Observation noise is random variation in measured or observed values that the underlying model does not explain. It can come from rounding errors, unmeasured differences between cases, or other unpredictable effects. Positive and negative noise are treated as equally likely, while large deviations are less likely than small ones. section 3.11 section 9.7
Noise matters because it changes how confidently a model should interpret data:
- Fitting every noisy fluctuation can produce overfitting and patterns that do not generalize. section 3.11
- A zero-mean normal model represents deviations that are equally likely to be positive or negative. section 3.13
- A larger variance represents a wider range of plausible measurement errors. section 9.7
House prices and CO₂ measurements
In a linear house-price model, noise slightly corrupts the relationship between area and price: y = mx + ε. Rounding in reported prices and house-specific features may explain ε. In a time-series model, a measured CO₂ value similarly equals the underlying latent state plus Gaussian observation noise. section 3.13 section 9.7
The observation equation shows how a measured value combines an underlying state with noise. section 9.7















