How Mathematics Improves The World | Knowing Where You Are When Every Sensor Is Slightly Wrong

State estimation, sensor fusion, Kalman filtering, covariance, prediction and measurement updates, observability, Extended Kalman Filters, Unscented Kalman Filters, particle filters, smoothing, factor graphs, navigation and uncertainty all belong to one practical problem: the world contains states we need to know but cannot measure perfectly. A spacecraft needs position and velocity. A robot needs pose. A car needs lane-relative location and motion. A phone wants orientation. Every sensor sees only part of the truth and adds noise, delay, bias, ambiguity or drift.

Sensor fusion is therefore not simply “average several sensors.” An IMU responds quickly and drifts. GNSS provides an absolute reference and can fail under blockage or multipath. Cameras provide dense geometry and depend on texture, lighting and correspondence. Wheel odometry is precise until tyres slip. Radar provides range and Doppler through conditions that trouble vision, but it produces clutter and ambiguous returns. A mathematically serious estimator carries a model of motion, a model of measurement and a model of uncertainty. It predicts what the hidden state should become, compares that prediction with evidence, and decides how much that evidence deserves to move the estimate.

The Kalman filter is the classic expression of that logic. NASA’s 2025 Navigation Filter Best Practices treats navigation filtering as a disciplined estimation problem involving dynamics, measurements, covariance, observability, measurement latency, bias modelling, sigma-point methods and consistency. On Artemis I, NASA used several Extended Kalman Filters to estimate translational and rotational vehicle states by propagating dynamics and IMU data, then updating state and covariance with GPS, optical navigation and star-tracker measurements. “Where are we?” is inseparable from “how uncertain are we, what evidence changed our mind, and when should we stop trusting ourselves?”

---

Quick Read

A state estimator infers variables that are not directly known from a combination of a system model and noisy observations. In a linear Kalman filter, the state is propagated through a transition model and covariance is propagated with it, enlarged by process noise. A new measurement produces an innovation: the difference between what the sensor reported and what the model predicted that sensor should report. The Kalman gain weights that innovation according to predicted state uncertainty and measurement uncertainty.

The filter does not simply “trust the best sensor.” It asks which directions of the state each sensor can observe, how errors are correlated, how old the measurement is, whether the assumed noise statistics agree with residuals, and whether the system model contains enough information to separate different hidden states. NASA’s navigation guidance emphasises observability, covariance management, measurement processing, latency, bias models and nonlinear effects because real filters fail when those details are treated as implementation trivia.

When models are nonlinear, engineers may linearise them in an Extended Kalman Filter, propagate sigma points in an Unscented Kalman Filter, or represent the state distribution with particles. Smoothers use future measurements to improve past estimates. Factor graphs and batch optimisation reformulate long trajectories as sparse global estimation problems. Different methods trade latency, computation, robustness, memory and assumptions.

The deepest boundary is simple: an estimate is not truth. A covariance matrix is not certainty because software prints six decimal places. If a model omits bias, noise is correlated, the system is unobservable, correspondence is wrong or a sensor is spoofed, an estimator can become confidently wrong. Good state estimation therefore includes diagnostics that test whether its own uncertainty claims remain believable.

One-sentence answer: Mathematics improves the world by turning imperfect, delayed and partial sensor readings into a continuously updated estimate of hidden physical state together with a quantitative account of uncertainty, making navigation, robotics, tracking and autonomous control possible even when no individual sensor knows the full answer.

---

1. The Real World Hides the Variables We Care About

A vehicle does not carry a sensor labelled “true position”. It carries devices that measure satellite signal timing, acceleration, angular rate, wheel rotation, image features, radar range, pressure and other proxies. Position, velocity and orientation are inferred from those measurements through models. The distinction sounds semantic and changes the whole problem.

The state is a compact collection of variables sufficient to describe the system for prediction. The measurement is what a sensor actually reports. A state can be physically meaningful and never be measured directly. The state can also include nuisance quantities—clock bias, sensor bias, scale factor, wind, calibration error—because estimating them helps recover the quantities we care about.

Once this is understood, sensor fusion stops being a hardware inventory. The question becomes inferential: what hidden state best explains the measurements and dynamics we have seen, and how uncertain should we remain?

2. A State Vector Is a Chosen Representation of Reality

An aircraft navigation state may contain position, velocity, attitude, gyro bias and accelerometer bias. A ground robot may add wheel-scale factors. A battery estimator may contain state of charge and internal resistance. A tracking system may include target position, velocity and turn rate.

We write a state vector x=[x₁,x₂,…,xₙ]ᵀ. That notation can make representation look neutral. It is not. Omit a slowly varying bias and the filter may force that bias into position or attitude. Add too many weakly observable parameters and covariance can become badly conditioned. Represent rotation with an unconstrained vector and the mathematics can leave the physical rotation manifold.

Estimator design begins before the Kalman equations. It begins with deciding what reality needs to be represented so that the measurements can explain it without pretending to know more than they do.

3. The Process Model Predicts How State Changes

A discrete linear process model writes xₖ=Fₖ₋₁xₖ₋₁+Bₖ₋₁uₖ₋₁+wₖ₋₁. F propagates state, u is a known control or measured input, B maps that input into state, and w represents unmodelled process effects.

For constant-velocity motion in one dimension, position advances by velocity times Δt while velocity remains approximately constant. For inertial navigation, acceleration and angular rate drive nonlinear equations of motion. For battery state estimation, current changes stored charge while electrochemical parameters shape voltage response.

No model is perfect. Tyres slip, winds change, thrusters misalign, drivers steer and physical parameters drift. Process noise exists because real dynamics contain variation the deterministic model does not capture. A useful estimator models enough physics to predict well and enough uncertainty to survive what it omitted.

4. The Measurement Model Predicts What a Sensor Should See

A linear measurement model writes zₖ=Hₖxₖ+vₖ. z is the measurement, H maps state into predicted sensor output and v is measurement noise. The estimator does not compare a sensor value directly with the whole state; it predicts what that particular sensor should read if the current state hypothesis were true.

A GNSS receiver observes quantities related to range and clock error. A star tracker observes directions in a body frame. A camera observes projected image coordinates. A wheel encoder observes rotations related to displacement. A magnetometer observes a field vector altered by orientation and local magnetic disturbances.

The measurement model is where physics, calibration and geometry enter the update. A perfect statistical algorithm using the wrong H or h(x) can be systematically wrong with great confidence.

5. Uncertainty Is Part of the Estimate, Not a Footnote

A state estimate without uncertainty is incomplete. Kalman-family estimators carry the state error covariance P=E[(x−x̂)(x−x̂)ᵀ]. Diagonal entries are error variances. Off-diagonal entries are cross-covariances describing how errors move together.

If position and velocity errors are correlated, a position measurement can change the velocity estimate even though the sensor did not measure velocity directly. The process model created that relationship through time. If two biases are strongly correlated, learning one can inform the other—or make them hard to distinguish.

Covariance therefore contains memory of information geometry. It tells the filter not only how uncertain it is, but where the uncertainty lies and how one piece of evidence can propagate through the state.

6. Covariance Ellipses Turn a Matrix Into Geometry

In two dimensions, a position covariance can be drawn as an ellipse. Eigenvectors define the principal axes. Eigenvalues determine their lengths for a chosen probability contour. A long thin ellipse means one direction is much better known than another.

This occurs naturally. One range measurement strongly constrains distance along a line of sight and weakly constrains transverse position. Two ranges from different directions intersect and shrink the uncertainty. A camera looking down a corridor may constrain sideways motion better than forward depth. Satellite geometry can make vertical GNSS uncertainty worse than horizontal.

Sensor fusion works because information has direction. A modest sensor that constrains an uncertain direction can be more valuable than an exquisitely precise sensor measuring something already well known.

7. Prediction Moves Both Knowledge and Ignorance Forward

The Kalman prediction step propagates x̂ₖ|ₖ₋₁=F x̂ₖ₋₁|ₖ₋₁+B u. It propagates covariance as Pₖ|ₖ₋₁=F Pₖ₋₁|ₖ₋₁Fᵀ+Q. Q is the process-noise covariance.

The mean follows the model. The covariance usually grows because unknown disturbances can accumulate. This is intellectually important. If time passes with no informative measurement, confidence should not stay frozen. A dead-reckoning robot moving through uncertainty should become less certain about where it is.

Prediction is therefore not only extrapolation. It is an explicit statement that motion creates ignorance as well as location change.

8. Process Noise Q Describes What the Model Does Not Explain

Q is often dismissed as a tuning matrix. That hides its meaning. It represents uncertainty injected into state evolution by disturbances or modelling simplifications. A constant-velocity model should admit random acceleration. An inertial bias model should admit bias drift. A target tracker should admit manoeuvres its kinematic model cannot predict exactly.

Choose Q too small and prediction becomes dogmatic: the filter assumes reality should obey the model and treats real motion as sensor error. Choose Q too large and the filter forgets useful dynamical continuity, becoming noisy and measurement-driven.

