Kevin Zhou Physics Olympiad Handouts

Problem Solving II: Data and Uncertainty

For a brief introduction to uncertainty analysis, see this handout, or for a more introductory take,
this handout and this comic. For some entertaining general discussion, see chapters I-5 and I-6 of
the Feynman lectures. There is a total of 77 points.

1 Basic Probability
Idea 1
If a quantity X has the probability distribution p(x), that means
Z b
the probability that a ≤ X ≤ b is p(x) dx.
a

In particular, the total probability has to sum to one, so

Z ∞
p(x) dx = 1.
−∞

Using the probability distribution, we can calculate expectation values, i.e. averages. For
example, the expectation value of X, also called the mean, is
Z ∞
⟨X⟩ = xp(x) dx
−∞

while the expectation value of an arbitrary function of X is

Z ∞
⟨f (X)⟩ = f (x)p(x) dx.
−∞

One especially important quantity is the variance of X, defined as

var X = ⟨X 2 ⟩ − ⟨X⟩2 .
√
The standard deviation is defined by σX = var X. It describes how “spread out” the
distribution of X is, and it will play an important role in uncertainty analysis.

[1] Problem 1. Suppose that x is a length. What are the dimensions of p(x), ⟨X⟩, var X, and σ?

Solution. Since p(x) dx is dimensionless, we have

[p(x)] = L−1

where L denotes length. Similarly,

[⟨X⟩] = L, [var X] = L2 , [σX ] = L.

1
 Kevin Zhou Physics Olympiad Handouts

Example 1

Trains arrive at a train station every 10 minutes. If I arrive at a random time, and X is the
number of minutes I have to wait, what is the standard deviation of X?

Solution
We see that X can be anywhere between 0 and 10, with all possibilities equally likely, so
(
1/(10 min) 0 ≤ x ≤ 10,
p(x) =
0 otherwise

where the denominator guarantees the total probability is 1. For the rest of this example,
we’ll suppress the units. We have
Z ∞ Z 10
x
⟨X⟩ = xp(x) dx = dx = 5
−∞ 0 10

which makes sense, as I should have to wait half the maximum time on average, and
Z ∞ Z 10 2
2 2 x 100
⟨X ⟩ = x p(x) dx = dx = .
−∞ 0 10 3

Then the standard deviation is

p 5
σX = ⟨X 2 ⟩ − ⟨X⟩2 = √ min.
3

[3] Problem 2. Consider an exponentially distributed quantity,

(
ae−ax x ≥ 0,
p(x) =
0 otherwise.

Verify that the total probability is 1, and compute the mean and standard deviation. To perform
the integrals, you will have to integrate by parts.
Solution. First, to check normalization,
Z ∞ Z ∞ Z ∞
−ax
p(x)dx = ae dx = e−u du = 1 − 0 = 1.
−∞ 0 0

Now, the mean can be evaluated using integration by parts,

Z ∞ Z ∞
∞ 1 ∞ 1
⟨x⟩ = xae−ax dx = −xe−ax + e−ax dx = 0 − e−ax = .
0 0 0 a 0 a
To calculate the standard deviation, we must evaluate
Z ∞ Z ∞
2 −ax 2 2
2
⟨x ⟩ = x ae dx = 0 + (2x)e−ax dx = ⟨x⟩ = 2 .
0 0 a a
We thus conclude r
p 2 1 1
σX = ⟨X 2 ⟩ − ⟨X⟩2 = 2
− 2 = .
a a a

2
 Kevin Zhou Physics Olympiad Handouts

[2] Problem 3. The purpose of subtracting ⟨X⟩2 in the variance is to make sure it doesn’t change
when a constant is added to x, since shifting something left or right on the number line shouldn’t
change its spread. Verify that for any constant c, var X = var(X + c).

Solution. We have
var(X + c) = ⟨(X + c)2 ⟩ − ⟨X + c⟩2 .
By the definition of the expectation value, we have

⟨A + B⟩ = ⟨A⟩ + ⟨B⟩, ⟨cA⟩ = c⟨A⟩

for any quantities A and B and any constant c. Thus,

var(X + c) = ⟨X 2 ⟩ + ⟨2Xc⟩ + ⟨c2 ⟩ − ⟨X⟩2 − 2⟨X⟩⟨c⟩ − ⟨c⟩2 = var X

as desired.

[3] Problem 4. We say X is normally distributed if

2
p(x) ∝ e−a(x−b) .

For simplicity, let’s shift X so that it’s centered about x = 0, so

2
p(x) ∝ e−ax .

You may use the result given in P1,

Z ∞ √
2
e−x dx = π.
−∞

Find the constant of proportionality in p(x), the mean, and the standard deviation.
2
Solution. Let p(x) = ke−ax . We fix the constant k by demanding normalization,
Z ∞ Z ∞
−ax2 k 2
ke dx = √ e−u du = 1.
−∞ −∞ a

Using the provided integral, we conclude

r
a
k= .
π
The mean is clearly zero, since the distribution is symmetric about that point. Thus, we have
r Z ∞ Z ∞
a 2 1 2
2
var X = ⟨X ⟩ = x2 e−ax dx = √ u2 e−u du.
π −∞ a π −∞

This remaining integral can be evaluated using integration by parts,

Z ∞ √
1 −u2 ∞ 1 ∞ −u2
Z
2 −u2 π
u e du = − ue + e du = 0 +
−∞ 2 −∞ 2 −∞ 2

from which we conclude √

1 π 1
var X = √ , σ=√ .
a π 2 2a

3
 Kevin Zhou Physics Olympiad Handouts

[2] Problem 5. If two random variables X1 and X2 are independent, then

⟨X1 X2 ⟩ = ⟨X1 ⟩⟨X2 ⟩.

Use this result to show that

var(X1 + X2 ) = var(X1 ) + var(X2 )

which implies that the standard deviation “adds in quadrature”,

q
σX1 +X2 = σX 2 + σ2 .
1 X2

This is an important result we’ll use many times below.

Solution. By definition, we have

var(X1 + X2 ) = ⟨(X1 + X2 )2 ⟩ − ⟨X1 + X2 ⟩2

Using the properties listed in problem 3,

var(X1 + X2 ) = ⟨X12 ⟩ + 2⟨X1 X2 ⟩ + ⟨X2 ⟩2 − ⟨X1 ⟩2 − 2⟨X1 ⟩⟨X2 ⟩ − ⟨X2 ⟩2

= var(X1 ) + var(X2 ) + 2(⟨X1 X2 ⟩ − ⟨X1 ⟩⟨X2 ⟩)

When X1 and X2 are independent, the last term vanishes, giving

var(X1 + X2 ) = var(X1 ) + var(X2 ).

