whatever) that passes through the slit region.
9.2. Cornu curvature **
We stated in the last remark in Section 9.5.3 that the Cornu spiral has the property
that the curvature at a given point is proportional to the arclength traversed (starting
at the origin) to that point. Prove this. Hint: Write down the x and y coordinates
associated with Eq. (51), and then find the “velocity” and “acceleration” vectors with
respect to u ≡ zmax , and then use a = v 2 /R.
9.7. SOLUTIONS 37 d sinγ

9.7 Solutions
9.1. Non-normal incidence θ
Fig. 54 shows how to obtain the distances from a given wavefront (the left one in the figure) γ d
to a distance screen. We see that the lower path is longer than the upper path by an amount
d sin θ, but also shorter by an amount d sin γ. So the difference in pathlengths is d(sin θ−sin γ). d sinθ
In the derivation in Section 9.2.1 for the γ = 0 case, the difference in pathlengths was d sin θ.
So the only modification we need to make in the γ 6= 0 case is the replacement of d sin θ in
Eq. (11) (and all subsequent equations) with d(sin θ − sin γ). So Eqs. (14) and (15) become
¡1 ¢ Figure 54
sin 2
N kd(sin θ − sin γ) sin(N α/2)
Atot (θ) = A(θ) ¡1 ¢ ≡ A(θ) , (56)
sin kd(sin θ − sin γ) sin(α/2)
2

where
2πd(sin θ − sin γ)
α ≡ kd(sin θ − sin γ) = . (57)
λ
As before, α is the phase difference between adjacent paths.
For small angles, we can use sin ² ≈ ² to write these results as
¡1 ¢
sin 2
N kd(θ − γ) sin(N α/2)
Atot (θ) = A(θ) ¡1 ¢ ≡ A(θ) , (58)
sin kd(θ − γ) sin(α/2)
2

where
2πd(θ − γ)
α ≡ kd(θ − γ) = . (59)
λ wall
The only difference between this result and the original γ = 0 result (for small θ) is that the screen
argument is θ − γ instead of θ. So the whole interference pattern is translated by an angle γ.
That is, it is centered around θ = γ instead of θ = 0, as we wanted to show.

Remark: The same result holds for the diffraction pattern from a wide slit, because this is simply
the limit of an N -slit setup, with N → ∞. But Fig. 55 gives another quick way of seeing why the
diffraction pattern is centered around the direction of the incident light. Imagine tilting the setup so
that the angle of the incident light is horizontal (so the wavefronts are vertical). Then the wall and
the screen are tilted. But these tilts are irrelevant (for small angles) because when we use Huygens
principle near the slit, the little wavelets are created simultaneously from points on the wavefronts,
and not in the slit. So the setup shown in Fig. 55 is equivalent to having the slit be vertical and
located where the rightmost wavefront is at this instant. (Technically, the width of this vertical slit
would be smaller by a factor of cos γ, but cos γ ≈ 1 for small γ.) And the tilt of the screen is irrelevant Figure 55
for small angles, because any distances along the screen are modified by at most a factor of cos γ. ♣
9.2. Cornu curvature
Writing the exponential in Eq. (51) in terms of trig functions tells us that the x and y
coordinates of the points on √
the spiral in the complex plane are given by (with a ≡ πD/λ,
and ignoring the factor of B0 D)
Z u Z u
x(u) = cos(az 2 ) dz, and y(u) = sin(az 2 ) dz. (60)
0 0

The “velocity” vector with respect to u is given by (dx/du, dy/du). But by the fundamental
theorem of calculus, these derivatives
√ are the values of the integrands evaluated at u. So we
have (up to an overall factor of B0 D)
³ ´ ¡ ¢
dx dy
, = cos(au2 ), sin(au2 ) . (61)
du du

The magnitude of this velocity vector is cos2 (au2 ) + sin2 (au2 ) = 1. So the speed is constant,
independent of the value of u. The total arclength from the origin is therefore simply u.
38 CHAPTER 9. INTERFERENCE AND DIFFRACTION

Since u is the upper limit on the z integral, and since z is proportional to the position y in
the slit (from z ≡ y/D), we’ve just shown that if the upper end of the slit is moved up at a
constant rate (the bottom end is held fixed at y = 0), then the corresponding point on the
Cornu spiral moves along the spiral at a constant rate. If you want, you can think of u as
the time that an object with constant speed has been moving.
The acceleration vector is the derivative of the velocity vector, which gives
µ ¶
d2 x d2 y ¡ ¢
, = − 2au sin(au2 ), 2au cos(au2 ) . (62)
du2 du2

The magnitude of this vector is 2au.

Now, the acceleration, speed, and radius of curvature are related by the usual expression,
a = v 2 /R. So we have R = v 2 /a, which gives R = (1)2 /(2au). The curvature is then
1/R = 2au. But u is the arclength, so we arrive at the desired result that the curvature is
proportional to the arclength. Note that since a ∝ 1/λ, we have R ∝ λ. So a small value of
λ yields a tightly wound (and hence small) spiral. This is consistent with the result in the
second remark in Section 9.5.3.
Chapter 10

Introduction to quantum
mechanics
David Morin, [email protected]

This chapter gives a brief introduction to quantum mechanics. Quantum mechanics can be
thought of roughly as the study of physics on very small length scales, although there are
also certain macroscopic systems it directly applies to. The descriptor “quantum” arises
because in contrast with classical mechanics, certain quantities take on only discrete values.
However, some quantities still take on continuous values, as we’ll see.
In quantum mechanics, particles have wavelike properties, and a particular wave equa-
tion, the Schrodinger equation, governs how these waves behave. The Schrodinger equation
is different in a few ways from the other wave equations we’ve seen in this book. But these
differences won’t keep us from applying all of our usual strategies for solving a wave equation
and dealing with the resulting solutions.
In some respect, quantum mechanics is just another example of a system governed by a
wave equation. In fact, we will find below that some quantum mechanical systems have exact
analogies to systems we’ve already studied in this book. So the results can be carried over,
with no modifications whatsoever needed. However, although it is fairly straightforward
to deal with the actual waves, there are many things about quantum mechanics that are a
combination of subtle, perplexing, and bizarre. To name a few: the measurement problem,
hidden variables along with Bell’s theorem, and wave-particle duality. You’ll learn all about
these in an actual course on quantum mechanics.
Even though there are many things that are highly confusing about quantum mechanics,
the nice thing is that it’s relatively easy to apply quantum mechanics to a physical system
to figure out how it behaves. There is fortunately no need to understand all of the subtleties
about quantum mechanics in order to use it. Of course, in most cases this isn’t the best
strategy to take; it’s usually not a good idea to blindly forge ahead with something if you
don’t understand what you’re actually working with. But this lack of understanding can
be forgiven in the case of quantum mechanics, because no one really understands it. (Well,
maybe a couple people do, but they’re few and far between.) If the world waited to use
quantum mechanics until it understood it, then we’d be stuck back in the 1920’s. The
bottom line is that quantum mechanics can be used to make predictions that are consistent
with experiment. It hasn’t failed us yet. So it would be foolish not to use it.
The main purpose of this chapter is to demonstrate how similar certain results in quan-
tum mechanics are to earlier results we’ve derived in the book. You actually know a good

1
2 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

deal of quantum mechanics already, whether you realize it or not.

The outline of this chapter is as follows. In Section 10.1 we give a brief history of the
development of quantum mechanics. In Section 10.2 we write down, after some motivation,
the Schrodinger wave equation, both the time-dependent and time-independent forms. In
Section 10.3 we discuss a number of examples. The most important thing to take away from
this section is that all of the examples we discuss have exact analogies in the string/spring
systems earlier in the book. So we technically won’t have to solve anything new here. All
the work has been done before. The only thing new that we’ll have to do is interpret the old
results. In Section 10.4 we discuss the uncertainty principle. As in Section 10.3, we’ll find
that we already did the necessary work earlier in the book. The uncertainty principle turns
out to be a direct consequence of a result from Fourier analysis. But the interpretation of
this result as an uncertainty principle has profound implications in quantum mechanics.