Good Q design comes from physics, empirical residuals and validation. It is the quantitative description of how quickly we allow reality to escape the model.

9. Measurement Noise R Describes Sensor Uncertainty

R is the measurement-noise covariance. A precise sensor has smaller variance; a noisy one larger variance. Multiple measurement channels can be correlated, creating off-diagonal terms. But R should describe the processed measurement in the actual operating environment—not merely a data-sheet headline.

GNSS error depends on satellite geometry and multipath. Camera feature accuracy depends on texture, blur and illumination. Radar range quality depends on signal strength and target structure. Magnetometer error depends on local fields. A nominal constant R can be useful and badly wrong in changing environments.

If R is too small, the filter believes the sensor too strongly. If R is too large, useful evidence is muted. Either can look smooth on a plot while degrading truth.

10. The Innovation Is the Surprise

When a measurement arrives, the estimator predicts the measurement ẑ=Hx̂⁻ or h(x̂⁻) and forms the innovation ν=z−ẑ. NASA navigation literature calls this pre-fit measurement difference a central part of the update and diagnostic process.

A small innovation means sensor and model agree. A large one may mean the state moved unexpectedly, the sensor is wrong, measurement uncertainty was underestimated, the measurement belongs to another target, calibration has shifted, or the model is missing dynamics.

The innovation is both evidence and accusation. It is evidence used to correct the state and an accusation that one or more assumptions may no longer fit reality.

11. Innovation Covariance Tells Us How Surprising Surprise Should Be

The expected innovation covariance is S=HPHᵀ+R in the linear case. It combines predicted state uncertainty projected into measurement space with measurement noise.

A ten-metre residual is alarming if the predicted standard deviation is half a metre and ordinary if uncertainty is fifty metres. Raw residual magnitude has no meaning without scale. Normalised residuals allow statistical gating and consistency checks.

This is a recurring virtue of probabilistic estimation: the system predicts not just what measurement it expects, but a distribution of plausible deviations. It can therefore distinguish ordinary surprise from statistically implausible surprise.

12. The Kalman Gain Is Dynamic Trust

The linear Kalman gain is K=PHᵀ(HPHᵀ+R)⁻¹. It is a matrix because one measurement can move several correlated state variables in different amounts.

If predicted uncertainty is large in a direction the measurement sees well and measurement noise is small, the gain becomes larger. If the predicted state is already certain or the sensor is noisy, the correction becomes smaller. Geometry changes H and therefore changes gain even when sensor noise stays constant.

The Kalman gain is not a permanent percentage saying “trust GPS 70%.” It is a continuously recomputed relationship among model uncertainty, sensor uncertainty and observability at that moment.

13. The Update Changes Both State and Confidence

The corrected estimate is x̂⁺=x̂⁻+Kν. Covariance is reduced in directions informed by the measurement. A numerically robust form writes P⁺=(I−KH)P⁻(I−KH)ᵀ+KRKᵀ.

The first equation says where the estimate moves. The second says how much uncertainty remains after seeing the evidence. A good measurement can move the state only a little and still reduce uncertainty greatly if it confirms the prediction. A surprising measurement can move the state substantially while leaving meaningful uncertainty if noise is high.

State and confidence are related and distinct outputs. An estimator should never report one while silently inventing the other.

14. A One-Dimensional Example Shows the Logic

Predicted position is 10 metres with variance 4 m². A sensor reports 13 metres with variance 9 m². The scalar gain is K=4/(4+9)≈0.308. The innovation is 3 metres, so updated position becomes about 10.92 metres.

The sensor does not drag the estimate to 13 because its variance is larger than prediction variance. If measurement variance approached zero, K would approach one. If measurement variance became enormous, K would approach zero.

This is uncertainty-weighted averaging. The full Kalman filter adds dynamics, matrices, cross-covariance and time. The simple example is useful because it reveals the governing intuition without hiding behind notation.

15. The Kalman Filter Is More Than an Average Because It Has Memory

A static weighted average combines estimates at one instant. A Kalman filter predicts through time. It remembers how state errors correlate and how old evidence should influence current belief.

A position measurement can update velocity because position and velocity error became correlated during propagation. A later attitude observation can update gyro bias because bias has been driving attitude error in a recognisable way. A range observation can reduce clock uncertainty if the measurement model couples them.

This temporal memory is why sensor fusion is not “take a mean of GPS and IMU.” The measurements inhabit different physical spaces and times. The state model provides the language in which they can become comparable evidence.

16. The Classical Optimality Claim Has Conditions

For linear dynamics and measurement models with appropriate Gaussian white-noise assumptions, the Kalman filter propagates the exact Gaussian posterior mean and covariance. Under those conditions it has minimum-variance properties and an elegant Bayesian interpretation.

Real navigation rarely satisfies every assumption exactly. NASA’s 2025 best-practices guidance treats the linear filter as essential foundation while acknowledging that practical navigation is typically nonlinear, biased, delayed and subject to model mismatch.

The mature conclusion is not that Kalman filtering stops working outside textbook conditions. It is that the word “optimal” should not be carried into new conditions without its assumptions attached. Approximation can be excellent; it remains approximation.

17. Observability Comes Before Algorithm Choice

Imagine estimating two hidden variables when every measurement depends only on their sum. Infinite pairs can explain the same data. No estimator can separate them without additional information. This is an observability failure.

For linear systems, observability matrices or Gramians reveal whether state can be reconstructed from measurements over time. In nonlinear systems, observability becomes local and more subtle. NASA’s navigation guidance emphasises it because a sophisticated filter cannot recover information that the sensor–dynamics combination never makes distinguishable.

Estimator design should therefore ask an information question before a software question: which state directions can our measurements actually see, under which trajectories and geometries?

18. Motion Can Create Observability

A stationary platform may be unable to distinguish one sensor bias from a small orientation error. Rotate it and the two hypotheses affect measurements differently. Accelerate and another scale factor becomes visible. Move past landmarks and camera geometry reveals depth.

Calibration manoeuvres exploit this intentionally. Aircraft and spacecraft execute motion that excites dynamics. Robots move through geometry that exposes camera–IMU alignment. Survey systems change viewpoint to improve geometric conditioning.

Sometimes the best way to improve estimation is not a more complicated filter. It is a better experiment—move the system so hidden parameters leave distinct fingerprints in the measurements.

19. Initialisation Is a Prior Whether You Admit It or Not

A recursive filter needs an initial estimate x̂₀ and covariance P₀. Tiny P₀ declares strong confidence. Large P₀ declares ignorance. Those choices affect early gain and can determine whether a nonlinear filter converges.

Bad initialisation is especially dangerous in EKFs. If attitude or position begins far from truth, local linearisation can be wrong before the filter has a chance to repair itself. The filter may become confident around the wrong solution.

Robust systems use alignment procedures, coarse sensors, batch initialisation or multiple hypotheses rather than pretending the filter can always bootstrap from arbitrary error. Prior knowledge should be explicit enough to challenge.

20. Consistency Means the Errors Behave Like the Filter Claims

An estimator can have a small average error and still be statistically inconsistent if its covariance is much too small. It can also be overly conservative, reporting huge uncertainty despite excellent accuracy.

When truth is available in simulation or test, Normalised Estimation Error Squared compares actual state error with P. Normalised Innovation Squared compares innovation with S. Across Monte Carlo trials these statistics can be compared with expected probability distributions.

A filter that says “±10 cm” while missing by five metres is not merely inaccurate. It has failed at self-knowledge. Calibration of uncertainty is part of estimation correctness because downstream decisions use confidence as well as mean state.

21. Innovation Whiteness Is a Health Test

If the model is well matched and noise assumptions are sensible, innovations should be approximately zero-mean, statistically scaled as predicted and not display unexplained temporal structure.

Persistent mean can indicate sensor bias or calibration error. Autocorrelation can reveal omitted dynamics or coloured noise. Innovation variance larger than predicted suggests Q or R is too small, outliers are entering, or the model is underestimating uncertainty. Variance too small may indicate unnecessary conservatism.

The residual therefore survives the update as diagnostic evidence. A good estimator keeps asking whether the pattern of its mistakes matches the story it tells about uncertainty.

22. Sensor Fusion Works Because Sensors Fail Differently

An IMU is fast and drifting. GNSS is absolute and vulnerable to blockage. Cameras are information-rich and sensitive to visibility. Lidar measures shape and can struggle in feature-poor geometry. Radar sees through darkness and rain better than cameras and returns clutter. Wheel encoders are precise on good traction and misleading during slip.

The strongest fusion pairs do not merely add redundancy. They add complementary failure structures. One sensor bridges another’s outage; another arrests long-term drift; another observes a state direction the first cannot see.

Adding two copies of the same flawed measurement can create little new information. Adding one differently informative measurement can transform observability. Diversity matters because error diversity becomes information.

23. Inertial Navigation Gives Speed at the Price of Drift

Gyroscopes measure angular rate and accelerometers measure specific force. Integrate angular rate to update attitude. Rotate acceleration into a navigation frame, handle gravity, integrate to velocity and integrate again to position.

