THE MACHINERY OF ERGODICITY
A Complete Guide to Time and Averages
How the Distinction That Governs Everything Actually Works
What follows is not advice.
It is not a risk management framework. Not a betting system. Not another essay about expected value dressed in physics clothing.
It is mechanism.
The actual machinery underneath the most consequential error in modern thought. The assumption that what happens on average across many people is what will happen to one person across time. The conflation that broke economics, distorted risk, and made entire populations misunderstand the games they are playing.
Most people live inside this error without seeing it. They compute expected values. They diversify portfolios. They evaluate bets by asking “what’s the average outcome?” They trust that the aggregate tells them something about their individual trajectory.
It does not.
This document is that seeing.
Nothing more.
What you do with it is your business.
PART ONE: THE HYPOTHESIS
The Question Boltzmann Asked
In 1871, Ludwig Boltzmann was trying to connect two worlds.
The world of individual molecules. Billions of them. Bouncing, colliding, exchanging energy in patterns too complex to track.
And the world of thermodynamics. Temperature. Pressure. Entropy. Macroscopic quantities that behave with stunning regularity.
The bridge between them required an assumption.
If you watch one molecule long enough, will its behavior over time look the same as the behavior of all molecules at a single instant?
Will the time average equal the ensemble average?
Boltzmann assumed yes.
He called this the ergodic hypothesis. From the Greek ergon (energy) and hodos (path). The energy path. The trajectory through every possible state.
The Two Averages
This is the distinction that governs everything.
THE TWO AVERAGES
┌────────────────────────────────────────────────────────┐
│ │
│ ENSEMBLE AVERAGE │
│ │
│ Look at many systems at one moment in time. │
│ Average across all of them. │
│ │
│ 1000 people flip a coin once. │
│ Average their outcomes. │
│ │
│ This is a snapshot across space. │
│ │
└────────────────────────────────────────────────────────┘
│
│ Are these equal?
▼
┌────────────────────────────────────────────────────────┐
│ │
│ TIME AVERAGE │
│ │
│ Watch one system across many moments. │
│ Average its trajectory. │
│ │
│ 1 person flips a coin 1000 times. │
│ Average their outcomes. │
│ │
│ This is a trajectory through time. │
│ │
└────────────────────────────────────────────────────────┘
When the two averages are equal, the system is ergodic.
When they are not, the system is non-ergodic.
Boltzmann assumed they were equal for gas molecules. For that specific system, with its specific dynamics, he was largely right. Molecules in a sealed box eventually explore every configuration. Given enough time, one molecule’s journey through phase space samples the same distribution as a snapshot of all molecules at once.
In 1931, George David Birkhoff proved this rigorously. The ergodic theorem. Under specific mathematical conditions, the time average converges to the ensemble average almost everywhere.
The theorem is correct.
The problem is that almost nothing in human life satisfies those conditions.
What Makes a System Ergodic
Three conditions must hold.
The system must be able to visit every possible state. No region of the state space is permanently walled off. No configuration is unreachable.
The system must have enough time. Enough for the trajectory to sample every region in proportion to its size. For gas molecules in a box, this happens quickly. For a macroscopic system returning to its exact initial state, the Poincare recurrence time exceeds the age of the universe.
There must be no absorbing barriers. No states that, once entered, cannot be exited.
CONDITIONS FOR ERGODICITY
┌────────────────────────────────────────────────────────┐
│ │
│ 1. FULL ACCESSIBILITY │
│ Every state reachable from every other state │
│ │
│ 2. SUFFICIENT TIME │
│ Trajectory samples all regions proportionally │
│ │
│ 3. NO ABSORBING BARRIERS │
│ No states you enter and never leave │
│ │
└────────────────────────────────────────────────────────┘
Remove any one condition:
Time average ≠ Ensemble average
The system is non-ergodic.
Gas molecules in thermal equilibrium satisfy all three. They bounce freely. They have effectively infinite time relative to their dynamics. And there is no molecular state equivalent to death.
Human systems satisfy none of them.
PART TWO: THE COIN TOSS
The Demonstration
This is the example that exposes the error.
A fair coin. Heads, your wealth increases by 50%. Tails, your wealth decreases by 40%.
Compute the expected value. The ensemble average.
0.5 x (+50%) + 0.5 x (-40%) = +5% per round.
Positive. A good bet. Take it every time. This is what every economics textbook says.
Now compute what actually happens to one person playing repeatedly.
Start with $100. One heads, one tails. In any order.
Heads then tails: $100 becomes $150 becomes $90.
Tails then heads: $100 becomes $60 becomes $90.
Either way: $90. A loss of 10%.
The expected value said +5% per round. The actual trajectory says -5.1% per round.