2 Uncertainty Propagation
In this section, we’ll establish the fundamental results needed to compute uncertainties.

Idea 2
When a physical quantity is measured in an experiment and reported as x±∆x, it is uncertain
what the true value of the quantity is. If the quantity has a probability distribution p(x),
then the reported uncertainty ∆x is essentially the standard deviation of p(x).

Remark
In practice, you’ll have to use intuition and experience to assign uncertainties for real mea-
surements. For example, if you’re using a clock that times only to the nearest second, you
might take ∆t = 0.5 s. If you’re using a good ruler, which has millimeter markings, you
might take ∆x = 0.5 mm. Of course, the ultimate test is the results: if you assigned the
uncertainties right, your final uncertainty should encompass the true result most (but not
all) of the time.

[2] Problem 6. Suppose x has uncertainty ∆x and y has uncertainty ∆y, where x and y are indepen-
dent. Explain why the uncertainty of x + y is
p
∆(x + y) = (∆x)2 + (∆y)2 .

This is called “addition in quadrature”. What is the uncertainty of x − y? How about x + x?

4
 Kevin Zhou Physics Olympiad Handouts

Solution. For independent variables, var(X

p 1 + X2 ) = var(X1 ) + var(X2 ). Since our uncertainties
represent the standard deviation, σX = var(X), we have
p
∆(x + y) = (∆x)2 + (∆y)2 .

For x − y = x + (−y), and ∆(−y) = ∆y, we get that ∆(x − y) = ∆(x + y). Finally, by linearity we
clearly have ∆(x + x) = 2∆x. (The formula above doesn’t apply, because x isn’t independent of x.)

Remark
Note how this differs from “high school” uncertainty analysis. In school, you might be told
to show uncertainty using significant figures, and when adding two things, to keep only the
figures that are significant in both of them. That corresponds to

∆(x + y) = max(∆x, ∆y)

which is an underestimate. Or, you might be told that the uncertainty needs to encapsulate
all the possible values, which implies that

∆(x + y) = ∆x + ∆y

which is an overestimate, since the uncertainties could cancel.

Example 2: F = ma 2016 25

Three students make measurements of the length of a 1.50 m rod. Each reports an uncertainty
estimate representing an independent random error applicable to the measurement.

• Alice performs a single measurement using a 2.0 m tape measure, to within 2 mm.

• Bob performs two measurements using a wooden meter stick, each to within 2 mm, which
he adds together.

• Christina performs two measurements using a machinist’s meter rule, each to within
1 mm, which she adds together.

Rank the measurements in order of their uncertainty.

Solution
√
The uncertainty in Alice’s measurement is 2 mm. The √ uncertainty in Bob’s is 2 2 mm by
quadrature, while the uncertainty in Christina’s is 2 mm by quadrature. So the lowest
uncertainty is Christina’s, followed by Alice’s, followed by Bob’s.

[1] Problem 7. Given N independent measurements of the same quantity with the same uncertainty, √
xi ± ∆x, find the uncertainty of their sum. Hence show the uncertainty of their average is ∆x/ N .
This result is extremely important, since repeating trials is one of the main ways to reduce
uncertainty. But it’s important to remember that the results derived above hold only for independent
measurements. For example, taking a single measurement, then averaging that single number with
itself 100 times certainly wouldn’t reduce the uncertainty at all!

5
 Kevin Zhou Physics Olympiad Handouts

Solution. The uncertainty of their sum ∆X can be found by adding in quadrature,

v
uN
uX √
∆X = t (∆xi )2 = N ∆x.
i=1

Therefore, the uncertainty of the average is

∆X ∆x
=√ .
N N

Idea 3
If x has uncertainty ∆x, and f (x) can be approximated by its tangent line, f (x′ ) ≈ f (x) +
(x′ − x)f ′ (x) within the region x ± ∆x, then f (x) has approximate uncertainty f ′ (x) ∆x.

√
[2] Problem 8. If x has uncertainty ∆x, find the uncertainties of x2 , x, 1/x, 1/x4 , log x, and ex .

Solution. Differentiate the functions and multiply by ∆x to find the uncertainties. The sign isn’t
important, since uncertainties are always positive. The results are:
√ ∆x ∆x
∆(x2 ) = 2x∆x ∆( x) = √ ∆(1/x) =
2 x x2
4∆x ∆x
∆(1/x4 ) = ∆(log(x)) = ∆(ex ) = ex ∆x
x5 x
[2] Problem 9. The tangent line approximation doesn’t always make sense. For example, suppose that
x is measured to be zero, up to uncertainty ∆x. Show that the above results for the uncertainties
√
of x2 and x give nonsensical results. What would be a more reasonable uncertainty to report?
√
Solution. The above uncertainties give 0, ∞ for the uncertainties of x2 and x respectively. Since
≪ x, 2 2
the uncertainties were found
2 2
√ with ∆x √ √ now with x ≪ ∆x, we can get (x + ∆x) − x =2
2x∆x √+ ∆x ≈ ∆x and x + ∆x − x ≈ ∆x. Thus, more reasonable uncertainties are (∆x)
and ∆x. There are numerical factors of order 1 because the shapes of the probability distributions
will be distorted, but we won’t worry about those, because we’re just looking to get a reasonable
result. (Of course, a professional would keep track of all these details.)

[2] Problem 10. Consider two quantities with independent uncertainties, x ± ∆x and y ± ∆y.

(a) Show that the uncertainty of xy is

s 2  2
∆x ∆y
∆(xy) = xy + .
x y

To do this, start by writing xy as exp(log x + log y).

(b) If we set x = y, then we find

s 
∆x 2 √

2 2
∆(x ) = x 2 = 2x∆x.
x

On the other hand, in a previous problem we found ∆(x2 ) = 2x∆x. Which result is correct?

6
 Kevin Zhou Physics Olympiad Handouts

(c) Find the uncertainty of x/y.

Solution. (a) We can write

xy = exp(log x + log y)
which implies
s 2  2
∆x ∆y
∆(xy) = exp(log x + log y)∆(log x + log y) = xy + .
x y

(b) The result that ∆(x2 ) = 2x∆x is correct, since the formula for ∆(xy) assumes x, y are
independent, which fails when we set y = x.

