A portal · the ground becomes a network

The Average Nobody Lives

An expected value is an average over the ensemble — the crowd of all the lives you might have led. You get exactly one. Whether that famous number describes your life is a separate mathematical question, and three layers already on this ground answer it three different ways: the average can be real and reserved for a vanishing few, infinite and unreachable by any amount of your own averaging, or exactly right and still the rarest thing that can happen to you. One fair coin is enough to build all three.

This is a portal. It takes three pieces already on the ground — each a checked, playable demonstration of one strange average — and supplies the single claim none of them states alone. Read side by side, they are not three curiosities about gambling. They are three independent failures of the same bridge: the bridge from a statement about a crowd of possible lives to a statement about the one life you actually get. Each failure has its own mechanism, its own theorem, and its own cure — and none of the three implies the others.

The three layers

The claim, in one line

What makes the three games below a genuine spine and not a grab-bag is this: they are the same coin. One stream of fair fifty-fifty flips — the most transparent random object there is — drives all three at once. The same toss that multiplies the first ledger, extends a run in the second, and moves the third one step. Three accountings of one night of chance, and the word average betrays each of them differently.

One coin, three ledgers — and a control

Deal the coin. Every flip is used three times over: ledger A multiplies your wealth by 1.5 on heads, 0.6 on tails; ledger B reads the same flips as runs — heads keep a St. Petersburg game alive, the first tails ends it on toss k and pays $2^k; ledger C steps a score +1/−1 and tracks who leads, one “night” per 200 flips. The fourth panel is the control: the plain average of the ±1 steps themselves — the one average here that your single life does converge to.

1 · The average that is real — and nobody’s

Ledger A is the plainest trap in decision theory. The coin is fair; heads multiplies your wealth by 1.5, tails by 0.6. The expected multiplier is ½·1.5 + ½·0.6 = 1.05 — a genuine, finite, correct +5% per flip. Average a million parallel players and their pooled wealth really does grow at that rate; after 100 flips the ensemble mean sits at ×131.5. The number is not a lie. It is the truth about the crowd.

But a single life multiplies, and what compounds down one path is the average of the logarithm: E[ln m] = ½·ln 1.5 + ½·ln 0.6 = −0.0527 — negative. One down-up pair is ×0.6×1.5 = ×0.9: every tails takes a bigger bite than the heads that “cancels” it restores. So while the ensemble grows +5% a flip, the path you are actually on decays at −5.13% a flip. The two growth rates don’t just differ — they point in opposite directions, and that is the definition of the trap: the process is non-ergodic. The time average is not the ensemble average.

Where did the +5% go? Into a tail of parallel worlds: the mean floats ≈25,500× above the median, held up by a vanishing fraction of impossibly lucky paths. The cure is Kelly’s (1956): bet a quarter of your wealth instead of all of it and the one-life growth rate turns positive (+0.62% a flip) — you rescue your time average by maximising E[ln], and betting even twice that fraction lands you exactly back at break-even. Run the full instrument, with Kelly’s dial →

2 · The average that cannot be reached

Perhaps, then, the fix is patience: live long enough, average your own games, and surely your running average finds the true mean. Ledger B is the three-century-old counterexample. Flip until the first tails; a first tails on toss k pays $2^k. Half of all games pay $2. A quarter pay $4. And the expected value is ½·2 + ¼·4 + ⅛·8 + … = 1 + 1 + 1 + … — infinite. Every textbook rule says pay any price to play; no one in three hundred years has agreed to more than a few dollars, and the refusers are right.

Here the bridge fails at the far end: the ensemble number exists only as a divergence, so the law of large numbers has nothing to deliver. Feller proved (1945) exactly what your lifetime looks like instead: the running average after n games doesn’t settle — it climbs, like log₂ n, forever, jolted upward each time a long run of heads lands and sagging between jackpots. There is no strong law behind it at all (Chow & Robbins, 1961): patience does not converge here. Watch panel B — the cyan line is your life’s average, and the faint curve it climbs along is log₂ of the games played.

Two honest ways to kill the paradox, both checked: cap the house — a bank holding $1,000,000 makes the fair price $20, and all the money on Earth makes it $49 — or do what Bernoulli did and average the logarithm of the payout, which prices the game at a homely geometric-mean $4. Note that move: it is the same logarithm that rescued ledger A. Play the game that refuses to average →

3 · The average that arrives — and still isn’t yours

So the mean failed you twice: once for being nobody’s, once for being infinite. Ledger C closes the last escape hatch, because here nothing goes wrong with the mean at all. Two players, fair coin, +1/−1, all night; count the fraction of the night each player spends in the lead. By symmetry the expected fraction is exactly ½ — not approximately, exactly. The variance is finite. Average over many nights and the sample mean marches obediently to ½; the law of large numbers works perfectly. Every alarm from games A and B is silent.

And the average night still never happens. The distribution of the lead-fraction is U-shaped — Lévy’s arcsine law — with all its weight piled at the extremes: the single most likely night is one player ahead the whole time, and a near-even split is the single least likely. In the continuous limit a night spent 45–55% ahead has probability 6.4%; a lopsided night (one player ahead ≥90% of the time) has probability 41% — better than six to one for the “unfair-looking” outcome, produced by a provably fair coin. The mean is the exact centre of a distribution that is empty in the centre.

There is a deeper twist, and the offline verifier walks it: stay in one unending game and track your own running lead-fraction. The step-average beside it (panel D) pins itself to zero like a good citizen — same path, same flips. The lead-fraction never settles at all: its spread across possible lives is as wide at half a million flips as at two thousand (the arcsine law is scale-free), so it converges to no constant, ever. Two averages of one life, one obedient, one lawless — the difference is not the coin but the functional. Watch ten thousand fair games pile into the U →

The control: an average you do live

If the portal stopped here it would be telling a half-truth — “averages lie” — and that is slop, not a claim. So take the same coin one more time and compute the most boring statistic it has: the running mean of the ±1 steps themselves. It converges. It converges to you: your one path’s step-average is pinned to 0 inside a shrinking ±1/√n funnel (panel D), exactly as the strong law promises. Additive increments, finite mean, finite variance, a statistic that concentrates: this is the regime where the ensemble number and your life agree — and it is the regime almost every everyday average lives in. The portal’s claim is not that the bridge is out. It is that the bridge has three load limits, each checkable before you cross:

Three failures, three mechanisms — and they are not the same failure

The temptation, having lined these up, is to blur them into one moral. The honest portal refuses: the point is that the three mechanisms are separable, and the verifier separates them. Game A has a finite, well-behaved mean — its failure is multiplicative dynamics, and capping nothing, changing nothing, only betting fractionally cures it. Game B is purely additive — no compounding anywhere — and its failure is the infinite mean: cap the payout at $1,024 and the verifier watches the law of large numbers snap obediently back into force. Game C has a finite mean, finite variance, additive steps and a working law of large numbers — and fails anyway, because convergence-of-the-average and typicality-of-the-average are different properties. Three games, three distinct broken parts, one word — “average” — covering for all of them.

The two cures are one cure

A last seam, and the oldest: game A’s cure (Kelly, 1956 — bet the fraction that maximises E[ln wealth]) and game B’s cure (Bernoulli, 1738 — price the game by E[ln payout], the geometric mean) are the same mathematical move, discovered two centuries apart: when lives multiply, average the logarithm, because the logarithm is what one life adds up. Game C needs no such repair — its number was never wrong. The only thing that has to move there is you: expect the lopsided night, and the fair coin stops looking rigged. Twice the cure is a formula; the third time the cure is a corrected expectation. That is the portal in miniature.

Sources