What is the worked model's equilibrium?

Twenty units under the stated update rule.

AI Glossary / Mathematics and optimization / Advanced

Dynamical system

Definition

A dynamical system describes how a state changes over time according to a rule. In a simple discrete-time model, the next state is obtained by applying a transition function to the current state: xₜ₊₁ = F(xₜ). Continuous-time models describe change through differential equations. Inputs, disturbances and observations can be included when the application requires them.

What it means

The state summarizes the information the model needs to describe subsequent evolution, together with any future inputs. It can be a single number or a vector containing quantities such as position, velocity, inventory or an algorithm's current parameters. The observed output need not expose the entire state.
The central question is temporal: given a starting condition and a rule, what happens next, and what happens after repeated evolution? A model may approach an equilibrium, oscillate or behave in more complicated ways. Calling something a dynamical system does not assert that it eventually settles down or that its model accurately represents the real process.

How it works

Choose the state variables, their units and the time convention. Specify an initial state and a transition rule. If an external input changes the transition, write it explicitly, for example xₜ₊₁ = F(xₜ, uₜ). An observation rule can separately describe what a sensor or reporting system sees.
Iterating a discrete rule produces a trajectory: a sequence of states. An equilibrium of an autonomous discrete system is a state x* satisfying F(x*) = x*. Starting there leaves the state unchanged. Convergence asks whether trajectories approach that state; stability asks how trajectories respond to nearby starting conditions. Those are related but distinct questions.
Random disturbances require a stochastic model or other explicit treatment of uncertainty. Numerical simulation also introduces choices such as time steps and precision. A calculated trajectory is a consequence of these assumptions, rather than independent evidence that the assumed rule is correct.

Three practical examples

A resource model tracks a retained quantity from one reporting period to the next, together with a new addition. Repeated updates reveal whether the model approaches a sustainable level or keeps increasing.
A recurrent neural network updates a hidden state as successive inputs arrive. The hidden state carries information across steps, while an output function converts that representation into a prediction. Useful temporal behavior still depends on training and evaluation.
An iterative optimization algorithm repeatedly updates its parameters. Treating the update as a dynamical system can help examine fixed points, convergence and sensitivity to the step size. An equilibrium of the update need not automatically be the desired global optimum.

Think of it this way

Think of a film built from a starting frame and a rule for generating each next frame. A snapshot shows where the system is; the transition rule describes how it moves. Two films can begin with the same visible picture but develop differently if their hidden states or transition rules differ. A plausible first frame alone cannot validate the whole sequence.

Worked example

Consider the invented discrete model xₜ₊₁ = 0.75xₜ + 5, with x₀ = 4. Interpret x as a nonnegative quantity measured at the same point in each reporting period. The rule retains three-quarters of the previous quantity and then adds five units.
The first update gives x₁ = 0.75 × 4 + 5 = 8. Continuing gives x₂ = 11, x₃ = 13.25 and x₄ = 14.9375. The quantity rises, but its increments decrease.
An equilibrium must satisfy x* = 0.75x* + 5. Rearranging yields 0.25x* = 5, so x* = 20. This is not the value reached after four steps. It is the level that reproduces itself under the rule: 0.75 × 20 + 5 = 20.
Subtracting 20 from both sides of the update gives xₜ₊₁ − 20 = 0.75(xₜ − 20). Repeated substitution yields xₜ − 20 = 0.75ᵗ(x₀ − 20). Because powers of 0.75 approach zero, the trajectory approaches 20. At step four, the difference is 0.75⁴ × (4 − 20) = −5.0625, agreeing with 14.9375 − 20.
The same calculation shows sensitivity to the initial state. Two trajectories following this rule have a difference that is multiplied by 0.75 at every step. Their initial separation shrinks. That conclusion follows from this particular update coefficient; it is not a property of every dynamical system.

Illustrative case study

In a fictional operations dashboard, a team uses the worked recurrence as a planning model. An analyst reports that the tracked quantity will be twenty units after four periods, confusing the equilibrium with the finite-horizon forecast. Recomputing the trajectory corrects the forecast to 14.9375 units.
The team then considers increasing the addition from five to six units while retaining the same coefficient. The new equilibrium satisfies x* = 0.75x* + 6 and is twenty-four. The previous trajectory and equilibrium no longer answer the revised scenario.
Before using either scenario operationally, the team checks what the retention fraction and period actually represent. It also asks whether the quantity can reasonably be fractional and whether capacity limits or changing additions were omitted. These questions concern model adequacy. Correct recurrence arithmetic alone cannot establish that the operational assumptions hold.

Common misconceptions and mistakes

A time series is a sequence of observations; a dynamical model additionally specifies how evolution is represented. Fitting observed values does not uniquely identify the underlying mechanism.
An equilibrium is not necessarily attracting. For example, xₜ₊₁ = 2xₜ has equilibrium zero, but any nonzero starting value grows in magnitude. Also, the discrete coefficient in a period-based recurrence cannot be reused unchanged after altering the length of that period without justification.

Business use

Dynamical models support scenario comparisons, capacity planning and analysis of feedback effects. Document the initial state, time unit, update rule, inputs and observation process. Distinguish a trajectory forecast from a long-run equilibrium, and test the model against appropriate observations before relying on either.

Check your understanding

What is x₂ in the example?

Answer: Eleven.

Why is twenty not the four-period forecast?

Answer: The trajectory has not reached its limiting equilibrium.

What changes when the addition becomes six?

Answer: The update rule and equilibrium; the new equilibrium is twenty-four.

Related terms

State
Trajectory
Fixed point
Stability
Recurrent neural network
state-transition function
equilibrium
recurrence
differential equation

Learn next

Compare stable and unstable recurrences, then add explicit inputs and observation rules to a state model.

State

Sources

Aalto Dictionary of Machine Learning, dynamical system: evolving state, discrete transition maps and continuous-time descriptions. https://aaltodictionaryofml.github.io/ADictML_Main.pdf The recurrence, equilibrium derivation and fictional dashboard case are original teaching illustrations.

Previous: Dynamic time warping Next: E-optimal design