THE DIVERGENCE
Expected value (ensemble): +5.0% per round
Time-average growth rate: -5.1% per round
┌────────────────────────────────────────────────────────┐
│ │
│ After 100 rounds: │
│ │
│ Mean wealth: $16,697 (pulled up by rare │
│ extreme winners) │
│ │
│ Median wealth: $0.52 (what actually happens │
│ to most people) │
│ │
│ Below start: 86% (lost money despite │
│ "positive" expected │
│ value) │
│ │
│ Top 1 person: ~70% of (nearly all wealth │
│ total concentrated in one │
│ wealth trajectory) │
│ │
└────────────────────────────────────────────────────────┘
After 1000 rounds, the mean wealth is $24. The median is effectively zero. Everyone goes broke. Except for a vanishingly rare trajectory that captures essentially all the wealth in the system.
The expected value was positive. Every individual went to ruin.
This is not a paradox. This is the mathematics of non-ergodicity.
Why the Averages Diverge
The mechanism is multiplication.
When your wealth changes by a percentage, the process is multiplicative. Each outcome multiplies the previous total. The sequence of returns compounds.
For additive processes, order does not matter. Gaining $50 then losing $40 gives the same result as losing $40 then gaining $50. The time average equals the ensemble average. The system is ergodic.
For multiplicative processes, order still does not matter in terms of final outcome. But the compounding does. Gaining 50% then losing 40% is 1.5 x 0.6 = 0.90. You lost 10% regardless of sequence.
The arithmetic mean of 1.5 and 0.6 is 1.05. Growth.
The geometric mean of 1.5 and 0.6 is 0.9487. Decline.
ADDITIVE VS MULTIPLICATIVE
ADDITIVE PROCESS:
┌────────────────────────────────────────────────────────┐
│ │
│ x(t+1) = x(t) + r │
│ │
│ Changes are independent of current state. │
│ +$50 is +$50 whether you have $100 or $10,000. │
│ │
│ Arithmetic mean governs. │
│ Time average = Ensemble average. │
│ ERGODIC. │
│ │
└────────────────────────────────────────────────────────┘
MULTIPLICATIVE PROCESS:
┌────────────────────────────────────────────────────────┐
│ │
│ x(t+1) = x(t) * r │
│ │
│ Changes scale with current state. │
│ +50% of $100 is $50. +50% of $10,000 is $5,000. │
│ │
│ Geometric mean governs. │
│ Time average < Ensemble average. │
│ NON-ERGODIC. │
│ │
└────────────────────────────────────────────────────────┘
The ensemble average is the arithmetic mean. It asks: across all possible parallel universes, what is the average outcome?
The time average is the geometric mean. It asks: for one person living through the sequence, what actually happens?
For multiplicative dynamics, the geometric mean is always less than or equal to the arithmetic mean. Always. This is a mathematical identity, not an approximation. The gap between them grows with volatility.
The Volatility Tax
In continuous mathematics, the relationship is precise.
For geometric Brownian motion, the standard model of multiplicative growth:
Ensemble-average growth rate: mu
Time-average growth rate: mu - sigma^2/2
The term sigma^2/2 is the volatility tax. It is always negative when volatility is nonzero. It represents the difference between what the ensemble promises and what time delivers.
THE VOLATILITY TAX
Growth
Rate
│
│ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ Ensemble average (mu)
│
│ ████████████████████████ Time average
│ (mu - sigma^2/2)
│
│ ◄──────────────►
│ sigma^2/2
│ (the volatility tax)
│
│ As volatility increases, the gap widens.
│ The ensemble looks better and better.
│ The individual does worse and worse.
│
└──────────────────────────────────────────────►
Volatility (sigma)
This term emerges from Ito’s lemma. It is not a correction factor someone invented. It falls out of the calculus of stochastic processes. It is the mathematical signature of non-ergodicity in any system with multiplicative dynamics and volatility.
A process can have positive expected value and negative time-average growth. The expected value says you should play. The time average says you will go broke. Both statements are mathematically correct. They are answering different questions.
The expected value answers: what happens to the average of many parallel copies?
The time average answers: what happens to you?
PART THREE: THE ABSORBING BARRIER
The State You Cannot Leave
Death is an absorbing barrier.
Bankruptcy is an absorbing barrier.
Extinction is an absorbing barrier.
An absorbing barrier is a state that, once entered, traps the system permanently. There is no recovery. No next round. No future trajectory.
This is why ergodicity fails for every living system.
THE ABSORBING BARRIER
Wealth
│
│ ╱╲ ╱╲ ╱╲
│ ╱ ╲ ╱ ╲ ╱╲ ╱ ╲
│ ╱ ╲╱ ╲╱ ╲╱ ╲
│ ╱ ╲
│╱ ╲
═════╪══════════════════════════╪═══════════════
$0 │ │
│ ABSORBING BARRIER │
│ │
│ No visit 29. │
│ No recovery. │
│ No next round. │
│ │
└──────────────────────────────────────►
Time
For a system to be ergodic, it must be able to visit every state and return from every state. An absorbing barrier removes this possibility. Once wealth reaches zero, no sequence of percentage gains can restore it. 50% of zero is zero. 500% of zero is zero. The game is over.
