Bertrand’s paradox explained: random is a procedure, not a complete probability model

Keep the geometry fixed while changing the measure

For a unit circle, an inscribed equilateral triangle has side length square root of three. A chord’s length depends on the distance of its midpoint from the centre: the chord exceeds that threshold when this distance is below one-half. Those geometric facts stay the same across the three protocols in the experiment.

The Bertrand chord board changes how the chord is sampled. It can choose random endpoints, a uniformly distributed radial midpoint distance, or a midpoint uniformly distributed over disk area. Jaynes’s discussion of the well-posed problem supplies primary context for the importance of a defined sampling question. The activity explicitly compares protocols rather than announcing that every use of random chord must refer to one uniquely intended physical mechanism.

Uniform radius and uniform area are not the same instruction

A ring of equal radial thickness covers more area near a circle’s edge than near its centre. Sampling the radius uniformly therefore does not sample locations uniformly over disk area. For uniform area, the midpoint radius is the square root of a uniform draw between zero and one.

That difference changes the fraction of midpoints within radius one-half. Uniform radius gives one-half; uniform area gives one-quarter because area scales with squared radius. Choosing random endpoints gives another distance distribution and the one-third answer. These probabilities are not rival arithmetic solutions to the same fully specified procedure. They are answers to distinct procedures that the informal wording left unresolved. Drawing more chords estimates the chosen procedure more precisely; it cannot decide which procedure the speaker originally meant.

Worked example: estimate the answer only after choosing the generator

The board uses pseudo-random teaching draws and a fixed geometric threshold. Compare its finite fractions with the exact probabilities without demanding exact agreement in a batch of one hundred.

1. Choose random endpoints and draw one hundred chords. Count the chords whose midpoint distance is below one-half and compare the realised long-chord fraction with one-third under this protocol.
2. Switch to uniform radius and generate a new batch under that measure. The reference probability becomes one-half even though the circle and length threshold have not changed.
3. Switch to uniform midpoint area and inspect the one-quarter reference. Explain the square-root radius transform before treating the new estimate as another observation of the earlier protocol.

A constant rate does not mean a regular timetable

The Poisson-arrival activity poses a different generation question. Independent exponential gaps create arrivals at a constant average rate. The count in a short interval fluctuates and the gaps are uneven, even though the rate parameter does not change. A scheduled arrival process would have a different structure.

MIT’s stochastic-process course notes provide university material on this process. In the board, two expected arrivals per minute imply twenty over ten minutes in expectation and a mean gap of one-half minute. They do not require twenty arrivals in each realised stream. State whether independent increments and a constant rate are plausible before using that model for demand, calls or visits with rush hours and dependence.

Questions to ask whenever a probability sounds obvious

A clear generator often resolves an apparent paradox faster than another formula. It also makes a simulation auditable: a reader should be able to reproduce the sampling rule, target event and conditioning information. If those differ between explanations, compare the models explicitly rather than averaging their answers or choosing the one that feels most symmetric.

• Define the random object and the exact procedure or probability measure used to generate it.
• Specify the event being counted and the units or geometry that define its threshold.
• Distinguish a uniform coordinate from a uniform area, volume or other transformed quantity.
• Keep expected values across repetitions separate from requirements on every realised sample path.
• Name the conditioning information before assuming that an elapsed wait changes the next event’s odds.

Fair paths need not spend equal time on each side

The arcsine occupation activity offers another useful surprise. Symmetric random walks can average near one-half positive occupation across many histories while individual walks spend much of the finite horizon on one side. Symmetry of an ensemble does not force each realised history to resemble the average history.

The memoryless waiting board similarly separates the initial rarity of a long unsuccessful wait from the unchanged next-attempt probability under independent geometric trials. Neither model should be applied by naming it alone. Check the generator, conditioning and question first. A good simulation makes those choices visible, and a good explanation preserves them when moving from a toy example to a real observation.