"I have been called $x_t$ by the forecaster, $s_t$ by the control engineer, and $z_t$ by the statistician, and all three were pointing at me. The least they could do was agree on a name before arguing about my dynamics."
A Latent State Tired of Living Under Three Aliases
Temporal AI is a confluence of three traditions that grew up apart. Time series statistics speaks of innovations and lag operators, control theory speaks of states and transition matrices, and machine learning speaks of hidden units and gradients. The same idea wears different symbols in each, and the same symbol means different things across them: $L$ is a lag operator to a statistician, a loss to a deep learning practitioner, and a lookback window to a forecaster. This appendix is the book's single home for notation. It fixes the typographic conventions used everywhere, presents one unified symbol table that every chapter defers to, reviews the linear algebra the derivations assume, and points onward to the probability refresher of Appendix B and the optimization and deep learning basics of Appendix C. When a chapter introduces a symbol, it links here; when two chapters seem to clash, this page is the arbiter.
1. Notation Conventions
The book follows one typographic rule throughout, and stating it once here lets every chapter stay terse. The rule distinguishes scalars, vectors, and matrices by typeface, marks time by subscript, and separates the random object from its realized value by context and decoration. A reader who internalizes this section can decode any equation in the book without a local legend.
Scalars are lowercase italic, as in $x$, $\gamma$, or $r_t$, and denote single numbers: a discount factor, a reward, one coordinate of a state. Vectors are lowercase bold, as in $\mathbf{x}$, $\mathbf{h}_t$, or $\boldsymbol{\theta}$, and are columns by default, so $\mathbf{x} \in \mathbb{R}^d$ stacks $d$ scalar components $x^{(1)}, \dots, x^{(d)}$. Matrices are uppercase bold, as in $\mathbf{A}$, $\mathbf{F}$, or $\mathbf{Q}$, with $\mathbf{A} \in \mathbb{R}^{m \times n}$ having $m$ rows and $n$ columns. Sets and spaces are uppercase blackboard or calligraphic: $\mathbb{R}$ for the reals, $\mathbb{R}^d$ for $d$-dimensional Euclidean space, $\mathcal{S}$ for a state space, $\mathcal{A}$ for an action space. Where a chapter discusses one scalar series and the bold weight would only add clutter, it writes $x_t$ in plain italic and says so; the table below records the canonical bold form regardless.
Time is always a subscript. A single observation at time $t$ is $x_t$, and a contiguous slice from time $1$ through time $T$ is written $x_{1:T} = (x_1, x_2, \dots, x_T)$, the colon reading as the word "through". A window of the most recent $L$ steps ending at $t$ is $x_{t-L+1:t}$. The same colon convention applies to states, actions, and rewards: $s_{1:T}$, $a_{0:T-1}$, $r_{1:T}$. When the book needs a whole sequence as one object, for instance as the input to a sequence model, it writes the slice $x_{1:T}$ or names the sequence $\mathbf{X} = (\mathbf{x}_1, \dots, \mathbf{x}_T)$ as a matrix whose columns are the per-step vectors. Discrete time indices run over the integers; the continuous-time models of Chapter 13 switch to $x(t)$ with a parenthesized real argument, and the appendix flags that change of notation where it occurs.
The distinction between a random variable and its realized value is carried by context and by decoration rather than by a separate alphabet, because the temporal literature is not consistent enough to support a strict capital-for-random convention without colliding with the matrix convention. The book writes $X_t$ for a random variable and $x_t$ for a particular value it takes only in the few probability-heavy passages where the distinction is load-bearing, and says so locally; elsewhere $x_t$ denotes both, with the surrounding text making clear whether an expectation or a number is meant. Estimates and predictions carry a hat: $\hat{x}_{t}$ is an estimate of the state at $t$, and $\hat{x}_{t+h \mid t}$ is the $h$-step-ahead forecast made using data through time $t$, the vertical bar reading "given". A tilde marks an error or a centered quantity, as in $\tilde{x}_t = x_t - \hat{x}_t$. A bar marks a sample mean, as in $\bar{x}$. A star marks an optimum, as in $\boldsymbol{\theta}^{\*}$ or the optimal value function $V^{\*}$. These five marks, hat for estimate, tilde for error, bar for mean, star for optimum, and the conditioning bar for "given", are used identically in every chapter.
Typeface encodes rank (italic scalar, bold-lower vector, bold-upper matrix), subscript encodes time ($x_t$, $x_{1:T}$), and decoration encodes role (hat for estimate, tilde for error, bar for mean, star for optimum, vertical bar for conditioning). Read those three axes and you can parse any expression in the book without a local key. Where the temporal traditions force a symbol clash, the unified table in Section 2 is the tie-breaker.