(c) We have
x
= exp(log x − log y)
y
and by a very similar calculation to part (a), we conclude
s
∆x 2 ∆y 2
  
x
∆(x/y) = + .
y x y

[2] Problem 11. A student launches a projectile with speed v = 5 ± 0.1 m/s in gravitational accelera-
tion g = 9.81 ± 0.01 m/s2 . The resulting range is d = 1.5 ± 0.02 m. Given that the launch angle was
less than 45◦ , find the launch angle, with uncertainty, assuming all uncertainties are independent.

Solution. From the projectile range equation d = v 2 sin(2θ)/g, we get

 
1 dg
θ = arcsin = 18.03◦ .
2 v2

To find the uncertainty, we write sin(2θ) = gd/v 2 . The left-hand side is

2 cos(2θ)∆θ

by the tangent line approximation. By the result of problem 10, the right-hand side is
s
∆d 2 ∆g 2 2∆v 2
    
dg
+ + = 0.0248
v2 d g v

Combining the results, we have

∆θ = 0.015 rad = 0.9◦
which means the final result should be written as

θ = 18.0◦ ± 0.9◦

where we removed a superfluous significant figure.

[2] Problem 12. Two physical quantities are related by y = xex .

(a) If x is measured to be 1.0 ± 0.1, find the resulting value of y, with uncertainty.

7
 Kevin Zhou Physics Olympiad Handouts

(b) If y is measured to be 2.0 ± 0.1, find the resulting value of x, with uncertainty.
Solution. (a) To find the central value of y, we plug in to get y = e = 2.7183. To find the error,
we use the tangent line approximation,
dy
= ex (x + 1)
dx
which gives us
∆y ≈ ex (x + 1)∆x = 0.54.
Thus, rounding to a reasonable number of significant figures, we have
y = 2.7 ± 0.5.
Note that it would be incorrect to apply the “addition in quadrature” rule for products,
s
∆x 2 ∆(ex ) 2
  
x
∆y = xe +
x ex
because x and ex aren’t independent.
(b) To find the central value of x, we solve the equation 2 = xe−x numerically. This can be done
using the method of iteration introduced in P1. That is, we have x = 2e−x , so by repeatedly
plugging 2e−Ans into the calculator, we get x = 0.8526.
Under the tangent line approximation,
∆y
∆x ≈ = 0.023.
ex (x+ 1)
Rounding to a reasonable number of significant figures, we conclude
x = 0.85 ± 0.02.
Idea 4
For practical computations, it is often useful to use relative uncertainties. The relative
uncertainty of x is ∆x/x, and can be expressed as a percentage.

[1] Problem 13. Some basic relative uncertainty results.

(a) Show that the relative uncertainty of the product or quotient of two quantities with independent
uncertainties is the square root of the sum of the squares of their relative uncertainties.
(b) Show that averaging
√ N independent trials as in problem 7 reduces the relative uncertainty by
a factor of N .
Solution. (a) Above we found that
s 2  2
∆x ∆y
∆(xy) = xy +
x y
Dividing both sides by xy gives
s 2  2
∆(xy) ∆x ∆y
= +
xy x y
which is the desired result.

8
Kevin Zhou Physics Olympiad Handouts

√ √
(b) We have ∆xN = ∆(Σx)/N = ∆x N /N = ∆x/ N . Then
∆xN ∆x 1
= √
x x N
as expected.

Remark
There are many situations where the rules above can’t be used. For example, consider the
uncertainty of x + y 2 /x, where x and y have independent uncertainties. You can calculate the
uncertainty of either term with the standard rules, but you can’t calculate the uncertainty
of their sum, because the terms are not independent (both contain x).

In these cases, you can use the multivariable equivalent of the tangent line approximation,
∂f ∂f
f (x′ , y ′ ) ≈ f (x, y) + (x′ − x) + (y ′ − y) .
∂x ∂y
Adding the two contributions to the uncertainty in quadrature gives
s 2  2
∂f ∂f
∆f = ∆x + ∆y .
∂x ∂y

This is the general rule that includes the rules you derived above as special cases. However, it
shouldn’t be necessary in Olympiad problems. If you run into such situations in an experiment,
often one of the uncertainties is much smaller, and can be neglected entirely.

Remark
As you saw in problem 9, the tangent line approximation can sometimes fail. The proper way
to handle situations like these would be to find the full probability distribution of the desired
quantity, rather than just describing it crudely with its standard deviation. However, this
can’t be done analytically except in the simplest of cases. So when professional physicists run
into situations like these, which are quite common, they often just numerically compute a few
million or billion values, starting with randomly drawn inputs each time, and use that to infer
the probability distribution. This technique is called Monte Carlo. It’s very powerful, but
certainly not needed for Olympiads! On Olympiads, you should just fall back to something
reasonable, such as taking the minimum and maximum possible values.

3 Using Uncertainties
Example 3: F = ma 2022 B21

Amora and Bronko are given a long, thin rectangle of sheet metal. (It has been machined very
precisely, so they can assume it is perfectly rectangular.) Using calipers, Amora measures
the width of the rectangle as 1 cm with 1% uncertainty. Using a tape measure, Bronko
independently measures its length as 100 cm with 0.1% uncertainty. What are the relative
uncertainties they should report for the area and the perimeter of the rectangle?

9
 Kevin Zhou Physics Olympiad Handouts

Solution
To compute the area, we multiply the two measurements, which means we add the relative
uncertainties in quadrature,
∆A p
= (1%)2 + (0.1%)2 ≈ 1%.
A
Note that in this case, the relative uncertainty of Bronko’s measurement is negligible; the rel-
ative uncertainty of the area is approximately the relative uncertainty of Alice’s measurement.

Computing the perimeter involves adding the measurements, which means the absolute
uncertainties are added in quadrature instead. These are 0.01 cm and 0.1 cm for Alice and
Bronko’s measurements, respectively, so the absolute uncertainty of Alice’s measurement
is negligible. Thus, the relative uncertainty of the perimeter is approximately the relative
uncertainty of Bronko’s measurement, 0.1%.

In simple Olympiad experiments, often only one uncertainty will really matter. This can
dramatically simplify calculations, but it might take a little thought to tell which one.

[3] Problem 14. 01^‚ Solve F = ma 2018 problems A12, A25, B19, and B25, and F = ma 2019
problems A16, B18, and B25. Make sure to strictly adhere to the total time. Since these are
F = ma problems, you don’t have to produce a writeup. If you find these questions difficult to
finish in the allotted time, go back and review the earlier material!
[2] Problem 15. Suppose the goal of an experiment is to measure the ratio T1 /T2 of the durations of
two physical processes, where T1 is about 15 seconds, and T2 is about 3 seconds. Also suppose your
stopwatch is only accurate to the nearest second. You have two minutes to perform measurements.
Assume each measurement is independent.
(a) Using your instinct, figure out whether it’s better to spend more total time measuring T1 ,
more total time measuring T2 , or an equal amount of time on both.

