State Uncertainty, Belief Updates and Filters

Why partial observability matters

Many business and engineering decisions depend on conditions that cannot be observed directly.

Machine health, customer intent, supplier reliability, true demand state and project completion risk are latent. We observe temperatures, orders, inspections, survey responses or progress reports that are imperfect. Treating the observation as the state can cause overreaction to noise and ignore prior evidence.

A partially observed model separates the hidden state from the observation process. The system model describes how the hidden state evolves after actions; the observation model describes how likely each signal is under each hidden state.

Belief states

A belief state is a probability distribution over hidden states given the full history of actions and observations.

The belief is a sufficient summary of history for many partially observable models. Decision rules can therefore act on the belief rather than storing every past observation. A belief of 70% degraded and 30% healthy is more informative than labelling the asset simply “degraded”.

The belief update has two stages: prediction through the state-transition model and correction using the new observation. This pattern appears across discrete Bayes filters, Kalman filters and particle filters.

Discrete Bayes filtering

For a finite hidden-state model, prediction sums over previous states and correction multiplies by observation likelihood.

Normalisation ensures the updated belief sums to one. The method is exact for the specified discrete model but can become expensive when the hidden state has many dimensions. Factorised or approximate representations may then be needed.

Kalman filtering for linear-Gaussian systems

When dynamics and observations are linear and uncertainty is Gaussian, the Kalman filter tracks mean and covariance efficiently.

The prediction step propagates the current mean through the linear dynamics and expands covariance according to process noise. The correction step compares the actual measurement with the predicted measurement. The Kalman gain weights the correction according to relative model and measurement uncertainty.

If measurement noise is small relative to prediction uncertainty, the filter responds strongly to the measurement. If the sensor is noisy, it moves less. This is a principled alternative to arbitrary smoothing, provided the linear-Gaussian assumptions are appropriate.

Extended Kalman filter

The extended filter handles nonlinear dynamics or observations by linearising them locally around the current estimate.

It propagates means through nonlinear functions and uses derivatives to approximate how covariance changes. This can work well when nonlinearity is modest and uncertainty is concentrated near the estimate. It can perform poorly when the distribution is highly nonlinear, multimodal or the local linear approximation is misleading.

Derivative implementation and numerical stability deserve testing. In production, monitor innovation statistics and covariance behaviour so divergence is detected rather than silently accepted.

Unscented Kalman filter

The unscented approach represents a Gaussian belief using a set of carefully chosen sigma points and passes them through nonlinear functions.

The transformed sigma points are recombined to approximate the resulting mean and covariance. This avoids explicit derivatives and can capture nonlinear transformation of mean and covariance more accurately than first-order linearisation in some problems.

It still represents the belief as approximately Gaussian. If the true state uncertainty has multiple distinct modes—such as either “healthy” or “severely damaged” with little probability between—the Gaussian approximation may be inappropriate.

Particle filtering

Particle filters represent the belief with weighted samples and can model nonlinear, non-Gaussian and multimodal distributions.

Particle count controls computational cost and approximation quality. Rare modes can disappear during resampling, so proposal design and rejuvenation methods may be needed. High-dimensional state spaces are especially challenging because many particles may be required to represent the posterior.

Worked example: condition-based maintenance

A generic machine has hidden health that degrades over time while vibration and inspection measurements are noisy.

A filter maintains the probability of health states. One unusual sensor spike does not automatically trigger replacement; it shifts the belief according to sensor reliability and prior condition. Repeated abnormal measurements can rapidly increase failure probability. Maintenance actions also change the transition model because repair resets or improves the hidden state.

The resulting belief can feed a sequential decision policy that weighs maintenance cost, downtime and failure risk. This is superior to treating every observation as ground truth, especially when sensor noise is material.

Filter validation

State estimators should be validated through innovation, calibration and downstream decision performance.

Check whether observed measurements fall within predicted distributions at the expected frequency. Compare estimates with ground truth when occasional inspection provides it. Simulate known state trajectories and test recovery after shocks. Most importantly, test whether the decisions made from the belief remain robust when model and sensor assumptions are wrong.

State-estimation operating controls

A technically correct method still needs an auditable operating translation.

A filter should expose its uncertainty to downstream users rather than output only a single best estimate. Dashboards and interfaces can show credible ranges, state probabilities or covariance indicators together with sensor-quality flags. Define what happens when observations are missing, contradictory or outside the calibrated range. If the estimator becomes uncertain, the correct response may be additional inspection, a conservative operating mode or human review rather than a forced precise state label. Keep separate alarms for an adverse estimated state and for loss of confidence in the estimator itself; both can require action, but for different reasons.

Application checklist

• Separate hidden state from noisy observations.
• Define transition and observation models explicitly.
• Maintain a belief distribution instead of a single state label when uncertainty matters.
• Use exact discrete filtering for manageable finite state spaces.
• Use Kalman filtering only when linear-Gaussian assumptions are reasonable.
• Select extended or unscented methods for appropriate nonlinear Gaussian approximations.
• Use particle filtering when multimodal or strongly non-Gaussian beliefs matter.
• Validate calibration, sensor assumptions and downstream policy performance.