Small biases accumulate. A constant accelerometer bias causes velocity error to grow roughly linearly and position error roughly quadratically. Gyro bias tilts the estimated gravity direction, introducing spurious horizontal acceleration and severe drift.

Inertial navigation is therefore an excellent short-term predictor because it is high rate and self-contained. It is a poor long-term absolute reference without aiding. The drift mechanism is not a surprise; it is exactly what the mathematics predicts from repeated integration.

24. Bias Belongs in the State When It Persists

If a gyro reports true rate plus a slowly changing bias b, treating b as fresh white noise every sample throws away temporal structure. Put b into the state and model it as constant, random walk or a first-order stochastic process.

External measurements can then infer the bias indirectly. A repeated attitude residual with a characteristic drift pattern tells the filter that orientation itself may not be the only problem. Position corrections over time can reveal accelerometer bias. Wheel–IMU disagreement can reveal scale factors.

NASA devotes significant attention to bias models because persistent error deserves persistent representation. An estimator can only learn a hidden quantity if the state gives that quantity somewhere to live.

25. GNSS Provides Absolute Information and Environmental Traps

GNSS uses precise signal timing, satellite ephemerides and clock information to estimate receiver position, velocity and time. In open sky it provides powerful absolute constraints that stop inertial drift.

Urban canyons create non-line-of-sight paths and multipath. Buildings block satellites. Atmospheric effects, receiver clock behaviour and poor satellite geometry alter uncertainty. A filter that treats every fix as independent Gaussian truth can follow a reflected signal directly into a wall.

Modern fusion therefore uses measurement-level quality, geometry, consistency gates and sometimes map or inertial context. Absolute reference does not mean unquestionable measurement. The estimator still needs evidence that the fix belongs to the world it thinks it inhabits.

26. Wheel Odometry Is Excellent Until Contact Physics Changes

Wheel encoders can estimate incremental motion precisely on predictable surfaces. A differential-drive robot turns left when right wheel distance exceeds left wheel distance. Cars derive wheel speed from rotation rates and tyre models.

Then the wheel slips. The encoder remains internally precise: it reports rotation correctly. The measurement model connecting rotation to ground displacement has failed. Ice, gravel, tyre deformation and wheel lift create exactly this problem.

Fusion with IMU, visual odometry or GNSS can detect disagreement. An estimator may increase wheel-odometry noise, estimate slip states or reject the measurement temporarily. Good fusion distinguishes sensor electronics from the physical assumption that makes the reading meaningful.

27. Cameras Measure Projection, Not Position

A camera fundamentally measures image intensity. Feature detectors or neural networks extract points, lines and semantic landmarks. Geometry then projects 3D structure through camera pose and calibration into pixels.

Monocular depth is ambiguous in a single frame without priors. Motion creates parallax. Stereo creates a baseline. Known landmarks anchor absolute pose. Reprojection error becomes a measurement residual.

Camera measurement functions are strongly nonlinear and association-dependent: the system must decide which feature in one frame corresponds to which feature later. Visual–inertial estimation therefore combines state estimation with data association, calibration and often sliding-window optimisation.

28. Lidar and Radar Observe Different Geometries

Lidar produces range measurements to surfaces and can create detailed 3D point clouds. Scan matching estimates motion by aligning current geometry with earlier scans or a map. Long featureless corridors, reflective surfaces, fog and dynamic objects can degrade that geometry.

Radar measures range, angle and often Doppler velocity, providing direct radial-motion information and working through conditions that reduce optical contrast. Radar returns are sparse and can contain clutter, multipath and ghost targets.

Fusing them is powerful because their observability differs. Lidar contributes precise shape. Radar contributes velocity and weather robustness. Vision contributes texture and semantics. The mathematics builds one hidden-state story from several measurement languages.

29. Multi-Rate Fusion Means Sensors Do Not Need to March Together

An IMU may run at 400 Hz, camera at 30 Hz, GNSS at 10 Hz and barometer at 20 Hz. Forcing them into artificial synchronous packets discards timing information.

One architecture predicts the state at the highest useful rate and performs an update whenever each measurement arrives. Covariance evolves between updates according to the process model. Fast sensors dominate short-term continuity; slower absolute sensors periodically reshape the estimate.

This makes timestamps part of the measurement. A very accurate sensor attached to the wrong time can be worse than a noisy one attached correctly. Estimation is geometry through time, not just geometry in space.

30. Latency and Out-of-Sequence Measurements Turn Estimation Into History Management

A camera frame captured at time t may finish processing 80 ms later. A cloud-derived map match may arrive after several newer IMU updates. NASA navigation guidance treats measurement latency explicitly because applying old evidence at the present epoch creates systematic error.

Systems can buffer past states, rewind and repropagate, derive out-of-sequence updates, or use fixed-lag smoothers and factor graphs where every factor attaches to its actual timestamp. Each method trades memory, computation and complexity.

The broader lesson is architectural: state estimation needs a time-aware data pipeline. Late evidence is not useless, but using it correctly requires remembering enough history to revise the story.

31. Correlated Sensors Can Double-Count Information

Suppose two modules output position estimates. Module B used Module A’s output internally. Fusing A and B as independent measurements counts the same evidence twice and shrinks covariance too far.

Centralised estimators preserve cross-covariance naturally when raw measurements enter one common state. Decentralised systems can lose that dependence. Covariance intersection offers a conservative fusion rule when cross-correlation is unknown, sacrificing some precision to avoid unjustified confidence.

Information provenance matters. Two numbers from two software boxes are not automatically two independent facts. Sensor fusion needs to know not only what estimates say, but what evidence they share.

32. Nonlinearity Creates the Extended Kalman Filter

Real measurement functions include division by depth, trigonometric rotations, range square roots and nonlinear dynamics. Write xₖ=f(xₖ₋₁,u)+w and zₖ=h(xₖ)+v. Gaussian distributions transformed through nonlinear functions stop being exactly Gaussian.

The EKF linearises f and h around the current estimate using Jacobians. It propagates the state through nonlinear functions and covariance through local first-order approximations. The method is efficient and extraordinarily successful when uncertainty is modest and local curvature is manageable.

Its weakness is circular: the linearisation depends on the estimated state, and the estimated state may be wrong. Large initial errors or strong nonlinearity can produce misleading Jacobians and self-reinforcing overconfidence.

33. Error-State Attitude Filters Respect Rotation Geometry

Orientation does not live in ordinary Euclidean space. Rotation matrices must remain orthonormal. Unit quaternions must retain unit norm and q and −q describe the same physical rotation.

Error-state or multiplicative EKFs keep the main attitude on the rotation group while estimating a small local rotation error in a three-dimensional tangent space. After correction, the small error is injected into the nominal orientation and reset.

NASA’s navigation best practices devote attention to multiplicative EKFs because representation geometry affects consistency. The estimator should not ask a quaternion to behave like four unconstrained Cartesian numbers merely because four components fit neatly into a vector.

34. The Unscented Kalman Filter Approximates Nonlinearity With Sigma Points

The UKF selects a structured set of sigma points whose weighted mean and covariance match the current Gaussian estimate. It propagates those points through the nonlinear function and reconstructs a transformed mean and covariance.

This often captures curvature better than first-order Jacobian linearisation and avoids analytic derivatives. It costs more function evaluations and still compresses the posterior back into mean and covariance, so multimodality is lost.

NASA’s guidance includes sigma-point methods because they are useful alternatives in nonlinear navigation. The UKF is not “EKF but always better.” It is a different approximation with different computational and numerical trade-offs.

35. Particle Filters Keep Multiple Hypotheses Alive

A Gaussian filter represents one central mode. Some localisation problems are genuinely multimodal: a robot may be in one of several visually identical corridors or a target may have taken one of several branches.

Particle filters represent the posterior using weighted samples. Propagate samples through process dynamics, weight them by measurement likelihood, and resample when effective sample size becomes small. They can represent skewed, multimodal and strongly nonlinear distributions.

The cost grows quickly with state dimension. Resampling can impoverish diversity. Proposal design matters. Particle filters buy distributional flexibility with computational population size.

36. Smoothing Lets Future Evidence Improve the Past

A real-time filter estimates xₖ using measurements only up to k. A smoother allows measurements from later times to refine earlier states. Future evidence can reveal what a past ambiguous measurement meant.

The Rauch–Tung–Striebel smoother performs a forward Kalman filter pass and a backward recursion in linear-Gaussian problems. Fixed-lag smoothers keep only a recent moving window, trading a controlled delay for improved accuracy and natural handling of late measurements.

NASA work has long used smoothing to improve reconstructed spacecraft state. The principle is general: inference after the fact is allowed to use evidence that real-time control did not yet possess.

37. Batch Estimation and Factor Graphs See a Whole Trajectory

Batch estimation collects many measurements and solves jointly for states or parameters by minimising weighted residuals. NASA navigation descriptions explain orbit determination in exactly this language: compare actual tracking data with computed measurements from a reference trajectory and adjust the trajectory to reduce residuals.