(b) To confirm this, qualitatively sketch the relative uncertainty of T1 /T2 as a function of the
fraction of time x spent measuring T1 , using explicit numeric examples if necessary.
Calculations of this sort are common when doing Olympiad experimental physics. You should be
able to do them instinctively, getting the ballpark right answer without explicit calculation.
Solution. (a) Since T2 is smaller, a single measurement of T2 has a much higher relative un-
certainty. Furthermore, T2 takes less time to measure. This means we definitely want more
distinct measurements of T2 than of T1 . As for how we split up the time, this is a bit harder to
judge, but intuitively because uncertainty adds in quadrature, taking a single measurement of
each makes T2 ’s uncertainty not 5 times as bad, but 25 times as bad. So T2 really completely
dominates the uncertainty here, and we should spent most of our time getting its uncertainty
down.
√
(b) We have ∆T ≈ 1 s and ∆Ti = ∆T / Ni , giving
s 2  2
T1 ∆T ∆T
∆(T1 /T2 ) = √ + √
T2 T1 N1 T2 N2

10
 Kevin Zhou Physics Olympiad Handouts

The total time Tt is constant, N1 T1 + N2 T2 = Tt where N1 T1 /Tt = x. We want to minimize

1 1
f (x) = + .
T1 x (1 − x)T2
The derivative is
1 1
f ′ (x) = − +
T1 x2 T2 (1 − x)2
and setting this to zero gives
x2 (1 − T1 /T2 ) − 2x + 1 = 0.
The smaller root is the desired one since x < 1, giving
p
1 − T1 /T2 1
x= = p
1 − T1 /T2 1 + T1 /T2
so we should spend 30% of our time measuring T1 . The graph of the uncertainty as a function
of x is concave up, with vertical asymptotes at x = 0 and x = 1.
[3] Problem 16. In the preliminary problem set, you measured g using a pendulum. If you didn’t do
uncertainty analysis for it, as we covered above, then you should go back and estimate uncertainties
more precisely. In this problem you’ll do a different experiment: you will estimate g by finding the
time needed for an object to roll down a ramp, with everything again made of household materials.
(a) Before starting, think about what the dominant sources of uncertainty will be, and how you
can design the experiment to minimize them. In particular, do you think the result will be
more or less precise than your pendulum experiment?
(b) Perform the experiment, taking at least ten independent measurements, and report the data
and results with uncertainty.
Solution. Our formula for g is
2ℓ(1 + β)
g=
t2 sin(θ)
where β = I/M R2 of the rolling object, and ℓ, t are the distance and time for the path. Let’s
assume you found a nice object, like a hollow can or a fully filled one, so that β is known relatively
precisely. Then the uncertainty is
s
2∆t 2
 2 
cos(θ)∆θ 2
 
∆ℓ
∆g = g + + .
t ℓ sin(θ)
Given the above, you definitely want a ramp as long as possible, and there’s a tradeoff with the
angle: if the angle is very large, t will be small so that the relative error on t will be large, while if
the angle is very small, the relative error on θ will be large. So in practice you want to choose a
moderately small, but not too small value of θ.
Some reasonable ballpark numbers are θ = (10 ± 1)◦ , and t = (3 ± 0.3) s, so you probably can’t
easily get an uncertainty smaller than a few percent. The overall result will be less precise than
the pendulum experiment, because for the pendulum there is no ∆θ term, and you can measure
N ≫ 1 periods in a single trial so that the relative error on t falls as 1/N . With the ramp, you√can
partially compensate by doing N separate trials, so that √ the relative error at best falls as 1/ N ,
which isn’t as good. It might not even be as good as 1/ N , because your uncertainties may not be
independent: you might systematically overestimate or underestimate the time or angle.

11
 Kevin Zhou Physics Olympiad Handouts

[3] Problem 17. [A] Consider N independent measurements of the same quantity, with results xi ±∆xi .
They can be combined into a single result by taking a weighted average. What is the optimal weighted
average, which minimizes the uncertainty?

Solution. Let the weights be wi , so we report the value

X
x= wi x i .
i

The uncertainty obeys X

(∆x)2 = wi2 (∆xi )2 .
i

A tempting but incorrect way to minimize this quantity to set the derivative with respect to wi
equal to zero. This doesn’t work because the solution is just w1 = . . . = wN = 0, which isn’t a
weighted average at all. To actually have a weighted average, we need the weights to sum to one,
X
wi = 1.
i

This is an optimization problem with a constraint, which can be solved with Lagrange multipliers.
However, for this particular problem, the constraint is simple enough to handle manually. Because
of the constraint, if one increases some weight, then one must decrease others. At the minimum,
the effect of increasing any weight infinitesimally and decreasing another the same amount must be
zero, as if it weren’t, we could just adjust those two weights to get a lower uncertainty. Setting the
change in the uncertainty due to adjusting wi and wj in this way to zero gives

0 = d(wi2 )(∆xi )2 + d(wj2 )(∆xj )2 = (2 dw)(−wi (∆xi )2 + wj (∆xj )2 ).

This tells us that wi ∝ 1/∆x2i , which means

1/(∆xi )2
wi = P 2
.
j 1/(∆xj )

Note that all measurements are included in the optimal average, no matter how bad they may be.

All of the examples above involve combining continuous quantities, so we’ll close this section with
some applications to “counting” experiments, which work slightly differently.

Remark
In this problem set, we have given rules for calculating the mean and standard deviation
of derived quantities. But in general, probability distributions can have all kinds of weird
features, which aren’t captured by those two numbers. The reason we focus on them anyway
is because of the central limit theorem, which roughly states that if we have many independent
random variables, the distribution of the sum will approach a normal distribution. As you
saw in problem 4, normal distributions are characterized entirely by their mean and standard
deviation, so we don’t lose any information by reporting only those two quantities.

12
 Kevin Zhou Physics Olympiad Handouts

Example 4

A fair coin is tossed 1000 times, and the number of heads is counted. If this process is
repeated many times, what is the standard deviation of the number of heads?

