Appendix A — Notation Covenant

Notation is a public interface between this book and its sibling. Core symbols retain their meanings; this volume adds symbols when the new diagnostic objects require them.

A.1 Inherited core

Object Notation Contract
Scalar \(x\) Italic lowercase unless a field convention requires otherwise
Vector \(\vect{x}\) Bold; a column on paper
Matrix or batched array \(\matr{A}\) Bold; shape completes the type
Aggregate objective \(\loss\) A local loss is \(\ell_i\)
Expectation \(\E\) Population and conditioning are stated
Variance \(\var\) Denominator convention is stated
Covariance \(\cov\) Sample and population objects are distinguished
Norm \(\norm{\cdot}\) The subscript names the norm when ambiguity matters
Definition \(:=\) Equality remains \(=\)
Approximation \(\approx\) Never used as silent equality
Algorithmic update \(\leftarrow\) Distinct from a mathematical identity

Code stores a batch by rows unless a chapter explicitly declares otherwise. Every non-obvious matrix product states dimensions first, and every reduction names its axes.

A.2 New in this volume

Object Notation First use
Centered sum of squares \(M_{2,n}\) Streaming state
Operator norm \(\norm{\matr{A}}_{\mathrm{op}}\) Random operators
Largest singular value \(\sigma_{\max}(\matr{A})\) Random operators
Smallest singular value \(\sigma_{\min}(\matr{A})\) Random operators
Frobenius norm \(\norm{\matr{A}}_{\mathrm{F}}\) Random operators
Epsilon-net \(\mathcal{N}_{\epsilon}\) Continuous extrema
Covering number \(N(K,d,\epsilon)\) Continuous extrema
Sub-Gaussian Orlicz norm \(\norm{X}_{\psi_2}\) Safe-zone concentration
Sub-exponential Orlicz norm \(\norm{X}_{\psi_1}\) Products and squares
Unit roundoff \(u\) Floating-point contract
Unit in the last place \(\operatorname{ulp}(x)\) Floating-point contract
Spectral condition number \(\kappa\) Quadratic modes
Spectral radius \(\rho(\matr{A})\) Quadratic modes
Sub-Gaussian MGF proxy scale \(\sigma\) Safe-zone concentration
Sample covariance \(\widehat{\matr{\Sigma}}\) Spectral nulls
Empirical spectral distribution \(\mu_{\matr{M}}\) Spectral nulls
Aspect ratio \(\gamma=n/m\) Spectral nulls
Marchenko–Pastur edges \(\lambda_-,\lambda_+\) Spectral nulls
Population spike \(\beta\) Spectral nulls
Effective rank \(r_{\mathrm{eff}}(\matr{\Sigma})\) Effective dimension
Stable rank \(\operatorname{sr}(\matr{A})\) Effective dimension
Jacobian \(\matr{J}_F(\vect{x})\) Reverse accumulation
Jacobian–vector product \(\matr{J}_F(\vect{x})\vect{v}\) Reverse accumulation
Vector–Jacobian product \(\vect{u}^{\mathsf T}\matr{J}_F(\vect{x})\) Reverse accumulation
Node cotangent \(\bar z_v\) Reverse accumulation
Realized Hessian \(\matr{H}\) Curvature objects
Generalized Gauss–Newton matrix \(\matr{G}\) Curvature objects
Model Fisher information \(\matr{F}\) Curvature objects
Empirical Fisher \(\widetilde{\matr{F}}\) Curvature objects
Model-curvature residual \(\matr{R}\) Curvature objects
Arithmetic intensity \(I=F/Q\) Runtime model
Machine balance \(I_*=P_{\max}/B\) Runtime model
Conditional noise-to-signal ratio \(\chi_k\) Stochastic regimes
Centering projector \(\matr{P}\) Coordinate-wise normalization
Nuclear norm \(\norm{\matr{A}}_*\) Matrix update geometry
Polar factor \(\operatorname{polar}(\matr{A})\) Matrix update geometry

The machine-readable source is contracts/notation-covenant.yml. A notation change is incomplete until that covenant, the HTML macros, the PDF macros, and this appendix agree.