Estimation under Noise
estimation, filtering, noise, Kalman, hidden state
1 Application Snapshot
In real systems, the quantity you want to control is often not measured directly.
What you receive instead is:
- a noisy sensor reading
- a partial projection of the state
- or a delayed, corrupted, or missing measurement
That is why many control loops are really:
measure -> estimate hidden state -> act
This page is the shortest bridge from abstract state-space language into the estimation layer that real systems need before feedback can become reliable.
2 Problem Setting
Suppose the system state is \(x_t\), the control is \(u_t\), and the measurement is \(y_t\).
A common discrete-time model is
\[ x_{t+1} = f(x_t, u_t, w_t), \qquad y_t = h(x_t) + v_t, \]
where:
- \(w_t\) is process noise or unmodeled disturbance
- \(v_t\) is measurement noise
The control problem is awkward because the controller often needs information about \(x_t\), but only gets to see \(y_t\).
So the system builds an estimate
\[ \hat{x}_t \]
from:
- the model
- the previous estimate
- the new measurement
That estimate then becomes the object the controller actually uses.
3 Why This Math Appears
This language reuses several math layers already on the site:
Probability and Statistics: noise models and conditional updating tell us how measurements should change beliefSignal Processing and Estimation: filtering is the load-bearing update rule for hidden-state inferenceControl and Dynamics: many controllers act on an estimated state rather than the true stateStochastic Control and Dynamic Programming: uncertainty changes the decision problem, not just the observation modelLinear Algebra: in linear-Gaussian settings, means and covariances evolve by matrix operations
So estimation under noise is not an optional add-on. It is often the missing middle layer between sensing and action.
4 Math Objects In Use
- hidden state \(x_t\)
- measurement \(y_t\)
- process noise \(w_t\)
- measurement noise \(v_t\)
- estimate \(\hat{x}_t\)
- prediction step from the model
- correction step from the new measurement
In linear-Gaussian form, the system often looks like
\[ x_{t+1} = Ax_t + Bu_t + w_t, \qquad y_t = Cx_t + v_t. \]
At first pass, the estimator tries to balance two sources of information:
- what the model predicts
- what the sensor just reported
5 A Small Worked Walkthrough
Imagine tracking a car on a straight road.
The hidden state is
\[ x_t = \begin{bmatrix} p_t \\ v_t \end{bmatrix}, \]
where:
- \(p_t\) is position
- \(v_t\) is velocity
Suppose the GPS measurement only reports position:
\[ y_t = p_t + v_t^{(\mathrm{noise})}. \]
Then:
- position is observed noisily
- velocity is not observed directly at all
If the controller needs velocity to brake smoothly or maintain spacing, raw GPS readings are not enough.
A useful estimator does two things repeatedly:
PredictionUse the motion model to guess where the car should be next.CorrectionUse the new GPS reading to adjust that guess.
That update loop is the systems-level meaning of filtering.
The key application lesson is simple:
- the sensor does not give the state
- the model does not guarantee the state
- the estimate is the compromise between them
6 Implementation or Computation Note
Three practical design questions appear quickly:
How much should I trust the model?A strong model can stabilize noisy sensing, but model mismatch can also mislead the filter.How much should I trust the sensor?High-noise measurements should not dominate every update.What exactly is the controller using?In many real loops, the action depends on \(\hat{x}_t\), not on \(x_t\).
Use these pages as the strongest follow-on support:
7 Failure Modes
- treating a noisy measurement as if it were the true state
- forgetting that unobserved coordinates can still matter for control
- assuming better sensors remove the need for a model
- assuming a good model removes the need to check measurement quality
- building a controller around estimates without asking how bad estimate error can become
8 Paper Bridge
- 16.322 / Stochastic Estimation and Control -
First pass- official MIT anchor for the estimation side of noisy dynamical systems. Checked2026-04-25. - EE278 / Introduction to Statistical Signal Processing -
Paper bridge- useful when you want the measurement-noise viewpoint to become more statistical and signal-processing flavored. Checked2026-04-25.
9 Sources and Further Reading
- 16.322 / Stochastic Estimation and Control -
First pass- official MIT course hub for hidden-state estimation under noise. Checked2026-04-25. - Lecture 11: Estimators / Observers -
First pass- compact official MIT notes showing the observer viewpoint directly. Checked2026-04-25. - Recitation 10: Observers -
Second pass- useful when you want the control-facing observer picture rather than a purely statistical one. Checked2026-04-25. - EE278 / Introduction to Statistical Signal Processing -
Second pass- official Stanford course anchor for filtering and noisy estimation. Checked2026-04-25. - AA273 / State Estimation and Filtering for Autonomous Systems -
Bridge outward- official Stanford course entry where estimation is central to autonomous systems. Checked2026-04-25. - Stats366 HMM notes -
Bridge outward- useful when you want a discrete hidden-state analogue of the same filtering story. Checked2026-04-25.