This is not a technical detail. This is the central fact.
The expected value calculation assumes you can play forever. That losses are temporary setbacks in an infinite sequence. That the law of large numbers will eventually deliver the arithmetic mean.
But if you can go broke, and going broke ends the game, then the law of large numbers does not apply to you. It applies to the ensemble. The collective. The population of parallel players.
You are not the ensemble. You are one trajectory. And your trajectory has a floor it cannot pass through.
The Casino Thought Experiment
One hundred people walk into a casino.
Each bets once. The house has a 1% edge. On average, across the group, the casino wins 1% of the total money wagered.
Ensemble thinking works here. Each person’s outcome is independent. The casino can compute its expected take with high confidence.
Now. One person walks into a casino. They bet 100 times. The house has the same 1% edge.
Is this the same situation?
It is not.
If bet 28 wipes them out, there is no bet 29. There is no bet 43, 67, or 99. The trajectory terminates. All future potential outcomes are annihilated by a single ruin event.
ENSEMBLE VS TRAJECTORY
┌─────────────────────────────┐ ┌─────────────────────────────┐
│ │ │ │
│ 100 PEOPLE │ │ 1 PERSON │
│ 1 BET EACH │ │ 100 BETS │
│ │ │ │
│ Each outcome independent. │ │ Each outcome dependent │
│ One person's ruin does │ │ on survival from the │
│ not affect others. │ │ previous bet. │
│ │ │ │
│ Ensemble average applies. │ │ Time average applies. │
│ Expected value is valid. │ │ Expected value misleads. │
│ │ │ │
│ Risk is statistical. │ │ Risk is existential. │
│ │ │ │
└─────────────────────────────┘ └─────────────────────────────┘
The confusion between these two situations is not a minor error. It is the foundational mistake underlying most of modern decision theory.
PART FOUR: THE ECONOMIC ERROR
Expected Value Is Not What You Get
In 1738, Daniel Bernoulli confronted the St. Petersburg paradox. A game with infinite expected value that no rational person would pay much to play. A coin is flipped until it lands heads. The payout is 2^n, where n is the number of flips. Expected value: infinite.
Bernoulli’s solution was to introduce utility. People don’t maximize money, he said. They maximize the logarithm of money. Diminishing marginal utility. This produces a finite valuation for the game.
For 275 years, economics built on this foundation. Expected utility theory. Rational agents maximizing the expected value of some utility function. The specific shape of the function was debated endlessly. But the framework was assumed.
In 2011, Ole Peters proposed something different.
The St. Petersburg paradox does not require utility functions at all.
It requires computing the right average.
The expected value (ensemble average) of the game is infinite. But the time-average growth rate is not. If you play the St. Petersburg game repeatedly, reinvesting your wealth, your time-average growth rate is finite and calculable. And it matches, exactly, Bernoulli’s logarithmic solution.
Same answer. Completely different reasoning.
Bernoulli said: “People have diminishing sensitivity to money.” A psychological claim.
Peters said: “The process is multiplicative and non-ergodic. The time average diverges from the ensemble average.” A mathematical claim.
TWO EXPLANATIONS, SAME RESULT
┌────────────────────────────────────────────────────────┐
│ │
│ BERNOULLI (1738) │
│ │
│ "People maximize expected utility, not expected │
│ money. The utility function is logarithmic." │
│ │
│ Framework: Psychology. Preference. Subjective │
│ value. Requires a utility function. │
│ │
│ Question answered: What do people prefer? │
│ │
└────────────────────────────────────────────────────────┘
│
│ Same numerical result
▼
┌────────────────────────────────────────────────────────┐
│ │
│ PETERS (2011) │
│ │
│ "The process is multiplicative. Compute the time │
│ average, not the ensemble average. The logarithm │
│ is the ergodicity transformation." │
│ │
│ Framework: Dynamics. Mathematics. No utility │
│ function needed. │
│ │
│ Question answered: What actually happens over time? │
│ │
└────────────────────────────────────────────────────────┘
The distinction matters because one framework requires calibrating a subjective psychological function for every decision context. The other requires identifying the dynamics of the process.
Risk aversion. Loss aversion. Probability weighting. The entire apparatus of behavioral economics that catalogs human “irrationality.”
Peters’ claim is that these are not irrational at all. They are dynamically optimal. People who appear to be “irrationally” risk-averse are correctly computing time-average growth rates for multiplicative processes. They are solving the right problem. Economists were solving the wrong one.
PART FIVE: THE KELLY CRITERION
Optimal Sizing
In 1956, John Kelly at Bell Labs derived a formula for optimal bet sizing. The question: given a favorable bet, how much of your wealth should you wager?
The answer comes from maximizing the time-average growth rate. Not the expected value.
For a simple bet with win probability p, loss probability q = 1-p, and odds b:
f* = (bp - q) / b
This is the Kelly fraction. The percentage of your wealth to bet.
