Time average versus ensemble average
A Socratic walk-through of time average versus ensemble average — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why can a bet with a favourable average outcome still ruin nearly everyone who keeps taking it?
Here is a coin game. Stake your whole wealth. Heads, it grows by 50 percent; tails, it shrinks by 40 percent. Fair coin, and you may play as often as you like.
The expected outcome of one round is half of 1.5 plus half of 0.6, which is 1.05. A five percent gain per round, forever. And yet if a thousand people play it a hundred times, almost every one ends up with a fraction of what they started with. Nothing in the arithmetic was wrong. So what was the "average" describing?
Reasoning it through
REASONING #Separate two questions that ordinary language runs together. If a thousand people each play one round, what is the average wealth across those thousand? That is an ensemble average, taken over a population. If one person plays a thousand rounds, what growth rate do they experience? That is a time average, taken along a single path. We usually assume these are the same number. Here they are not even the same sign.
Trace one path. Play twice and get one head and one tail — the most typical outcome there is. Your wealth is multiplied by 1.5 and by 0.6, and 1.5 x 0.6 = 0.9. You are down ten percent, and the order made no difference, because multiplication commutes. Meanwhile the ensemble mean after two rounds is 1.05 squared, or 1.1025. The same two rounds leave the typical player at 0.9 and the population average at 1.10.
Why the gap? Because wealth compounds multiplicatively, and what accumulates along a path is not the sum of the outcomes but their product. The right per-round summary of a product is the geometric mean: the square root of (1.5 x 0.6) = the square root of 0.9 = 0.9487. Each round, the typical trajectory is multiplied by about 0.949 — a loss of 5.13 percent per round, sitting inside a game whose ensemble mean gains 5 percent per round.
Run it out to a hundred rounds and the two numbers separate absurdly. The ensemble mean is 1.05 to the power 100, about 131.5 times the stake. The median trajectory is 0.9487 to the power 100, about 0.0052 — roughly one part in 194 of what you started with. Same game, same hundred rounds.
If nearly everyone is ruined, who is holding the 131.5? Work out how many heads merely break you even: you need k heads out of 100 with 1.5 to the k, times 0.6 to the (100 − k), at least 1. Taking logarithms, k x 0.4055 + (100 − k) x (−0.5108) ≥ 0, so k ≥ 55.7, meaning 56 heads. The chance of 56 or more heads in a hundred fair flips is about 13.6 percent. Roughly seven players in eight end below their stake, and the ensemble mean is carried by the sliver of paths with a large excess of heads — paths whose wealth is astronomical and whose probability is minute.
That is the whole phenomenon. The mean exists, it is finite, it is correct, and it describes a population that will never be your experience, because you only ever get one path.
The analogy
THE ANALOGY #Think of a thousand hikers crossing a range where every step climbs 50 percent of your current altitude or drops it by 40 percent — proportional, not fixed. Ask the helicopter pilot for the party's average altitude and it rises steadily all afternoon. Ask any individual hiker and almost all are far below where they started, while a handful sit implausibly high on one ridge, carrying the average alone.
hikers can see each other and turn around, so the party would notice and stop, whereas the point of the multiplicative gamble is that each player observes only their own descent and cannot infer from it that the population mean is climbing — the divergence is invisible from inside a single path.
Clarifying the model
THE MODEL #Several corrections keep this honest.
This is not the heavy-tail problem. The neighbouring walk-through on unstable averages concerns distributions whose mean is hard to estimate or does not exist. Nothing of that kind happens here: wealth after any finite number of rounds has a perfectly finite mean, and with enough players you could estimate it to any precision you liked. The failure is different in kind — the estimate would be correct and still describe no one. That property, time average not equal to ensemble average, is what "non-ergodic" means.
It is the multiplication, not the odds, that does the damage. Change the game to a fixed stake: win 50 pounds or lose 40 pounds on a fixed 100-pound bet, with your wealth carried alongside. Now outcomes add rather than multiply, the time average and the ensemble average coincide at +5 per round, and the bet is simply good. Additivity restores ergodicity. This is the limiting case that shows the culprit is compounding.
The bet is not unplayable — the sizing was wrong. Stop betting everything. Wager a fraction f of wealth each round and the per-round log growth is half of ln(1 + 0.5*f*) plus half of ln(1 − 0.4*f*). Setting the derivative to zero gives 0.25/(1 + 0.5*f*) = 0.2/(1 − 0.4*f*), so 0.25 − 0.1*f* = 0.2 + 0.1*f*, and f = 0.25. Betting a quarter of your wealth each round yields a log growth of +0.0062, about 0.62 percent per round — genuinely positive, genuinely compounding. The favourable average was real; extracting it required not staking everything on each draw.
What is old and what is contested. The mathematics here — geometric versus arithmetic means, the log-optimal bet size — is uncontroversial and predates modern finance. The stronger programme reframing economic rationality around time averages rather than expected utility is a live argument, not settled fact, and I flag it rather than assert it.
Where the bell curve went. The central limit theorem still applies here, but to the logarithm of wealth, which is a sum of independent steps. So log wealth is approximately normal, wealth itself approximately lognormal, and a lognormal's mean sits well above its median. Non-ergodicity is what a right-skewed distribution feels like from inside.
A picture of it
THE PICTURE #How to readThe bands carry players, not money — a thousand of them, splitting five hundred each way at the first toss — through two rounds of the coin game, and every node label gives the wealth multiple that group is standing on. Read the right column by width: 750 of the 1000 players finish at 0.9 times their stake or below, and only 250 finish above it. Then read it by arithmetic: 250 x 2.25 plus 500 x 0.9 plus 250 x 0.36 comes to 1102.5, exactly the promised 1.05 squared per player. The single thin-but-tall band at 2.25 is where the whole ensemble average lives.
What became clearer
WHAT CLEARED #An expected value is an average across possible worlds at one instant. A growth rate is an average along one world across time. For anything that adds up they coincide and nobody needs the distinction. For anything that compounds they can point in opposite directions, and the mean stops summarising a typical experience and becomes a report on the luckiest paths. The coin game was never a good bet at full stake; it was a good bet only for a population, and no individual is ever a population.
Where to go next
ONWARD #- The Kelly criterion in general form, and why leveraged funds decay even when their underlying index rises.
- Insurance and pooling, which are attempts to buy access to the ensemble average from inside a single path.
Key terms
TERMS #| Term | What it means |
|---|---|
| Ensemble average | the mean over many parallel instances at a fixed time; time average — the long-run average along one trajectory. |
| Ergodic | the property that these two coincide; multiplicative wealth dynamics are not ergodic. |
| Geometric mean | the *n*th root of a product, the correct per-round summary of compounding growth. |
| Kelly criterion | the bet size maximising the time-average growth rate, found by maximising expected log wealth. |
Every term the collection defines is gathered in the glossary.