10.1 A brief history

Before discussing the Schrodinger wave equation, let’s take a brief (and by no means com-
prehensive) look at the historical timeline of how quantum mechanics came about. The
actual history is of course never as clean as an outline like this suggests, but we can at least
get a general idea of how things proceeded.
1900 (Planck): Max Planck proposed that light with frequency ν is emitted in quantized
lumps of energy that come in integral multiples of the quantity,

E = hν = h̄ω (1)
where h ≈ 6.63 · 10−34 J · s is Planck’s constant, and h̄ ≡ h/2π = 1.06 · 10−34 J · s.
The frequency ν of light is generally very large (on the order of 1015 s−1 for the visible
spectrum), but the smallness of h wins out, so the hν unit of energy is very small (at least on
an everyday energy scale). The energy is therefore essentially continuous for most purposes.
However, a puzzle in late 19th-century physics was the blackbody radiation problem. In a
nutshell, the issue was that the classical (continuous) theory of light predicted that certain
objects would radiate an infinite amount of energy, which of course can’t be correct. Planck’s
hypothesis of quantized radiation not only got rid of the problem of the infinity, but also
correctly predicted the shape of the power curve as a function of temperature.
The results that we derived for electromagnetic waves in Chapter 8 are still true. In
particular, the energy flux is given by the Poynting vector in Eq. 8.47. And E = pc for
a light. Planck’s hypothesis simply adds the information of how many lumps of energy a
wave contains. Although strictly speaking, Planck initially thought that the quantization
was only a function of the emission process and not inherent to the light itself.
1905 (Einstein): Albert Einstein stated that the quantization was in fact inherent to the
light, and that the lumps can be interpreted as particles, which we now call “photons.” This
proposal was a result of his work on the photoelectric effect, which deals with the absorption
of light and the emission of elections from a material.
We know from Chapter 8 that E = pc for a light wave. (This relation also follows from
Einstein’s 1905 work on relativity, where he showed that E = pc for any massless particle,
an example of which is a photon.) And we also know that ω = ck for a light wave. So
Planck’s E = h̄ω relation becomes
E = h̄ω =⇒ pc = h̄(ck) =⇒ p = h̄k (2)

This result relates the momentum of a photon to the wavenumber of the wave it is associated
with.
10.1. A BRIEF HISTORY 3

1913 (Bohr): Niels Bohr stated that electrons in atoms have wavelike properties. This
correctly explained a few things about hydrogen, in particular the quantized energy levels
that were known.
1924 (de Broglie): Louis de Broglie proposed that all particles are associated with waves,
where the frequency and wavenumber of the wave are given by the same relations we found
above for photons, namely E = h̄ω and p = h̄k. The larger E and p are, the larger ω
and k are. Even for small E and p that are typical of a photon, ω and k are very large
because h̄ is so small. So any everyday-sized particle with large (in comparison) energy and
momentum values will have extremely large ω and k values. This (among other reasons)
makes it virtually impossible to observe the wave nature of macroscopic amounts of matter.
This proposal (that E = h̄ω and p = h̄k also hold for massive particles) was a big step,
because many things that are true for photons are not true for massive (and nonrelativistic)
particles. For example, E = pc (and hence ω = ck) holds only for massless particles (we’ll
see below how ω and k are related for massive particles). But the proposal was a reasonable
one to try. And it turned out to be correct, in view of the fact that the resulting predictions
agree with experiments.
The fact that any particle has a wave associated with it leads to the so-called wave-
particle duality. Are things particles, or waves, or both? Well, it depends what you’re doing
with them. Sometimes things behave like waves, sometimes they behave like particles. A
vaguely true statement is that things behave like waves until a measurement takes place,
at which point they behave like particles. However, approximately one million things are
left unaddressed in that sentence. The wave-particle duality is one of the things that few
people, if any, understand about quantum mechanics.
1925 (Heisenberg): Werner Heisenberg formulated a version of quantum mechanics that
made use of matrix mechanics. We won’t deal with this matrix formulation (it’s rather
difficult), but instead with the following wave formulation due to Schrodinger (this is a
waves book, after all).
1926 (Schrodinger): Erwin Schrodinger formulated a version of quantum mechanics that
was based on waves. He wrote down a wave equation (the so-called Schrodinger equation)
that governs how the waves evolve in space and time. We’ll deal with this equation in depth
below. Even though the equation is correct, the correct interpretation of what the wave
actually meant was still missing. Initially Schrodinger thought (incorrectly) that the wave
represented the charge density.
1926 (Born): Max Born correctly interpreted Schrodinger’s wave as a probability am-
plitude. By “amplitude” we mean that the wave must be squared to obtain the desired
probability. More precisely, since the wave (as we’ll see) is in general complex, we need to
square its absolute value. This yields the probability of finding a particle at a given location
(assuming that the wave is written as a function of x).
This probability isn’t a consequence of ignorance, as is the case with virtually every
other example of probability you’re familiar with. For example, in a coin toss, if you
know everything about the initial motion of the coin (velocity, angular velocity), along
with all external influences (air currents, nature of the floor it lands on, etc.), then you
can predict which side will land facing up. Quantum mechanical probabilities aren’t like
this. They aren’t a consequence of missing information. The probabilities are truly random,
and there is no further information (so-called “hidden variables”) that will make things un-
random. The topic of hidden variables includes various theorems (such as Bell’s theorem)
and experimental results that you will learn about in a quantum mechanics course.
4 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

1926 (Dirac): Paul Dirac showed that Heisenberg’s and Schrodinger’s versions of quantum
mechanics were equivalent, in that they could both be derived from a more general version
of quantum mechanics.

10.2 The Schrodinger equation

In this section we’ll give a “derivation” of the Schrodinger equation. Our starting point will
be the classical nonrelativistic expression for the energy of a particle, which is the sum of
the kinetic and potential energies. We’ll assume as usual that the potential is a function of
only x. We have
1 p2
E = K + V = mv 2 + V (x) = + V (x). (3)
2 2m
We’ll now invoke de Broglie’s claim that all particles can be represented as waves with
frequency ω and wavenumber k, and that E = h̄ω and p = h̄k. This turns the expression
for the energy into
h̄2 k 2
h̄ω = + V (x). (4)
2m
A wave with frequency ω and wavenumber k can be written as usual as ψ(x, t) = Aei(kx−ωt)
(the convention is to put a minus sign in front of the ωt). In 3-D we would have ψ(r, t) =
Aei(k·r−ωt) , but let’s just deal with 1-D. We now note that
∂ψ ∂ψ
= −iωψ =⇒ ωψ = i , and
∂t ∂t
∂2ψ ∂2ψ
= −k 2 ψ =⇒ k2 ψ = − 2 . (5)
∂x2 ∂x
If we multiply the energy equation in Eq. (4) by ψ, and then plug in these relations, we
obtain
h̄2 2 ∂ψ −h̄2 ∂ 2 ψ
h̄(ωψ) = (k ψ) + V (x)ψ =⇒ ih̄ = · +Vψ (6)
2m ∂t 2m ∂x2
This is the time-dependent Schrodinger equation. If we put the x and t arguments back in,
the equation takes the form,

∂ψ(x, t) −h̄2 ∂ 2 ψ(x, t)

ih̄ = · + V (x)ψ(x, t). (7)
∂t 2m ∂x2
In 3-D, the x dependence turns into dependence on all three coordinates (x, y, z), and the
∂ 2 ψ/∂x2 term becomes ∇2 ψ (the sum of the second derivatives). Remember that Born’s
(correct) interpretation of ψ(x) is that |ψ(x)|2 gives the probability of finding the particle
at position x.
Having successfully produced the time-dependent Schrodinger equation, we should ask:
Did the above reasoning actually prove that the Schrodinger equation is valid? No, it didn’t,
for three reasons.