Solution
Consider one trial of 1000 tosses. The number of heads is X = X1 + X2 + . . . + X1000 , where
(
1 heads on toss i
Xi = .
0 tails on toss i

Of course, the mean of each of these variables is ⟨Xi ⟩ = 0.5, so that the mean of X is 500. In
addition, the Xi are independent of each other, so the variances add. The variance of each
one of them is
1 1 1
var Xi = ⟨Xi2 ⟩ − ⟨Xi ⟩2 = − = .
2 4 4
Thus, the standard deviation of the number of heads is
√ p
var X = 1000/4 ≈ 16.

So getting 520 heads would not be surprising, but if you got 550, you might be justified in
suspecting the coin isn’t fair. (Also, the number of heads is very close to normally distributed,
by the central limit theorem mentioned above.) To check whether you understand p this, you
can redo it with a general probability p of getting heads, where you should get 1000 p(1 − p).

[3] Problem 18. At any moment, a Geiger counter can click, indicating that it has detected a particle
of radiation. Suppose that there is an independent probability α dt of clicking at each infinitesimal
time interval dt. Let the number of clicks observed in a total time T be X.

(a) Find the expected value and standard deviation of X, and thereby compute its relative
uncertainty. (Hint: split the total time into many tiny time intervals, and let Xi be the
P
number of clicks in interval i, so X = i Xi .)

(b) Using a Geiger counter on a sample, you hear 197 clicks in 5 minutes of operation. Estimate
the activity α of the sample (i.e. the expected clicks per second), with uncertainty. If you
measure for longer, how does the uncertainty reduce over time?

(c) Now suppose that for a different sample, N = 0 after 5 minutes. Estimate the activity α of
the sample (i.e. the expected clicks per second), with a reasonable uncertainty. If you measure
for longer, and continue to hear no clicks, how does the uncertainty reduce over time?

Solution. (a) There are N = T /dt time intervals. Using the hint and applying linearity of
expectation, X
⟨X⟩ = ⟨Xi ⟩ = N (α dt) = αT.
i

Since the Xi are independent, their variances add. The variance of Xi is

⟨Xi2 ⟩ − ⟨Xi ⟩2 = α dt − (α dt)2 = α dt.

13
 Kevin Zhou Physics Olympiad Handouts

Thus, by adding the variances, we have

var X = αT
√ √
so the standard deviation is ∆X = αT . The relative uncertainty is ∆X/⟨X⟩ = 1/ αT .
(b) Applying the formulas above, we estimate
197
α= = 0.66 s−1
T
with an uncertainty of r
α α
∆α = √ = = 0.05 s−1 .
αT T
√
The uncertainty falls as 1/ T . Note that this is very similar to previous results we’ve found,
√
where the uncertainty falls as 1/ n where n is the number of trials. In some sense, each
instant of time we wait is another trial here.
(c) Of course, we estimate α = 0, but then the formulas above imply ∆α = 0 and hence that we
are absolutely certain α = 0, which is absurd. (If you don’t think that’s absurd, note that
the same result would have occurred if we had heard zero clicks in an arbitrarily short time
interval, such as a nanosecond.)
This is a case where the basic rules of uncertainty propagation break down, and we need
to think. The point of giving an uncertainty is to indicate the range of parameter values
compatible with the data we observed. Now, the probability of having no clicks in time T
is e−αT . If αT ≫ 1, then it would be very unlikely to have no clicks, so we can rule out
α ≫ 1/T . But if αT ≲ 1, this isn’t unlikely at all. Thus, your uncertainty window should be
α ∈ [0, c/T ] where c is an order-one number, whose value depends on √ the specific statistical
procedure you use. (Note that the upper bound falls as 1/T , not 1/ T .)
[4] Problem 19. [A] This problem extends problem 18 to derive some canonical results.
(a) Let λ = αT . Find the probability p(X = k) of hearing exactly k clicks in terms of λ and k.
(b) To check your result, show that the sum of the p(X = k) is equal to one.
(c) ⋆ In the limit λ ≫ 1, show that the probabilities p(X = k) approach that of a normal
distribution with the mean and standard deviation calculated in problem 18, thereby providing
an example of the central limit theorem at work. This is a rather involved calculation, which
will use many of the techniques from P1. It will also require Stirling’s approximation,
√  n n
n! ≈ 2πn
e
for n ≫ 1, which we mentioned in P1. (Hint: because the relative uncertainty falls as λ
increases, start by writing k = λ(1 + δ) for |δ| ≪ 1, and expand in powers of δ. Be careful not
to drop too many terms, as δ is small, but λδ isn’t.)
P
Solution. (a) Following the notation of problem 18, we have X = i Xi , and we get k clicks if
precisely k of the Xi are equal to 1. Thus,
Nk λk −λ
 
N
p(X = k) = (α dt)k (1 − α dt)N −k ≈ (α dt)k (1 − α dt)N = e .
k k! k!
This is known as the Poisson distribution.

14
 Kevin Zhou Physics Olympiad Handouts

(b) This follows from the Taylor series of the exponential,

∞ ∞
X X λk
p(X = k) = e−λ = 1.
k!
k=0 k=0

(c) Using Stirling’s approximation, we have

 k
1 λe
p(X = k) = √ e−λ
2πk k
 λ(1+δ)
1 e
=p e−λ
2πλ(1 + δ) 1+δ
1
≈√ eδλ (1 + δ)−λ(1+δ)
2πλ
where we used the fact that δ ≪ 1.
Now we need to use a technique from P1. Letting the final term be equal to 1/y, we have

δ2 δ2λ
 
3
log y = λ(1 + δ) log(1 + δ) = λ(1 + δ) δ − + O(δ ) = δλ + + O(δ 3 λ).
2 2

In P1, we only expanded up to the first term, but here we need to keep the order δ 2 term.
The reason is we want an approximation that works for √the whole peak of the probability
distribution,
√ and we know it has relative uncertainty 1/ λ, which means we need to take
δ ∼ 1/ λ. That implies that δ 2 λ is of order one and cannot be dropped, but δ 3 λ is small and
can be dropped. Anyway, plugging this in, we find
1 2 1 2
p(X = k) ≈ √ e−δ λ/2 = √ e−(k−λ) /2λ
2πλ 2πλ

01mƒ
which is precisely a normal distribution with the appropriate mean and standard deviation.

[3] Problem 20. IPhO 2023, problem 1, parts A, B, and D.3. A short derivation of the key features
of Brownian motion. It requires only the ideas of this problem set, and some basic mechanics.