Factor graphs express the same idea structurally. State variables become nodes; measurements and dynamics become factors connecting only the variables they constrain. The resulting least-squares problem is sparse, allowing large trajectories to be optimised efficiently.

Modern robotics often uses factor graphs because delayed constraints, loop closures and multi-sensor calibration fit naturally into a revisable trajectory rather than a purely forward recursion.

38. Robust Estimation Begins by Admitting Outliers Exist

Gaussian models assign rapidly decreasing probability to large residuals. Real sensors produce gross outliers: a mismatched image feature, reflected GNSS signal, radar ghost or corrupted packet.

Innovation gating rejects measurements whose normalised residual is implausibly large. Robust losses such as Huber functions reduce the influence of large residuals gradually. Mixture models can explicitly represent inlier and outlier populations. RANSAC searches for a geometrically consistent subset before estimation.

The aim is not to ignore surprising evidence automatically. It is to prevent one bad datum from dragging a recursive estimator so far that every good datum afterward appears wrong.

39. Data Association Is Estimation’s Hidden Discrete Problem

A radar sees five detections and tracks four vehicles. Which detection belongs to which target? A camera sees hundreds of points across frames. Which point is the same physical landmark?

Kalman equations assume the measurement model is known. Wrong association violates that assumption catastrophically. Gating narrows plausible matches. Nearest-neighbour methods choose locally. Joint probabilistic data association averages among plausible correspondences. Multiple-hypothesis tracking keeps alternative association histories alive.

Many “filter failures” are actually association failures. Continuous state estimation can only be as good as the discrete identity decisions that decide which evidence belongs to which state.

40. Multi-Target Tracking Couples Association and Motion

Tracking several objects adds birth, death, missed detections and false alarms. A track can disappear behind an occluder and reappear. Two targets can cross and become hard to distinguish.

Motion models predict where each target should be. Measurement likelihood scores possible assignments. Track-management logic decides when to initiate or delete tracks. More advanced random-finite-set filters represent uncertain target number mathematically.

The general principle mirrors all sensor fusion: prediction narrows interpretation of evidence, evidence corrects prediction, and uncertainty determines when identity should remain unresolved rather than forced prematurely.

41. SLAM Estimates the Map and the Observer Together

Simultaneous localisation and mapping asks a difficult circular question. To build a map, the robot needs to know where it is. To know where it is, the robot benefits from a map.

SLAM treats robot poses and landmarks as one coupled estimation problem. Motion factors connect successive poses. Sensor factors connect poses to landmarks. Loop closure recognises that the robot has returned to a previously seen place, imposing a long-range constraint that can correct accumulated drift.

This is why SLAM often benefits from factor graphs and smoothing. A new loop closure can improve the entire historical trajectory, not merely the current state.

42. Gauge Freedom Means Some Absolute Quantities Cannot Be Estimated at All

A purely relative SLAM problem may determine the shape of a map perfectly while remaining free to translate or rotate the entire map globally. Monocular vision may also have scale ambiguity without additional information.

These are not numerical failures. They are gauge freedoms: transformations that leave every measurement unchanged. The Hessian or information matrix becomes rank deficient in those directions.

Systems fix a reference frame, add an absolute measurement, or explicitly handle null spaces. Observability theory tells us when absolute certainty would be mathematically dishonest because the measurements never contained that information.

43. The Information Matrix Is Covariance Turned Inside Out

For a Gaussian distribution, the information matrix Λ=P⁻¹ and information vector η=Λx̂ provide an alternative representation. Independent information contributions can add naturally in information form.

Factor-graph normal equations and Fisher information connect to this perspective. Large information in a direction means small uncertainty. Zero information reveals an unobservable direction.

The Cramér–Rao bound gives a theoretical lower bound on covariance for unbiased estimators under a probabilistic model. Such bounds help separate algorithmic weakness from fundamental information scarcity: sometimes no estimator can do much better because the data simply do not contain enough information.

44. Coordinate Frames Are Part of the Mathematics

Sensor measurements arrive in frames: camera, IMU, vehicle body, Earth-centred, local navigation, map. A vector expressed in one frame cannot be added casually to a vector in another.

Rigid transforms connect frames. Rotations must be applied in the correct direction and order. Quaternions and rotation matrices require consistent conventions. Left-handed and right-handed axes, degrees and radians, north-east-down and east-north-up can all produce plausible-looking catastrophic errors.

Many estimation bugs are not advanced probability failures. They are frame errors. Formal naming and transformation discipline are part of high-performance Mathematics because representation mistakes can overwhelm any statistical sophistication downstream.

45. Numerical Stability Is a Safety Property for Covariance

A covariance matrix should remain symmetric and positive semidefinite. Floating-point roundoff, ill-conditioning and repeated subtractive updates can violate those properties in code.

Joseph-form updates, square-root filters, Cholesky factors and UD factorisation improve numerical robustness. Solvers prefer stable decompositions over explicit matrix inversion. Scaling state variables prevents one dimension measured in nanoseconds and another in kilometres from destroying conditioning.

The mathematics of estimation therefore includes numerical linear algebra. A theoretically correct covariance recursion implemented carelessly can create negative variances and unstable gains—an algebraic proof defeated by finite-precision reality.

46. Tuning Q and R Should Be Model Identification, Not Superstition

Teams sometimes tune filters by sliding Q and R until plots “look good”. That can hide modelling errors. A better process begins from sensor characterisation, physical disturbance models and residual statistics.

Allan variance can characterise inertial noise processes. Static and dynamic experiments estimate sensor covariance. Innovation sequences reveal mismatch. Monte Carlo simulation tests consistency. Adaptive methods can change noise estimates online when environment varies.

Tuning remains necessary because real systems are imperfectly known. The goal is to make tuning evidence-based: every covariance parameter should correspond to a plausible source of uncertainty and be validated against data rather than becoming an unexplained knob.

47. Model Mismatch Is More Dangerous Than Noise

Random noise broadens uncertainty. Systematic model mismatch moves the centre. A filter can survive large zero-mean noise if R is honest. A tiny unmodelled bias can accumulate into large error while covariance shrinks.

Examples include wrong wheel radius, camera focal length error, lever-arm offsets between sensors, magnetic disturbance, constant latency, wrong gravity model or a target manoeuvre absent from dynamics.

This is why validation asks not only “how noisy is the sensor?” but “what systematic structure remains in residuals?” Persistent structure is evidence the model is incomplete. Estimation quality depends more on honest modelling than on pretending every error is random.

48. Filter Divergence Is Often Confidence Running Away From Truth

Divergence occurs when estimated state error grows while the filter fails to recognise it. In nonlinear filters, bad linearisation, wrong process models, inconsistent covariance or outlier acceptance can create this feedback loop.

Once covariance becomes too small, good measurements receive low gain because the filter trusts itself more than evidence. Innovations grow, gates begin rejecting measurements, and the estimator becomes isolated from reality.

Recovery may require covariance inflation, reinitialisation, alternative hypotheses or a separate integrity monitor that detects the estimator has left its valid region. A filter should have a plan for losing confidence, not only gaining it.

49. Sensor Dropout Should Increase Uncertainty, Not Freeze the Last Answer

GNSS disappears in a tunnel. Camera loses features in darkness. Radar is blocked. A well-designed estimator continues prediction using remaining sensors while covariance grows in newly unobserved directions.

This covariance growth is useful. Downstream control can reduce speed, widen safety margins or request human intervention as uncertainty crosses thresholds. When the missing sensor returns, the larger covariance allows strong correction.

A system that simply holds the last position while the world moves is not estimating. A system that dead-reckons and admits increasing uncertainty is behaving mathematically honestly.

50. Fault Detection and Isolation Turn Residuals Into Alarms

Redundant sensors make it possible to detect disagreement. Innovations can be tested against expected distributions. Sudden bias shifts, stuck sensors and outliers create residual patterns that trigger fault hypotheses.

Fault detection asks whether something is wrong. Isolation asks which component is responsible. Reconfiguration decides how to continue after excluding or downweighting the suspect source.

Navigation integrity systems use this logic because accuracy alone is insufficient. A safety-critical system needs a bound on the probability that an undetected estimation error exceeds a protection limit. Knowing when not to trust the estimate becomes part of the product.

51. Spoofing and Adversarial Measurements Turn Estimation Into Security

A spoofed GNSS signal can be internally smooth and statistically plausible while deliberately false. An adversarial visual pattern can perturb feature matching. A compromised sensor can report measurements designed to remain just inside ordinary gates.

Resilience therefore uses cross-sensor consistency, cryptographic provenance where available, physical constraints, map plausibility and independent integrity channels. No single probabilistic gate is a security proof.

Sensor fusion helps because an attacker must fool several differently structured sources consistently. It also creates new attack surfaces because the estimator becomes a central trust engine. Estimation architecture belongs inside cybersecurity design.

52. Autonomous Cars Are Moving Estimation Laboratories

An autonomous vehicle must estimate its own pose, surrounding object states, road geometry and sometimes hidden intent. GNSS, IMU, cameras, lidar, radar and wheel encoders contribute at different rates and scales.

Urban canyons damage GNSS. Rain affects lidar and cameras. Radar sees velocity well. Lane markings constrain lateral pose. HD maps provide prior geometry. Dynamic objects complicate feature selection. Every kilometre is a changing sensor-quality experiment.

The vehicle therefore needs not one magical fusion algorithm but a hierarchy: calibration, localisation, object tracking, map alignment, integrity monitoring and fallback behaviour. Mathematics converts the sensor stack into a continuously revised model of where the car is and what surrounds it.

53. Artemis I Shows Kalman Filtering at Planetary Scale

NASA’s Artemis I navigation system used multiple EKFs for different mission phases. Atmospheric navigation estimated vehicle position, velocity and attitude during ascent and entry. Other filters handled Earth-orbit, cislunar translational states and attitude.

The filters propagated state using dynamics models and IMU measurements, then updated state and covariance with GPS where available, optical measurements of Earth or Moon in cislunar flight, and star-tracker observations for attitude.

The example matters because it dispels the idea that Kalman filtering is a classroom trick. The same predict–measure–update logic supports navigation where error can grow across hundreds of thousands of kilometres and sensors change with mission phase.

54. Drones and Robots Need Fast Local Truth, Not Only Global Position

A drone may receive GNSS at ten hertz and needs attitude corrections hundreds of times per second. The controller cannot wait for global position. It needs a fast local estimate built from IMU and aided by slower absolute sensors.

Indoor robots may have no GNSS at all. Visual–inertial odometry, lidar SLAM, wheel encoders and maps create local consistency. A warehouse robot may care more about centimetre-level repeatability relative to shelves than globally precise latitude and longitude.

“Where am I?” is application-specific. Estimation should optimise the reference frame, latency and uncertainty the controller actually needs, not chase a globally impressive metric irrelevant to the task.

55. Phones and Headsets Fuse Tiny Sensors Into Stable Experience

A smartphone or AR headset combines gyroscopes, accelerometers, magnetometers, cameras and sometimes depth sensors. Human perception is extraordinarily sensitive to latency and inconsistency: a virtual object that lags head motion by tens of milliseconds feels wrong.

Fast inertial prediction keeps orientation responsive. Vision corrects drift. Gravity constrains tilt. Magnetic or visual references constrain heading. The estimator often predicts slightly into the future to compensate display latency.

The user sees none of the covariance matrices. They see a virtual object that appears nailed to the table instead of swimming around. Mathematics disappears into perceptual stability.

56. State Estimation Extends Beyond Navigation

Industrial systems estimate temperatures and loads that are hard to measure directly. Power grids estimate electrical state from distributed meters. Battery-management systems estimate state of charge and health. Medical monitoring can fuse physiological sensors. Weather and ocean models assimilate observations into evolving dynamical models.

The specific equations differ, but the architecture repeats: hidden state, process model, observation model, uncertainty, prediction, correction and validation.

This is why Kalman’s idea became much larger than navigation. It is a general mathematical language for reasoning when reality evolves continuously and evidence arrives imperfectly.

57. Worked Example: Fusing a Fast Drifting Sensor With a Slow Absolute Sensor

Imagine a cart moving along a line. A wheel sensor reports displacement every 10 ms with tiny short-term noise and a slowly changing scale bias. An external beacon reports absolute position once per second with larger noise.

If we use only the beacon, motion looks noisy and delayed. If we use only the wheel sensor, the trajectory is smooth and drifts. Put position, velocity and wheel-scale bias into one state. Between beacons, wheel increments propagate position and covariance. When a beacon arrives, its residual corrects position and—through accumulated cross-covariance—nudges the bias estimate.

The fused estimate is smooth because high-rate odometry carries local motion and stable because the absolute beacon prevents unbounded drift. The filter has not found a perfect sensor. It has built a better instrument from two imperfect ones.

58. Worked Example: Estimating Bias From Residual Shape

A stationary IMU reports a small constant angular rate. Integrating it makes attitude drift steadily. An external reference measures orientation every few seconds. If bias is not in the state, each external update snaps attitude back and drift begins again.

Add gyro bias b to the state. The repeated external residuals now contain information about b because a persistent positive bias creates predictable directional attitude error. The covariance develops correlation between attitude and bias. Updates gradually move the bias estimate toward the true offset.

Once bias is estimated, prediction improves between external updates. The filter learned not merely “where orientation should be” but “why it kept drifting away”. That is the difference between correction and diagnosis.

59. Worked Example: A Nonlinear Range Measurement

A robot at unknown position (x,y) measures range r to a beacon at (bₓ,bᵧ): r=√((x−bₓ)²+(y−bᵧ)²)+v. One range defines a circle, not a unique position.

An EKF computes the Jacobian of range with respect to x and y at the current estimate. The measurement constrains the radial direction strongly and tangential direction weakly. A second beacon from another direction changes the geometry and can collapse the uncertainty ellipse.

If the initial estimate is on the wrong side and uncertainty is large, linearisation may be poor. A UKF or particle filter can handle that geometry differently. The example shows how nonlinear measurement, observability and uncertainty geometry interact in one simple problem.

60. Worked Example: What Happens During a GNSS Outage

A car enters a tunnel. GNSS stops. The estimator continues integrating IMU and wheel odometry. Position covariance grows along directions affected by inertial bias and wheel uncertainty. Lane geometry or map constraints may keep lateral uncertainty bounded while longitudinal uncertainty grows.

Thirty seconds later GNSS returns. The filter should not cling to its dead-reckoned position with tiny covariance. The outage should have increased uncertainty enough that the absolute measurement can correct accumulated drift strongly, subject to integrity checks.

If GNSS returns with a gross multipath error, innovation gating or map consistency may reject it. Robust estimation therefore needs both humility during outage and scepticism during recovery.

61. Validation Requires Truth Data, Simulation and Deliberate Failure

Filter validation should not end with one smooth trajectory plot. Use high-grade reference systems or controlled ground truth. Run Monte Carlo simulations with known states and noise to test NEES and NIS consistency. Inject sensor biases, dropouts, latency and outliers. Vary initial conditions.

Hardware-in-the-loop testing checks timing and software paths. Replay logs let teams compare algorithm versions against identical measurements. Field testing exposes environmental errors absent from simulation. Failure cases deserve their own benchmark set.

A filter is not validated because it worked once. It is validated when its accuracy, uncertainty and failure behaviour remain explainable across the conditions it claims to handle.

62. Primary Mathematics: Estimation Begins With Weighted Evidence

Primary students can understand the core idea without matrices. Two thermometers disagree. One is usually more reliable. How should we combine them? A child can compute an average, then see why an ordinary average treats both equally even when one is much noisier.

Graphs show a prediction and a new measurement. Error bars show uncertainty. Repeated measurements show random variation. A simple moving object introduces position, speed and time.

The educational job is not to teach Kalman equations early. It is to build the habit that measurements are evidence with quality, and that a good answer includes how certain we should be.

63. Secondary and Advanced Mathematics: The Filter Reveals Why We Learn Matrices and Probability

Secondary students add algebra, functions, trigonometry, vectors, coordinate geometry and probability. Covariance becomes a way to describe spread and correlation. Motion models become matrices. Sensor geometry becomes trigonometry. Residuals become statistical tests.

Advanced study adds multivariate Gaussian distributions, eigenvalues, Jacobians, differential equations, optimisation, Bayesian inference and numerical linear algebra. Factor graphs connect estimation to sparse optimisation. Lie groups handle rotations. Particle filters connect inference to Monte Carlo methods.

State estimation is a remarkable demonstration that abstract school topics eventually cooperate inside one machine trying to answer a very ordinary question: where am I?

64. Why This Improves the World

• It makes imperfect sensors useful together. Complementary error structures create estimates no single device can sustain.
• It gives uncertainty operational meaning. Controllers and safety systems can react not only to the estimate but to confidence.
• It keeps systems working through temporary blindness. Prediction bridges outages while covariance records the growing cost.
• It enables modern navigation. Spacecraft, aircraft, cars, drones, robots and phones rely on continuous state inference.
• It makes hidden biases learnable. Persistent residual structure becomes information about calibration and model error.
• It creates diagnostic self-awareness. Innovations, consistency tests and integrity monitors reveal when assumptions are breaking.
• It connects real-time control with offline knowledge. Filters, smoothers and factor graphs let systems trade latency for accuracy intelligently.

The deeper contribution is epistemic. State estimation gives machines a disciplined way to say, “This is what I currently believe about the world, this is how uncertain that belief is, and this is the evidence that would make me change my mind.”

65. What Mathematics Does Not Do

A Kalman filter does not create information the sensors never measured. A covariance matrix does not guarantee calibration. An EKF does not make severe nonlinearity linear. A smooth trajectory does not prove accuracy. More sensors do not guarantee better fusion when their errors are correlated or their calibration is wrong.

State estimation does not protect automatically against spoofing, bad data association, incorrect frames, time-sync errors or omitted biases. It does not know that a measurement model became invalid because tyres began slipping unless the system contains enough evidence or logic to detect it.

And an estimator never possesses “the true state” in the philosophical sense. It maintains a model-conditioned belief supported by measurements. Its intellectual strength comes from making that dependence explicit.

Frequently Asked Questions

What is state estimation?

State estimation infers hidden system variables such as position, velocity, orientation or bias from models and imperfect measurements, usually together with a quantitative representation of uncertainty.

What is sensor fusion?

Sensor fusion combines information from multiple sensors or estimators so complementary measurements constrain a common hidden state. Good fusion accounts for timing, measurement models, uncertainty and correlation rather than simply averaging outputs.

What is a Kalman filter?

A Kalman filter is a recursive state estimator for linear dynamical systems with probabilistic uncertainty. It alternates prediction through a process model with measurement correction weighted by covariance.

What does covariance mean?

Covariance describes expected state-estimation error magnitude and correlation. Its diagonal entries are variances; off-diagonal entries show how errors in different state components are related.

What is an Extended Kalman Filter?

An EKF applies Kalman-style estimation to nonlinear systems by linearising process and measurement models around the current estimate using Jacobians.

What is an Unscented Kalman Filter?

A UKF propagates a structured set of sigma points through nonlinear functions and reconstructs the transformed mean and covariance instead of using first-order Jacobian linearisation.

When are particle filters useful?

Particle filters are useful when the state distribution is strongly nonlinear, non-Gaussian or multimodal. They represent the posterior using weighted samples at greater computational cost.

What is observability?

Observability describes whether the hidden state can be distinguished from available measurements over time. An unobservable state direction cannot be estimated uniquely regardless of algorithm sophistication.

Why can a filter become confidently wrong?

If model error, bias, correlation, nonlinearity or bad measurements are underestimated, covariance can shrink while real error grows. The estimator then downweights corrective evidence and can diverge.

What is the difference between filtering and smoothing?

Filtering estimates current state using measurements up to the present. Smoothing also uses future measurements to refine past state estimates, improving accuracy when latency is acceptable.

Sources and Further Reading

• NASA Technical Reports Server, Navigation Filter Best Practices, 2025 — current NASA guidance on EKFs, covariance, measurement processing, bias modelling, sigma-point methods and navigation-filter design.
• NASA Technical Reports Server, Extended Kalman Filter Performance on the Artemis-1 Mission — application of multiple EKFs using IMU, GPS, optical navigation and star-tracker measurements.
• NASA Science, Basics of Space Flight: Navigation — orbit determination, residuals and batch estimation.
• NASA NTRS, The Kalman Filter — state prediction, covariance propagation, innovation and measurement update equations.

Continue Through eduKateSG

Continue with How Mathematics Works. Compare this article with Giving Every Place on Earth an Address, where GNSS geometry turns timing into position, and with Sending a Spacecraft to Somewhere That Is Moving, where navigation state feeds trajectory control.

For the boundary between probability and search effort, read Searching the Right Place First When Someone Is Missing. Both fields update belief under imperfect evidence; state estimation tracks a dynamical hidden state while Bayesian search allocates action across uncertain locations.

Final Thought: Navigation Is Not Knowing; It Is Knowing How to Update

The IMU says the vehicle kept moving.

The wheel says slightly less.

The camera sees the landmark farther left than predicted.

GNSS has disappeared behind concrete.

The estimator predicts.

Uncertainty grows.

A measurement arrives late.

The model translates it into the state’s language.

Covariance decides how much it deserves to matter.

Residuals ask whether the entire story still makes sense.

Then another tenth of a second passes and the world changes again.

Mathematics improves the world here not by giving machines perfect senses.

It gives them a disciplined way to remain useful when every sense is slightly wrong.

---

Advanced Field Manual | Information, Geometry and Calibration

The main article explains the estimator as an operating system for uncertain sensors. This advanced field manual keeps going into the information geometry, calibration and robustness work that usually decides whether a real filter survives deployment. These sections are intentionally technical: they are the reason this article earns longform scale rather than stopping at the textbook predict–update loop.

66. Information Gain Explains Why Some Measurements Are Worth More Than Others

A measurement is valuable when it reduces uncertainty in state directions that matter. In Gaussian problems, covariance volume can be summarised through determinants. A measurement that shrinks det(P) substantially provides large global information gain; a measurement that only refines an already precise direction may have little practical value.

This perspective helps sensor scheduling. If a spacecraft can point one camera at only one landmark, choose the observation predicted to reduce uncertainty most. If a robot can activate an expensive lidar only intermittently, trigger it when covariance has grown in directions the cheaper sensors cannot constrain.

Information-aware sensing turns estimation into active perception. The system does not merely wait for measurements. It chooses actions that make future state easier to know.

67. Entropy Gives Uncertainty a Volume

For a multivariate Gaussian, differential entropy depends on log det(P). This connects covariance to an information-theoretic measure of uncertainty volume. Lower determinant means a tighter ellipsoid and lower entropy, though two covariances with the same determinant can distribute uncertainty differently across directions.

Entropy is useful when planning measurements because it gives one scalar objective for expected uncertainty reduction. But safety problems often care about specific directions rather than total volume. A navigation system may tolerate horizontal uncertainty and care intensely about vertical clearance.

No scalar metric replaces the full covariance. It compresses it for one decision. Choosing the compression is part of the engineering objective.

68. Mutual Information Measures How Much a Sensor Can Teach About the State

Mutual information between a prospective measurement and the hidden state quantifies expected reduction in uncertainty before seeing the actual measurement value. It can rank candidate viewpoints, landmarks or sensor activations.

A camera view containing many distant collinear features may add less pose information than a slightly longer route revealing nearby features at diverse depths. A range measurement from a geometry already duplicated by other beacons may be less useful than one from a new direction.

This is the mathematical version of a practical instinct: do not ask another question that tells you what you already know. Seek measurements whose likely answers separate the remaining hypotheses.

69. Sensor Placement Is an Estimation Problem Before Any Sensor Is Switched On

Where should beacons, cameras or monitoring stations be installed? Placement determines observability and covariance before the estimator begins. Sensors clustered in one region can leave another state direction weakly constrained. Wide baselines can improve triangulation; poor geometry can amplify measurement noise.

Facility placement for sensing often optimises Fisher information, expected covariance or worst-case uncertainty. Constraints include cost, line of sight, maintenance and failure resilience.

The estimator cannot repair fundamentally bad sensor geometry after deployment. Sometimes the best algorithmic upgrade is moving a sensor two metres, adding one new viewpoint or changing a trajectory so the geometry becomes informative.

70. Fisher Information Links Measurement Sensitivity to Achievable Precision

Fisher information describes how sensitively the probability of measurements changes with the state. In regular estimation problems, the inverse Fisher information appears in the Cramér–Rao lower bound for unbiased estimators.

If two very different state values produce almost identical measurement distributions, information is low and variance must remain large. If tiny state changes create clear measurement changes relative to noise, information is high.

This provides a benchmark independent of one specific filtering algorithm. When a state estimate is poor, ask whether the algorithm is inefficient or the sensing problem itself contains little information. The lower bound distinguishes computational weakness from physical ambiguity.

71. Dilution of Precision Is Geometry Converted Into Error Amplification

GNSS makes geometry visible through dilution-of-precision measures. Satellites spread widely across the sky constrain position better than satellites clustered in one region because their line-of-sight vectors span space more effectively.

The same principle appears in triangulation, acoustic localisation and landmark-based navigation. Measurement noise is transformed through the inverse geometry matrix into state covariance. Poor geometry amplifies noise.

This is why a sensor can report unchanged precision while the final navigation error changes dramatically. Measurement quality is not only a property of the sensor. It is a property of sensor, state and geometry together.

72. Experiment Design and Estimation Meet in Calibration Manoeuvres

To estimate a parameter, excite it. A camera–IMU calibration performed while the device barely rotates may leave rotational alignment weakly observable. A wheel-radius calibration driven only straight may fail to separate left and right scale errors. An aircraft bias calibration may require manoeuvres that create distinct acceleration and rotation signatures.

Optimal experiment design chooses inputs or trajectories that maximise information about unknown parameters while respecting safety and operational constraints.

The lesson is important for students: measurement is not always passive. Sometimes we learn by choosing how to perturb the system. The right movement can make hidden parameters mathematically visible.

73. Extrinsic Calibration Estimates How Sensors Are Mounted Relative to One Another

A camera and IMU may be bolted to the same device and still have different coordinate origins and orientations. Their rigid transform—rotation and translation—is called extrinsic calibration.

A few degrees of rotational error can create large apparent motion disagreement. A lever arm matters during rotation because sensors at different locations experience different translational velocity and acceleration.

Extrinsics can be measured mechanically, calibrated offline or estimated online as state variables. Online estimation is powerful when mounting changes and dangerous when motion does not excite enough geometry. Calibration parameters deserve observability analysis like any other hidden state.

74. Time-Offset Calibration Is Often as Important as Spatial Calibration

Two perfectly calibrated sensors can disagree because their clocks do. If one camera timestamp lags the IMU by 20 ms during rapid rotation, projected features appear inconsistent with predicted attitude.

Estimate a time offset parameter and the model becomes h(x(t+δt)). Motion creates sensitivity to δt. Without sufficient motion, the time offset may remain unobservable.

High-dynamic systems often need hardware synchronisation plus software verification of latency. A calibration report that lists millimetres and arcminutes but ignores milliseconds is incomplete.

75. Lever Arms Couple Rotation to Translation

Place a GNSS antenna one metre from a vehicle’s rotation centre. During yaw, the antenna traces an arc even if the centre remains nearly stationary. An IMU elsewhere experiences rotational acceleration terms.

Measurement models must transform sensor locations through the rigid-body kinematics. Ignoring lever arms can look like unexplained velocity or acceleration bias during turns.

This is a small geometric detail with large estimator consequences. Physical layout belongs in the mathematical model because “the vehicle position” and “the sensor position” are different states connected by motion.

76. Scale Factors Turn Multiplicative Error Into Hidden State

A wheel encoder may report 1.01 times true distance because effective tyre radius differs from nominal. An accelerometer may have gain error. A camera focal-length estimate can be slightly wrong.

Scale error is multiplicative, not additive. A one-percent error looks small over one metre and becomes ten metres over a kilometre of dead reckoning.

Including scale factors in the state lets external measurements estimate them when motion excites the parameter. The estimator then improves future propagation rather than repeatedly correcting the symptom after it accumulates.

77. Temperature Turns Calibration Into a Function

Sensor bias and scale can depend on temperature. A calibration performed at 20°C may not remain valid in an engine bay at 60°C or at high altitude in cold conditions.

One strategy uses laboratory calibration tables. Another augments bias models with temperature dependence. Another lets online estimation absorb residual variation with process noise.

The best approach depends on repeatability and observability. The important idea is that “sensor calibration” may be a surface over environment rather than one permanent number.

78. Adaptive Q and R Try to Learn When the World Changes

A fixed R tuned in clear weather may underestimate camera error in rain. A fixed Q tuned for cruising may underestimate manoeuvre uncertainty during aggressive turns.

Adaptive filters infer changing noise levels from innovation statistics, operating mode or learned context. Covariance matching compares observed residual variance with predicted S and adjusts Q or R. Multiple-model approaches switch among pre-tuned regimes.

Adaptation can improve robustness and also destabilise a filter if it interprets model error as sensor noise indiscriminately. Learning uncertainty is itself an estimation problem and needs constraints.

79. Innovation Covariance Matching Is a Diagnostic Before It Is a Tuning Algorithm

If predicted innovation covariance S repeatedly differs from empirical innovation covariance, something in P, Q, R or the measurement model is inconsistent.

Blindly scaling R until the numbers match may hide process-model error. A bias can inflate residual variance. Unmodelled correlation can make residuals non-white. Outliers can dominate sample statistics.

Use covariance matching first as a question: why is observed surprise different from predicted surprise? Tuning should follow diagnosis, not replace it.

80. Interacting Multiple-Model Filters Represent Several Motion Stories at Once

A target may cruise straight, turn or accelerate. One process model cannot describe every mode well without enormous Q. Interacting Multiple Model, IMM, estimators maintain several filters—perhaps constant velocity, coordinated turn and acceleration—along with probabilities for each mode.

State estimates mix according to mode probabilities, filters update separately, and evidence shifts probability among models. The system can adapt rapidly when behaviour changes while retaining a precise model during steady motion.

This is a structured alternative to saying “the target might do anything”. Several simple hypotheses can outperform one vague model when mode changes have meaningful physical interpretation.

81. Heavy-Tailed Noise Makes Gaussian Confidence Fragile

Real residual distributions often contain more extreme values than a Gaussian predicts. Multipath, feature mismatches and intermittent interference create heavy tails.

Student-t measurement models, mixture distributions and robust M-estimators give large residuals less catastrophic influence. Instead of treating every measurement as either perfect Gaussian inlier or total rejection, robust likelihoods create gradual downweighting.

The trade-off is efficiency under truly Gaussian noise versus resilience to contamination. Robust statistics accepts slightly less ideal-case precision to avoid enormous real-world failures.

82. M-Estimators Replace Squared Error With a More Cautious Penalty

Least squares penalises residual r with r². That gives extreme residuals enormous influence. A Huber loss is quadratic near zero and approximately linear for large residuals. Cauchy and Tukey losses suppress extremes more strongly.

In factor-graph optimisation, robust losses can sit on camera or lidar factors so one mismatch does not rotate the entire trajectory. In recursive filters, robust update rules modify effective measurement weight.

No loss function knows whether a surprising measurement is a genuine manoeuvre or a bad datum. Robustness manages influence; system context still decides interpretation.

83. H-Infinity Filtering Asks for Robust Error Bounds Instead of Exact Noise Statistics

Kalman filtering is probabilistic: it relies on noise covariances and distributional assumptions. H∞ filtering takes a more worst-case control perspective, seeking to bound how much disturbance energy can amplify into estimation error.

This can be attractive when noise statistics are poorly known and robustness matters more than optimal Gaussian performance. The resulting estimator can be conservative.

The contrast is instructive. One method asks “what is optimal under this probability model?” Another asks “what error amplification can I guarantee against a bounded class of disturbances?” Estimator choice is therefore partly a philosophy of uncertainty.

84. The Right Estimator Depends on the Shape of Uncertainty, Not Fashion

Linear Kalman filter for linear Gaussian structure. EKF for efficient local nonlinear approximation. UKF when sigma-point propagation is attractive. Particle filter when multimodality matters. Factor graph when long histories, loop closure and delayed constraints justify optimisation. Robust filters when outliers or model uncertainty dominate.

The wrong question is “which estimator is best?” The right questions are: what is the state dimension, what are the nonlinearities, how multimodal is the posterior, how much latency is allowed, what computation is available, which assumptions can be validated, and what failure mode matters most?

Algorithm selection should follow the geometry of the problem, not the novelty of the method.

85. Active Estimation Closes the Loop Between Knowing and Acting

An autonomous system can choose actions partly to improve its estimate. A drone can circle a landmark to improve depth geometry. A robot can rotate to calibrate IMU bias. A telescope can point at a calibration source. A tracking radar can schedule a beam where uncertainty is growing fastest.

This couples control and estimation. Actions change the physical state and the information available about that state. In partially observable decision problems, the belief state itself becomes part of planning.

Mathematics then answers a deeper question than “where am I?” It asks “what should I do next so that, after acting, I will know enough to act better?”

Advanced Field Manual | Integrity, Learning and the Edge of Trust

86. Marginalisation Compresses Old States Into a Prior

A sliding-window estimator cannot keep every past state forever. To remove old variables without discarding their information, it marginalises them. In Gaussian problems, the Schur complement folds their effect into a prior factor on variables that remain.

This compression is powerful and subtle. If a nonlinear factor was linearised before marginalisation, the resulting prior remembers that old linearisation point. Later relinearisation of remaining states can create inconsistency because the prior cannot be relinearised against variables that no longer exist.

Sliding-window visual-inertial systems therefore manage marginalisation carefully. Memory efficiency has an inference cost: once history is compressed, some flexibility to reinterpret the past is gone.

87. The Schur Complement Is the Algebra Behind Eliminating Nuisance Variables

Partition a linear system into variables we want to keep and variables we want to eliminate. Solving the eliminated block analytically produces a smaller system for the retained variables through the Schur complement.

Bundle adjustment uses this idea to eliminate thousands of landmark variables before solving for camera poses. Factor graphs use variable elimination orders to exploit sparsity. Marginalisation in fixed-lag smoothers uses the same algebra to convert old states into dense priors.

One matrix identity therefore connects numerical linear algebra to practical estimation architecture. Mathematics improves speed not only by inventing faster hardware but by choosing which unknowns need to remain explicit.

88. IMU Preintegration Compresses Hundreds of Samples Into One Motion Factor

Between two camera frames, an IMU can produce dozens or hundreds of measurements. Re-integrating all of them from scratch every time an optimiser changes the starting pose would be expensive.

IMU preintegration combines the high-rate measurements into a relative motion constraint with associated covariance and bias sensitivity. When the optimisation adjusts pose or bias slightly, the preintegrated factor can be corrected efficiently without replaying every raw sample.

This is a beautiful example of mathematical compression preserving the information relevant to a slower optimisation layer. The raw stream disappears; its integrated effect survives as one structured factor.

89. Relinearisation Is Why Smoothing Can Beat a Forward EKF

An EKF linearises a measurement once at the current estimate and then moves on. If that estimate later changes substantially, the old update was based on a Jacobian evaluated at a state we no longer believe.

Nonlinear smoothing retains factors and can relinearise them around improved state estimates. A loop closure arriving minutes later can trigger a new solution in which old camera observations are interpreted from better poses.

The improvement costs computation and latency. It is another version of the general trade-off: real-time filtering commits early; smoothing keeps the past negotiable longer.

90. EKF-SLAM Can Become Inconsistent When Linearisation Invents Information

In SLAM, global translation and rotation may be unobservable. Certain EKF linearisations can accidentally break those symmetries and make the filter believe it learned absolute information that the measurements never supplied.