2. The Unified Notation Table
This is the centerpiece of the appendix. The table below collects the recurring symbols of the book and gives each one canonical meaning, drawn so that machine learning, control theory, and statistics can coexist on the same page. Where a single letter is overloaded across the three traditions, the table lists each meaning on its own row and the prose beneath reconciles the clash explicitly, naming which chapter uses which sense. Every chapter that introduces one of these symbols links back to the row here, so this table is the canonical reference and the chapters carry only local reminders.
| Symbol | Meaning | First used |
|---|---|---|
| $x_t$ | Observed input or signal at time $t$; the series being modeled or fed to a model | Chapter 1 |
| $y_t$ | Target or output at time $t$; the quantity forecast or predicted, distinct from the input $x_t$ in supervised settings | Chapter 1 |
| $z_t$, $s_t$ | Latent or hidden state at time $t$; $z_t$ in statistics and generative models, $s_t$ in control and reinforcement learning. The unobserved variable that summarizes the past | Chapter 7 |
| $h_t$ | Hidden state of a recurrent network at time $t$; the learned analogue of the filter state $z_t$ | Chapter 10 |
| $a_t$ | Action taken at time $t$ by an agent or controller | Chapter 22 |
| $r_t$ | Reward received at time $t$; the scalar signal an agent maximizes | Chapter 22 |
| $\gamma$ | Discount factor, $\gamma \in [0,1)$; the weight on future rewards in a return | Chapter 22 |
| $\boldsymbol{\theta}$ | Model parameters (vector); regression coefficients, network weights, or policy parameters depending on context | Chapter 5 |
| $\theta$ (scalar) | Moving-average coefficient in ARMA models; distinct from the parameter vector $\boldsymbol{\theta}$ (see reconciliation below) | Chapter 5 |
| $\phi$ | Autoregressive coefficient in ARMA models | Chapter 5 |
| $\varepsilon_t$ | Innovation, shock, or white-noise term at time $t$; the unpredictable part of the series, also the Kalman innovation | Chapter 3 |
| $\mathbf{F}$, $\mathbf{H}$ | State-transition matrix and observation matrix of a linear-Gaussian state-space model | Chapter 7 |
| $\mathbf{Q}$, $\mathbf{R}$ | Process-noise covariance and measurement-noise covariance in the Kalman filter (see reconciliation: $\mathbf{R}$ also names the action-value reward model nowhere here, and $\mathbf{Q}$ collides with the $Q$-function below) | Chapter 7 |
| $\lambda$ | Intensity of a point process, an eigenvalue, or a regularization weight, disambiguated by context | Chapter 4 |
| $\pi$ | Policy mapping states to actions or action distributions, $\pi(a \mid s)$; never the constant $3.14159\ldots$, which the book writes out | Chapter 22 |
| $V$, $Q$ | State-value function $V(s)$ and action-value function $Q(s,a)$ in reinforcement learning | Chapter 22 |
| $L$ (operator) | Lag (backshift) operator: $L\,x_t = x_{t-1}$ (see reconciliation below) | Chapter 5 |
| $\mathcal{L}$ (loss) | Loss or objective function minimized during training; calligraphic to separate it from the lag operator | Chapter 9 |
| $L$ (lookback) | Lookback or context-window length; the number of past steps a model conditions on | Chapter 2 |
| $H$ (horizon) | Forecast horizon; the number of future steps predicted. Distinct from the observation matrix $\mathbf{H}$ by weight and context | Chapter 2 |
| $p(\cdot)$, $p(\cdot \mid \cdot)$ | Probability density or mass function and conditional density; the generative model of the data | Chapter 3 |
| $\mathbb{E}[\cdot]$ | Expectation; $\mathbb{E}_{p}[\cdot]$ when the distribution must be named | Chapter 3 |
| $\mathrm{Var}[\cdot]$, $\mathrm{Cov}[\cdot,\cdot]$ | Variance of a random variable and covariance between two | Chapter 3 |
| $\mathrm{KL}(p \,\|\, q)$ | Kullback-Leibler divergence from $q$ to $p$; the asymmetric gap between two distributions | Chapter 17 |
| $\nabla$, $\nabla_{\boldsymbol{\theta}}$ | Gradient operator; $\nabla_{\boldsymbol{\theta}}\mathcal{L}$ is the gradient of the loss with respect to the parameters | Chapter 9 |
The table refers four times to a reconciliation, and the book resolves each clash by a fixed rule rather than by hoping context suffices. The most frequent collision is the letter $L$. As a lag operator it is plain italic and acts on a time-indexed quantity, so $L\,x_t = x_{t-1}$ and a polynomial in $L$ defines an ARMA model; this is the meaning throughout Chapter 5 and Chapter 6. As a training loss the book writes the calligraphic $\mathcal{L}$, never the plain $L$, so the deep learning chapters from Chapter 9 onward carry no ambiguity. As a lookback length $L$ is a positive integer counting time steps and appears as a subscript range bound such as $x_{t-L+1:t}$, never as an operator applied to a symbol; the data and forecasting chapters use it this way. The three never appear in the same equation, and the typeface (italic versus calligraphic) plus the grammatical role (operator versus integer count) keeps them apart on sight.
