Bertrand’s Paradox: When “Random” Isn’t a Real Answer

In 1889, the French mathematician Joseph Bertrand posed what looked like a harmless geometry question:

It sounds clean. Precise. Mathematical. But there is no single correct answer. Depending on how you interpret “at random,” the probability is: $\frac{1}{4}$, $\frac{1}{3}$, or $\frac{1}{2}$. And all three answers can be derived rigorously.

Who Was Bertrand?

What Is the Problem Exactly?

Why This Matters

Bertrand’s paradox reveals something interesting: Probability does not live in geometry alone. It lives in:

$$ (\text{sample space}) + (\text{a probability measure}) $$

If you do not specify the measure — how chords are generated — the problem is incomplete. Different generation methods impose different probability distributions. Different distributions produce different answers. No contradiction or mistake, just different assumptions.

Why Should We Care

Bertrand’s paradox shows up anytime someone says:

• “We sampled uniformly.”
• “We ran random simulations.”
• “We chose random test cases.”

Uniform over what? Random with respect to which measure? Those hidden modeling choices shape conclusions. Bertrand’s paradox forces us to confront an uncomfortable truth: There is no such thing as neutral randomness. There are only assumptions — explicit or implicit. And once you make them explicit, the paradox disappears.

Okay, But Which Method Is “Right”?

Now that we’ve seen three different answers — $\frac{1}{3}$, $\frac{1}{2}$, and $\frac{1}{4}$ — your brain probably wants closure.

So let’s slow down and examine what actually changed.

The geometry didn’t change.
The circle didn’t change.
The triangle didn’t change.

Only the procedure changed. And the procedure is the probability distribution.

Let’s Make the Geometry Explicit

Assume the circle has radius $R$. The side length of the inscribed equilateral triangle is:

$$ s = \sqrt{3} R $$

Now consider a chord whose midpoint is a distance $r$ from the center. The length of that chord is:

$$ \ell = 2\sqrt{R^2 - r^2} $$

We want:

$$ \ell > \sqrt{3} R $$

Substitute:

$$ 2\sqrt{R^2 - r^2} > \sqrt{3} R $$

Square both sides:

$$ 4(R^2 - r^2) > 3R^2 $$

$$ 4R^2 - 4r^2 > 3R^2 $$

$$ R^2 > 4r^2 $$

$$ r < \frac{R}{2} $$

So the chord is longer than the triangle side exactly when its midpoint lies within a circle of radius $R/2$. That’s the geometric heart of the puzzle. Everything now depends on how we distribute midpoints.

Why the Three Answers Happen

Case 1: Uniform Midpoints in the Disk

If midpoints are uniformly distributed in area:

• Total area = $\pi R^2$
• Favorable area = $\pi (R/2)^2$

So:

$$ P = \frac{\pi (R/2)^2}{\pi R^2} = \frac{1}{4} $$

Case 2: Uniform Distance Along a Radius

If we:

• Choose a random direction,
• Then choose $r$ uniformly in $[0, R]$,

Then:

$$ P = \frac{R/2}{R} = \frac{1}{2} $$

Because now we’re weighting radius linearly, not by area.

Notice something subtle: Uniform in area means density proportional to $r$. Uniform in radius means density constant in $r$. Those are different measures.

Case 3: Uniform Endpoints on Circumference

This implicitly induces yet another distribution on $r$.

If you derive it carefully, the density of $r$ becomes:

$$ f(r) = \frac{2}{\pi \sqrt{R^2 - r^2}} $$

Integrating over $r < R/2$ gives:

$$ P = \frac{1}{3} $$

Different mechanism. Different induced measure.

Same geometry.

The Core Insight

“Random chord” is not a well-defined object.

It is shorthand for:

A chord drawn according to some probability measure on the space of chords.

Different parameterizations create different uniformity assumptions.

Uniform in:

• angle
• radius
• midpoint area
• endpoints

are all different probability spaces.

The Deeper Mathematical Issue

Probability is not invariant under reparameterization.

If:

• $X$ is uniform in $[0,1]$,
• and $Y = X^2$,

then $Y$ is not uniform.

Uniformity is coordinate-dependent.

Bertrand’s paradox is what happens when you say:

“Uniform” — but fail to specify in which coordinates.

A Mental Model That Helps

Think of it this way.

Imagine three robots:

• Robot A picks two random edge points.
• Robot B picks a random radius and distance.
• Robot C picks a random midpoint in area.

You press “run simulation.”

All three claim:

“We generated 10 million random chords.”

They disagree on the answer. But none are incorrect. They just implemented different measures.

Why This Matters Beyond Geometry

This puzzle shows up anytime someone says:

• “We sampled uniformly.”
• “We ran unbiased simulations.”
• “We chose random test cases.”

Those words are meaningless without a measure.

Uniform over:

• parameters?
• exposure?
• physical space?
• time?
• scenarios weighted by frequency?

Each choice encodes assumptions. Bertrand’s paradox is the geometry version of:

So How Do We Decide Which Is Correct?

We impose a physical principle.

For example:

• If chords are created by dropping random sticks onto a circle, midpoint-uniform may be justified.
• If chords arise from random angles, endpoint-uniform may be justified.
• If chords are constructed by choosing random radius offsets, then radius-uniform applies.

The “correct” answer depends on the generative process. No generative process → no unique probability.

A Slightly Philosophical Ending

Bertrand’s paradox teaches a powerful lesson: