19  Notation and Conventions

Every chapter in this book reuses the symbols on this page rather than redefining its own. When a chapter needs a symbol that isn’t here yet, it introduces it locally and it gets added here. If two chapters ever seem to use the same letter for two different things, that’s a bug in the book, not a feature of the mathematics — please open an issue.

19.1 The learning problem

Symbol Meaning
\(x \in \mathcal{X}\) an input / feature vector
\(y \in \mathcal{Y}\) a target: a label, a real value, a rank, a token, a preference — whatever \(\mathcal{Y}\) is for the problem at hand
\(\theta \in \Theta\) the parameters of a model
\(f_\theta\) the model / predictor, \(f_\theta : \mathcal{X} \to \mathcal{Z}\)
\(z = f_\theta(x)\) the model’s raw output: a real number, a vector of logits, a score — before any link function is applied
\(\hat y\) a point prediction, when one is being made (as opposed to a distribution)
\(P\) the true, unknown data-generating distribution over \((X, Y)\)
\(D = \{(x_i, y_i)\}_{i=1}^n\) an observed dataset, drawn i.i.d. from \(P\)

19.2 Loss, risk, and likelihood — kept visually distinct on purpose

This book is explicit about a distinction that informal usage often blurs: a loss is a penalty you choose to define what a good prediction means; a likelihood is a probability the model itself assigns to data that already happened. They are related — Chapter 2 is the whole story of how one becomes the other — but conflating their notation would quietly conflate the concepts too. So:

Symbol Meaning
\(\ell(\theta; x, y)\), or \(\ell(y, \hat y)\) the per-example loss — lowercase script \(\ell\), always
\(R(\theta) = \mathbb{E}_{(X,Y)\sim P}[\ell(\theta; X, Y)]\) the expected risk (population risk) — what you actually want to minimize
\(\hat R(\theta) = \frac{1}{n}\sum_{i=1}^n \ell(\theta; x_i, y_i)\) the empirical risk — what you can actually compute, from \(D\)
\(\mathcal{L}(\theta) = p(D \mid \theta) = \prod_i p(y_i \mid x_i; \theta)\) the likelihood — uppercase calligraphic \(\mathcal{L}\), always. A function of \(\theta\) with the data fixed; not a probability distribution over \(\theta\) and not, by itself, a loss
\(\log \mathcal{L}(\theta)\) the log-likelihood
\(\mathrm{NLL}(\theta) = -\log \mathcal{L}(\theta) = -\sum_i \log p(y_i \mid x_i;\theta)\) the negative log-likelihood. Chapter 2 derives that minimizing this is a particular, well-motivated choice of \(\hat R(\theta)\) — but the two symbols are never used interchangeably before that equivalence has been earned
\(p(\theta)\), \(p(\theta \mid D)\) prior and posterior over parameters (used from Chapter 3 onward)

19.3 Probability, information, and distributions

Symbol Meaning
\(p(y \mid x; \theta)\), or \(p_\theta(y \mid x)\) the conditional distribution the model defines over \(y\) given \(x\)
\(\sigma(z) = \dfrac{1}{1+e^{-z}}\) the logistic sigmoid
\(\mathrm{softmax}(z)_k = \dfrac{e^{z_k}}{\sum_j e^{z_j}}\) the softmax, mapping a logit vector to a categorical distribution
\(H(P) = -\sum_x P(x)\log P(x)\) the entropy of \(P\)
\(H(P, Q) = -\sum_x P(x) \log Q(x)\) the cross-entropy of \(Q\) relative to \(P\)
\(D_{\mathrm{KL}}(P \| Q) = \sum_x P(x) \log \frac{P(x)}{Q(x)}\) the Kullback–Leibler divergence from \(Q\) to \(P\)
\(\mathcal{N}(\mu, \sigma^2)\), \(\mathrm{Laplace}(\mu, b)\), \(\mathrm{Bernoulli}(p)\), \(\mathrm{Categorical}(p)\) the standard named distributions, used exactly as their conventional parameterizations

All logarithms are natural (\(\ln\)) unless a chapter says otherwise; this matters only for the numerical scale of NLL/entropy figures (base-2 gives “bits” instead of “nats”), never for any derivation’s structure.

19.4 Decision theory (from Chapter 8 onward)

Symbol Meaning
\(a \in \mathcal{A}\) a decision / action — often, but not always, the same space as \(\mathcal{Y}\)
\(u(a, y)\) a utility; \(\ell(a,y) = -u(a,y)\) up to convention
conditional risk of \(a\) at \(x\) \(\mathbb{E}_{Y \mid X=x}[\ell(a, Y)]\)
Bayes-optimal decision \(a^\star(x) = \arg\min_a \mathbb{E}_{Y\mid X=x}[\ell(a, Y)]\)

19.5 Ranking (Chapters 1013)

Symbol Meaning
\(s_i = f_\theta(x_i)\) the model’s score for item \(i\)
\(i \succ j\) item \(i\) is preferred to / ranked above item \(j\)
\(\pi\) a permutation (a full ranking) of a set of items
\(\mathrm{NDCG}, \mathrm{DCG}, \mathrm{MRR}, \mathrm{MAP}\) ranking-quality metrics, defined where first used

19.6 Preference and sequence objectives (Chapters 1516)

Symbol Meaning
\(x_{<t} = (x_1, \dots, x_{t-1})\) a token sequence’s prefix before position \(t\)
\(y^+ \succ y^-\) \(y^+\) is the preferred completion in a pairwise preference pair
\(\pi_\theta\), \(\pi_{\mathrm{ref}}\) a trainable policy and a fixed reference policy over completions — always subscripted; bare \(\pi\) is reserved for a ranking permutation (above) and never denotes a policy
\(r(x,y)\) a scalar reward for completion \(y\) given prompt \(x\)

19.7 Recurring visual conventions

Two devices repeat in every chapter that introduces an objective:

Origin tags. Immediately after a loss is derived, a short callout names where it actually came from, using exactly one of six labels: derived from likelihood, derived from Bayesian/MAP reasoning, derived from decision theory, surrogate/relaxation, heuristic/design choice, or hybrid. This is not decoration — it’s the book’s running answer to “is this a fact about the world, or a choice I’m making?”

Definition boxes. A named, citable formula (e.g. the RankNet loss, the Bradley–Terry model) gets a callout-note box titled “Definition — …”. Boxes are for landmarks worth finding again on a re-read, not for every equation.

See CONVENTIONS.md in the repository root for the full authoring template these chapters follow.