Appendix J. Notation and Glossary

This appendix collects the symbols and terms used throughout the book in one place. The notation table fixes the meaning of each symbol so that a formula pulled from Chapter 4 reads the same way as one from Chapter 18. The glossary gives short, working definitions of the vocabulary that recurs across parts. Neither section is a substitute for the chapter that introduces a concept in full; both are meant for quick reference while you read or build.

J.1 Notation

The book keeps a single convention: a physical scene has a hidden state \(s\), a sensor turns that state into an observation \(y\) or a signal \(x\), and an estimator recovers an estimate \(\hat{s}\) of the state. The measurement model that ties them together is written

$$ x = h(s) + \eta, $$

where \(h\) is the (possibly nonlinear) sensor response and \(\eta\) is measurement noise. When we reason under uncertainty we track a belief \(p(s \mid x)\), the posterior over the state given what the sensor reported. The table below lists the symbols in the order you will most often meet them.

SymbolMeaning
\(x(t)\)Continuous-time signal, for example a raw voltage or pressure trace
\(x[n]\)Discrete-time (sampled) signal, index \(n = 0, 1, 2, \dots\)
\(t\)Time (seconds); \(n\) is its discrete sample index
\(f_s\)Sampling rate in samples per second (hertz)
\(s\)Hidden state of the world we want to know (position, class, health)
\(y\)Observation or label produced from the sensor stream
\(h(\cdot)\)Measurement or observation model mapping state to signal
\(\eta\)Measurement noise, often modeled as zero-mean Gaussian
\(\hat{s}\)Estimate of the state produced by an estimator or network
\(p(s \mid x)\)Belief: posterior probability of state \(s\) given signal \(x\)
\(p(x \mid s)\)Likelihood of the signal under a state hypothesis
\(\theta\)Learnable model parameters (weights and biases)
\(f_\theta(\cdot)\)A model (network) with parameters \(\theta\)
\(\mathcal{L}\)Loss (objective) minimized during training
\(\mathcal{D}\)Dataset of examples \(\{(x_i, y_i)\}_{i=1}^{N}\)
\(N\)Number of examples, or number of samples in a window
\(\mathbb{E}[\cdot]\)Expectation over the indicated distribution
\(\sigma\)Standard deviation, or a nonlinearity when written \(\sigma(\cdot)\)
\(\mathbf{K}\)Kalman gain in a recursive state estimator
\(\|\cdot\|\)Norm, defaulting to the Euclidean (\(\ell_2\)) norm
\(\nabla_\theta\)Gradient with respect to the parameters \(\theta\)

A few habits keep the notation unambiguous. Bracketed indices such as \(x[n]\) always mean a sampled signal; parenthesized arguments such as \(x(t)\) always mean a continuous one. A hat marks an estimate, so \(\hat{s}\) is never the true state \(s\). Bold symbols such as \(\mathbf{K}\) denote vectors or matrices, while plain italics denote scalars. Where a chapter needs an extra symbol it defines it locally, but it never reassigns one from this table.

J.2 Glossary

Definitions below are deliberately short. Each term has a home chapter where it is developed with examples; use this list to jog your memory, not to learn a concept for the first time.

Aliasing
The false low-frequency content that appears when a signal is sampled below twice its highest frequency (the Nyquist rate). Once present it cannot be removed, so it is prevented with an anti-aliasing filter before sampling.
Bias
A systematic offset between what a sensor or estimator reports and the truth. Distinct from noise, which averages out; bias does not.
Calibration
Adjusting a sensor or model so its outputs match a trusted reference. For probabilistic models it also means making predicted confidences match observed accuracy.
Conformal prediction
A distribution-free method that turns any point predictor into prediction sets or intervals with a guaranteed coverage rate, using a held-out calibration split.
Domain shift
A change between the data a model was trained on and the data it meets in deployment, for example a new sensor, season, or site. It is the usual cause of silent accuracy loss.
Drift
A slow change over time in a sensor's response or in the data distribution, such as a gas sensor whose baseline creeps as it ages. Corrected by recalibration or adaptation.
Foundation model
A large model pretrained on broad data that serves as a reusable starting point, fine-tuned or prompted for many downstream sensing tasks rather than trained from scratch.
Inertial measurement unit (IMU)
A package of accelerometers and gyroscopes (often a magnetometer) that reports motion. Central to dead reckoning, gesture sensing, and fusion with cameras or GPS.
Kalman filter
A recursive estimator that fuses a motion model with noisy measurements to track a state, updating the estimate \(\hat{s}\) and its uncertainty through a predict step and a correct step weighted by the Kalman gain.
Leakage
When information from the test set or the future sneaks into training, inflating reported accuracy. Common in sensor data through overlapping windows or per-subject splits done wrong.
Micro-Doppler
The small frequency shifts a radar sees from the moving parts of a target, such as swinging arms or spinning rotor blades, used to classify what is moving, not just that something moves.
Nyquist rate
Twice the highest frequency present in a signal; the minimum sampling rate that captures it without aliasing.
Occupancy
A representation of which regions of space are filled, free, or unknown. Occupancy grids and occupancy networks let a robot reason about where it can move.
Odometry
Estimating change in position over time from onboard sensors (wheels, an IMU, or a camera). Accurate over short spans but subject to accumulating drift.
Point cloud
A set of 3D points, typically from lidar or depth cameras, describing surfaces in a scene without an imposed grid.
Remaining useful life (RUL)
The predicted time or cycles a machine or component has left before failure, a core target of predictive maintenance from vibration, temperature, and current signals.
Sensor fusion
Combining measurements from several sensors so the joint estimate is better than any one alone, exploiting complementary strengths (radar range plus camera semantics, for instance).
Self-supervised learning
Learning useful representations from unlabeled data by solving a pretext task the data supplies its own answer to, such as predicting a masked segment of a signal. Well suited to sensing, where raw data is plentiful and labels are scarce.
Signal-to-noise ratio (SNR)
The ratio of signal power to noise power, usually in decibels. It sets a ceiling on how well any downstream model can perform.
SLAM (simultaneous localization and mapping)
Building a map of an unknown environment while at the same time tracking the sensor's own pose within it, a chicken-and-egg problem solved by joint estimation.
Spectrogram
A time-frequency image showing how a signal's frequency content changes over time, produced by a short-time Fourier transform and widely used as a network input.
Test-time adaptation
Adjusting a deployed model to the current data stream without labels, for example by updating normalization statistics, to counter domain shift on the fly.
TinyML
Machine learning that runs on microcontrollers and other severely constrained devices, with kilobytes of memory and milliwatts of power, so inference happens at the sensor.
Time of flight (ToF)
Measuring distance by the time a pulse of light or sound takes to travel to a target and back. The principle behind lidar, sonar, and many depth cameras.
Uncertainty (aleatoric and epistemic)
Aleatoric uncertainty is irreducible noise in the data; epistemic uncertainty reflects what the model does not know and shrinks with more data. Separating them guides where to collect more or trust less.
Window
A fixed-length slice of a streaming signal, the basic unit fed to a model. Choices of length and overlap trade latency against context.
World model
A learned internal model that predicts how a scene evolves, letting an agent imagine the consequences of actions before taking them. Increasingly used to give perception a predictive, physics-aware backbone.

If a symbol or term you meet in the text is missing here, check the index and the chapter that introduces it; the definitions in the running text are always the authoritative ones, and this appendix simply gathers them for convenience.