Ergodic vs Non-Ergodic: The One Distinction

In the intro we watched a single coin — pay you +50% on heads, dock you −40% on tails — split into two contradictory verdicts. The ensemble average said “grow your money.” The time average said “lose it.” Both were computed correctly. This lesson names the one property that decides which of them is telling the truth about your future, and hands you a test you can run before you bet a cent.

The two averages, in one breath

Keep both pictures in your head, because everything hangs on the difference.

• The ensemble average is a crowd at one instant: line up a million people, each plays the game once, and you average their results side by side. It answers “if everyone tried this simultaneously, what would the mean outcome be?”
• The time average is one person along their path: you play the same game over and over, and each round compounds on whatever the last round left you. It answers “if I keep doing this, what rate does my own wealth actually grow at?”

For some games these two numbers are identical. For others they are wildly, dangerously different. The name for “they’re identical” is ergodic. The name for “they diverge” is non-ergodic. That’s the whole distinction — the rest of this lesson makes it precise.

Ergodic, defined

Analogy. Picture a thoroughly shuffled deck that you re-shuffle after every draw. Draw one card a thousand times, and the fraction of hearts you see converges to exactly 25% — the same fraction you’d get by fanning the whole deck out and counting hearts in a single glance. Your long trip through time samples the whole space, so your personal average matches the snapshot of everything at once.

Origin. The word comes from statistical mechanics. In the 1870s the physicist Ludwig Boltzmann needed to relate the time-average behaviour of a single gas molecule careening around a box to the average over all possible states of the gas at one instant. He coined ergodic for the happy case where a single trajectory, given long enough, eventually explores the entire space of states — so that following one molecule through time gives you the same answer as photographing the whole gas at once. Time average equals ensemble average.

Precise definition. A process is ergodic when its time average equals its ensemble average. One player’s long-run experience converges to the crowd’s instantaneous mean; the snapshot of the many is a preview of the one.

Worked example. A casino booking one dollar per spin on a slightly favourable-to-the-house bet takes millions of tiny, independent wagers. No single spin can move the house’s fortune by much, and the results simply add up. Over a million spins the house’s realised return per bet converges to the theoretical edge — say 2.7% on European roulette — which is exactly the ensemble average across a million gamblers each betting once. The casino’s time average and the ensemble average coincide. That coincidence is precisely why the house can quote you an “expected value” and actually collect it.

Pitfall. Ergodicity is a property of the process, not a reward for having lots of data. A million observations of a non-ergodic process still won’t make its ensemble average describe your path. Sample size is not the cure.

When to use it

Reach for the ensemble average — trust expected value directly — only when the process genuinely resamples the same fixed spread each round and no single step can carry you somewhere you can’t return from. Small, additive, bounded, endlessly repeatable bets from a deep-pocketed position live here.

Non-ergodic, defined

Analogy. Now the deck is not reshuffled — every card you draw is removed, and one particular card (“you go bust”) ends the game the instant you hit it. Your path is no longer a fair sample of the full deck; it’s a single, order-dependent sequence that can terminate. Where you go next depends entirely on where you already are.

Origin. The idea was dragged into economics by the physicist Ole Peters of the London Mathematical Laboratory, building on much older insights from Daniel Bernoulli (1738) and John Kelly (1956). Peters’ framing is a single sharp question: is one person gambling 100 days in a row the same problem as 100 people each gambling for one day? For a multiplicative game — one where each round scales your wealth rather than adding to it — the answer is no. The 100 separate people include a few enormous winners that drag the crowd’s average up, but the one person compounding day after day gets multiplied down toward zero. Same game, opposite fates.

Precise definition. A process is non-ergodic when its time average diverges from its ensemble average. Your path is not a sample of the crowd’s spread; it is its own compounding sequence, and one bad enough step can end it — after which no later “average” can rescue you, because you’re no longer playing.

Worked example. Take the flagship coin: +50% on heads, −40% on tails, betting your whole balance each round. The ensemble average of a single round is $0.5 \times 1.5 + 0.5 \times 0.6 = 1.05$, a cheerful +5%. But the time average — the per-round growth rate you’d actually live — is $\sqrt{1.5 \times 0.6} = \sqrt{0.9} \approx 0.949$, a −5.1% bleed every round. The crowd’s mean climbs while the typical player’s wealth decays. That gap is non-ergodicity, and lesson 02 dissects this exact coin flip in full.

