Prediction begins with an act of compression. A moving body becomes a state estimate: position, velocity, perhaps attitude, mass, thrust, and sensor bias. Beside it sits a covariance matrix describing the estimated uncertainty of each component and the correlations among them.1 The world is not reduced because it is simple. It is reduced because an equation requires something finite to carry forward.

Consider a satellite in low Earth orbit. A position measurement fixes a point along its track, with an uncertainty around it. The next forecast must also estimate atmospheric drag. A small error in density changes the inferred drag; over successive orbits, that error becomes a growing displacement along the track. Between tracking passes, the estimate and its uncertainty must both move forward.

State estimate and correlated uncertainty A state vector and covariance matrix lead to an uncertainty ellipse whose tilt represents correlation between two measured quantities. State / estimate x̂ = r v b STATE COMPONENTS P = σ²ᵣσᵣᵥσᵣᵦ σᵥᵣσ²ᵥσᵥᵦ σᵦᵣσᵦᵥσ²ᵦ UNCERTAINTY STRUCTURE state component 1 state component 2 CORRELATED DIRECTION The estimate is a point. The covariance describes the scale and structure of uncertainty around it.
Figure 1The estimate is a point; the covariance describes the scale and correlation of uncertainty around it. Adapted from the state-estimation framework formalized by Kalman.

In an ideal deterministic system, a state x evolves according to a rule. Supply the initial condition and the flow map supplies the future. This is the compact mathematical form of Laplace’s old intuition: if the present were known completely, uncertainty would be an imperfection of the observer rather than a property of the world.

Actual state propagation is less serene. Initial conditions are measured rather than revealed. Models omit forces. Numerical integration replaces continuous evolution with finite steps. Atmospheric drag varies with solar activity; radiation pressure depends on geometry and reflectance; sensors carry biases that may be estimated but never removed absolutely. The dynamical law may be deterministic while every usable prediction remains a distribution.

Law is not forecast

Determinism and predictability answer different questions. Determinism concerns whether a state follows uniquely from an earlier state. Predictability concerns whether a finite observer can distinguish that state far enough into the future to be useful. Chaos opens a distance between them.

For systems with sensitive dependence on initial conditions, nearby trajectories separate approximately exponentially along unstable directions. A positive Lyapunov exponent quantifies that local rate of divergence. If the initial uncertainty has scale δ0, then δ(t) ≈ δ0eλt gives a local estimate of its growth while the linear approximation remains valid. Better measurements postpone the loss of resolution; they do not abolish it. Each additional digit purchases only a finite interval. Lorenz’s atmospheric model made this separation between deterministic law and long-range prediction mathematically unmistakable.2

Weather forecasting makes the limit operational. The atmosphere is represented by deterministic fluid equations coupled to parameterizations for processes that cannot be resolved at the grid scale. Forecast centers therefore evolve ensembles of perturbed initial states, often with perturbations to model physics as well. Their widening spread is not evidence that the atmosphere has ceased to obey equations. It records how rapidly the measured present stops selecting one forecast with useful precision.3

Ensemble trajectories approaching a prediction horizon Nearby initial states evolve into a widening family of trajectories. A marked horizon indicates where their spread becomes operationally large. Ensemble / evolution HORIZON NEARBY STATES GROWING SPREAD one realized path perturbed ensemble members time →
Figure 2An ensemble begins with nearby states and widens as their consequences accumulate. The marked horizon is operational, not a boundary where physical law ends. Methodological source: ECMWF ensemble forecasting.

Why randomness enters an equation

A stochastic model does not carry a single metaphysical meaning. In a stochastic differential equation, deterministic drift may be joined by diffusion driven by an idealized random process.

The random term can represent unresolved collisions, environmental forcing, measurement error, or a deliberately coarse description of many microscopic degrees of freedom. In other settings it may be proposed as fundamental. The same mathematical form can therefore encode ignorance, approximation, or ontology.

Statistical mechanics is built upon this distinction. A gas follows microscopic dynamics, yet pressure and temperature become tractable only after individual motions are replaced by distributions. Brownian motion was modeled stochastically before molecular impacts could be followed one by one. The model succeeds not because the hidden collisions are causeless, but because their aggregate has a stable probabilistic structure.

Quantum probability raises a further question about interpretation: unitary accounts, hidden-variable theories, and physical-collapse models do not assign randomness the same role.4,5,6 An orbit forecast does not have to settle that dispute. In the filter, the random term is process noise: an allowance for accelerations the model leaves out, sized so the covariance grows about as fast as the real errors do.

The horizon

At the next tracking pass, the satellite is measured again. The new observation can expose a bad drag estimate, narrow the covariance, and move the predicted track. Nothing about that correction requires the orbit to have become lawless in the interval. The forecast carried an incomplete estimate of the forces acting on it.

In a chaotic system, the interval before another observation is needed can become much shorter. The practical question is how long a model can preserve the distinctions its user needs: which side of a storm, which patch of ground, which window for a maneuver. A trajectory remains useful only while its uncertainty fits inside that decision.

References

  1. R. E. Kalman, “A New Approach to Linear Filtering and Prediction Problems”, Journal of Basic Engineering 82, 1960. ↩
  2. Edward N. Lorenz, “Deterministic Nonperiodic Flow”, Journal of the Atmospheric Sciences 20, 1963. ↩
  3. European Centre for Medium-Range Weather Forecasts, “Medium-range forecasts”, ensemble methodology and documentation. ↩
  4. Hugh Everett III, “Relative State Formulation of Quantum Mechanics”, Reviews of Modern Physics 29, 1957. ↩
  5. David Bohm, “A Suggested Interpretation of the Quantum Theory in Terms of Hidden Variables I”, Physical Review 85, 1952. ↩
  6. Angelo Bassi et al., “Models of Wave-function Collapse, Underlying Theories, and Experimental Tests”, Reviews of Modern Physics 85, 2013. ↩