The result is inconsistency: covariance becomes too small in gauge directions. Observability-constrained and invariant filtering methods were developed partly to preserve the correct unobservable subspace.

This failure is profound because no individual sensor is wrong. The error comes from the approximation method changing the information structure of the problem. An estimator can manufacture certainty mathematically if its linearisation violates the underlying geometry.

91. Invariant Filtering Uses Symmetry as Part of the Estimator

Many navigation systems evolve on Lie groups such as SE(3), the group of rigid-body poses. Invariant EKFs define estimation error in ways aligned with group symmetries, producing error dynamics that can be less dependent on the estimated trajectory and better preserve observability structure.

The details are advanced, but the principle is accessible: if the physical problem has a symmetry, build that symmetry into the error representation instead of hoping a generic Euclidean approximation respects it accidentally.

Geometry is not decoration around probability. It can determine which probabilistic approximation remains consistent.

92. Learned Measurement Models Can Help and Can Hide Failure

Neural networks can estimate depth, optical flow, object detections or uncertainty from raw sensor data. Those outputs can become measurements inside a classical estimator.

This hybrid architecture is attractive because learning handles complex perception while state estimation provides temporal structure and uncertainty propagation. But learned measurements may be miscalibrated, correlated across frames and brittle outside training distribution.

The filter should not assume a neural network’s confidence score is a statistically valid covariance. Learned uncertainty needs calibration and out-of-distribution testing like any other sensor model.

93. A Neural Network Can Learn Q, R or Residual Corrections

One research direction lets machine learning adapt process noise, measurement noise or model residuals from context. A vehicle can learn that wheel odometry is less reliable on gravel. A camera model can predict feature uncertainty from blur and texture.

This can make a filter responsive to conditions difficult to hand-model. It can also create a hidden feedback loop: the learned module changes covariance, covariance changes state, state changes features presented to the learned module.

Hybrid systems need validation at the combined-system level. A learned covariance is useful only if downstream statistical tests confirm that its confidence matches real error.

94. Differentiable Filters Make Estimation Part of End-to-End Learning

Kalman-style operations can be implemented differentiably so gradients flow through prediction and update during neural-network training. The network can learn dynamics, observation models or noise parameters while the filter imposes a structured probabilistic state architecture.

This blends model-based and data-driven reasoning. The learned component does not need to rediscover from scratch that uncertainty should propagate through time; the estimator supplies that inductive bias.

The danger is opacity. A system trained end-to-end can fit a dataset while violating physical meaning in its hidden state. Interpretability and out-of-distribution validation remain necessary.

95. Out-of-Distribution Conditions Are Estimation’s Unspoken Exam

A filter tuned on sunny suburban roads enters snow. A drone calibrated in still air encounters rotor wash near a wall. A phone model trained on handheld motion is mounted rigidly to machinery.

Noise distributions, biases and dynamics can all change. Innovation statistics may drift before state error becomes obvious. A robust system monitors context and residual health rather than assuming training conditions persist forever.

Generalisation in estimation is not only an AI issue. Every probabilistic model has a domain. Deployment asks whether the current world still belongs to it.

96. Integrity Is the Probability That the Error Is Safely Bounded

Accuracy asks how close an estimate usually is. Integrity asks whether the system can bound dangerous error and alert when that bound cannot be trusted.

Safety-critical navigation may compute protection levels designed to bound position error with very high probability under defined fault models. If the protection level exceeds an operational limit, the system should not continue claiming the navigation solution is safe for that task.

This is a crucial maturation of uncertainty. Confidence is not merely a descriptive error bar. It becomes an operational contract deciding whether automation may proceed.

97. RAIM Shows How Redundancy Can Test Itself

Receiver Autonomous Integrity Monitoring in GNSS uses redundant satellite measurements to test consistency and, in more advanced forms, identify faulty measurements.

If there are more measurements than minimally required for the state, the extra redundancy creates residual degrees of freedom. Inconsistent measurements can produce test statistics exceeding thresholds.

The principle generalises far beyond aviation navigation: redundancy is valuable not only because it improves average accuracy, but because it lets the system ask whether its measurements agree with one another.

98. Protection Levels Turn Covariance Into a Decision Boundary

Suppose an automated landing system can tolerate ten metres of horizontal error. A navigation estimate with two-metre RMS error may sound excellent, but rare undetected faults can dominate safety risk.

A protection level incorporates statistical uncertainty and fault hypotheses to produce a conservative bound. If the bound is twelve metres, the system should declare the navigation service unavailable for that landing requirement even if the estimated position looks smooth.

This distinction—estimate versus guaranteed-enough bound—is why integrity engineering is more demanding than ordinary accuracy benchmarking.

99. Common-Mode Failure Defeats Naive Redundancy

Three GNSS receivers fed by the same spoofed signal are not three independent navigation sources. Two cameras blinded by the same fog are not independent. Redundant IMUs on one vibrating mount can share mechanical disturbance.

Reliability improves most when redundancy spans different failure mechanisms. Sensor diversity, independent power, separate processing paths and different physical principles reduce common-mode risk.

Probability calculations that assume independence can be dangerously optimistic when failure is correlated. Redundancy should be designed around causality, not component count.

100. Graceful Degradation Is Better Than Pretending Nothing Happened

A navigation system loses lidar. Instead of maintaining the same performance claim, it may continue with camera and IMU while widening uncertainty. Lose camera next and the system may reduce speed or return control to a human.

Graceful degradation maps estimator health to operational capability. Different sensor subsets have different observability and uncertainty growth. The system should know which tasks remain safe under each configuration.

This turns fusion architecture into resilience architecture. Failure need not mean immediate collapse, but it should mean honest change in what the system claims it can do.

101. Fallback Modes Should Be Designed Before the Primary Estimator Fails

When a sophisticated estimator diverges, the system may need a simpler independent fallback: raw GNSS with conservative limits, attitude from a separate sensor, low-speed odometry or a safe-stop controller.

The fallback should not share every failure mechanism with the primary path. Otherwise a single bug or spoofed measurement can defeat both.

High-assurance estimation therefore asks not only “how accurate is our best solution?” but “what is our simplest defensible state estimate after we discover the best solution is no longer trustworthy?”

102. Simulation Needs a Truth Model More Detailed Than the Filter Model

If simulation uses exactly the same equations and noise assumptions as the estimator, validation becomes circular. The filter is being tested in a universe designed to make it correct.

A stronger truth model includes higher-order dynamics, realistic sensor bias, latency, dropouts, correlation, environmental effects and faults absent from the estimator. Then the estimator must survive model mismatch closer to real deployment.

Simulation should challenge assumptions, not merely confirm implementation. The goal is to discover which simplifications remain safe and where the estimator’s confidence becomes optimistic.

103. Monte Carlo Testing Measures Statistical Honesty

Run one simulation and the result may be luck. Run thousands with random initial error, process noise, measurement noise and failures. Compare actual error distributions with predicted covariance.

NEES and NIS confidence intervals reveal whether the estimator is systematically overconfident or conservative. Failure rate reveals tail behaviour hidden by RMS accuracy. Scenario stratification shows which operating regimes dominate risk.

State estimation is probabilistic, so validation should be probabilistic too. One pretty trajectory cannot validate a probability model.

104. Replay Testing Makes Algorithm Comparison Fair

Record raw sensor data from a difficult drive, flight or robot mission. Replay the identical stream through old and new estimator versions. Differences now come from software rather than environment.

Replay supports regression testing and rapid diagnosis. Add metadata marking tunnels, multipath zones, feature loss and aggressive manoeuvres. Compare not only mean error but covariance calibration, fault flags and recovery time.

The limitation is closed-loop interaction: a new estimator in real operation might change control actions and therefore future measurements. Replay is powerful evidence and not a replacement for live system testing.

105. A Professional Estimator Has a Failure Library

Maintain named scenarios: GNSS multipath, total GNSS outage, wheel slip, camera glare, feature-poor corridor, lidar rain, magnetometer disturbance, IMU saturation, time-sync jump, duplicated measurement, stale measurement, incorrect association, sensor spoofing, unmodelled manoeuvre.

Each scenario should specify expected estimator behaviour: error growth, covariance response, alarms, rejected sensors, fallback state and recovery.

This turns operational history into a reusable test suite. A mature estimator is not defined only by how it performs when everything works. It is defined by how predictably it behaves when the world stops matching the assumptions.

Closing Return: The Best Estimate Knows the Shape of Its Ignorance

A sensor gives a number.

A second sensor disagrees.

The model predicts something between them.

Covariance asks which disagreement is plausible.

Observability asks whether the hidden states can be separated at all.

Calibration asks whether the sensors speak the coordinate language the equations assume.

Integrity asks whether the remaining uncertainty is safe enough for the next action.

Then the world changes and the estimate has to begin again.

The remarkable achievement is not that Mathematics tells a machine exactly where it is.

It tells the machine how to combine what it knows, what it merely suspects and what it should still be afraid of being wrong about.