Ergodicity is a property of a system where the average outcome you’d get by watching one individual over a long time equals the average outcome you’d get by looking at many individuals at a single moment. When a system is ergodic, these two types of averages converge. When it’s not, they can diverge wildly, and that difference has profound consequences in fields ranging from physics to economics to psychology.
Two Ways to Calculate an Average
The concept becomes clearest with a concrete picture. Imagine you want to know the “average” return of a particular investment strategy. You could gather 1,000 people who all follow the strategy and check how they’re doing right now. That’s an ensemble average: a snapshot across many parallel cases at one point in time. Alternatively, you could follow a single person using the strategy for 1,000 days and average their results over that entire period. That’s a time average: one trajectory tracked through time.
If the investment is ergodic, both numbers will match. It doesn’t matter whether you look across people or across time. But many real-world processes aren’t ergodic. A coin-flip betting game where you gain 50% on heads and lose 40% on tails has a positive ensemble average (the “expected value” across all players looks good), yet the time average for any single player trends toward ruin. Most individual players go broke even though the group average keeps rising, because a few extreme winners pull the average up while the majority quietly lose. That gap between the two averages is exactly what non-ergodicity looks like in practice.
Where the Idea Came From
The concept originated in 19th-century physics. Ludwig Boltzmann, working on the foundations of thermodynamics, needed a way to connect the behavior of individual gas molecules to the large-scale properties of a gas like temperature and pressure. His ergodic hypothesis proposed that a single gas molecule, given enough time, would visit every possible state available to it. If true, you could predict the behavior of the whole gas by tracking one molecule long enough, rather than needing to observe every molecule at once. Boltzmann used this idea alongside several other tools in his lifelong effort to ground thermodynamics in the statistics of molecular motion.
The hypothesis was later formalized in the 1930s by the mathematician George David Birkhoff, whose ergodic theorem gave precise mathematical conditions under which time averages and ensemble averages are guaranteed to converge. In the formal definition, a system is ergodic if it can’t be split into two or more separate subsystems that never interact. More precisely, any subset of the system that remains unchanged as the system evolves must either contain everything or nothing. There’s no room for the system to get “stuck” in one region of its possible states.
Why It Matters in Economics
For most of the 20th century, mainstream economics relied on expected utility theory, which essentially uses ensemble averages to model how people should make decisions. The assumption is that a rational person evaluates a gamble by averaging across all possible outcomes, weighting each by its probability. But physicist Ole Peters has argued that this framework misses something fundamental: people don’t live across parallel universes. They live one life, through time.
Peters’ work, known as ergodicity economics, starts from the observation that wealth dynamics are typically non-ergodic. The expected value of a gamble (the ensemble average) can look positive while the time-average growth rate for any individual player is negative. His alternative model proposes that people maximize their time-average growth rate of wealth rather than the expected value of some psychological utility function. Interestingly, this approach produces a mathematical correspondence where what economists have traditionally called a “utility function” turns out to map directly onto what physicists call an “ergodicity transformation,” the mathematical operation needed to convert a non-ergodic process into an ergodic one. This reframing predicts that a person’s risk preferences should depend heavily on the specific dynamics of how their wealth grows and shrinks, not just on abstract psychological attitudes toward risk.
The Problem With Group Averages in Psychology
Psychology has its own ergodicity problem. Most research studies measure a group of people at one or a few time points, calculate group-level patterns, and then assume those patterns describe what happens within any individual person over time. A study might find that across 500 people, higher stress correlates with worse sleep. The implicit assumption is that this also holds within you personally: on days when your stress goes up, your sleep gets worse.
But that inference only works if the underlying process is ergodic. A 2019 paper in the Proceedings of the National Academy of Sciences examined this directly and found that ergodicity is “the exception in psychological data.” The processes generating human behavior tend to be heterogeneous across individuals (your stress-sleep relationship differs from mine) and unstable over time (your own relationship between stress and sleep may shift across months or years). Both of these properties violate the conditions needed for ergodicity. The authors of the original analysis called non-ergodicity a “threat to human subjects research” because it means group-level findings may not generalize to the individual level at all. What’s true of the average person may not be true of any particular person.
Ergodicity in Biological Evolution
Evolution offers a particularly rich example. Researchers have drawn an exact mathematical correspondence between the standard population genetics model (the Wright-Fisher process) and the statistical mechanics framework Boltzmann originally developed for gases. Random mutations at the genetic level create a kind of diffusion through the space of possible body plans and structures at the phenotype level, much like how molecules bouncing around sample different energy states.
This analogy raises a practical question: has evolution had enough time to “explore” the space of possibilities? Two classes of biological structures emerge from the analysis. The first class has an ergodicity timescale (the time needed to sample all relevant possibilities) that exceeds the actual time evolution has had to work. These structures haven’t finished exploring their options, and there’s no reason to expect them to converge on an optimal solution. The second class is more interesting: structures whose ergodicity timescale is shorter than available evolutionary time. These have effectively sampled all the options within their constraints, and you’d expect to see convergence toward similar solutions across unrelated species. This is why eyes, for instance, evolved independently dozens of times across the animal kingdom. The evolutionary “search” for light-sensing organs was ergodic: given enough time, different lineages found similar answers.
Ergodic vs. Stationary
One common point of confusion is the difference between a stationary process and an ergodic one. A process is stationary if its statistical properties don’t change over time. The probabilities governing the system today are the same as they’ll be tomorrow. Ergodicity is a stronger condition. A process can be stationary without being ergodic. Picture two coins, one fair and one double-headed, selected at random and then flipped forever. The process is stationary (the statistical rules aren’t changing) but not ergodic, because your time average depends entirely on which coin you happened to get. The ensemble average across all coin-flippers would show 75% heads, but no individual flipper would ever see that rate. They’d see either 50% or 100%.
Every ergodic process is stationary, but not every stationary process is ergodic. The distinction matters because stationarity is often easier to test for, and people sometimes assume it’s sufficient. It isn’t. Ergodicity additionally requires that the system can’t get trapped in subsets of its possible states, that over time, a single trajectory is representative of the whole.
Practical Applications
Beyond these theoretical domains, ergodicity shows up in everyday engineering. In queueing theory, which governs everything from call centers to web servers, ergodicity determines whether a queue reaches a stable state. If a queueing system is ergodic, the distribution of waiting times converges to a steady pattern as time goes on, regardless of starting conditions. This makes it possible to predict average wait times and plan staffing levels. If the system isn’t ergodic (because arrivals outpace service capacity, for example), waiting times grow without bound and no stable prediction is possible.
The core lesson of ergodicity is deceptively simple: the average across a group is not always the same as the average across time for one member of that group. Recognizing when this gap exists changes how you interpret statistics, evaluate risks, and understand whether general findings apply to your specific situation.