4 Data Analysis
Idea 5
All data analysis for the USAPhO and IPhO can be done using extremely basic methods.
Sometimes, it suffices to just calculate a value based on a single data point, or by cleverly
using a pair of data points. When this isn’t enough, you’ll have to do graphical data analysis,
which will usually correspond to drawing a line and measuring its slope and intercept. This
is quite limited compared to modern statistical tools, but also can be surprisingly powerful.

15
 Kevin Zhou Physics Olympiad Handouts

Example 5

The activity of a radioactive substance obeys A(t) = A0 e−t/τ . Using measurements of t and
A(t), plot a line to find A0 and τ .

Solution
To handle exponential relationships, take the logarithm of both sides for

log A(t) = log A0 − t/τ.

Then a plot of log A(t) vs. t has slope −1/τ and y-intercept log A0 .

[1] Problem 21. For a power law y = αxn where y and x are measured, what line can be plotted to
find α and n?

Solution. We have
log(y) = log(αxn ) = log(α) + n log(x).
Thus, if we plot log y against log x, the slope will be n and the y-intercept will be log(α).

[2] Problem 22. The rate R of electron emission from a solid in an electric field E is

R = βe−E/E0

for some constants β and E0 . The particular form is because the effect is due to quantum tunneling,
and you will derive it in X2.

(a) If E and R are measured, what line can be plotted to find β and E0 ?

(b) Your answer for part (a) should have formally incorrect dimensions, by the standards of P1.
This often happens when one takes logarithms. What’s going on? If the dimensions are wrong,
how can the result be right?

(c) Suppose both β and E0 have 1% uncertainty. For small E, which is more important for the
uncertainty of R? What about for large E? Around where is the crossover point?

Solution. (a) Take the natural log of the equation to get

E
log R = − + log β.
E0

Plotting E on the x-axis and log R on the y-axis will give a line with slope − E10 and y-intercept
log β.

(b) This gets into the details of what it even means to plot data. As a simpler case, consider the
relationship y = kx where y and x both have units of energy. We can plot y versus x to find
the slope k, but in reality, you can’t actually plot a dimensionful quantity: what would it
even mean to move your pencil a distance of “3.7 J” on a page? Instead, we write y and x as
dimensionless multiples of a standard unit of energy. That is, we are actually plotting
y x
=k
E0 E0

16
 Kevin Zhou Physics Olympiad Handouts

where E0 is some unit of energy, which is typically 1 J. But we don’t bother to write E0
explicitly because this step is kind of obvious.
Exactly the same thing is going on in this problem, but it looks strange because logarithms
have the property log(xy) = log x + log y. Both R and β have units of rate, so define a unit
of rate R0 and subtract log R0 from both sides to get an equation with correct dimensions,
R E β
log =− + log .
R0 E0 R0
This reflects what we actually do when constructing a log plot, though it is usually left implicit.

(c) The uncertainty in β alone always gives a 1% uncertainty in R. But the uncertainty in R
due to the uncertainty in E0 depends on the value of E. For E ≪ E0 , we can expand the
exponential as (1 − E/E0 ), and in this case the uncertainty in E0 does almost nothing at all,
so the uncertainty in β dominates. For E ≫ E0 , the reverse is true. By dimensional analysis,
the crossover must be around E ∼ E0 .

Example 6

Suppose that y and x are related nonlinearly, as

y = bx + ax2 .

For example, this could model the force due to a non-Hookean spring. Using measurements
of x and y, plot a line to find a and b.

Solution
If we divide by x, we find
y
= ax + b.
x
Therefore, we can plot y/x versus x, which gives a line with slope a and intercept b. More
generally, we can plot a line whenever we can rearrange a given relation into the form

(known) = (unknown)(known) + (unknown)

where all four terms can be arbitrarily complicated. In this way, it is possible to turn a lot
of very nonlinear relations into lines.

[3] Problem 23. Some more examples of finding lines to plot.

(a) Suppose that you are given points (x, y) that lie on a circle centered at (a, 0) with radius r.
What line can be plotted to find a and r?

(b) Consider an Atwood’s machine with masses m and M > m. The acceleration of the machine
is measured as a function of M . However, since the pulley has mass, it slows the acceleration
of the Atwood’s machine, so that
M −m
a= g.
M + m + δm

17
 Kevin Zhou Physics Olympiad Handouts

Find a line that can be plotted to find g and δm, assuming m, M , and a are known. This is
an example of how plotting a line can separate out a systematic error, i.e. the value of δm,
which would be impossible if only one value of M were used.
(c) Suppose an object is undergoing simple harmonic motion with amplitude A and angular
frequency ω. Given measurements of the position x and velocity v, what line can be plotted
to find A and ω?
Solution. (a) The equation of the circle is
(x − a)2 + y 2 = r2 , y 2 + x2 = 2ax + r2 − a2
Plotting y 2 + x2 vs. x will give a slope of 2a and a y-intercept of r2 − a2 . Combining the two
pieces of information yields a and r.
(b) The equation can be slightly rearranged to give
M −m M + m δm
= + .
a g g
Therefore, a plot of (M − m)/a vs. M + m has slope 1/g and y-intercept δm/g.
(c) By conservation of energy, A2 = x2 + v 2 /ω 2 , so
x2 = A2 − v 2 /ω 2 .

01W
Thus, a plot of x2 vs. v 2 has y-intercept A2 and slope −1/ω 2 .

01W
[3] Problem 24. USAPhO 2012, problem A2. (This one requires basic thermodynamics.)

01c‚
[3] Problem 25. USAPhO 2011, problem A2.

[3] Problem 26. INPhO 2018, problem 7. (This one requires basic fluid dynamics.)
Solution. See the official solutions here.
[3] Problem 27 (USAPhO 2024). An experimentalist drives a series RLC circuit with an sinusoidal
voltage V (t) = V0 cos ωt. In E6, you will learn how to show that the voltage across the capacitor,
in the steady state, oscillates with amplitude
V0
Vc = p 2
(1 − ω /ω0 )2 + (ω/ω0 Q)2
2

where ω0 is the resonant angular frequency and Q is the circuit’s quality factor. The experimentalist
takes the following data near the resonance, for a fixed value of V0 :
ω (rad/s) 133.0 133.5 134.0 134.5 135.0 135.5 136.0 136.5 137.0
Vc (Volts) 3.64 4.76 6.52 8.53 8.18 6.06 4.44 3.42 2.75
Find the values of ω0 and Q as accurately as possible. Uncertainty analysis is not required. (Hint:
this is the trickiest data analysis problem in the history of the USAPhO. It can be solved by
drawing lines, but such a method is relatively inefficient. It is better to carefully approximate the
given formula, and to consider just a few data points at a time.)

