Lesson 1 handed you the distinction: a process is ergodic when the time average equals the ensemble average, and non-ergodic when they part ways. That was the map. This lesson is the territory — one coin, worked out to the last decimal, until the split between the two averages stops being a definition you can recite and becomes a thing you can feel. It is the emotional centre of the course for a simple reason: it is the friendliest-looking bet in the world, and it will quietly ruin nearly everyone who takes it.
Before you read — take a guess
Start with $100 and play the fair coin: heads multiplies your money by 1.5, tails by 0.6, and you always stake everything. After 100 rounds with exactly 50 heads and 50 tails, roughly how much money do you have?
The coin, restated exactly
Here is the whole gamble in one breath. The coin is fair — a genuine 50/50, nothing loaded. On heads, your wealth grows by 50%, i.e. it multiplies by 1.5. On tails, your wealth shrinks by 40%, i.e. it multiplies by 0.6. Odds honest, edge real, no trick.
Compute the ensemble expected value the way every statistics course teaches it — average the multiplier across the two equally likely outcomes:
A flat, delicious +5% per round. Expected value says: play forever, grow without bound. If you handed this to a casino they would beg to be on your side of the table. Hold onto that number — $1.05 on the dollar, every round, on average — because everything that follows is the story of how that honest, positive average leads almost everyone straight to ruin.
When to reach for this: any time a bet is framed as “positive expected value, so obviously take it,” pause and ask what kind of outcome you just averaged. Averaging a multiplier (a percentage move on your whole stake) is the tell that you may be in non-ergodic territory — and that the +5% might belong to a crowd you’ll never join.
Live the sequence, don’t average the crowd
The ensemble EV imagines a thousand copies of you playing once, side by side, and averages across them. But you don’t get a thousand copies. You get one run, and each round multiplies whatever the last round left you. So watch what multiplication actually does over a sequence.
Take one heads then one tails. Your wealth is now multiplied by 1.5 × 0.6 = 0.9 — you are down 10% after one win and one loss. And here is the quietly devastating part: order does not save you. Tails-then-heads gives 0.6 × 1.5 = 0.9 as well, because multiplication commutes. Every balanced pair of flips — and over a long fair run, roughly half your flips pair off like this — shaves a tenth off everything you own. Start at $100 and let the pairs stack up:
| Round | Sequence so far | Ensemble average says | Your compounding path |
|---|---|---|---|
| 0 | — | $100 | $100 |
| 2 | H, T | $110.25 | $90.00 |
| 10 | 5×H, 5×T | ≈ $163 | ≈ $59 |
| 40 | 20×H, 20×T | ≈ $704 | ≈ $12 |
| 100 | 50×H, 50×T | ≈ $13,150 | ≈ $0.59 |
Read that table across, then down, and let it sink in. The two columns start at the same $100 and then sprint in opposite directions. The ensemble average — 1.05 raised to the round number — climbs from $100 to over thirteen thousand dollars. Your actual path — 0.9 raised to the number of pairs — collapses from $100 to fifty-nine cents. Same coin. Same 100 rounds. One column is describing a crowd; the other is describing you.
You play the coin and get heads, then tails. A friend plays and gets tails, then heads. Both of you staked everything each round. Who has more money after those two flips?
The geometric mean: the growth rate you actually live
The ensemble EV (+5%) answers “what’s the average wealth across the crowd?” To find the rate you compound at, you need a different average — the geometric mean of the per-round multipliers, which is the honest per-round growth factor of a multiplicative process.
