liveBook · Manning

Observation model

As explained in Grokking Bayes

An observation process specifies how measured data are generated from an underlying system state, including observation noise.

An observation model connects an unobserved system state to the data we measure. In a time series, the latent state may represent the true atmospheric CO₂ level, while each recorded value is that level plus measurement noise. section 9.2

A simple observation model assumes that measurement noise is Gaussian with zero mean: errors are equally likely to be positive or negative, and large errors are less likely than small ones. The noise variance R controls how reliable the measurements are. section 9.7

𝑦ᵢ = 𝑥ᵢ+ 𝑟,

The equation states that the observed value yᵢ equals the latent state xᵢ plus observation noise r. section 9.7

The observation model plays several roles in sequential inference:

  • It supplies the likelihood that links each latent state xᵢ to its observed datum yᵢ. section 9.4
  • It helps a filter update its predicted state after a new measurement arrives. section 9.3
  • It represents measurement quality through R; for annual CO₂ averages, R is set to 0.05 of the empirical variance. section 9.7

Annual CO₂ measurements

Suppose xᵢ is the true CO₂ concentration in year i and yᵢ is the recorded annual average. The model treats yᵢ as xᵢ plus random measurement error. With a linear dynamic model and Gaussian observation noise, the resulting filter is a Kalman filter. section 9.12 section 9.7

How Grokking Bayes covers it

How deep it goes
Defines a Gaussian-noise equation, explains variance R, and connects the result to Kalman filtering.
Examples
Uses annual CO₂ concentration measurements, Gaussian observation noise, and a NumPy variance calculation for R.
What it assumes
Assumes a latent-state and state-transition formulation while introducing measurement noise and likelihood-based inference.
What you can do after
Apply it via a Kalman filter by specifying measurement noise and updating latent-state beliefs with observed data.
How it connects
Links latent state, state transition model, likelihood, Bayesian filter, and Kalman filter.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.