5 Estimation
Estimation is a useful skill for checking the answers to real-world problems.

18
Kevin Zhou Physics Olympiad Handouts

Example 7

Estimate the circumference of the Earth.

Solution
If you know that the United States is 3,000 miles wide, and there is a time zone difference of
three hours between California and New York, then a reasonable estimate is 24,000 miles.
Or, if you know the factoid that light can go about seven times around the Earth in a second,
then a reasonable estimate is (3/7) × 108 m ≈ 4 × 107 m.

Let’s check these results are compatible. There are about 5 miles in 8 kilometers, a fact
you can get by remembering how your car’s speedometer looks, or by noting that 3 feet are
about 1 meter. Then 4 × 104 km ≈ (5/8) × 4 × 104 mi = 2.5 × 104 mi, so the two results are
compatible. There are probably at least a hundred more ways to perform this estimation.

Example 8

Estimate the density of air, and compare this to the density of water.

Solution
We can directly use the ideal gas law, P V = nRT . The density is ρ = µn/V where µ is the
mass of one mole of air, so
µP
ρ= .
RT
Atmospheric pressure is about 105 Pa, typical temperatures are about 300 K, and air is mostly
N2 , which has a molar mass of µ = 28 g/mol, so

(0.028)(105 ) kg kg
ρ= 3
≈ 1 3.
(8.3)(300) m m

The density of water is, almost by definition,

kg
ρw ≈ 103 .
m3
Most liquids and solids have densities within an order of magnitude of this, since in all
cases the atoms are packed close together. Evidently, air molecules are about a factor of
(103 )1/3 = 10 times further apart than typical water molecules.

Example 9

Estimate how much useful power you can produce in a short burst.

19
 Kevin Zhou Physics Olympiad Handouts

Solution
This is a bit tricky to test, because most exercises just burn energy against air resistance or
friction, which is hard to estimate. However, a task that directly performs work is useful. I
weigh about 75 kg and can run up a 3 m high staircase in around 3 s, so

P = mgv = (75)(10)(3/3) W ≈ 750 W.

This is a typical max power output, while typical steady state power outputs are several
times smaller, and the corresponding numbers for elite athletes are several times larger.

For the below questions, feel free to look up specific numbers if you’re stuck. In all cases, an answer
to the nearest order of magnitude is good enough.
[3] Problem 28. Some questions about light energy.
(a) Estimate the number of photons emitted per second by a standard light bulb. (The energy of
a photon is E = hf , and the frequency of a photon is related to the wavelength by c = f λ.)
(b) The Sun supplies power of intensity 1400 W/m2 to the Earth. The nearest star is about 4
light years away. Assuming this star is similar to the Sun, about how many of its photons
enter your eye per second?
Solution. Before we continue, it’s important to note that for estimation questions, one should only
expect an answer to within an order of magnitude. Some teachers tweak their example calculations
until they give almost exactly the right answer. This makes them look brilliant, but it’s deceptive,
because then when the student tries to do the same, their results will be much further off. So to
combat this, in all solutions here, we’ve just presented our very first, simplest guesses. They can be
up to an order of magnitude off from the real numbers, so if your numbers are within two orders of
magnitude of ours, you’re fine!
(a) We can estimate a standard light bulb to have around 50 W of power. The power P = N E
where N is the number of photons emitted per second, and the wavelength of visible light is
from 400 − 700 nm (we can use 500). Then
Pλ
N= ≈ 1020 photons/s
hc
(b) 1 AU is about 1.5 × 1011 m. (If you forget, you can use something like GMS /r2 = (2π/T )2 r,
where T is one year and MS ≈ 2 × 1030 kg). 1 light year is c × 1 year ≈ 9.5 × 1015 m. Then
the intensity from the star is reduced by a factor of (1 AU/4 ly)2 due to the inverse square
law, so I ≈ 3.5 × 10−7 W/m2 . The area of a human pupil depends on the light conditions, but
is roughly πr2 = π(5 mm)2 . Then the number of photons that enter it per second is P λ/hc,
which gives N ≈ 107 photons/s. That’s plenty, so it’s very easy to see such a star at night,
while it might be difficult during the day because of the background light from the Sun.
[2] Problem 29. Estimate the radius of the largest asteroid you could jump off of, and never return.
p
Solution. The escape velocity is v = 2GM/R, and we will assume a uniform spherical asteroid
with density ρ. Rock is probably a few times denser than water, so ρ ≈ 3 × 103 kg/m3 and
√
M ≈ 43 πρR3 . Humans can jump around half a meter, which determines v = 2gh. Thus
2G 4
2gh = πρR3 .
R 3

20
 Kevin Zhou Physics Olympiad Handouts

Since g ≈ π 2 in SI units, this simplifies to

s
3πh
R≈ ≈ 2 km.
4Gρ

[4] Problem 30. Some questions about energy.

(a) Estimate the digestible energy content of a stick of butter. (A calorie is about 4000 J, and is
also the energy needed to raise the temperature of a kilogram of water by 1 K.)

(b) Estimate the rate at which your body burns energy when at rest.

(c) Estimate the rate at which a human being radiates energy. (The Stefan–Boltzmann law states
that the radiation power per unit area from a blackbody is σT 4 , where σ = 5.7×10−8 W/m2 K4 .)
Is radiation a significant source of energy loss for a human being, or is it negligible?

(d) A human being develops mild hypothermia, with their core body temperature dropping by
5 ◦ F. Neglecting any heat transfer with the environment, estimate the number of calories
required to raise their temperature back to normal.

Now let’s verify the energy content of the butter microscopically. This will be a very rough estimate,
so expect answers to be only within two orders of magnitude.

(e) A chemical bond typically involves two electrons, and a characteristic atomic separation
distance of one angstrom, r ∼ 10−10 m. Estimate the binding energy of one chemical bond.

(f) The fats in butter are digested by inputting energy to break the bonds in the molecules, then
harvesting energy by combining the atoms into CO2 and H2 O, which have somewhat more
stable bonds.

Estimate the energy content of a kilogram of butter. How close is this to the true result?

Solution. (a) A typical diet is 2000 Calories per day, where the “Calorie” used to describe
nutritional facts is actually a kilocalorie, which is about 4000 J. A stick of butter has about
800 Calories, giving an energy of 3 × 106 J.

