Step 04 · Spectral & Models
Ergodicity Notes
Every definition of a stationary process imagines an endless collection of records, as though we could run the same experiment again and again. Ergodicity is the promise that one sufficiently long record can stand in for all of them.
The problem it solves
When we speak of the mean of a process, we imagine averaging across the many records the process might have produced at one fixed moment. In practice, we possess exactly one record, and we average it across time instead. These two averages are different constructions, and there is no automatic reason for them to agree. Ergodicity is the condition under which they do.
Memory and the mixing of a record
A record is ergodic when values far apart in time are effectively unrelated, so that the single long run keeps visiting every part of the process's behaviour. The practical guise of this condition is the decay of autocorrelation. If the autocorrelation falls toward zero as the displacement grows, then distant parts of the record carry little shared information, and the time average settles toward the process mean. If the autocorrelation never falls, the record is stuck in one region and the time average is a poor guide to anything beyond it.
| Signal | What it suggests |
|---|---|
| autocorrelation decaying to zero | time averages settle |
| a broad spread of visited values | the record explores its range |
| spectral density finite at the slow end | no persistent drift |
| half-record summaries in close agreement | the record is well mixed |
A quiet but central assumption
Much of the compendium quotes averages, spreads and spectra computed from a single record. Every one of those quotations leans on the ergodic hope that the record is long enough and well mixed enough for the quotation to mean something beyond itself. The hope is not a decoration on the mathematics; it is the load-bearing beam, and naming it plainly is the honest way to keep it in view.
The burden of a short record
Ergodicity is a statement about long records. A short record, even from a well-behaved process, may wander without visiting the full range of behaviour, and its time average may sit far from the process mean by chance alone. This is why the compendium pairs every quoted average with a warning about the length and the mixing of the record behind it.
Putting the four steps together
Follow the thread and the shape of the whole compendium appears. The curve gives a shape; the moments summarise it; the matrix records how two quantities share it; the series stretches it through time; the spectrum spreads the variance across frequencies; the recipes compress the memory into parameters; and ergodicity promises that one long record may speak for the process. Each step leans on the one before it, and the promise holds only if the earlier steps were read with care.
A reader who reaches this page has travelled the full rail. The final step gathers the notation in one place, states the editorial aims behind the compendium, and offers a way to reach the desk that keeps it.
The patience of a long record
Ergodicity rewards patience in a way that few ideas in the compendium do. A short record may wander without visiting the full range of behaviour, and its time average may sit far from the process mean through chance alone. As the record lengthens, the wandering accumulates and the average settles, but only if the record keeps mixing rather than settling into one region. The promise is real, and it is a promise about length.
A reader should therefore treat every quoted average as a claim that leans on the length of the record behind it. Where the record is long and the autocorrelation decays quickly, the claim is strong. Where the record is short or the memory is long, the claim is weak, and the honest response is to say so rather than to quote the average as though it settled the question. Naming the weakness is part of the work, not a retreat from it.