1. The reasoning is based on de Broglie’s assumption that there is a wave associated with
every particle, and also on the assumption that the ω and k of the wave are related to
E and p via Planck’s constant in Eqs. (1) and (2). We had to accept these assumptions
on faith.

2. Said in a different way, it is impossible to actually prove anything in physics. All we

can do is make an educated guess at a theory, and then do experiments to try to show
10.2. THE SCHRODINGER EQUATION 5

that the theory is consistent with the real world. The more experiments we do, the
more comfortable we are that the theory is a good one. But we can never be absolutely
sure that we have the correct theory. In fact, odds are that it’s simply the limiting
case of a more correct theory.

3. The Schrodinger equation actually isn’t valid, so there’s certainly no way that we
proved it. Consistent with the above point concerning limiting cases, the quantum
theory based on Schrodinger’s equation is just a limiting theory of a more correct one,
which happens to be quantum field theory (which unifies quantum mechanics with
special relativity). This is turn must be a limiting theory of yet another more correct
one, because it doesn’t incorporate gravity. Eventually there will be one theory that
covers everything (although this point can be debated), but we’re definitely not there
yet.

Due to the “i” that appears in Eq. (6), ψ(x) is complex. And in contrast with waves in
classical mechanics, the entire complex function now matters in quantum mechanics. We
won’t be taking the real part in the end. Up to this point in the book, the use of complex
functions was simply a matter of convenience, because it is easier to work with exponentials
than trig functions. Only the real part mattered (or imaginary part – take your pick, but not
both). But in quantum mechanics the whole complex wavefunction is relevant. However,
the theory is structured in such a way that anything you might want to measure (position,
momentum, energy, etc.) will always turn out to be a real quantity. This is a necessary
feature of any valid theory, of course, because you’re not going to go out and measure a
distance of 2 + 5i meters, or pay an electrical bill of 17 + 6i kilowatt hours.
As mentioned in the introduction to this chapter, there is an endless number of difficult
questions about quantum mechanics that can be discussed. But in this short introduction
to the subject, let’s just accept Schrodinger’s equation as valid, and see where it takes us.

Solving the equation

If we put aside the profound implications of the Schrodinger equation and regard it as
simply a mathematical equation, then it’s just another wave equation. We already know
the solution, of course, because we used the function ψ(x, t) = Aei(kx−ωt) to produce Eqs.
(5) and (6) in the first place. But let’s pretend that we don’t know this, and let’s solve the
Schrodinger equation as if we were given it out of the blue.
As always, we’ll guess an exponential solution. If we first look at exponential behavior
in the time coordinate, our guess is ψ(x, t) = e−iωt f (x) (the minus sign here is convention).
Plugging this into Eq. (7) and canceling the e−iωt yields

h̄2 ∂ 2 f (x)
h̄ωf (x) = − + V (x)f (x). (8)
2m ∂x2
But from Eq. (1), we have h̄ω = E. And we’ll now replace f (x) with ψ(x). This might
cause a little confusion, since we’ve already used ψ to denote the entire wavefunction ψ(x, t).
However, it is general convention to also use the letter ψ to denote the spatial part. So we
now have
h̄2 ∂ 2 ψ(x)
E ψ(x) = − + V (x)ψ(x) (9)
2m ∂x2
This is called the time-independent Schrodinger equation. This equation is more restrictive
than the original time-dependent Schrodinger equation, because it assumes that the parti-
cle/wave has a definite energy (that is, a definite ω). In general, a particle can be in a state
that is the superposition of states with various definite energies, just like the motion of a
6 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

string can be the superposition of various normal modes with definite ω’s. The same rea-
soning applies here as with all the other waves we’ve discussed: From Fourier analysis and
from the linearity of the Schrodinger equation, we can build up any general wavefunction
from ones with specific energies. Because of this, it suffices to consider the time-independent
Schrodinger equation. The solutions for that equation form a basis for all possible solutions.1
Continuing with our standard strategy of guessing exponentials, we’ll let ψ(x) = Aeikx .
Plugging this into Eq. (9) and canceling the eikx gives (going back to the h̄ω instead of E)

h̄2 h̄2 k 2
h̄ω = − (−k 2 ) + V (x) =⇒ h̄ω = + V (x). (10)
2m 2m
This is simply Eq. (4), so we’ve ended up back where we started, as expected. However, our
goal here was to show how the Schrodinger equation can be solved from scratch, without
knowing where it came from.
Eq. (10) is (sort of) a dispersion relation. If V (x) is a constant C in a given region, then
the relation between ω and k (namely ω = h̄k 2 /2m + C) is independent of x, so we have
a nice sinusoidal wavefunction (or exponential, if k is imaginary). However, if V (x) isn’t
constant, then the wavefunction isn’t characterized by a unique wavenumber. So a function
of the form eikx doesn’t work as a solution for ψ(x). (A Fourier superposition can certainly
work, since any function can be expressed that way, but a single eikx by itself doesn’t work.)
This is similar to the case where the density of a string isn’t constant. We don’t obtain
sinusoidal waves there either.

10.3 Examples
In order to solve for the wavefunction ψ(x) in the time-independent Schrodinger equation
in Eq. (9), we need to be given the potential energy V (x). So let’s now do some examples
with particular functions V (x).

10.3.1 Constant potential

The simplest example is where we have a constant potential, V (x) = V0 in a given region.
Plugging ψ(x) = Aeikx into Eq. (9) then gives
s
h̄2 k 2 2m(E − V0 )
E= + V0 =⇒ k = . (11)
2m h̄2

(We’ve taken the positive square root here. We’ll throw in the minus sign by hand to obtain
the other solution, in the discussion below.) k is a constant, and its real/imaginary nature
depends on the relation between E and V0 . If E > V0 , then k is real, so we have oscillatory
solutions,
ψ(x) = Aeikx + Be−ikx . (12)
But if E < V0 , thenpk is imaginary, so we have exponentially growing or decaying solutions.
If we let κ ≡ |k| = 2m(V0 − E)/h̄, then ψ(x) takes the form,

ψ(x) = Aeκx + Ba−κx . (13)

We see that it is possible for ψ(x) to be nonzero in a region where E < V0 . Since ψ(x) is
the probability amplitude, this implies that it is possible to have a particle with E < V0 .
1 The “time-dependent” and “time-independent” qualifiers are a bit of a pain to keep saying, so we usually

just say “the Schrodinger equation,” and it’s generally clear from the context which one we mean.
10.3. EXAMPLES 7

This isn’t possible classically, and it is one of the many ways in which quantum mechanics
diverges from classical mechanics. We’ll talk more about this when we discuss the finite
square well in Section 10.3.3.
If E = V0 , then this is the one case where the strategy of guessing an exponential function
doesn’t work. But if we go back to Eq. (9) we see that E = V0 implies ∂ 2 ψ/∂x2 = 0, which
in turn implies that ψ is a linear function,
ψ(x) = Ax + B. (14)
In all of these cases, the full wavefunction (including the time dependence) for a particle
with a specific value of E is given by

ψ(x, t) = e−iωt ψ(x) = e−iEt/h̄ ψ(x) (15)