For the 50/50 coin toss with +50%/-40%:
f* = (1.25 x 0.5 - 0.5) / 1.25 = 0.10
Optimal bet: 10% of your wealth per round.
Not all of it. Not half. Ten percent.
GROWTH RATE VS BET SIZE
Time-average
growth rate
│
│ ┌──────┐
│ / \
MAX │ / \ ← Kelly fraction (f*)
│ / \
│ / \
│ / \
0% │────/────────────────────\──────────────
│ / \
│ / \
NEG │/ \
│ \
└──────────────────────────────────────────►
0% f* 2f* 100%
Fraction of wealth bet
Below Kelly: Suboptimal but safe.
At Kelly: Maximum growth.
Above Kelly: Growth declines.
At 2x Kelly: Zero growth. No better than not playing.
Above 2x: Negative growth. Guaranteed ruin.
Leo Breiman proved in 1961 that the Kelly criterion is equivalent to maximizing the geometric growth rate of wealth. This is identical to maximizing the time-average growth rate of a multiplicative process.
The Kelly criterion is ergodicity economics in practice. It is the formula that corrects for non-ergodicity by computing what actually happens to one trajectory over time, rather than what happens on average across many parallel trajectories.
Bet more than Kelly and you will go broke. Not probably. Certainly. Given enough time.
The expected value of over-betting is positive. The time-average growth rate is negative.
Same error. Same mechanism. Same divergence between what the ensemble promises and what time delivers.
PART SIX: BROKEN ERGODICITY IN PHYSICS
When Phase Space Fragments
Return to physics. Where the concept was born.
A system’s phase space is the set of all possible configurations. For gas molecules in a box, phase space is the set of all possible positions and momenta for every molecule. The ergodic hypothesis says the system will eventually visit every region of this space.
Sometimes it does not.
In 1982, R.G. Palmer formalized what he called broken ergodicity. The phenomenon where a system’s phase space fragments into regions that are mutually inaccessible on any practical timescale. The system gets trapped. It cannot explore.
BROKEN ERGODICITY
ERGODIC SYSTEM:
┌────────────────────────────────────────────────────────┐
│ │
│ ╱╲ ╱╲ ╱╲ ╱╲ ╱╲ ╱╲ │
│ ╱ ╲ ╱ ╲╱ ╲ ╱ ╲╱ ╲ ╱ ╲ │
│ ╱ ╲╱ ╲ ╱ ╲╱ ╲ │
│ ╱ ╲╱ ╲ │
│ ╱ Trajectory visits all regions ╲ │
│ │
└────────────────────────────────────────────────────────┘
NON-ERGODIC SYSTEM:
┌──────────────────────┐ ║ ┌──────────────────────────┐
│ │ ║ │ │
│ ╱╲ ╱╲ │ ║ │ │
│ ╱ ╲╱ ╲ │ ║ │ Inaccessible │
│ ╱ ╲ │ ║ │ region. │
│ ╱ Trapped ╲ │ ║ │ │
│ here │ ║ │ System cannot │
│ │ ║ │ reach this. │
│ │ ║ │ │
└──────────────────────┘ ║ └──────────────────────────┘
║
BARRIER
The canonical example is a spin glass. A disordered magnetic material where competing interactions create frustration. The energy landscape has many local minima separated by enormous barriers. The system falls into one valley and cannot climb out. It never reaches equilibrium. It never explores. It is stuck.
Giorgio Parisi received the 2021 Nobel Prize in Physics for describing the mathematical structure of this trapping. His replica symmetry breaking solution shows that the phase space of a spin glass shatters into an ultrametric hierarchy of trapped states. Ergodicity does not just weaken. It breaks.
Glass transitions exhibit the same phenomenon. Cool a liquid fast enough and it cannot find its crystalline ground state. It gets trapped in an amorphous configuration. Not equilibrium. Not ergodic. Just stuck. On timescales that exceed the patience of any experimentalist.
The Ergodic Hierarchy
Ergodicity is not binary. There is a spectrum.