The second collision is $\theta$. In ARMA notation the scalar $\theta$ (often $\theta_1, \theta_2, \dots$) is a moving-average coefficient, a fixed convention inherited from the time series literature and used in Chapter 5. In machine learning the bold $\boldsymbol{\theta}$ is the full parameter vector of a model, network, or policy. The book keeps the weight distinction strict: a bare italic $\theta_i$ is always an MA coefficient, and a bold $\boldsymbol{\theta}$ is always a parameter vector, so the two cannot be confused even when an ARMA model is itself being trained and both senses are nearby. The third collision is $\lambda$, which serves as a point-process intensity in Chapter 18, as an eigenvalue in the spectral and stability arguments of Chapter 4, and as a regularization weight in several training objectives; these live in different chapters and different equations, and each occurrence is named in its local prose. The fourth, $\mathbf{Q}$, is the Kalman process-noise covariance in Chapter 7 and the action-value function $Q(s,a)$ in Chapter 22; the first is a bold matrix and the second a plain-italic function of two arguments, so weight and signature separate them. The book never relies on the reader guessing: where any of these symbols appears, the surrounding sentence states its sense.
Spare a thought for $\pi$. In these pages it is a policy, $\pi(a \mid s)$, almost everywhere, which means the one place a reader might expect the circle constant, $\pi \approx 3.14159$, the book spells it out or hides it inside a Gaussian normalizer so the symbol stays free for decision making. A small sacrifice of tradition buys a large saving of confusion.
3. Linear Algebra Essentials
The derivations in the book, above all the Kalman filter of Chapter 7, the attention mechanism of Chapter 12, and the structured state-space models of Chapter 13, lean on a compact set of linear algebra facts. This section states them in the book's notation so the chapters can use them without re-deriving. It is a refresher and a reference, not a first course; readers meeting these ideas for the first time should pair it with a linear algebra text.
A vector $\mathbf{x} \in \mathbb{R}^d$ is a column of $d$ reals, and the inner product of two vectors of the same length is $\langle \mathbf{x}, \mathbf{y} \rangle = \mathbf{x}^{\top}\mathbf{y} = \sum_{i=1}^{d} x_i y_i$, a scalar that measures alignment and underlies every attention score and every projection in the book. The Euclidean norm is $\|\mathbf{x}\|_2 = \sqrt{\mathbf{x}^{\top}\mathbf{x}}$, the $\ell_1$ norm is $\|\mathbf{x}\|_1 = \sum_i |x_i|$, and the general $\ell_p$ norm is $\|\mathbf{x}\|_p = \left(\sum_i |x_i|^p\right)^{1/p}$; the $\ell_2$ norm measures forecast error and the $\ell_1$ norm drives sparse regularization. A matrix $\mathbf{A} \in \mathbb{R}^{m \times n}$ maps $\mathbb{R}^n$ to $\mathbb{R}^m$ by $\mathbf{y} = \mathbf{A}\mathbf{x}$, its transpose is $\mathbf{A}^{\top}$, and a square matrix is invertible when its determinant $\det(\mathbf{A})$ is nonzero, with $\mathbf{A}^{-1}\mathbf{A} = \mathbf{I}$ for the identity $\mathbf{I}$.
The eigen-decomposition writes a square matrix $\mathbf{A}$ in terms of vectors it merely rescales. A nonzero $\mathbf{v}$ with $\mathbf{A}\mathbf{v} = \lambda \mathbf{v}$ is an eigenvector with eigenvalue $\lambda$, and when $\mathbf{A}$ has a full set of independent eigenvectors it factors as $\mathbf{A} = \mathbf{V}\boldsymbol{\Lambda}\mathbf{V}^{-1}$, with $\boldsymbol{\Lambda}$ diagonal holding the eigenvalues. The eigenvalues govern the stability of a linear recurrence $\mathbf{x}_{t+1} = \mathbf{A}\mathbf{x}_t$: the iteration stays bounded exactly when every eigenvalue satisfies $|\lambda| \le 1$, the spectral-radius condition behind both stationarity in Chapter 5 and the vanishing-gradient analysis of recurrent networks in Chapter 10. The singular value decomposition generalizes this to any rectangular matrix, factoring $\mathbf{A} = \mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\top}$ with $\mathbf{U}$ and $\mathbf{V}$ orthogonal and $\boldsymbol{\Sigma}$ diagonal holding the nonnegative singular values; the largest singular value is the operator norm, the spectral norm that bounds how much a layer can amplify its input.