Again, we’re using the letter ψ to stand for two different functions here, but the meaning of
each is clear from the number of arguments. Any general wavefunction is built up from a
superposition of the states in Eq. (15) with different values of E, just as the general motion
of a string is built of from various normal modes with different frequencies ω. The fact that a
particle can be in a superposition of states with different energies is another instance where
quantum mechanics diverges from classical mechanics. (Of course, it’s easy for classical
waves to be in a superposition of normal modes with different energies, by Fourier analysis.)
The above E > V0 and E < V0 cases correspond, respectively, to being above or below
the cutoff frequency in the string/spring system we discussed in Section 6.2.2. We have an
oscillatory solution if E (or ω) is above a particular value, and an exponential solution if
E (or ω) is below a particular value. The two setups (quantum mechanical with constant
V0 , and string/spring with springs present everywhere) are exactly analogous to each other.
The spatial parts of the solutions are exactly the same (well, before taking the real part
in the string/spring case). The frequencies, however, are different, because the dispersion
relations are different (h̄ω = h̄2 k 2 /2m + V0 and ω 2 = c2 k 2 + ωs2 , respectively). But this
affects only the rate of oscillation, and not the shape of the function.
The above results hold for any particular region where V (x) is constant. What if the
region extends from, say, x = 0 to x = +∞? If E > V0 , the oscillatory solutions are fine,
even though they’re not normalizable. That is, the integral of |ψ|2 is infinite (at least for any
nonzero coefficient in ψ; if the coefficient were zero, then we wouldn’t have a particle). So
we can’t make the total probability equal to 1. However, this is fine. The interpretation is
that we simply have a stream of particles extending to infinity. We shouldn’t be too worried
about this divergence, because when dealing with traveling waves on a string (for example,
when discussing reflection and transmission coefficients) we assumed that the sinusiodal
waves extended to ±∞, which of course is impossible in reality.
If E < V0 , then the fact that x = +∞ is in the given region implies that the coefficient
A in Eq. (13) must be zero, because otherwise ψ would diverge as x → ∞. So we are left
with only the Ba−κx term. (It’s one thing to have the integral of |ψ|2 diverge, as it did in
the previous paragraph. It’s another thing to have the integral diverge and be dominated
by values at large x. There is then zero probability of finding the particle at a finite value
of x.) If the region where E < V0 is actually the entire x axis, from −∞ to ∞, then the B
coefficient in Eq. (13) must also be zero. So ψ(x) = 0 for all x. In other words, there is no
allowed wavefunction. It is impossible to have a particle with E < V0 everywhere.

10.3.2 Infinite square well

Consider the potential energy,
½
0 (0 ≤ x ≤ L)
V (x) = (16)
∞ (x < 0 or x > L).
 8 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

This is called an “infinite square well,” and it is shown in Fig. 1. The “square” part of the
V= V= name comes from the right-angled corners and not from the actual shape, since it’s a very
8

8 (infinitely) tall rectangle. This setup is also called a “particle in a box” (a 1-D box), because
the particle can freely move around inside a given region, but has zero probability of leaving
V=0 the region, just like a box. So ψ(x) = 0 outside the box.
-a a
The particle does indeed have zero chance of being found outside the region 0 ≤ x ≤ L.
Figure 1 Intuitively, this is reasonable, because the particle would have to climb the infinitely high
potential cliff at the side of the box. Mathematically, this can be derived rigorously, and
we’ll do this below when we discuss the finite square well.
We’ll assume E > 0, because the E < 0 case makes E < V0 everywhere, which isn’t
possible, as we mentioned above. Inside the well, we have V (x) = 0, so this is a special
case of the constant potential discussed above. We therefore have the oscillatory solution
in Eq. (12) (since E > 0), which we will find more convenient here to write in terms of trig
functions,
√
h̄2 k 2 2mE
ψ(x) = A cos kx + B sin kx, where E = =⇒ k = . (17)
2m h̄
The coefficients A and B may be complex.
We now claim that ψ must be continuous at the boundaries at x = 0 and x = L. When
dealing with, say, waves on a string, it was obvious that the function ψ(x) representing
the transverse position must be continuous, because otherwise the string would have a
break in it. But it isn’t so obvious with the quantum-mechanical ψ. There doesn’t seem
to be anything horribly wrong with having a discontinuous probability distribution, since
probability isn’t an actual object. However, it is indeed true that the probability distribution
is continuous in this case (and in any other case that isn’t pathological). For now, let’s just
assume that this is true, but we’ll justify it below when we discuss the finite square well.
Since ψ(x) = 0 outside the box, continuity of ψ(x) at x = 0 quickly gives A cos(0) +
B sin(0) = 0 =⇒ A = 0. Continuity at x = L then gives B sin kL = 0 =⇒ kL = nπ, where
n is an integer. So k = nπ/L, and the solution for ψ(x) is ψ(x) = B sin(nπx/L). The full
solution including the time dependence is given by Eq. (15) as

³ nπx ´ h̄2 k 2 n2 π 2 h̄2

ψ(x, t) = Be−iEt/h̄ sin where E= = (18)
L 2m 2mL2

We see that the energies are quantized (that is, they can take on only discrete values) and
indexed by the integer n. The string setup that is analogous to the infinite square well is
a string with fixed ends, which we discussed in Chapter 4 (see Section 4.5.2). In both of
these setups, the boundary conditions yield the same result that an integral number of half
n=4 wavelengths fit into the region. So the k values take the same form, k = nπ/L.
The dispersion relation, however, is different. It was simply ω = ck for waves on a
string, whereas it is h̄ω = h̄2 k 2 /2m for the V (x) = 0 region of the infinite well. But as
n=3 in the above case of the constant potential, this difference affects only the rate at which
the waves oscillate in time. It does’t affect the spatial shape, which is determined by the
wavenumber k. The wavefunctions for the lowest four energies are shown in Fig. 2 (the
n=2
vertical separation between the curves is meaningless). These look exactly like the normal
modes in the “both ends fixed” case in Fig. 24 in Chapter 4.
n=1
x=0 x=L

Figure 2
 Energy in units of
π2h2/2mL2
10.3. EXAMPLES 9
E
16 n=4
The corresponding energies are shown in Fig. 3. Since E ∝ ω = (h̄2 /2m)k 2 ∝ n2 , the
gap between the energies grows as n increases. Note that the energies in the case of a string
are also proportional to n2 , because although ω = ck ∝ n, the energy is proportional to ω 2
(because the time derivative in Eq. (4.50) brings down a factor of ω). So Figs. 2 and 3 both
√
apply to both systems. The difference between the systems is that a string has ω ∝ E,
where as the quantum mechanical system has ω ∝ E. 9 n=3
There is no n = 0 state, because from Eq. (18) this would make ψ be identically zero.
That wouldn’t be much of a state, because the probability would be zero everywhere. The
lack of a n = 0 state is consistent with the uncertainty principle (see Section 10.4 below), n=2
4
because such a state would have ∆x∆p = 0 (since ∆x < L, and ∆p = 0 because n = 0 =⇒
k = 0 =⇒ p = h̄k = 0), which would violate the principle.
1 n=1

10.3.3 Finite square well Figure 3

Things get more complicated if we have a finite potential well. For future convenience, we’ll
V = V0
let x = 0 be located at the center of the well. If we label the ends as ±a, then V (x) is given
by
½
0 (|x| ≤ a) V=0
V (x) = (19) -a a
V0 (|x| > a).
This is shown in Fig. 4. Given V0 , there are two basic possibilities for the energy E: Figure 4

• E >p V0 (unbound state): From Eq. (11),√ the wavenumber k takes thep general form
of 2m(E − V (x))/h̄. This equals 2mE/h̄ inside the well and 2m(E − V0 )/h̄ E
outside. k is therefore real everywhere, so ψ(x) is an oscillatory function both inside V0
and outside the well. k is larger inside the well, so the wavelength is shorter there. A
possible wavefunction might look something like the one in Fig. 5. It is customary to -a a
draw the ψ(x) function on top of the E line, although this technically has no meaning
because ψ and E have different units. Figure 5
The wavefunction extends infinitely on both direction, so the particle can be anywhere.
Hence the name “unbound state.” We’ve drawn an even-function standing wave in
Fig. 5, although in general we’re concerned with traveling waves for unbound states.
These are obtained from superpositions of the standing waves, with a phase thrown
in the time dependence. For traveling waves, the relative sizes of ψ(x) in the different
regions depend on the specifics of how the problem is set up.
√
• p
0 < E < V0 (bound state): The wavenumber k still equals 2mE/h̄ inside the well and
2m(E − V0 )/h̄ outside, but now that latter value is imaginary. So ψ is an oscillatory
function inside the well, but an exponential function outside. Furthermore, it must
be an exponentially decaying function outside, because otherwise it would diverge at
x = ±∞. Since the particle has an exponentially small probability of being found
far away from the well, we call this a “bound state.” We’ll talk more below about
the strange fact that the probability is nonzero in the region outside the well, where
E < V (x).
There is also the third case were E = V0 , but this can be obtained as the limit of the
other two cases (more easily as the limit of the bound-state case). The fourth case,
E < 0, isn’t allowed, as we discussed at the end of Section 10.3.1.