For our coin, over one heads-and-one-tails, the geometric mean multiplier is the square root of their product:
That is a multiplier below one: about −5.1% per round of time-average growth. Line the two averages up side by side and stare at the sign flip:
| Average | What it measures | Per-round result | Verdict |
|---|---|---|---|
| Ensemble (arithmetic) mean | Wealth averaged across the crowd | 1.05 → +5% | “Get rich” |
| Time (geometric) mean | Growth rate of one compounding path | 0.949 → ≈ −5.1% | “Go broke” |
Same coin, opposite signs. This is the whole non-ergodic drama in two rows. The arithmetic mean is the average of the outcomes; the geometric mean is the average of the growth, and for anything that compounds, growth is what you live in. Note too that the geometric mean is always ≤ the arithmetic mean (they’re equal only when every outcome is identical) — the gap between them is precisely the “volatility tax” that multiplicative bets levy on you.
For the +50%/−40% coin, the arithmetic (ensemble) mean multiplier is 1.05 and the geometric (time-average) multiplier is √0.9 ≈ 0.949. What does the gap between these two numbers represent?
Where the +5% hides
If the typical player decays at −5% a round, where does the +5% ensemble average actually live? It has to be somewhere — the arithmetic is airtight. The answer is the single most important picture in this course.
After many rounds, the crowd’s wealth doesn’t spread out evenly. Because wealth is built by multiplying random factors, its distribution is lognormal — and wildly right-skewed. The overwhelming majority of players cluster near zero (a run dominated by pairs, decaying at −5%/round), while a vanishing fraction of players who happened to flip mostly heads become astronomically rich. A player who got, say, 65 heads out of 100 doesn’t end up modestly ahead — they end up with a fortune large enough, and multiplied across enough near-copies, to single-handedly haul the arithmetic mean of the whole crowd up into five figures.
So the +5% is real, but it is unclaimable by the typical person. It is an average propped up by a handful of lottery-winner paths that almost no individual walks. The number that describes you is not the mean but the median — the middle player — and the median is decaying toward zero.
Mean ≫ median is the signature of a non-ergodic trap
When the arithmetic mean sits far above the median, you’re looking at a right-skewed, multiplicative process where a few explosive winners carry the average and the typical player is being ground down. The mean is honest — it just belongs to the crowd. The median is what you should plan your life around. Any time someone quotes an “average return” without a median beside it, ask where the skew is hiding.
After 100 rounds of the coin across thousands of players, the ARITHMETIC MEAN final wealth is about $13,000 but the MEDIAN final wealth is about 59 cents. Which reading is correct?
Order doesn’t matter — so you can’t outsmart it
Because multiplication commutes, the sequence of your flips is powerless to change your final wealth. This is worth pausing on, because it quietly kills every “just be clever about timing” escape plan.
Start with $100. Suppose over some stretch you get exactly one heads and one tails, in some order.
- Heads first: $100 × 1.5 = $150, then × 0.6 = $90.
- Tails first: $100 × 0.6 = $60, then × 1.5 = $90.
Identical. And it generalises: with any fixed bag of heads and tails, your final wealth is $100 × 1.5^(#heads) × 0.6^(#tails), and multiplication doesn’t care what order you apply the factors in. So there is no clever ordering, no “wait for a hot streak,” no timing trick that turns this coin profitable. The count of heads and tails is your entire destiny, and for a fair coin that count marches you toward 50/50 — straight into the −5% pairs.
The one thing that is under your control is not which flips you take or when, but how much of your wealth you stake on each flip. Stake everything and every pair costs you 10%. Stake a small fraction, and each flip barely moves your base — the −40% can no longer gut you, and the geometric mean of your actual per-round outcomes can climb back above 1.0. That single lever, sized correctly, is Kelly betting — the subject of lesson 04, and the only real way out of this coin.
Run it yourself: watch the crowd and the path split
Enough arithmetic — go feel it. Below is the same coin, made draggable, run across 400 parallel players at once. One line is the ensemble average (the mean wealth of the whole crowd); the other is the typical player (the median person, living the sequence over time). Do two experiments:
- Run the default coin first — heads ×1.5, tails ×0.6, fair, full stake. Watch the ensemble line rocket upward while the typical line collapses toward zero, and note the “wiped out” share: the fraction of players busted. This is the table above, rendered in motion.