A symmetric matrix $\mathbf{A} = \mathbf{A}^{\top}$ is positive semidefinite when $\mathbf{x}^{\top}\mathbf{A}\mathbf{x} \ge 0$ for every $\mathbf{x}$, and positive definite when the inequality is strict for every nonzero $\mathbf{x}$. Covariance matrices are positive semidefinite by construction, which is why the Kalman filter can carry a covariance $\mathbf{P}_t$ and trust it to stay valid, and positive definiteness is the condition for a quadratic form to have a unique minimum, the fact that makes least-squares and Gaussian log-likelihoods well behaved. The book uses a small set of matrix-calculus identities to differentiate these forms. For the gradient of a linear form, $\nabla_{\mathbf{x}}\,(\mathbf{a}^{\top}\mathbf{x}) = \mathbf{a}$; for a quadratic form with symmetric $\mathbf{A}$, $\nabla_{\mathbf{x}}\,(\mathbf{x}^{\top}\mathbf{A}\mathbf{x}) = 2\mathbf{A}\mathbf{x}$; for a squared residual, $\nabla_{\boldsymbol{\theta}}\,\|\mathbf{y} - \mathbf{X}\boldsymbol{\theta}\|_2^2 = -2\mathbf{X}^{\top}(\mathbf{y} - \mathbf{X}\boldsymbol{\theta})$, whose zero gives the normal equations $\mathbf{X}^{\top}\mathbf{X}\boldsymbol{\theta} = \mathbf{X}^{\top}\mathbf{y}$ of ordinary least squares. These three identities, applied repeatedly, generate the closed-form estimators of Part II and the gradient updates of Part III.
One linear algebra idea recurs across the book more than any other: the eigenvalues of a transition matrix decide whether a recurrence forgets or explodes. They set the stationarity region of an autoregressive model, the convergence of a Kalman filter to its steady state, the vanishing and exploding gradients of a recurrent network, and the long-range memory of a structured state-space model. When a later chapter says a system is stable, it is making a claim about $|\lambda| \le 1$.
4. Pointers to the Companion Appendices
This appendix fixes notation and reviews linear algebra; two companion appendices carry the rest of the mathematical background, and the book defers to them rather than repeating their content. The probability and statistics the chapters assume, random variables and their distributions, conditional and joint densities, Bayes' rule, the Gaussian and its conjugacy, estimators and their bias and variance, hypothesis tests, and the maximum-likelihood and Bayesian inference machinery that powers the filters of Chapter 7 and the probabilistic forecasts of Chapter 19, live in Appendix B: Probability and Statistics Refresher. The optimization and deep learning basics, gradient descent and its stochastic and adaptive variants, convexity, backpropagation, the standard layers and activation functions, regularization, and the training loop that every model in Part III and beyond is fit by, live in Appendix C: Optimization and Deep Learning Basics.
The three appendices are meant to be read together as the book's mathematical preamble, and they share the notation fixed here so a symbol means the same thing in all three. When a chapter introduces a notion that belongs to one of these refreshers, it links to the relevant appendix on first use; when it introduces a symbol, it links to the unified table of Section 2 above. A reader who wants to see where any symbol first earns its keep can follow the "First used" column of that table into the chapter, or scan the full chapter sequence through the Table of Contents, which threads the notation forward from the foundations of Part I to the sequential decision making of Part VI and the deployed systems of the closing parts.
The 2024 to 2026 wave of temporal foundation models, Chronos, TimesFM, Moirai, Lag-Llama, and MOMENT among them, has sharpened a practical problem this appendix only papers over: the field has no agreed notation for a model that is simultaneously a forecaster, a control prior, and a sequence transformer. Papers in this line variously call the same context window the lookback $L$, the context length, or the patch sequence, and the same horizon the prediction length, the forecast horizon $H$, or the rollout. A reader moving between these papers and this book should expect to translate, and the unified table here is offered partly as a Rosetta stone for exactly that traffic. Watch for community efforts to standardize temporal model notation; until they land, the table in Section 2 is the book's contribution to the peace.