In both of these cases, the complete solution for ψ(x) involves solving the boundary
conditions at x = ±a. The procedure is the same for both cases, but let’s concentrate on
the bound-state case here. The boundary conditions are given by the following theorem.
10 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

Theorem 10.1 If V (x) is everywhere finite (which is the case for the finite square well),
then both ψ(x) and ψ 0 (x) are everywhere continuous.

Proof: If we solve for ψ 00 in Eq. (9), we see that ψ 00 is always finite (because V (x) is always
finite). This implies two things. First, it implies that ψ 0 must be continuous, because if ψ 0
were discontinuous at a given point, then its derivative ψ 00 would be infinite there (because
ψ 0 would make a finite jump over zero distance). So half of the theorem is proved.
Second, the finiteness of ψ 00 implies that ψ 0 must also be finite everywhere, because if
ψ were infinite at a given point (excluding x = ±∞), then its derivative ψ 00 would also be
0

infinite there (because ψ 0 would make an infinite jump over a finite distance).
Now, since ψ 0 is finite everywhere, we can repeat the same reasoning with ψ 0 and ψ that
we used with ψ 00 and ψ 0 in the first paragraph above: Since ψ 0 is always finite, we know
that ψ must be continuous. So the other half of the theorem is also proved.
Having proved this theorem, let’s outline the general strategy for solving for ψ in the
E < V0 case. The actual task of going through the calculation is left for Problem 10.2. The
calculation is made much easier with the help of Problem 10.1 which states that only even
and odd functions need to be considered. p
If we let k ≡ iκ outside the well, then we have κ = 2m(V0 − E)/h̄, which is real and
positive since E < V0 . The general forms of the wavefunctions in the left, middle, and right
regions are

x < −a : ψ1 (x) = A1 eκx + B1 e−κx ,

−a < x < a : ψ2 (x) = A2 eikx + B2 e−ikx ,
x>a: ψ3 (x) = A3 eκx + B3 e−κx , (20)

where √ p
2mE 2m(V0 − E)
k= , and κ= . (21)
h̄ h̄
We’ve given only the x dependence in these wavefunctions. To obtain the full wavefunction
ψ(x, t), all of these waves are multiplied by the same function of t, namely e−iωt = e−iEt/h̄ .
We now need to solve for various quantities. How many unknowns do we have, and how
many equations/facts do we have? We have seven unknowns: A1 , A2 , A3 , B1 , B2 , B3 , and
E (which appears in k and κ). And we have seven facts:

• Four boundary conditions at x = ±a, namely continuity of ψ and ψ 0 at both points.

• Two boundary conditions at x = ±∞, namely ψ = 0 in both cases.

R∞
• One normalization condition, namely −∞ |ψ|2 dx = 1.
As we mentioned at the end of Section 10.3.1, the boundary conditions at ±∞ quickly
tell us that B1 and A3 equal zero. Also, in most cases we’re not concerned with the overall
normalization constant (the usual goal is to find E), so we can ignore the normalization
condition and just find all the other constants in terms of, say, A1 . So were’re down to
four equations (the four boundary conditions at x = ±a), and four unknowns (A2 , B2 , B3 ,
and E). Furthermore, the even/odd trick discussed in Problem 10.1 cuts things down by a
factor of 2, so we’re down to two equations and two unknowns (the energy E, along with
one of the coefficients), which is quite manageable. The details are left for Problem 10.2,
but let’s get a rough idea here of what the wavefunctions look like.
 V0
10.3. EXAMPLES 11
E5

It turns out that the energies and states are again discrete and can be labeled by an
integer n, just as in the infinite-well case. However, the energies don’t take the simple form
in Eq. (18), although they approximately do if the well is deep. Fig. 6shows the five states for E4
a well of a particular depth V0 . We’ve drawn each wave relative to the line that represents
the energy En . Both ψ and ψ 0 are continuous at x = ±a, and ψ goes to 0 at x = ±∞.
We’ve chosen the various parameters (one of which is the depth) so that there are exactly E3
five states (see Problem 10.2 for the details on this). The deeper the well, the more states
there are.
E2
Consistent with Eq. (20), ψ is indeed oscillatory inside the well (that is, the curvature
is toward the x axis), and exponential decaying outside the well (the curvature is away E1
from the x axis). As E increases, Eq. (21) tells us that k increases (so the wiggles inside -a a V=E=0
the well have shorter wavelengths), and also that κ decreases (so the exponential decay is
slower). These facts are evident in Fig. 6. The exact details of the waves depend on various Figure 6
parameters, but the number of bumps equals n.

Explanation of the quantized energy levels

The most important thing to note about these states is that they are discrete. In the infinite-
well case, this discreteness was clear because an integral number of half wavelengths needed
to fit into the well (because ψ = 0 at the boundaries). The discreteness isn’t so obvious in
the finite-well case (because ψ 6= 0 at the boundaries), but it is still reasonably easy to see.
There are two ways to understand it. First, the not-so-enlightening way is to note that we
initially had 7 equation and 7 unknowns. So all the unknowns, including E, are determined.
There may be different discrete solutions, but at least we know that we can’t just pick a
random value for E and expect it to work.
Second, the more physical and enlightening way is the following. If we choose a random
value of E, it probably won’t yield an allowable function ψ, and here’s why. Pick an arbitrary
value of the coefficient A1 in Eq. (20), and set B1 = 0. So we have an exponentially decaying
function in the left region, which behaves properly at x = −∞. Since E determines κ, we
know everything about the A1 eκx function.
Now pick a point x0 in the left region, and imagine marching rightward on the x axis.
We claim that all of the subsequent values of ψ are completely determined, all the way up
to x = +∞. This is clear in the left region, because we know what the function is. But it
is also true in the middle and right regions, because we know the values of ψ, ψ 0 , and ψ 00
at any given point, so we can recursively find these three values at the “next” point as we
march along. More precisely: (a) from the definition of the derivative, the values of ψ and
ψ 0 at a given point yield the value of ψ at the next point, (b) likewise, the values of ψ 0 and
ψ 00 at a given point yield the value of ψ 0 at the next point, and (c) finally, the value of ψ at
a given point yields the value of ψ 00 at that point, via the Schrodinger equation, Eq. (9). So
we can recursively march all the way up to x = +∞, and the entire function is determined.
There is no freedom whatsoever in what the function turns out to be.
A particular choice of E and A1 might yield the first (top) function shown in Fig. 7. It has
the correct exponential and oscillatory nature in the left and middle regions, respectively.
But in the right region it apparently has an exponentially growing piece. Because it diverges
at x = +∞, this function isn’t an allowable one. So it is impossible for the energy to take
on the value that we chose.
We can try to remedy the divergence at x = +∞ by increasing the value of E. This will -a a
make ψ oscillate quicker inside the well, so that it encounters the x = a boundary with a
negative slope. We then might end up with the second function shown in Fig. 7. We still Figure 7
have the divergence at x = +∞, so again we have an invalid ψ. If we increase E a little
more, by precisely the right amount, then we’ll end up with the third function shown in Fig.
 12 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

7. The growing exponential term is now absent from the right region, so we have success.
This function is allowable, as is the associated energy.
If we increase E a little more, then we’ll end up with something like the fourth (bottom)
function in Fig. 7. This function diverges, but now in the negative direction. So we’re back
to an invalid ψ. If we continue to increase E, then eventually we’ll end up with the situation
where there is now (approximately) one additional half oscillation, and the function again
decays to zero at x = +∞. And so on and so forth with additional half-oscillations, although
eventually E will become larger than V0 , in which case we won’t have a bound state anymore.
(We can also decrease E from the value in the top plot in Fig. 7. We will encounter ψ’s
with two bumps and then one bump. The latter will have the lowest possible energy.) This
reasoning makes it clear why only discrete values of E are allowed. Only special values of
E make the coefficient A3 in Eq. (20) be zero. Other values of E produce an exponentially
growing piece in the right region, resulting in a non-normalizable wavefunction.

