Is state space modeling always linear?

No. Linear Gaussian models are an important tractable case within a wider family.

Is the state directly observed?

Sometimes partly, but the formulation often uses latent components inferred from measurements.

Many formulations can propagate predictions through gaps, although uncertainty and the reason observations are missing still matter.

AI Glossary / Time series and signal processing / Advanced

State space modeling

Definition

State space modeling describes a changing system using an underlying state and equations connecting that state to observations. A transition equation explains how the state evolves, while an observation equation explains how measurements arise from it. In time-series analysis, this framework can represent hidden levels, trends, seasonal components and other dynamic quantities together with uncertainty.

What it means

A measured value need not equal the quantity a decision maker wants to understand. A sensor reading may contain measurement noise, while the actual quantity also changes between readings. State space modeling separates those two ideas explicitly. Its state is a model-based representation, not necessarily a directly observable physical object. The assumptions determine whether the resulting estimates and forecasts are useful.

How it works

A common linear formulation uses x_t = F x_(t-1) + w_t for state evolution and y_t = H x_t + v_t for observations. Here F describes dynamics, H maps states to measurements, and w_t and v_t represent process and observation disturbances. Inputs, offsets and time-varying matrices can be added.
With Gaussian initial uncertainty and suitable independent Gaussian disturbances, the Kalman filter computes state estimates and covariances recursively. Filtering uses observations available through the current time; smoothing revises past states using later observations. Forecasting propagates the state forward. More general nonlinear or non-Gaussian state space models may require approximate inference or different algorithms, so the framework is broader than one filter.

Three practical examples

A monitoring system estimates an underlying temperature from noisy readings. Process noise allows the temperature to change, while measurement noise describes imperfect readings. These uncertainties play different roles.
A sales model represents an evolving level and seasonal components. The observed sales series provides evidence about those latent quantities, rather than revealing them without error.
A tracking system uses position and velocity as a state while observing position only. Dynamics connect successive times so the model can infer motion from multiple measurements. These examples describe possible formulations, not guaranteed accuracy improvements.

Think of it this way

Imagine following a vehicle through occasional, blurry photographs. One rule describes how the vehicle can move; another describes how a photograph relates to its actual location. Combining them gives a better-defined estimation problem than treating every photograph as an exact position or assuming the vehicle cannot move between pictures.

Worked example

Construct a scalar local-level model: x_t = x_(t-1) + w_t and y_t = x_t + v_t. Assume independent zero-mean Gaussian process and measurement disturbances with variances one and nine, respectively, and independence from the previous state uncertainty. Treat these variances as known teaching assumptions, not estimated results.
Before the next transition, the filtered state mean is 100 and its variance is four. The predicted mean remains 100, while predicted state variance becomes four plus one = five. A new observation is 114, so the innovation, observed minus predicted value, is 14. Its variance is five plus nine = 14.
The Kalman gain is predicted state variance divided by innovation variance: K = 5/14. The updated mean is 100 + (5/14) times 14 = 105. Updated state variance is (1 - 5/14) times five = 45/14, approximately 3.2143. The model moves toward the reading without treating it as exact. This is a conditional mean and variance under the stated model, not a claim that the true state equals 105.
For the following period, before any new observation, the state mean remains 105 and its variance becomes 45/14 + 1 = 59/14. The next observation's forecast variance is larger: 59/14 + 9 = 185/14, approximately 13.2143. Confusing state uncertainty with observation uncertainty would omit the sensor's measurement variation.
If the following reading is missing, the model can retain the prediction without a measurement update. It does not replace the missing reading with zero. After three transitions with no intervening observations, state variance becomes 45/14 + 3 = 87/14, while observation forecast variance is 213/14. The increasing uncertainty follows from the stipulated random-walk dynamics.

Illustrative case study

Fictional Harbor Instruments uses the constructed update to review a dashboard. An operator expects the displayed level to jump immediately from 100 to the latest reading of 114. The analyst shows that the assumed measurement variance of nine makes a partial update appropriate within this model.
The team labels the estimate and raw reading separately. It also distinguishes intervals for the latent level from intervals for future readings and marks gaps where only prediction was possible. Before operational use, it plans to examine whether the noise assumptions and dynamics fit the actual sensor process. This scenario illustrates model interpretation; it does not validate the chosen variances or a real monitoring system.

Common misconceptions and mistakes

A state space representation does not automatically make a model accurate. Too little process noise can make estimates resist real changes, while too little observation noise can make them follow unreliable readings too closely. Unknown variances still require estimation or justification. A smoothed historical estimate uses later evidence and should not be presented as information available to an earlier live decision. Also, different mathematical representations can describe the same observable process, so latent-state labels deserve care.

Business use

Specify the quantity being estimated, available measurements, dynamics and sources of uncertainty before choosing software. Compare forecasts against simple baselines and inspect innovations for remaining patterns. Record how missing observations, initialization and parameter estimation are handled. If decisions use uncertainty, verify whether the reported interval concerns the state or a future observation. Separating these quantities can prevent a seemingly precise dashboard from communicating more certainty than the model supports.

Check your understanding

Why is the next observation's variance 185/14 rather than 59/14?

Answer: The latter is predicted state variance. The observation adds independent measurement variance nine, giving 59/14 + 126/14 = 185/14.

Related terms

State space
Latent variable
Kalman filter
Smoothing
transition equation
observation equation
process noise
measurement noise

Learn next

Repeat the first update with measurement variance one instead of nine. Predicted state variance remains five, so the gain becomes 5/6, updated mean becomes 111.6667 and updated variance becomes 5/6. Explain why the model now gives the new reading more weight.

Sources

Statsmodels, Time Series Analysis by State Space Methods: https://www.statsmodels.org/stable/statespace.html Glossary of Forecasting Terms, State space modeling: https://robjhyndman.com/mwh3/FG4.pdf The scalar model, variance choices, calculations and Harbor Instruments scenario are original teaching constructions. Parameters are stipulated rather than fitted, and no empirical accuracy claim is made.

Previous: State-space landscape Next: State-space search