THE ERGODIC HIERARCHY
┌────────────────────────────────────────────────────────┐
│ │
│ LEVEL 1: ERGODIC │
│ Time average = ensemble average. │
│ Weakest mixing. Slowest decorrelation. │
│ │
└────────────────────────────────────────────────────────┘
│ stronger ▼
┌────────────────────────────────────────────────────────┐
│ │
│ LEVEL 2: WEAK MIXING │
│ Correlations decay in the mean. │
│ Past influences future on average but not │
│ at every step. │
│ │
└────────────────────────────────────────────────────────┘
│ stronger ▼
┌────────────────────────────────────────────────────────┐
│ │
│ LEVEL 3: STRONG MIXING │
│ Correlations decay absolutely. │
│ Any two events become independent as time │
│ separation grows. The cocktail analogy. │
│ Past becomes irrelevant. │
│ │
└────────────────────────────────────────────────────────┘
│ stronger ▼
┌────────────────────────────────────────────────────────┐
│ │
│ LEVEL 4: KOLMOGOROV (K-SYSTEM) │
│ Positive Kolmogorov-Sinai entropy. │
│ Information production at every timescale. │
│ Entire coarse-grained past becomes irrelevant. │
│ Chaos. │
│ │
└────────────────────────────────────────────────────────┘
│ stronger ▼
┌────────────────────────────────────────────────────────┐
│ │
│ LEVEL 5: BERNOULLI │
│ Completely unpredictable. │
│ Equivalent to independent coin tosses. │
│ No information from the past helps predict │
│ the future. Maximum randomness. │
│ │
└────────────────────────────────────────────────────────┘
Each level is strictly stronger than the one above. Bernoulli implies Kolmogorov implies strong mixing implies weak mixing implies ergodic. But not the reverse.
Most real systems do not even reach level one.
PART SEVEN: ENTROPY AND INFORMATION
The Tombstone Equation
On Boltzmann’s grave in Vienna, an equation is carved.
S = k log W
Entropy equals the Boltzmann constant times the logarithm of the number of microstates.
This equation connects two worlds. The thermodynamic world of entropy, temperature, and the second law. And the statistical world of probability, counting, and information.
The connection runs through ergodicity.
If a system is ergodic, the time it spends in a given macrostate is proportional to the number of microstates that realize it. The most probable macrostate is the one with the most microstates. This is the equilibrium state. The state of maximum entropy.
The second law of thermodynamics says entropy increases. Boltzmann showed why. Not because of some mystical force driving toward disorder. Because there are overwhelmingly more disordered microstates than ordered ones. An ergodic system exploring its phase space will spend almost all its time in high-entropy states simply because there are more of them.
ENTROPY AND PHASE SPACE
Microstates
(W)
│
│████████████████████████████████ ← High entropy
│████████████████████████████████ (disordered,
│████████████████████████████████ many microstates)
│
│████████████████ ← Medium entropy
│████████████████
│
│████ ← Low entropy
│████ (ordered,
│████ few microstates)
│
└──────────────────────────────────────────────────
Macrostates
An ergodic system visits each microstate equally.
It spends more time in high-entropy macrostates
because there are more microstates there.
Entropy increase is not a force.
It is statistics.
But if the system is not ergodic, this reasoning breaks. A non-ergodic system does not visit all microstates. It gets trapped. The connection between time spent in a state and the number of microstates realizing that state dissolves. The second law still holds on average, across the ensemble. But for any individual trajectory, the system may remain in a low-entropy state indefinitely.
The Kolmogorov-Sinai entropy measures something different. The rate of information production by the dynamics. How quickly the system generates unpredictability. K-systems (level 4 in the ergodic hierarchy) have positive KS entropy. They produce information. They mix. They forget their past. This is the information-theoretic signature of strong ergodicity.
Irreversibility
The Poincare recurrence theorem says that any bounded, measure-preserving system will eventually return arbitrarily close to its initial state.
This creates a tension with the second law.
If the system will eventually return, how can entropy always increase?
The resolution is time. For a macroscopic system, the recurrence time exceeds 10^(10^23) years. The universe is approximately 1.4 x 10^10 years old. Recurrence is mathematically certain and physically irrelevant. The system is technically ergodic. Practically, it will never return.
This is the gap between mathematical ergodicity and physical relevance. A system can satisfy the ergodic hypothesis in the infinite-time limit while behaving as completely non-ergodic on every timescale that matters.
PART EIGHT: WEALTH AND COOPERATION
How Inequality Emerges
Take a population of identical agents. Same starting wealth. Same skill. Same effort. Subject them to multiplicative dynamics with random shocks.
No one cheats. No one is smarter. No one works harder.
Wealth concentrates anyway.
This is the mathematical consequence of non-ergodicity in multiplicative systems. Under geometric Brownian motion, the distribution of wealth across agents widens without bound. The mean grows. The median stagnates or declines. A vanishing fraction of agents captures a growing fraction of total wealth.
WEALTH DISTRIBUTION UNDER MULTIPLICATIVE DYNAMICS
Time = 0
┌────────────────────────────────────────────────────────┐
│ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ │
│ Equal starting wealth │
└────────────────────────────────────────────────────────┘
Time = 50
┌────────────────────────────────────────────────────────┐
│ █ █ █ █ █ █ █ █ █ █ █ ██ ██ ██ ███ ████ █████████ │
│ Most decline. A few grow. │
└────────────────────────────────────────────────────────┘
Time = 200
┌────────────────────────────────────────────────────────┐
│ . . . . . . . . . . . . . . . . . . █████████████████│
│ Nearly all at zero. One has everything. │
└────────────────────────────────────────────────────────┘
No differences in ability.
No exploitation.
Pure mathematics of multiplicative randomness.