The terminology for this bound-state setup is that the particle is “trapped” in the well.
However, we have discovered the strange fact that the particle has a nonzero probability
of being found outside the well where E < V0 , because ψ is nonzero there. Strange, but
true. Classically, it is impossible for a particle to have E < V . But classical mechanics isn’t
correct. It usually is, but not always. And this is one of the places where it fails.
p What happens if V0 → ∞, so that the finite well approaches an infinite well? κ equals
2m(V0 − E)/h̄, so it approaches ∞ as V0 → ∞. The exponential decay outside the well
is therefore infinitely quick. In the case of a very large by finite V0 , Fig. 8 shows what the
first two states look like. As V0 increases further, the region of the exponential decay gets
-a a
smaller and smaller. Eventually you can’t tell the difference between the plots in Fig. 8 and
Figure 8 the bottom two plots in Fig. 2 for the infinite well. Technically the decay for the finite well
always extends to x = ±∞, because the exponential function never actually equals zero.
But it’s essentially zero after a very short distance.
We see that as V0 → ∞, ψ approaches a function that is still continuous, but has a
discontinuity in its first derivative. So if we consider the infinite well in Section 10.3.2 to
be the limit of a finite well as V0 → ∞ (which is the most reasonable way to consider it),
then the present discussion justifies our assumption in Section 10.3.2 that ψ was continuous.
And furthermore it justifies why we didn’t also assume that ψ 0 was continuous.
As with the infinite square well, the finite square well also has a direct analogy with a
setup involving a string. Recall the discussion of the string/spring system in Section 6.2.2,
involving evanescent waves and a low-frequency cutoff. Consider the system shown in Fig.
9, where the springs extend to x = ±∞. From Section 6.2.2, we know that if the frequency
is below the cutoff frequency, then we have a sinusoidal wave in the middle region, but
an evanescent wave (that is, an exponentially decaying wave) in the side regions. This is
exactly what we have in the quantum-mechanical finite-well setup. In both setups the most
general forms of the waves (as functions of x) in the different regions are given by Eq. (20)
(but we need to take the real part in the string/spring case). And the boundary conditions
are the same: continuity of ψ and ψ 0 at the boundaries, and ψ = 0 at x = ±∞. If you’ve
solved for one ψ(x), you’ve solved for the other.

(no springs)

-a a

Figure 9
10.3. EXAMPLES 13

However, as with the other examples we’ve studied, the dispersion relation in the string/spring
system is different from the relation in the quantum-mechanical system. In particular, in
the quantum case, the wave equations and dispersion relations inside and outside the well
are
∂ψ −h̄2 ∂ 2 ψ h̄2 k 2
Inside : ih̄ = · 2
=⇒ h̄ω = ,
∂t 2m ∂x 2m
∂ψ −h̄2 ∂ 2 ψ h̄2 k 2
Outside : ih̄ = · + V 0 ψ =⇒ h̄ω = + V0 . (22)
∂t 2m ∂x2 2m
And for the string/spring system we have

∂2ψ 2
2∂ ψ
Middle : = c =⇒ ω 2 = c2 k 2 ,
∂t2 ∂x2
∂2ψ ∂2ψ
Sides : 2
= c2 2 + ωs2 ψ =⇒ ω 2 = c2 k 2 + ωs2 . (23)
∂t ∂x
But the differences in these equations don’t affect the shape of the waves, because the shape
is determined by the wavenumbers k and κ, where κ is the exponential decay constant. ω is
irrelevant
p for the shape; it determines only how fastpthe wave oscillates in time. κ is given
by κ = 2m(V0 − h̄ω)/h̄ in the quantum case, and (ωs2 − ω 2 )/c in the string/spring case.
So in both cases, k and κ are related by an equation of the form, k 2 + κ2 = A, where A is
a constant that equals 2mV0 /h̄2 in the quantum case, and ωs2 /c2 in the string/spring case.
Note that ω doesn’t appear in A in either case. When you solve Problem 10.2, you will
see that the shape of the wave is determined by the boundary conditions, along with the
k 2 + κ2 = A equation. So the shape is indeed independent of ω.

10.3.4 Tunneling
Consider the string/spring system that is the “opposite” of the system shown in Fig. 9. So
we now have springs in the middle region, and a normal string in the outer regions, as shown
in Fig. 10.

-a a

Figure 10

If a rightward-traveling wave comes in from the left, and if the frequency ω is less than
the cutoff frequency ωs , then we will have an evanescent wave in the middle region. And we
will also have rightward-traveling wave in the right region (but no leftward-traveling wave,
because we’re not sending anything in from the right). So the waves in the three regions
take the general form,

x < −a : ψ1 (x) = A1 eikx + B1 e−ikx ,

−a < x < a : ψ2 (x) = A2 eκx + B2 e−κx ,
x>a: ψ3 (x) = A3 eikx . (24)

We’ve given only the x dependence here. All of these waves are multiplied by the same
function of t, namely e−iωt . The ratio B1 /A1 is the reflection coefficient, and the ratio A3 /A1
 14 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

is the transmission coefficient. Note that we can have both the exponentially growing and
decaying terms in the middle region, because the boundary conditions at ±∞ don’t apply.
The frequency ω can take on any value. It need not take on certain discrete values as
it did in the “opposite” string/spring system analogous to the finite well above. There are
two ways of seeing why this is true. First, it makes physical sense, because you are free to
wiggle the string at whatever ω you want, when creating the incoming wave, A1 ei(kx−ωt) . 2
Second, it makes mathematical sense if we count the number of unknowns and equations.
In the finite-well case, we had seven unknowns: A1 , A2 , A3 , B1 , B2 , B3 , and E (or equiva-
lently ω); see Eqs. (20) and (21). And we had seven equations: Six boundary conditions and
the normalization condition. So we could solve for everything. In the present case, we have
six unknowns: A1 , A2 , A3 , B1 , B2 , and ω (which determines k and κ). And we have four
equations: two boundary conditions at each of x = ±a (there are no boundary conditions at
x = ±∞). Since we have two more unknowns than equations, there will be two unknowns
that we can’t solve for. We can take these to be A1 and ω. In other words, we are free to
give the incoming wave whatever amplitude and frequency we want it to have, which makes
sense.
We can therefore solve for the reflection and transmission coefficients, B1 /A1 and A3 /A1
in terms of ω (and any other parameters in the setup). The calculation is rather messy,
so we won’t do it here (see Problem [to be added]). The situation isn’t symmetric, since
we’re throwing things in from the left, so we can’t make easy use of the even/odd trick from
Problem 10.1.
The important point to note is that the coefficient A3 is not equal to zero. This is
believable, because the string in the middle region will move at least a little, so the right
end of it will wiggle the string in the right region, creating a rightward-traveling A3 ei(kx−ωt)
wave. Intuitively, if the springs are weak, or if the width 2a of the middle region is small,
V = V0 then virtually all of the wave will make it through, so A3 ≈ A1 . In the other extreme, if the
springs are stiff, or if the width 2a of the middle region is large, then A3 will be small, but
it will always be nonzero. Increasing the stiffness of the springs and/or increasing a should
V=0 make A3 smaller, but it should just be a matter of degree, and A3 should asymptotically
-a a
approach zero. There isn’t any reason why A3 should suddenly exactly equal zero.
Figure 11 Let’s now consider the analogous quantum system shown in Fig. 11. V (x) equals zero
except for the −a < x < a region where it equals V0 . If we send a particle (or technically
a stream of particles) in from the left, and if the energy is less than the height V0 of the
bump, then we have exactly the same situation as with the string/spring system. We have
sinusoidal waves in the outer regions, and an evanescent wave in the middle region. The
wavefunctions are again given by Eq. (24). The boundary conditions are also the same, so
the resulting wavefunctions are exactly the same. (As with the finite-well case, the dispersion
relations, and hence the ω’s, are different, but this doesn’t affect the shape of the waves.)
So we again reach the conclusion that A3 is nonzero.
However, while this conclusion is quite intuitive in the string/spring setup, it is quite
bizarre and highly counterintuitive in the quantum-mechanical setup. Classically, if a par-
ticle has E < V0 , then it has zero chance of making it over the potential-energy bump. If
a stream of particles comes in from the left, all of them are reflected, and none of them
make it through. But not so in quantum mechanics. Since A3 is nonzero, there is a nonzero
probability that a particle can make it through. This phenomenon is known as tunneling.
With macroscopic particles, the probability is prohibitively small. If you roll a ball toward
2 We can’t use the same reasoning in the case of the finite well, because there would be a kink in the

