The butterfly, in practice
Sensitivity to initial conditions is a measured property with a timescale attached, not a metaphor.
The chaos isn't poetic. It's a number, and it tells you exactly how long you have.
What Lorenz actually showed
In the early 1960s, Edward Lorenz was running a primitive atmospheric simulation on an LGP-30 computer at MIT in Cambridge, Massachusetts. He re-entered a value mid-run to save time, typing three decimal places where the machine stored six. The output diverged wildly from the original. The atmosphere — even a toy mathematical version of it — was so sensitive to initial conditions that a rounding error of less than one part in a thousand was enough to produce a completely different future. The butterfly effect ↗ followed as a metaphor, but the finding underneath it was precise and quantitative: small errors in the starting state grow at a measurable rate.
That rate is governed by what mathematicians call the Lyapunov exponent — a number describing how quickly nearby trajectories in a system diverge. For the real atmosphere, decades of verification work have put the practical limit of deterministic forecasting somewhere in the range of two to three weeks. Beyond that, errors in the initial state have typically grown large enough to swamp the signal. The metaphor makes it sound whimsical; the exponent makes it a hard engineering constraint.
Why the initial state is always wrong
Every forecast begins with data assimilation: the process of combining millions of observations — radiosondes, satellites, aircraft, buoys — with a short-range model forecast to produce the best possible estimate of the atmosphere's present state. That estimate is never exact. Observations have instrument error. Coverage is uneven; vast stretches of the southern oceans have almost no conventional data. The model itself introduces error when it smooths the observed values onto its computational grid. What enters the model as the "initial conditions" is already a compromise.
Sensitivity to initial conditions is a measured property with a timescale attached, not a metaphor.
The errors that matter most are not the large, obvious ones — those are caught quickly. It is the small, coherent errors in dynamically active regions that are dangerous: a slightly misplaced jet stream trough, a surface pressure analysis a few hectopascals off. These seed the divergence that Lorenz described. In an unstable flow, such as a developing mid-latitude cyclone, the error doubles in roughly a day or two. In a quieter pattern, it may take longer. The atmosphere itself sets the pace, which is why a forecast can be excellent for twelve days in one week and mediocre for five in another.
What the ensemble does with this
The operational answer to Lorenz's problem is the ensemble: running the model not once but dozens of times from slightly different starting states, each within the plausible range of initial-condition uncertainty. The European Centre for Medium-Range Weather Forecasts in Reading, England, pioneered operational ensemble forecasting in the early 1990s, and ECMWF's ensemble system ↗ now runs fifty-one members. Where those fifty-one members agree, the forecast is confident. Where they spread, the atmosphere is telling you something useful: this region, at this time, is sensitive, and the honest answer is a range, not a number.
What this piece turns on
- Lyapunov exponentthe mathematical rate at which nearby trajectories diverge; governs the ~two-week forecast limit
- Ensemble spreadthe practical proxy for initial-condition sensitivity; wide spread = high sensitivity at that lead time
- Data assimilation errorthe inevitable imprecision in the starting state that seeds divergence
This is sensitivity to initial conditions made operational. The spread of the ensemble is not an admission of failure — it is a direct measurement of how much the atmosphere's future, at a given moment, depends on what you cannot know precisely about its present. A wide spread at day five is information. It tells a forecaster to weight shorter-range updates heavily and to distrust any single deterministic answer.
The limit is real, not a ceiling to raise
More data and faster computers have pushed the useful forecast horizon steadily outward — a modern five-day forecast is roughly as accurate as a three-day forecast was in the 1980s, a genuine achievement. But the two-to-three-week limit is not a technological ceiling waiting to be raised by better hardware. It is a property of the atmosphere itself, intrinsic and mathematically bounded. Lorenz understood this. The practical consequence is not pessimism but precision: knowing where the limit is tells you exactly how to spend your uncertainty budget, and the ensemble is the instrument you spend it with.
Chronology
- Early 1960sLorenz discovers sensitive dependence on initial conditions at MIT
- Early 1990sECMWF introduces operational ensemble forecasting
- 1980s to presentfive-day forecast accuracy has roughly caught up to what three-day accuracy was forty years ago
Elsewhere in Where it fails
And why the failure has a shape. Everything in this section.
- Where it fails, and why the failure has a shapeLongForecast error is not random: it grows fastest where the atmosphere is least stable, which is why some situations are predictable for days and others for hours.
- The Grid, and What Falls BetweenLongA model divides the atmosphere into boxes, and anything smaller than a box — a shower, a hill — has to be represented rather than resolved.
- EnsemblesMediumRunning the model many times from slightly different starting points converts a single answer into a spread, which is the honest output.
- VerificationShortA forecast is only as good as the record of how it did, and the scoring is its own small discipline.