Empirical analysis of US wealth data by Berman and Peters found that the dynamics exhibit negative reallocation. Wealth flows from poorer to richer on average. Under such conditions, no stationary distribution exists. Inequality does not converge to some natural level. It grows without limit.
This is not a claim about fairness or policy. It is a mathematical result. Multiplicative dynamics with no reallocation or with negative reallocation produce unbounded concentration. The system never reaches equilibrium because there is no equilibrium to reach.
Why Cooperation Emerges
Here is the surprising result.
Peters and Adamou showed that wealth sharing between agents facing multiplicative risk is a positive-sum game. Not zero-sum. Positive-sum.
Two agents independently facing the +50%/-40% coin toss both decline at -5.1% per round. If they pool their wealth and share outcomes equally, their individual time-average growth rates increase. Both benefit. Neither sacrifices.
This is not altruism. It is not reciprocity. It is not kin selection. It is not iterated game theory.
It is mathematics.
COOPERATION FROM DYNAMICS
INDEPENDENT:
┌──────────────────────────┐ ┌──────────────────────────┐
│ │ │ │
│ Agent A │ │ Agent B │
│ Growth: -5.1%/round │ │ Growth: -5.1%/round │
│ Fate: ruin │ │ Fate: ruin │
│ │ │ │
└──────────────────────────┘ └──────────────────────────┘
POOLED:
┌──────────────────────────────────────────────────────────┐
│ │
│ Agents A + B share outcomes │
│ Growth: > -5.1%/round (closer to 0 or positive) │
│ Volatility reduced by diversification. │
│ Volatility tax reduced. │
│ Both agents better off. │
│ │
└──────────────────────────────────────────────────────────┘
Mechanism: Pooling reduces sigma.
Reduced sigma reduces sigma^2/2.
Reduced volatility tax increases time-average growth.
Cooperation is not a moral choice.
It is a volatility reduction strategy.
The mechanism is the volatility tax. sigma^2/2. By pooling, agents reduce the effective volatility of their individual wealth trajectories. Lower volatility means a smaller gap between the ensemble average and the time average. The time-average growth rate increases for everyone.
Insurance works the same way. Not because of risk aversion. Not because of diminishing marginal utility. Because sharing reduces the volatility tax on multiplicative dynamics. Peters and Adamou titled their paper plainly: “Insurance makes wealth grow faster.”
Cooperation is not a puzzle to be explained by invoking altruism or reciprocity. In a non-ergodic world with multiplicative dynamics, cooperation is the dynamically rational response. It emerges from the mathematics without any psychological assumption at all.
PART NINE: THE CONSTRAINTS
When Ergodicity Holds
Not everything is non-ergodic.
Gas molecules in thermal equilibrium are effectively ergodic. The dynamics are fast relative to observation timescales. The phase space is connected. There are no absorbing barriers.
Short-term additive processes can be ergodic. If each outcome adds a fixed amount (not a percentage) and there is no ruin threshold, the time average converges to the ensemble average.
Highly liquid markets with no leverage can approximate ergodicity for specific quantities over specific timeframes. If you cannot go broke, if your position is small relative to your wealth, if the dynamics are approximately additive over the period in question, the ensemble average is a reasonable guide.
THE ERGODICITY SPECTRUM
◄───────────────────────────────────────────────────────►
STRONGLY STRONGLY
ERGODIC NON-ERGODIC
Gas molecules Short-term Leveraged Biological
in equilibrium additive investing evolution
bets
Shuffled cards Small bets Career Single
relative trajectories human life
Coin tosses to wealth
(additive Startup Extinction
payoff) founding events
│ │
│ Time average Time average │
│ ≈ ensemble ≠ ensemble │
│ average average │
│ │
The question is never “is this system ergodic?” in the abstract.
The question is: “over the timescale and for the quantity I care about, does the time average converge to the ensemble average?”
The Timescale Problem
A system can be theoretically ergodic but practically non-ergodic.
The Poincare recurrence time for a macroscopic system is longer than the age of the universe. In principle, the system will return. In practice, it never will.
A career lasts 40 years. A human life lasts 80. A species lasts a few million. The universe is 14 billion years old.
If the time required for ergodic convergence exceeds the observation window, the system is functionally non-ergodic. The mathematics says the average will converge eventually. Eventually is longer than you have.
TIMESCALE AND EFFECTIVE ERGODICITY
Convergence
Time
│
│████████████████████████████████████████ ← Poincare
│ recurrence
│ (~10^10^23 years)
│
│████████████████████ ← Wealth distribution
│ convergence
│ (if it converges at all)
│
│██████████ ← Market return convergence
│ (~decades)
│
│████ ← Molecular equilibration
│ (~nanoseconds)
│
└──────────────────────────────────────────────────
│
│ Human Universe
│ lifetime age
│ (80 yrs) (14B yrs)
If the convergence time exceeds your lifetime, the ensemble average is not your future. It is a description of a population you happen to be a member of. A population whose average experience you will almost certainly not have.
