Published 13 September 2026

Borel’s paradox

A uniformly selected point lies exactly on a line of area zero. What is the distribution along that line? The ordinary conditional-probability ratio cannot answer: both its numerator and denominator are zero. A limiting procedure can give an answer, but the way its neighborhoods shrink matters.

The classical Borel–Kolmogorov discussion uses a sphere and a great circle. Here a square makes the same issue visible with elementary area calculations. Let X and Y be independent uniform coordinates on [−1,1]. We compare two observations that become arbitrarily close to the horizontal line Y = 0.

Narrow two different neighborhoods

The strip retains |Y| < ε. Its vertical thickness is 2ε at every horizontal position. The wedge retains |Y| < ε|X|. Its thickness is 2ε|X|, narrowing to zero at the center and widening toward the left and right edges. Both concentrate near the same line as ε decreases, but they weight positions along it differently.

The target is the central interval |X| < a. At a = 1/2, the strip gives conditional probability 1/2, whereas the wedge gives 1/4. These values hold for every positive ε up to one in the displayed model, so their limits remain different.

Count strip area

The whole square has area four. The strip has width two and height 2ε, hence probability ε. Its part with |X| < a has area 4aε. Dividing by the strip area gives a.

Equivalently, the conditional density of X is 1/2 throughout [−1,1]. The strip gives every horizontal location the same vertical opportunity to be selected. Shrinking ε reduces the probability of selection, but does not change those relative opportunities.

Count wedge area

At horizontal coordinate x, the wedge’s vertical thickness is 2ε|x|. Integrating from −1 to one gives area 2ε and probability ε/2. Integrating only from −a to a gives area 2εa². Dividing yields a².

The corresponding conditional density of X is |x| on [−1,1], which integrates to one. It favors locations near the outer edges. Both conditional distributions have mean X equal to zero by symmetry, yet they disagree on central probability and on the expected absolute distance from zero: one half for the strip and two thirds for the wedge.

Which variable is observed?

For X different from zero, define Z = Y/X. The wedge is the event |Z| < ε, while the strip is |Y| < ε. At zero tolerance, Z = 0 and Y = 0 describe the same line except at the origin, a null difference. But observing Y to a small tolerance and observing Z to a small tolerance define different experiments.

Changing variables consistently does not change the answer to one fixed positive-probability event. The difference here arises because the two tolerances correspond to different events before taking a limit. A label such as “uniform along the line” already chooses additional structure; it is not forced by uniform area in the surrounding square.

What conditional densities still allow

This phenomenon does not make ordinary continuous conditional densities unusable. A model can specify a conditioning random variable, a measurement procedure, or a regular conditional distribution. Such conditional distributions are determined only almost everywhere in the conditioning variable, so values at a particular null outcome can need an additional convention or continuity choice.

The diagrams reuse one seeded point cloud. At narrow widths it may contain no selected points despite a positive exact neighborhood probability. An empty empirical denominator is reported explicitly. Increasing the sample can improve a frequency estimate, but it cannot choose between the two observation protocols.

Reference and next step

Michael Rescorla, Some Epistemological Ramifications of the Borel-Kolmogorov Paradox discusses the sphere and simpler coordinate examples. The strip-and-wedge derivation here states its own neighborhoods explicitly. Compare Bertrand chords, where different measures arise before any conditioning takes place.