- Then drag the bet fraction down toward 20% and run again. Same coin, same odds — but now the typical path bends from a slide into ruin into a genuine climb. You didn’t fix the coin; you fixed how much of yourself you staked on each flip.
Ergodicity engine
Run the +5% coin and watch it ruin the typical player
A multiplicative coin-flip, played two ways at once. The ENSEMBLE line is the average wealth across 400 parallel players; the TYPICAL line is the single median player living the sequence over time. Set the gamble, then run it. With the default coin the ensemble average rockets up while the typical player is quietly ground toward zero — the same bet, two opposite fates. Shrink the bet fraction and watch the typical path finally turn upward.
Set the gamble and press run. Watch the ensemble average and the typical player split apart.
Keep dragging the bet fraction back and forth across the point where the typical player flips from ruin to growth. That crossover isn’t a technicality — it’s the whole survival logic of the course, and you’re watching it live.
The pitfall: “the law of large numbers guarantees I win”
Here is the trap that catches even careful people. “Sure, any single round is random,” the reasoning goes, “but the law of large numbers says that over enough rounds, my results converge to the expected value — +5% a round. So if I just play long enough, I’m guaranteed to come out ahead.”
It sounds airtight. It is exactly backwards. The law of large numbers describes what happens to the average across many independent players as the number of players grows — the ensemble average converges to the EV. It says nothing about what happens to one player’s compounding wealth as the number of rounds grows. For a single path, more rounds don’t rescue you; they doom you, because they give the −5% geometric drag more time to grind. The LLN is a statement about the crowd getting more predictable, not about you getting richer. Confusing “the average of many players stabilises” with “my own path trends up” is the precise error that makes a +5% coin feel like free money.
The one-line correction
The law of large numbers stabilises the ensemble average as you add players. It does not pull a single time average toward the ensemble EV as you add rounds. In a non-ergodic process, playing longer doesn’t converge you to +5% — it converges you to zero.
When to use this model
You will meet the flagship coin wearing a thousand disguises. Learn to spot the shape rather than the story: a multiplicative bet announces itself whenever the moves are percentages of your whole stake — where this round’s outcome resets the base for next round, and a big enough loss can’t be recovered by a symmetric gain.
- Reach for the time average / geometric mean when: returns compound on your whole bankroll (all-in trades, leveraged positions, a business you can’t diversify away from, your one career), a single bad round can be irreversible, and the swings are large. Here the ensemble EV is quietly computing someone else’s future.
- The ordinary arithmetic average is fine when: the bet is a small, bounded side-stake whose outcome adds to (rather than multiplies) your total wealth, no single round can wipe you out, and you can genuinely play many independent copies. That additive-vs-multiplicative fork is exactly where lesson 03 goes next.
The reflex to build: before you accept any “+EV, obviously take it,” ask “is this a percentage of my whole stake, and can a loss end the game?” If yes, the geometric mean — not the arithmetic one — is the number that maps your future.
Recap
Why does the ensemble average of the +50%/−40% coin end up around $13,000 after 100 rounds while the typical player ends up around 59 cents?
Check your answer to continue.
What you now own
You’ve worked the flagship coin end to end. The ensemble average climbs to $13,000 because a vanishing fraction of lucky paths hauls the mean up; the typical (median) player decays to pocket change at the geometric-mean rate of about −5%/round; order can’t save you because multiplication commutes; and the law of large numbers describes the crowd, not your path. Mean ≫ median is the fingerprint you’ll now recognise everywhere. The only escape — betting less — is waiting in lesson 04.
Next up: lesson 03, Additive vs Multiplicative — the fork beneath everything you just saw. Why is averaging perfectly safe for a small, bounded side-bet but fatal for your whole bankroll? Because one adds to your wealth and the other multiplies it — and that single difference decides whether the two averages agree or tear apart.