PART TEN: THE TWO MODES
Ergodic and Non-Ergodic Thinking
Every decision framework implicitly assumes one or the other.
════════════════════════════════════════════════════════════
MODE A: ERGODIC THINKING (ENSEMBLE)
Assumption: What happens on average is what will happen
to me, given enough time.
Calculation: Expected value. Arithmetic mean.
Probability-weighted outcomes.
Valid when:
• Process is additive
• No absorbing barriers
• Timescale is short relative to convergence
• Outcomes are independent of current state
Failure mode:
Treats every bet as if you have infinite lives.
════════════════════════════════════════════════════════════
MODE B: NON-ERGODIC THINKING (TRAJECTORY)
Assumption: What happens on average across many people
is not what will happen to me across time.
Calculation: Time-average growth rate. Geometric mean.
Kelly criterion.
Valid when:
• Process is multiplicative
• Absorbing barriers exist (ruin, death)
• Timescale is long relative to convergence
• Outcomes scale with current state
Failure mode:
Treats every bet as if it's the only one that matters.
════════════════════════════════════════════════════════════
The error is not in either mode. Both are correct within their domain.
The error is applying ensemble thinking to non-ergodic situations.
Computing expected value for a multiplicative process. Asking “what’s the average outcome?” when the average is driven by extreme outliers that one trajectory will never experience. Treating the population statistic as a personal prediction.
This error is embedded in standard economics, standard finance, standard decision theory, and most of the frameworks people use to evaluate risk.
The correction is not to avoid risk. It is to compute the right quantity. The time-average growth rate. The geometric mean. The quantity that describes what actually happens to one system evolving through one sequence of events in one irreversible lifetime.
PART ELEVEN: THE COMPLETE PICTURE
The Unified Framework
Everything connects.
THE COMPLETE ERGODICITY FRAMEWORK
┌──────────────────────────────────────────────────────────┐
│ │
│ THE CORE QUESTION │
│ │
│ Does the time average equal the ensemble average? │
│ │
└──────────────────────────────────────────────────────────┘
│
│
┌───────────────┼───────────────┐
│ │ │
▼ ▼ ▼
┌──────────────────┐ ┌─────────────┐ ┌──────────────────┐
│ │ │ │ │ │
│ DYNAMICS │ │ BARRIERS │ │ TIMESCALE │
│ │ │ │ │ │
│ Additive or │ │ Absorbing │ │ Convergence │
│ multiplicative? │ │ states? │ │ vs observation │
│ │ │ Ruin? │ │ window? │
│ This determines │ │ Death? │ │ │
│ which average │ │ │ │ Infinite time │
│ governs. │ │ This breaks │ │ fixes many │
│ │ │ ergodicity │ │ things. You │
│ │ │ absolutely. │ │ do not have │
│ │ │ │ │ infinite time. │
│ │ │ │ │ │
└──────────────────┘ └─────────────┘ └──────────────────┘
│ │ │
└───────────────┼───────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ │
│ IMPLICATIONS │
│ │
│ Wealth inequality is mathematical, not moral. │
│ Cooperation is dynamical, not psychological. │
│ Risk aversion is rational, not biased. │
│ Survival is primary, not secondary. │
│ The ensemble is not your future. │
│ │
└──────────────────────────────────────────────────────────┘
Ergodicity is not a property of outcomes. It is a property of processes.
The same bet can be ergodic or non-ergodic depending on how it compounds. The same game can be safe or lethal depending on whether ruin exists. The same system can be ergodic over nanoseconds and non-ergodic over centuries.
The question is never “is this risky?” in some absolute sense.
The question is: “for this process, with these dynamics, on this timescale, for this one irreversible trajectory, does the ensemble average describe what will actually happen?”
Usually it does not.
The Deepest Implication
You are not an ensemble.
You are one trajectory through one life. Unrepeatable. Irreversible. Bounded by death.
The average outcome across all possible versions of your life is mathematically interesting and personally meaningless. You will not experience the average. You will experience the one path you actually walk.
The expected value of a career choice across all possible people is not what your career will look like. The average return of an investment strategy across all possible market histories is not what your returns will be. The mean outcome of a health decision across a population is not what will happen to your body.
Every time someone says “on average” and uses it to make a decision about a single, irreversible, non-repeatable trajectory, they are importing an ergodic assumption into a non-ergodic world.
The machinery does not care whether you understand it. Multiplicative dynamics compound. Absorbing barriers absorb. The volatility tax extracts. The ensemble average diverges from the time average exactly as the mathematics predicts.
Understanding changes nothing about the machinery.
It changes which average you compute.
And that changes everything about the trajectory.