(b) The power is

2000 Cal
≈ 100 W.
86400 s
This is used to maintain homeostasis in your body, and eventually is exhausted as heat.

(c) A rough estimate for the surface area of the human body is A ≈ 1 m2 . Human skin is on the
order of 300 K, so P ≃ AσT 4 ≈ 500 W.
This is much too high, as it can’t possible be higher than (b). The point is that the radiation
output by the human body is almost completely cancelled by the radiation input by the
environment, which is at almost the same temperature (in absolute terms). For example,

21
Kevin Zhou Physics Olympiad Handouts

in typical indoor conditions, the environment is at 70◦ F and human skin is at 90◦ F, for a
difference of about 10 K. So the power is smaller by a factor of 1 − (290/300)4 = 0.13, giving
a reasonable 65 W. It’s still a significant contribution, but not unreasonably large. Of course,
in practice it’s further reduced by wearing clothes, which trap a layer of insulating air and
cloth near you.

(d) 5◦ F is 5 × 5/9◦ C ≈ 3◦ C. Now we use Q = mcT , and since humans are mostly water, we’ll
approximate the specific heat to be the same as water. We take a human mass of 70 kg. Since
the “Calorie” used to describe food is actually a kilocalorie (amount of energy needed to raise
1 kg by 1◦ C), we just need ∼ 200 Calories to raise our temperature back to normal.
A cute, related estimate is the caloric cost of drinking ice-cold water. If you drink eight cups
of water a day, that’s 2 kg of water. Warming it from 0◦ C to ∼ 35◦ C takes ∼ 70 calories.

(e) A basic estimate for the binding energy is

e2
E∼ ∼ 2 × 10−18 J.
4πϵ0 r
As a check, this is about 10 eV, and the binding energy of hydrogen is about 13.6 eV (one of
those classic numbers you should remember), so this is in the right ballpark. Of course, the
energy is actually negative, even though electrons repel, because it’s due to how the electrons
are attracted to the nuclei. We can, however, very roughly estimate this negative energy using
the positive energy of repulsion e2 /4πϵ0 r because all energy scales in the problem should be
roughly similar.
Actually, in reality the answer should be about an order of magnitude lower, for two reasons.
The first is simply that atomic separations are a bit bigger, but this is cancelled by the fact
that the nuclei have charge Zi > 1. The main issue is that covalent bonds are a bit more
subtle.
Naively, you could say that a covalent bond is attractive because the electrons in one atom are
attracted to the nuclei of the other. But this is too naive, because at least parametrically, it’s
cancelled out by the repulsion of the nuclei with each other, and the repulsion of the electrons
with each other, as all four of these terms are of order ±e2 /4πϵ0 r. Covalent bonds are stable
because the electron orbitals can deform a bit, so that the negative contributions end up a
bit bigger than the positive ones. So e2 /4πϵ0 r isn’t really an estimate for the binding energy,
but for the sizes of terms which mostly cancel out to give the binding energy, which is why
the real answer is about 10 times smaller.

(f) Fats are mostly carbon. As a very rough estimate let’s say that the carbon atoms end up in
bonds that are twice as stable as before, so the energy released per carbon atom is on the
order of magnitude of what we found in part (e). Then

energy C atoms kilograms −1 12 g −1

   
energy −18
= ∼ (2 × 10 J) NA = 108 J/kg.
kilogram C atom mole mole mole
For comparison, the energy of one gram of fat is 9 calories, so the true answer is
J
(9)(4000)(1000) = 3.6 × 107 J/kg
kg
which is not too far off!

22
 Kevin Zhou Physics Olympiad Handouts

[2] Problem 31 (Povey). When human beings lose weight, most of it is by exhalation of carbon.
About 20% of the air in the atmosphere is oxygen. When we breathe in and then out, about 25%
of the oxygen is converted to carbon dioxide.

(a) Estimate the mass of air contained in a single breath.

(b) Estimate the amount of weight we lose every day by breathing alone.

Solution. (a) If I don’t take a deep breath, I can barely blow up a crushed plastic water bottle
(holds half a liter of volume), so I would estimate the volume in a single breath to be around
0.5 L. From chemistry class (or ideal gas law: n = P V /RT ), we know that mole of gas takes
up 22.4 liters of volume at STP (our body temperature, 310 K isn’t that much more than 273
K but we can just use 22.4 × 310/273 ≈ 25 L). Most of the air is nitrogen (N2 ) with molecular
mass 28 g/mol (oxygen, O2 , is 32 which is pretty close). Then one breath should have a mass
of 0.5 L/25 L/mol × 28 g/mol ≈ 0.6 g.

(b) By counting, we can estimate humans to breathe around 10 to 15 times a minute, so let’s
use 12.5, giving around 20,000 breaths in a day. In each breath, 20% of the air is oxygen,
and 25% of the oxygen is convert to carbon dioxide, for a net fraction of 5%. Carbon dioxide
(CO2 ) has a molecular mass of 48 g/mol, and oxygen is 32. Thus we lose a proportion of
(48/32 − 1) × 0.05 = 0.5 = 0.025 of the mass of the air we breathe in every day, which is about
0.3 kg. Most of the (non-water) mass of the food we eat leaves this way.

[2] Problem 32 (Insight). How long a line can you write with a pencil?

Solution. Graphene, a layer of carbons arranged in a hexagonal way, famously can be made from
using scotch tape to extract a few layers of graphite from pencil markings. It’ll take plenty of tries
to erase pencil from paper with tape probably (but progress is definitely noticeable), so we can
estimate there to be around 100 layers of the hexagonal carbon from graphite.
We will assume that the line is drawn with the pencil perfectly vertical and the lead not sharpened.
The diameter of the lead is around 2 mm, and the mass of a pencil should be around 2-10 grams,
so the mass of the lead is on the order of 1 g. Assuming that the lead is almost all made out of
carbon, we can estimate how many carbon atoms it has, and the surface density of carbon atoms.
The carbons are spread apart in a hexagonal fashion with a characteristic distance of r ≈ 10−10
m, and the centers of 3 adjacent hexagons will have a carbon atom at its center, giving a spacing
of 1 carbon atom every r2 square meters. There should be
1
6.022 × 1023 atoms/mole × 1 g × ≈ 5 × 1022 atoms of C
12 g/mol

Thus that gives around 500 m2 of a single layer of carbon, so around 5 m2 of lead usage. The line
will be approximately a rectangle with area dℓ, where d is the diameter of 2 mm.
Thus the pencil line should be around 2.5 km long. One can find other estimates of the order
50 km, i.e. a spread of an order of magnitude. The precise result within this order of magnitude of
course depends on the details of the pencil.

23