AI Glossary / AI and machine learning / Advanced
Linear Dynamical System
Definition
A Linear Dynamical System is a model for a hidden state that evolves smoothly over time, with the next state a linear function of the current one plus noise, and observations a linear readout of the state, the structure behind the Kalman filter.
What it means
Many business quantities cannot be observed directly: true demand beneath noisy sales, a machine's real wear beneath vibrating sensors, a customer's underlying engagement beneath sporadic activity. The Linear Dynamical System models the hidden level separately from the noisy sightings.
The state evolves by a linear rule: today's hidden value is a linear combination of yesterday's, plus a dose of randomness. Observations are another linear step: what you measure equals a linear readout of the state, plus measurement noise.
This structure supports the Kalman filter, the celebrated algorithm that updates the best estimate of the hidden state as each observation arrives, weighing new evidence against the prediction by how noisy each is.
The model is the ancestor of modern time-series machine learning: smooth hidden evolution, noisy observations, principled updating. Its linearity is both the tractability and the ceiling, since violently non-linear dynamics break the mathematics.
For a business, you meet it inside forecasting, tracking, sensor fusion and econometric tools. The durable idea is the split between the true state and the noisy sighting, and the discipline of updating beliefs as evidence arrives.
How it works
Define the hidden state: the true quantity of interest, such as underlying demand or machine wear, with its linear evolution rule.
Define the observation model: how measurements read the state, with honest noise levels for both evolution and measurement.
Run the Kalman filter: predict the next state, observe, and update the estimate by weighing prediction against observation by their uncertainties.
Extract smoothed histories for analysis and filtered now-estimates for operations, and forecasts by rolling the evolution rule forward.
Re-fit the noise levels and rules when the process changes, because the filter's wisdom is only as good as its noise model.
Three practical examples
- A demand planner whose Kalman filter tracks true underlying demand through promotion spikes and supply gaps in the sales feed.
- An industrial monitor fusing three noisy vibration sensors into one wear estimate, updating with every reading.
- A central bank style nowcast: many noisy indicators, one hidden estimate of the economy's current state, revised as each release lands.
Think of it this way
Think of watching a swimmer underwater through a rippling surface. You never see the swimmer exactly, only wobbling glimpses, but your brain tracks the real position by combining each glimpse with where the swimmer must plausibly be. The filter is that tracking, formalized.
The limit: the tracker believes the swimmer moves like a swimmer. If the hidden process jumps, a regime change, a break, the linear evolution rule lags, and the estimate trails reality until the model is told the water changed.
Worked example
demand beneath the noise
Suppose a grocer's sales data is battered by promotions, stockouts and weather, while purchasing needs true underlying demand. The figures are illustrative.
A Linear Dynamical System separates the two: the hidden demand state evolves smoothly, and sales are noisy sightings distorted by known events.
Purchase quantities set from the filtered estimate cut both stockouts and waste by double-digit percentages.
The model also timestamps regime changes: when the filtered trend turns, the buyer sees it weeks before the raw sales line admits it.
Illustrative case study
A fictional logistics firm fused its fleet trackers with a Kalman filter tuned on summer data, and the first autumn, with its fogs and delays, produced estimates the dispatchers learned to ignore within a fortnight.
The noise model no longer matched reality, so the filter weighed evidence wrongly, and nobody had been assigned to re-tune it.
The invented lesson: a filter is a theory about noise. When the world's noise changes, re-tune the theory, or the tracker becomes a very precise liar.
Common misconceptions and mistakes
- It is just a moving average. A moving average smooths the past; the filter carries an explicit hidden state with an evolution rule, which enables principled prediction and uncertainty.
- Linear means it cannot handle real data. Real data is often locally linear enough, and the filter remains the workhorse of tracking and fusion worldwide.
- The filter removes the need for judgment. Its estimates are only as good as the noise assumptions, which are human choices requiring maintenance.
Business use
Use the hidden-state split for any noisy operational signal: model the true level and the sightings separately.
Assign ownership of the noise model: re-fit evolution and measurement noise on a schedule and after process changes.
Track the filter's uncertainty, not just its estimate: a widening band is the model saying the world no longer fits.
People also ask
What is the Kalman filter?
The algorithm that updates the best estimate of the hidden state as observations arrive, weighing prediction and evidence by their uncertainties.
Where is it used?
Navigation, tracking, forecasting, sensor fusion and econometrics, anywhere a true state hides beneath noisy measurements.
What breaks it?
Dynamics or noise that are violently non-linear, and regime changes the linear evolution rule cannot anticipate.
Check your understanding
Why does a Linear Dynamical System separate the hidden state from the observations?
Answer: Because the true quantity evolves smoothly while measurements are noisy glimpses of it, and keeping the two apart allows principled updating that weighs each sighting against the prediction by how much each can be trusted.
Related terms
Learn next
AI for Business Owners explains how to read a smoothed estimate and know when to doubt it: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Limited Memory AI - Limited memory AI is an informal category used to describe AI systems that use past information, learned patterns or recent observations when producing an...
- Linde - Buzo - Gray algorithm - The Linde - Buzo - Gray algorithm, or LBG algorithm, is an iterative method for designing a vector quantizer.
- Linear - Linear describes a relationship that follows a straight line: double the input and you double the output's contribution, with no curves, thresholds or...
- Linear predictor function - A linear predictor function forms a score by adding weighted input features, usually with an intercept.
- Linear separability - Linear separability is the property that two classes of points can be separated perfectly by a linear decision boundary in the chosen feature space.
- Local case-control sampling - Local case-control sampling is a subsampling method for fitting logistic regression to large datasets.
Sources
[S13] Minka, Statistical Learning/Pattern Recognition Glossary. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Linear discriminant analysis Next: Linear filter