Bertrand's Paradox: What Does "At Random" Mean?
One question, three impeccable arguments, three different answers. The paradox is not a flaw in probability — it is the discipline's founding lesson, and by the end of this subchapter you will be unable to write the words "at random" without finishing the sentence.
In this subchapter
6.4.1The question
Draw a chord of a circle "at random". What is the probability that it is longer than the side of the inscribed equilateral triangle?
Fix the setup geometrically: circle of radius r; the inscribed equilateral triangle has side √3·r. The useful reduction is to a threshold on the midpoint: a chord is longer than the triangle's side iff its midpoint lies within distance r/2 of the center (chord length is 2√(r² − d²) with d the midpoint's distance from center; setting this above √3 r gives d < r/2). So the question is really: for a "random" chord, what is the chance its midpoint falls in the inner disk of radius r/2? And the answer depends entirely on what "random chord" distributes.
6.4.2Three answers, each proved
Pick two points independently and uniformly on the circumference; join them. By rotational symmetry fix the first point; the second point's position, uniform on the arc, is the only randomness. Mark the triangle's vertices at angles 0°, 120°, 240° from the first point. The chord is longer than the side iff the second point lands on the arc between the two far vertices — the middle third of the circle: an arc of 120° out of 360°. P = 1/3. The computation is a genuine uniform-measure calculation on the circle — a special case of the product construction of 6.1.
Pick a radius uniformly from all directions (uniform angle), then a point uniformly along it, and draw the chord through that point perpendicular to the radius. The chord's length depends only on the point's distance d from center: it is long iff d < r/2, and the point is uniform on [0, r]: P = 1/2. Also a perfectly specified uniform computation — uniform angle times uniform distance.
Pick the chord's midpoint uniformly in the interior disk (uniform area measure, λ/πr² — the measure of 3.1 on the plane, normalized), and draw the perpendicular chord there. Long iff the midpoint falls in the disk of radius r/2: probability (π(r/2)²)/(πr²) = 1/4. P = 1/4. Again: a completely specified, correctly computed probability.
6.4.3Resolution: three questions, three measures
Each of the three constructions defines a genuine probability measure on the set of chords, each computation is correct, and the three measures are different. There is no contradiction, and no unique answer: the phrase "a random chord" does not determine a probability question. A probability question is well-posed only when the probability space — outcome set, σ-algebra, and above all the measure — is specified.
It suffices to exhibit one event with three different probabilities — the very event of the paradox does: its probabilities are 1/3, 1/2, 1/4 under the three measures. That is a proof of difference, not merely an assertion: if the measures were equal, every event — including this one — would have one probability. The three underlying constructions are each legitimate applications of the product-measure theorem (6.1 Theorem 1): e.g. the endpoint measure is uniform-on-circle ⊗ uniform-on-circle pushed forward by the map "join the pair"; the midpoint measure is normalized area pushed forward by "perpendicular chord through the point". Different pushforwards (5.4) of different initial measures — the surprise would be if they agreed. ∎
The tempting escape: "just use the symmetric measure". All three constructions are rotation-symmetric — each treats all directions identically. The difference lies in a subtler symmetry: what happens to the measure under maps of the chord space. The endpoint measure is invariant under rotations only; the midpoint (area) measure is invariant under rotations and reflections, and under any area-preserving map of the disk; the radius-point measure is invariant under rotations but under a different group of chord-transformations. "Symmetric" is not one property; it is a family of properties indexed by a group of transformations, and naming the group is part of naming the measure. This is the precise content behind the paradox's discomfort: the words of the question underdetermine the mathematics, and only the mathematics computes.
6.4.4Invariance: which measure matches which symmetry
- Uniform on the circle (arc-length measure): invariant under all rotations — the unique rotation-invariant probability on the boundary. This is the measure of "a random direction", used everywhere in the course (the Cauchy generator of 5.5, the angles of Brownian directions).
- Normalized area on the disk: invariant under all rotations and all rigid motions of the plane — the physically natural measure for "a random point in the region".
- Uniform on a radius: a hybrid — natural when a physical process first picks a direction, then travels along it (a photon's path, a drilling rig's bore).
Which is "right"? The question is malformed. The right question is: which one models the physical process you are actually analyzing — and that is an empirical or design question, not a mathematical one. Mathematics, having been handed a measure, computes flawlessly; it cannot choose your measure for you.
6.4.5The modeling discipline
"A random time": uniform over the day? over business hours? weighted by trading activity? Each is a different measure on the same clock; any statistic computed "at a random time" changes with the choice — and papers that leave the choice implicit have simply hidden their result inside it. "A random price path": uniform over what — increments? endpoints? drawdown shapes? The space of paths admits many natural measures (Chapter 8's tools will make this precise), and a "typical path" claim is a measure claim. "A random user/data point": sampled by frequency? by account? by session? — sampling frames are measures, and selection effects are exactly Bertrand's paradox wearing a lanyard. The paradox is 130 years old and re-enacted daily in every data team.
Every probability statement in applied work should pass this test: can you name the probability space? Outcome set, σ-algebra (usually implicit and fine), and the measure — concretely enough that two competent readers would construct the same one. "At random", "typical", "on average", "the market's distribution": each is either a pointer to a specified measure or an unresolved ambiguity. The whole apparatus of Chapters 1–6 exists so that this test can always be applied; Bertrand's paradox is where it earns its keep.
6.4.6Chapter 6, closed
Chapter 6 built the multi-variable world: independence as multiplication on product spaces (6.1), iterated computation under Fubini–Tonelli (6.2), the long-run logic of Borel–Cantelli (6.3), and the discipline that probability questions are measure questions (6.4). What remains — the payoff of the entire course — is the asymptotic: what happens when you average, repeatedly, forever. Chapter 7 assembles the machinery of convergence (modes, laws of large numbers, the central limit theorem), and Chapter 8 completes the theory with conditioning. Every theorem from here forward runs on the exact infrastructure now in place: σ-algebras as information, products as independence, and measures as the sole arbiter of "at random".
There is no such thing as a random chord — only a chord drawn from a measure you can name. Bertrand's three answers are the course's thesis in a single picture.
Chapter 7 asks the question every practitioner believes they already understand: what does it mean for random variables to converge? Almost surely, in probability, in Lp, in distribution — a zoo of modes, a strict hierarchy, and the famous counterexamples (the typewriter sequence among them) that keep the distinctions honest.