A playable instrument · Pattern seam
Ahead the Whole Game
Toss a fair coin all night, one point to the winner each throw. You would guess the lead swaps hands often and each player is ahead about half the time. Both guesses are wrong — and not by a little. The most likely thing a fair game does is let one player lead almost the entire way.
Two players, a fair coin. Heads is a point for A, tails a point for B. After each toss, whoever has more points so far is in the lead. Play a few thousand tosses and ask a simple question: what fraction of the game was A in the lead?
Almost everyone expects the answer to hover near 50% — a fair coin, a fair game, the lead trading back and forth. The truth is the opposite. That answer, near a half, is the single least likely outcome. Watch one game first.
One fair game · 50 tosses
Most single games hug one side. A game that spends near half its time on each side is the rare one — you may have to flip many to see it.
Run it a dozen times. You will see games where A leads the whole way, games where B does, games that flip early and then commit to one side — but very few that split the time evenly. The lead changed hands only a handful of times, not constantly. Neither of those is an accident of the few games you watched. It is a law.
Ten thousand fair games
Here is the same fair game — 200 tosses — played over and over. Each game contributes one bar: the fraction of the game A spent in the lead, sorted into twenty columns from 0% (A never led) to 100% (A led throughout). If the naïve intuition were right, the bars would pile up in the middle. Watch where they actually go.
The distribution over many fair games
The gold curve is the exact law — density = 1 / (π·√(x(1−x))) — drawn straight from the formula, not fitted. Your bars converge onto it as you play more games.
The shape is a U. The bars tower at the two ends and sag in the middle: the most common outcome by far is that one player was ahead the whole time, and the rarest is a roughly even split. This is the arcsine law, first pinned down for coin games by William Feller in a chapter he built precisely because the results are so contrary to intuition, and proved in the continuous limit by Paul Lévy in 1939.
Toggle the buttons on the instrument. The last time the lead changed and the moment A's score hit its highest point follow the same U-shaped law — three different questions, one curve. The last lead change tends to fall near the very start or the very end, almost never in the middle; a fair game usually settles its leader early and then never seriously reconsiders.
But the coin was fair
That is the unsettling part. Nothing here is rigged. There is no momentum, no hot hand, no player "taking control." The coin has no memory and no bias. The lopsidedness is a property of pure symmetric randomness — the walk wanders away from zero and, once out, tends to stay out, because returning to a tie is itself rare. Ties thin out fast: in a game of N tosses the number of times the score returns to level grows only like √(2N/π) — so in a million tosses you expect fewer than 800 ties, and the fraction of the game spent near even shrinks toward nothing.
Which raises the honest question: is any of this about fairness, or only about this exact, drift-free coin? Test it. Bias the coin and watch the law break.
Bias the coin — the law is a fact about the fair game
At exactly 0.50 the two ends are equal — the arcsine U. Nudge the coin even slightly and the symmetry that builds the U shatters: the likelier player is soon ahead essentially the whole time, and the distribution piles into one corner. The arcsine law lives only at the knife-edge of a genuinely fair coin.
So it is not a statement about competition or fairness in the loose sense. It is a sharp statement about the driftless symmetric random walk: the one built from a coin with no edge at all. Give either side the faintest real advantage and the paradox dissolves into the boring, expected answer — the better player leads.
Where this shows up
The pure-chance component of any back-and-forth contest behaves this way. On an election night where the count is essentially a coin-flip between two candidates, one is likely to appear ahead for most of the evening — not because of momentum, but because that is what a near-tied random count does. A gambler at a fair game who feels the whole night was against them is very often right about the pattern and wrong about the cause: being behind almost the entire session is a common outcome of a perfectly fair game, not evidence of a crooked one.
Real elections, sports, and markets carry genuine skill and drift on top of the noise, so this is a claim about the random part, not the whole. Naming that boundary is the point of the biased-coin instrument above: the clean arcsine picture is exactly as true as the coin is fair.
The check. Every number on this page is recomputed from first principles in research/arcsine-law/verify.mjs — 47/47 green.
- ✓ The exact law — #games with 2k tosses in the lead = C(2k,k)·C(2N−2k,N−k) — proved by brute-force enumeration of all 22N games for N up to 11 (4,194,304 games).
- ✓ The last lead change obeys the identical law; the peak position matches it by Sparre Andersen's equivalence — both proved as exact identities by enumeration.
- ✓ The bars are tallest at both ends and shortest in the middle; the exact discrete law converges onto Lévy's arcsine curve as N grows (max deviation 0.078 → 0.0003).
- ✓ Balanced ≈ 6.4%, lopsided ≈ 41%; ties grow like √(2N/π); a biased coin is provably not arcsine.
The convention, named
"In the lead" counts the tosses whose path-segment lies above the level line (the standard Feller / Chung–Feller count; because consecutive scores differ by one, every segment is strictly above or strictly below — there is nothing to fudge). A slightly different count — the number of moments the score is strictly positive — differs by a vanishing fraction and shares the same arcsine limit. The histogram draws the segment count; the drift instrument uses it too.