string where you grab it (in the middle region). In the present setup, it doesn’t matter if there’s a kink
very far to the left. (If we wiggled the string in the finite-well case very far to the left, then the evanescent
wave generated would die out well before the middle region. So essentially no sinusoidal wave would be
generated.)
10.4. UNCERTAINTY PRINCIPLE 15

a hill, and if it doesn’t reach the top, then you will (essentially) never see it appear on the
other side of the hill and roll down. But on the atomic level there are countless examples
and applications of tunneling, including alpha decay, flash memory, and scanning tunneling
microscopes.

10.4 Uncertainty principle

The uncertainty principle (often prefixed with “Heisenberg,” after the physicist who first
formulated it) makes a statement about the uncertainties in the position and momentum
of a particle. There are actually many different uncertainty principles, applying to many
different pairs (always pairs) of variables, but the most common one involves the position
and momentum. This uncertainty principle says that the product of the uncertainties in x
and p can never be smaller than a certain value. More precisely,

h̄
∆x∆p ≥ (25)
2

This tells us that if we know one of x or p very well, then we must be rather unsure of the
other.
Remark: Although we won’t be concerned here with the precise definition of the uncertainties, ∆x
and ∆p, we’ll give the official definition for completeness. The uncertainty in a variable is defined to
be the standard deviation of the variable, which in turn is defined by (with an “overline” denoting
the average)
q
∆x = (x − x)2 (26)
In words: to find ∆x, you first calculate the average value, x. Then you calculate the average
value of the squares of the distances from x. And then you take the square root of the result.
In general, the probability distributions for x and p are continuous ones, so the averages in Eq.
(26) involve calculating some integrals. But to illustrate the procedure, let’s look at the simpler
case of a discrete distribution. Let’s say we have three equally possible values of x: 1, 2, and
6. Then the average value is x = 3, and the squares of the differences from x are (1 − 3)2 = 4,
(2 − 3)2 =p1, and (6 − 3)2 = 9. The average value of these squares is 14/3, and the square root
of this is 14/3 =⇒ ∆x ≈ 2.16. But as we mentioned, the precise definition of the uncertainties
won’t be important here. For the present purposes, we can just take the uncertainty to be the
rough spread of the bump of the probability distribution of x or p. ♣

The quickest way to rigorously prove the uncertainty principle makes use of what are
called quantum-mechanical operators, which you will learn about when you take an actual ψ(x)
course on quantum mechanics. But if you want to understand physically what’s going on
with the uncertainty principle, the best way is to invoke a few results from Chapter 3
on Fourier analysis. It turns out that the uncertainly principle is simply a fancy way of
interpreting a certain fact from Fourier analysis.
Consider the wavefunction ψ(x) shown in Fig. 12. This represents a particle that is
localized to within roughly the spread of the bump. Our Fourier result in Eq. (3.43) says x
that we can write (we’ll use the letter φ here instead of C)
Figure 12
Z ∞ Z ∞
1
ψ(x) = φ(k)eikx dk, where φ(k) = ψ(x)e−ikx dx. (27)
−∞ 2π −∞

φ(k) is the Fourier transform of ψ(x), and it tells us how much of ψ(x) is made up of the
function eikx with a particular value of k.
16 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

Eq. (27) is a mathematical statement that need not have anything to do with physics.
But now let’s introduce some physics. We know from de Broglie’s proposal in Section 10.1
that the function eikx corresponds to a momentum of p = h̄k. Therefore, since φ(k) tells
us how much of ψ(x) is made up of the function eikx with a particular value of k, we see
that φ(k) (or technically |φ(k)|2 , once it is normalized to 1) gives the probability that the
particle has a momentum of p = h̄k.
Recall now from Section 3.4.1 that if a Gaussian function ψ(x) is narrow, then the (also
Gaussian) function φ(k) is wide, and vice versa. This opposing behavior of the widths is
also true in general for the other pairs of non-Gaussian functions we dealt with (exponen-
tial/Lorentzian, square-wave/sinc). If a function is narrow (wide), then its Fourier transform
is wide (narrow). This can be proved rigorously, but due to all the examples we gave in
Chapter 3, we’ll just accept it as true here.
Therefore, since φ(k) is the Fourier transform of ψ(x), we see that if we’re very sure about
the position (so ψ(x) is narrow), then we’re very unsure about the momentum (because φ(k)
is wide). Basically, if we want to construct a thin spike, then we need to use many different
eikx functions to build it up.
If you do things rigorously, you will obtain the h̄/2 term on the righthand side of Eq.
(25). But even from our above qualitative argument, this term is believable, because we
saw in Section 3.4 that the product of the rough widths of a plot and its Fourier transform
is always of order 1. That is, ∆x∆k ≈ 1. And since p = h̄k, we obtain ∆x∆p ≈ h̄. So
the exact result of ∆x∆p ≥ h̄/2 is plausible. It turns out that if x (and hence also p) is
a Gaussian function, then the lower bound of h̄/2 is achieved. But for any other function,
∆x∆p is strictly larger than h̄/2.
A common physical interpretation of the uncertainty principle is the following. If you
want to measure the position of a particle, the most reasonable way to do this is to look at
it. This involves shining photons and observing the ones that bounce off in your direction.
Since photons carry momentum (which is given by p = E/c = hν/c), they will transfer some
random fraction of this momentum to the particle when they hit it. You will therefore be
uncertain of the particle’s resulting momentum. Furthermore, if you want to narrow down
the position, you will need to use a shorter wavelength of light, because it is impossible
to resolve the position of a particle to more accuracy than the order of the wavelength
of the light you’re using. (Trying to do so would be like trying accurately measure the
width of a penny by using a ruler whose smallest markings are inches.) Therefore, if you
want to decrease ∆x, then you need to decrease the wavelength λ, which means increasing
the frequency ν (since ν = c/λ), which means increasing the photon’s momentum (since
p = E/c = hν/c). This then increases ∆p. So a small ∆x implies a large ∆p, consistent
with the uncertainty principle.
However, although this physical reasoning makes the uncertainty principle a little more
believable, it can be slightly misleading, because you might think that it is possible to
conjure up a clever way of measuring a position that doesn’t use light and doesn’t disturb
the particle. The uncertainty principle states that this isn’t possible. It states that no
matter how clever you are, you won’t be able to beat the ∆x∆p ≥ h̄/2 lower bound.
The uncertainty principle isn’t a measurement-specific result. Rather, it is a mathematical
consequence of the wave nature of matter with a little Fourier analysis thrown in.
10.5. PROBLEMS 17

