Random Needs a Rule

Two careful mathematicians can calculate the probability of the same-looking event and get different answers. Neither has to be bad at math. They may have quietly chosen different definitions of random.

This is the lesson behind Bertrand's paradox, presented by French mathematician Joseph Bertrand in his 1889 book Calcul des probabilités.

1. The mechanism

Draw a circle. Fit an equilateral triangle inside it so all three corners touch the circumference. Now draw a random chord, a straight segment joining two points on the circle.

What is the probability that the chord is longer than one side of the triangle?

Bertrand showed why that question is incomplete. At least three defensible procedures produce three different answers:

• Random endpoints: Pick two points independently and uniformly around the circumference; join them. Probability: 1/3.
• Random distance: Pick a direction, choose a point uniformly along a radius, then draw a chord perpendicular to that radius. Probability: 1/2.
• Random midpoint: Pick a point uniformly by area inside the disk and make it the chord's midpoint. Probability: 1/4.

The geometry has not changed. The rule for generating a chord has.

Here is why the last two differ. A chord beats the triangle's side exactly when its midpoint is less than half the circle's radius from the center. The inner half of a radius occupies half its length, but a disk of half the radius occupies only one-quarter of the area. Uniform over length is not uniform over area.

The first answer comes from a different rule: with one endpoint fixed, only a third of possible uniformly chosen second endpoints produce a sufficiently long chord.

This is not three answers to one fully specified probability problem. It is three different probability models hidden inside one vague instruction.

2. Outside business: a mathematical trap

Imagine asking students to draw a chord “at random” for a classroom demonstration. One student picks two points on the circle's edge. Another closes their eyes and points somewhere inside the disk, using that location as the midpoint. Both genuinely randomize. Their resulting collections of chords will differ systematically even after many trials.

Collecting more observations does not erase the difference. With enough trials, each method becomes more precise about its own rule. The ambiguity is upstream of arithmetic.

3. Apply it to a beverage decision

Consider a hypothetical packaging study. You randomly select 20 stores and interview the same number of shoppers in each. A colleague instead draws shoppers randomly from the chain's transaction records.

Both can be legitimate random samples, but they weight stores differently. The first treats stores roughly equally. The second gives higher-volume stores more representation. The results may diverge even with careful execution.

Before deciding whether to spend on a portfolio redesign, define the commercial question. Are you measuring problems at the typical store or among the typical transaction-weighted shoppers? Neither is always right. They serve different decisions, and the target population, selection probabilities and weighting must match that decision.

4. Challenge the assumption

“We randomized the sample, so the result must be representative.”

Random selection is powerful, not magical. It can reduce selection bias relative to an appropriate sampling frame, but it does not automatically repair a wrong frame, nonresponse, or a mismatch between the population sampled and the population you want to understand.

A fair counterargument: we shouldn't turn Bertrand's paradox into an excuse to distrust statistics. Once the sampling mechanism and target are specified, probability can give precise, useful answers. The lesson is to make the rule explicit, not to abandon measurement.

5. Your mental operator

Before trusting a percentage, ask: Randomly selected from what? By which procedure? Weighted how? To answer whose question?

6. Reasoning exercise: answer before reading on

In a hypothetical library system, Branch A processes 900 loans, 90 of them overdue. Branch B processes 100 loans, 30 overdue. What overdue percentage would you expect if you first pick a branch with equal probability and then pick a random loan within it? What if you pick one loan uniformly from all 1,000 loans?

Worked answer: The branch-first method gives (10% + 30%) / 2 = 20%. The loan-first method gives (90 + 30) / (900 + 100) = 12%. Neither calculation is wrong. Each answers a question defined by a different selection rule. The figures are invented for practice, not real library data.

Keep this: “Random” is not a complete method. It is a promise that still needs a procedure.

Sources

• John D. Norton, University of Pittsburgh, Bertrand's Geometrical Paradoxes: Proofs.
• Internet Encyclopedia of Philosophy, Probability and Induction.
• Wolfram MathWorld, Bertrand's Problem.
• ETH-Bibliothek Zürich, Bertrand's Calcul des probabilités (1889).

Educational editorial content. Hypothetical examples are labeled in the original edition.