CITATIONS
Foundational Mathematics
Birkhoff’s Ergodic Theorem
Birkhoff, G.D. (1931). “Proof of the Ergodic Theorem.” Proceedings of the National Academy of Sciences, 17(12), 656-660. https://www.pnas.org/doi/10.1073/pnas.17.2.656
Moore, C.C. (2015). “Ergodic theorem, ergodic theory, and statistical mechanics.” PNAS, 112(7), 1907-1911. https://pmc.ncbi.nlm.nih.gov/articles/PMC4343160/
The Ergodic Hierarchy
“The Ergodic Hierarchy.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/ergodic-hierarchy/
Statistical Mechanics and Thermodynamics
Boltzmann’s Ergodic Hypothesis
“Boltzmann’s Work in Statistical Physics.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/statphys-Boltzmann/
Lee, M.H. (2006). “Why does Boltzmann’s ergodic hypothesis work and when does it fail.” Physica A, 365(1), 150-154. https://arxiv.org/abs/cond-mat/0510509
Broken Ergodicity
Palmer, R.G. (1982). “Broken Ergodicity.” Advances in Physics, 31(6), 669-735. https://www.tandfonline.com/doi/abs/10.1080/00018738200101438
Ergodicity Economics
Core Papers
Peters, O. (2019). “The ergodicity problem in economics.” Nature Physics, 15, 1216-1221. https://www.nature.com/articles/s41567-019-0732-0
Peters, O. & Gell-Mann, M. (2016). “Evaluating gambles using dynamics.” Chaos, 26, 023103. https://pubs.aip.org/aip/cha/article/26/2/023103/134886/Evaluating-gambles-using-dynamics
Peters, O. (2011). “The time resolution of the St Petersburg paradox.” Philosophical Transactions of the Royal Society A, 369, 4913-4931. https://royalsocietypublishing.org/doi/10.1098/rsta.2011.0065
Peters, O. (2011). “Optimal leverage from non-ergodicity.” Quantitative Finance, 11(11), 1593-1602. https://arxiv.org/pdf/0902.2965
Cooperation and Insurance
Peters, O. & Adamou, A. (2015). “Insurance makes wealth grow faster.” arXiv:1507.04655. https://arxiv.org/pdf/1507.04655
Peters, O. & Adamou, A. “The ergodicity solution of the cooperation puzzle.” https://www.researchgate.net/publication/360794480_The_ergodicity_solution_of_the_cooperation_puzzle
Wealth Inequality
Berman, Y. & Peters, O. “Wealth Inequality and the Ergodic Hypothesis: Evidence from the United States.” https://www.semanticscholar.org/paper/Wealth-Inequality-and-the-Ergodic-Hypothesis:-from-Berman-Peters/8fd46c69bdaef5d7fea4665c7a232846f60289eb
Optimal Bet Sizing
Kelly Criterion
Kelly, J.L. (1956). “A New Interpretation of Information Rate.” Bell System Technical Journal, 35(4), 917-926.
Breiman, L. (1961). “Optimal gambling systems for favorable games.” Fourth Berkeley Symposium on Mathematical Statistics and Probability, 1, 65-78.
Risk and Absorbing Barriers
Practical Applications
Taleb, N.N. (2018). “The Logic of Risk Taking.” Skin in the Game: Hidden Asymmetries in Daily Life. Random House. https://medium.com/incerto/the-logic-of-risk-taking-107bf41029d3
Experimental Evidence
Human Decision-Making
Meder, D. et al. (2021). “Ergodicity-breaking reveals time optimal decision making in humans.” PLOS Computational Biology. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8454984/
Evolution and Biology
Ergodic Limits
Dingle, K. et al. (2015). “Are there ergodic limits to evolution?” Interface Focus, 5(6). https://royalsocietypublishing.org/doi/full/10.1098/rsfs.2015.0041
Information Theory
Entropy Connections
“On the Connections of Generalized Entropies With Shannon and Kolmogorov-Sinai Entropies.” Entropy, 16(7), 3732-3753. https://mdpi.com/1099-4300/16/7/3732/htm
Document compiled from foundational mathematics, statistical mechanics, ergodicity economics, and applied risk theory.
Related Machineries
- THE MACHINERY OF ENTROPY. Entropy is the equilibrium destination of ergodic systems. Ergodicity is the property that determines whether a system actually reaches that destination.
- THE MACHINERY OF EQUILIBRIUM. Equilibrium requires ergodicity. Non-ergodic systems get trapped far from equilibrium, in metastable states that persist indefinitely.
- THE MACHINERY OF PHASE TRANSITIONS. Phase transitions are points where ergodicity breaks. The system’s phase space fragments and previously accessible states become unreachable.
- THE MACHINERY OF SCALING_LAWS. Scaling laws describe how quantities change across scales. In non-ergodic systems, the scaling of ensemble averages diverges from the scaling of time averages.
- THE MACHINERY OF PATH DEPENDENCE. Path dependence is what non-ergodicity looks like in practice. When ensemble and time averages diverge, the specific sequence of events determines the outcome. Path dependence is the mechanism. Non-ergodicity is the condition.