10.5 Problems
10.1. Even and odd functions *
Show that if the potential energy is an even function of x (that is, V (−x) = V (x)),
then when solving the time-independent Schrodinger equation, it suffices to consider
even and odd functions ψ (that is, ones for which ψ(−x) = ψ(x) or ψ(−x) = −ψ(x)).
By “suffices,” we mean that any solution can be written as a linear combination of
even and/or odd solutions. Hint: Replace x with −x in the Schrodinger equation.
10.2. Finite square well **
Consider the finite square well discussed in Section 10.3.3 and shown in Fig. 4. From
Problem 10.1 we know that we need only consider even and odd functions for ψ. Let’s
consider the even functions here (Exercise [to be added] covers the odd functions).
(a) If ψ(x) is even, write down the most general possible form of the functions ψ2 (x)
and ψ3 (x) in Eq. (20).
(b) Apply the boundary conditions at a, and show that (ka) tan(ka) = κa. We could
cancel a factor of a here, but it’s easier to work with the dimensionless quantities,
ka and κa.
(c) From the definitions of k and κ in Eq. (21), it quickly follows that (ka)2 +(κa)2 =
2mV0 a2 /h̄2 . Show graphically how to obtain the solutions for k and κ (and hence
E) by drawing the appropriate curves in the κa vs. ka plane and finding their
intersections.
18 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

10.6 Solutions
10.1. Even and odd functions
Assume that ψ(x) is a solution to the time-independent Schrodinger equation,

h̄2 ∂ 2 ψ(x)
Eψ(x) = − + V (x)ψ(x). (28)
2m ∂x2
If we replace x with −x (or if that seems like a fishy step, you can define z ≡ −x, and then
later on relabel the letter z with the letter x), then we have

h̄2 ∂ 2 ψ(−x)
Eψ(−x) = − + V (−x)ψ(−x). (29)
2m ∂(−x)2

Using the given fact that V (−x) = V (x), along with (−x)2 = x2 , we obtain

h̄2 ∂ 2 ψ(−x)
Eψ(−x) = − + V (x)ψ(−x). (30)
2m ∂x2
This tells us that if ψ(x) is a solution to the Schrodinger equation with potential V (x) and
total energy E, then ψ(−x) is also a solution with the same V (x) and E. And since the
Schrodinger equation is linear, this means that any linear combination of the functions ψ(x)
and ψ(−x) is also a solution for the same V (x) and E. In particular, the combinations,
³ ´ ³ ´
1 1
ψeven (x) = ψ(x) + ψ(−x) and ψodd (x) = ψ(x) − ψ(−x) (31)
2 2
are solutions. But these¡ two function are even
¢ and odd, as you can verify by replacing x with
−x. And since ψ(x) = ψeven (x) + ψodd (x) /2, any solution with a particular energy can be
written as a linear combination of the even and odd solutions with that energy. It turns out
that in many cases there is only one solution with a given energy, so it is either even or odd
(so one of the functions in Eq. (31) is identically zero).
10.2. Finite square well

(a) Since ψ2 (x) is even, only the cos kx parts of the exponentials in Eq. (20) survive. So
ψ2 (x) = A cos kx. In the righthand region, the boundary condition at infinity tells us
that the eκx term can’t exist, so ψ3 (x) = Be−κx . And then in the lefthand region we
have ψ1 (x) = Beκx , but this ends up giving redundant information.
(b) k and κ are unknowns, but they aren’t independent, because E determines both of
them through Eq. (21). ψ1 and ψ2 therefore contain three unknowns, namely A, B, and
E. And we have three equations, namely the two boundary conditions at x = a (the
boundary conditions at x = −a are redundant, given that we are using an even function
by construction), and the normalization condition. However, since our goal is to find
only k and κ (and from these, E), we won’t need to use the normalization condition.
The boundary conditions at x = a (continuity of the wavefunction and its slope) are

ψ2 (a) = ψ3 (a) =⇒ A cos ka = Be−κa ,

ψ20 (a) = ψ30 (a) =⇒ −kA sin ka = −κBe−κa . (32)

Dividing these gives

k tan ka = κ =⇒ (ka) tan(ka) = κa, (33)

as desired.
(c) If we square and add the relations in Eq. (21), the E’s cancel, and we end up with (after
multiplying through by a2 )

(ka)2 + (κa)2 = 2mV0 a2 /h̄2 . (34)

10.6. SOLUTIONS 19

Eqs. (33) and (34) are two equations in two unknowns, k and κ (or equivalently ka and
κa). They can’t be solved analytically, so if we want to obtain precise values we have
to solve them numerically (after being given the values of a, m, V0 , and h̄). However,
to get a general idea of what the solutions look like, it is far more instructive to solve
the equations graphically than numerically. Our goal is to find the specific pairs of k
and κ values that simultaneously satisfy Eqs. (33) and (34). These pairs can be found
by drawing the curves represented by Eqs. (33) and (34) in the κa vs. ka plane and
finding their intersections. (We’re working with ka and κa because these quantities
are dimensionless.) These two curves are shown in Fig. 13. Eq. (33) is a series of
(ka) tan(ka) curves, and Eq. (34) is a circle. The intersections are shown by the dots.
The (k, κ) pairs associated with these dots give the desired solutions.

2mV 0a2
______
(ka)2 + (κa)2 = 2
h
κa κa = (ka)tan(ka)

14

12

10

0 ka
π 2π 3π 4π

Figure 13

For a given value of a, the (ka) tan(ka) curves are located at fixed positions, but the size
of the circle depends on the value of mV0 a2 /h̄2 . And the size of the circle determines
how many solutions there are. Let’s look at the cases of very small and very large
circles. Given the values of m, a, and h̄, this is equivalent to looking at the cases of
very small V0 and very large V0 .

Small V0 (shallow well): There is always at least one even solution, no matter
how small V0 is, because the circle always intersects at least the first of the (ka) tan(ka)
curves, even if the radius is very small. (The same can’t be said for the odd states, as
you will see in Exercise [to be added].) In the small-V0 limit, the circle is close to the
origin, so ka and κa are small (more precisely, these dimensionless quantities are much
smaller than 1). We can therefore use tan(ka) ≈ ka to write Eq. (33) as k2 a2 ≈ κa. ψ(x)
This says that κa is a second-order small quantity, much smaller that ka. (In other
words, the slope of the (ka) tan(ka) curve equals zero at the origin.) p The (κa)2 term
in Eq. (34) is therefore negligible, so our solution for k and κ is k ≈ 2mV0 /h̄2 and
κ ≈ 0.
x
The plot of the wave looks something like the function shown in Fig. 14. Since κ ≈ 0, -a a
the exponential
p part of the function hardly decays, so it’s almost a straight line. Also,
(small V0)
ka ≈ 2mV0 a2 /h̄2 , which is the radius of the circle in Fig. 13, which we are assuming
is small. Therefore, since 2ka is the phase change inside the well, the wave has hardly
any net curvature in the middle region, so it’s roughly a straight line there too. So Figure 14
we end up with a nearly straight line, with a slight negative curvature in the middle,
20 CHAPTER 10. INTRODUCTION TO QUANTUM MECHANICS

and a slight positive curvature on the sides. Note that since κ → 0, virtually all of the
area under the curve (or technically under the |ψ|2 curve) lies outside the well. So the
particle has essentially zero probability of actually being found inside the well. This is
consistent with the fact that the energy E is very small (because it is squeezed between
0 and the very small V0 ), which means that the wavefunction should be nearly the same
as the wavefunction for a very small E that is barely larger than V0 . But in that case
we have a free particle which can be anywhere from −∞ to ∞.

Large V0 (deep well): If V0 is large, then the circle in Fig. 13 is very large, so
there are many solutions for k and κ. This makes sense, because the well is very deep,
so it should roughly approximate an infinite square well, which we know from Section
10.3.2 has an infinite number of possible states. From Fig. 13, a large V0 (and hence
a large circle) implies that the solutions for ka are roughly equal to π/2, 3π/2, 5π/2,
etc. (except for the solutions near the right end of the circle), because the (ka) tan(ka)
curves asymptotically approach the vertical dotted lines in the figure, which are located
at the odd multiples of π/2. This means that the total phase of the cosine in the well is
approximately k(2a) = π, 3π, 5π, etc. So the number of wavelengths that fit into the well
is approximately 1/2, 3/2, 5/2, etc. This makes sense, because we approximately have
an infinite well, and from Fig. 3 you can see that these are the numbers of wavelengths
that fit into an infinite well for the even functions (the ones with odd values of n).