Pitfall. “But the expected value is positive!” is true and irrelevant. In a non-ergodic game the expected value is the crowd’s number, propped up by a vanishing sliver of astronomically lucky paths you almost certainly won’t be on.

When to use it

Assume non-ergodicity — and reason with the time average — whenever wealth compounds and a step can be ruinous. That covers most of real life, which is exactly why this distinction is worth a whole course.

The test: does the step multiply, or add?

Here is the single question that sorts almost any process. Does one step’s effect depend on where you currently are?

• If a step adds a fixed amount regardless of your current level — win $10, lose $8 — the effect is path-independent. These games tend to be ergodic: trust the ensemble mean.
• If a step multiplies your current level — win +50%, lose −40% — the effect is path-dependent, because the same percentage lands as a bigger or smaller dollar swing depending on where you stand. These games tend to be non-ergodic: live the time average. And any game that can zero you out is non-ergodic no matter what, because zero is a trap you never climb out of.

Shorthand: additive, bounded, non-ruinous, repeatable → usually ergodic. Multiplicative and/or capable of ruin → non-ergodic.

Fully worked mini-example. Compare two coins, each 50/50, each played on a $100 starting balance.

The additive coin pays +$10 on heads, −$8 on tails — a flat dollar amount either way. The ensemble average of one round is $0.5 \times (+10) + 0.5 \times (-8) = +1$ dollar. And because the dollars just accumulate — heads then tails leaves you at $100 + 10 - 8 = 102$, exactly as tails then heads would — the long-run time average is also +$1 per round. The two averages agree. Ergodic.

The multiplicative coin pays +50% on heads, −40% on tails. The ensemble average of one round is +5% (shown above). But heads then tails takes $100 to $150 to $90, and tails then heads takes $100 to $60 to $90 — either order leaves you below where you started. The time average is −5.1% per round. The two averages disagree, sharply. Non-ergodic.

Pitfall. Percentages are the tell. The instant a game is quoted in percent of your current wealth rather than fixed dollars, you should suspect multiplicative dynamics and check the time average before believing any “expected return.”

When to use it

Run the test on any recurring decision before you accept its advertised average: ask whether the outcome adds to your pile or scales it, and whether a bad enough run can take you out. That one question routes you to the correct average.

Sort each process by which average to trust

Run the additive-vs-multiplicative test on each of these and drop it in the right bucket.

The master error: assuming ergodicity by default

Here is the trap the entire course exists to defuse. Most statistics textbooks compute an expected value — the ensemble average — and quietly hand it to you as though it were your future. That silent assumption is fine for the ergodic processes those textbooks grew up on (measurement error, coin tosses for fixed stakes, sampling from a fixed population). It is catastrophically wrong for the non-ergodic processes that dominate real economic life.

Worked illustration. A fund advertises a “+5% expected return per period,” which is a perfectly correct ensemble average of the +50%/−40% coin above. An investor who compounds their whole stake at those odds does not earn +5%; they lose about 5% per period and trend toward zero, while the fund’s brochure was never technically lying — it just quoted the crowd’s number for a game you play alone.

Pitfall. The error is invisible precisely because it’s an omission. Nobody writes “we assume ergodicity here”; they simply write $E[X]$ and move on. Treating a non-ergodic process as ergodic — believing the ensemble average is your destiny — is the master error, and every later lesson is a variation on catching it.

When to use it

Whenever someone quotes you an expected value for something that compounds — a return, a growth rate, a survival prospect — pause and ask: is this process ergodic? If a step multiplies or can ruin you, quietly replace their ensemble number with the time average before you act on it.

Where this shows up: anything that compounds

The distinction earns its keep because non-ergodic dynamics are the default wherever a quantity compounds and a step can be ruinous:

• Wealth — you spend and reinvest from one evolving balance, and $0 is absorbing.
• A portfolio — returns multiply, leverage sharpens the multiplication, and a margin call is a ruin state.
• A firm — revenues compound but a single insolvency ends the story.
• A career — reputation and skill build on each other, yet one disqualifying blow-up can reset you to zero.
• Health — good and bad states feed forward, and death is the ultimate absorbing barrier.

In every one of these you are one player walking one path through time, not a crowd being averaged. So the number that governs you is the time average — and the ensemble mean you were handed is, at best, someone else’s future.