The number \(A=y^{\mathsf T}Ky/(N\lVert K\rVert_F)\) comparing a Gram matrix \(K\) with the outer product of the labels; \(A=1\) for \(K\propto yy^{\mathsf T}\). A cheap diagnostic computed before training (Chapter 8).
Amplitude encoding
Storing a vector \(x\) in the amplitudes of a state, \(\lvert\psi(x)\rangle=\lVert x\rVert^{-1}\sum_jx_j\lvert j\rangle\), using \(\lceil\log_2d\rceil\) qubits. Loses the norm of \(x\); generic preparation costs \(O(2^n)\) gates (Chapter 5).
Angle encoding
Writing each feature into a single-qubit rotation angle, \(\bigotimes_jR_Y(x_j)\lvert0\rangle\). Depth one, a product state, kernel \(\prod_j\cos^2\frac{x_j-x_j'}2\) (Chapter 5).
Ansatz
The trainable circuit \(U(\theta)\) of a variational model: rotations with adjustable angles plus fixed entangling gates (Chapter 9).
Barren plateau
A loss landscape that is exponentially flat in the number of qubits: \(\operatorname{Var}[\partial_\theta C]\sim2^{-n}\) for a deep 2-design ansatz and a global cost (Chapter 12).
Basis encoding
Mapping a bit string \(x\) to the computational basis state \(\lvert x\rangle\); the induced kernel is \(\delta_{xx'}\) (Chapter 5).
Bloch sphere
The surface of the unit ball that represents the pure states of one qubit, by a polar angle \(\theta\) and an azimuth \(\varphi\) (Chapter 2).
Born rule
Measuring \(\lvert\psi\rangle\) in an orthonormal basis \(\{\lvert\phi_k\rangle\}\) gives outcome \(k\) with probability \(\lvert\langle\phi_k\vert\psi\rangle\rvert^2\) (Chapter 2).
CNOT
The controlled-NOT gate \(\lvert a,b\rangle\to\lvert a,a\oplus b\rangle\); the basic entangling gate, computing parity (Chapter 3).
Data re-uploading
Feeding the input into the circuit at every layer. A single upload gives a first-order trigonometric model; \(L\) uploads give frequencies up to \(L\) (Chapter 5).
Density matrix
The operator \(\rho=\sum_ip_i\lvert\psi_i\rangle\langle\psi_i\rvert\) describing a statistical ensemble: Hermitian, positive semidefinite, of unit trace. Pure iff \(\operatorname{tr}\rho^2=1\) (Chapter 2).
Dequantisation
Replacing a claimed quantum speed-up by a classical algorithm of comparable efficiency, for example by sampling or by classical surrogate models (Chapter 12).
Empirical risk
The average loss on the training examples, \(\hat R(f)=\frac1N\sum_i\ell(f(x_i),y_i)\); to be distinguished from the true risk (Chapter 4).
Entanglement
A property of a multi-qubit state that cannot be written as a product of single-qubit states, for example \((\lvert00\rangle+\lvert11\rangle)/\sqrt2\) (Chapter 2).
Expectation value
The average outcome \(\langle O\rangle=\operatorname{tr}(\rho O)\) of an observable; estimated from \(M\) shots with standard error at most \(1/\sqrt M\) (Chapter 2).
Expressibility
How uniformly an ansatz covers state space, measured as the divergence of its fidelity distribution from the Haar one (Chapter 9).
Feature map
A function \(\phi\) from the data into an inner-product space in which the model is linear. In QML, the encoding \(x\mapsto\lvert\varphi(x)\rangle\) (Chapters 4 and 5).
Fidelity kernel
The quantum kernel \(k(x,x')=\lvert\langle\varphi(x)\vert\varphi(x')\rangle\rvert^2=\operatorname{tr}(\rho(x)\rho(x'))\) (Chapter 6).
Generalisation gap
The difference \(R(\hat f)-\hat R(\hat f)\) between true and empirical risk (Chapters 4 and 12).
Gram matrix
The matrix \(K_{ij}=k(x_i,x_j)\) of all pairwise kernel values of the training set (Chapter 4).
Hilbert space
A complex vector space with an inner product, complete in the induced norm; \(\mathbb{C}^d\) in finite dimension (Chapter 2).
Hybrid loop
The training scheme in which a quantum device evaluates a model and a classical computer updates its parameters (Chapter 1).
Inversion test
Estimating a kernel entry by applying \(U(x)\) then \(U(x')^\dagger\) and counting the all-zeros outcome, whose probability is \(k(x,x')\) (Chapter 7).
Kernel
A function \(k(x,x')=\langle\phi(x),\phi(x')\rangle\) evaluated without forming the feature vectors; valid iff every Gram matrix is positive semidefinite (Chapter 4).
Observable
A Hermitian operator whose eigenvalues are the possible outcomes of a measurement (Chapter 2).
Parameter-shift rule
For a Pauli rotation, \(f'(\theta)=\tfrac12\big[f(\theta+\frac\pi2)-f(\theta-\frac\pi2)\big]\), exactly (Chapter 10).
Partial trace
The operation \(\rho_A=\operatorname{tr}_B\rho_{AB}\) that gives the state of one subsystem; it turns half of an entangled pure state into a mixed state (Chapter 2).
Positive semidefinite (PSD)
A symmetric matrix \(K\) with \(c^{\mathsf T}Kc\ge0\) for all \(c\); the condition for a Gram matrix to come from a feature map (Chapter 4).
Post-selection
Keeping only the runs in which a measured qubit gave a chosen outcome; used by the distance-based classifier at the cost of discarded shots (Chapter 13).
Product state
A state of \(n\) qubits that factorises into single-qubit states; needs only \(O(n)\) numbers and is classically simulable (Chapter 2).
QCNN
A quantum convolutional neural network with shared two-qubit convolution unitaries and pooling by mid-circuit measurement; provably free of barren plateaus (Chapter 13).
Representer theorem
For regularised empirical risk minimisation in a kernel's feature space, the minimiser lies in the span of the training data (Chapter 4).
Shot
One run of a circuit ending in a measurement; each shot returns a single outcome, so expectation values must be estimated from many (Chapter 2).
SPSA
Simultaneous perturbation stochastic approximation: a gradient estimate from two perturbed evaluations, whatever the number of parameters (Chapter 10).
Support vector
A training point with non-zero weight \(\alpha_i\) in the SVM dual; only support vectors enter the decision function (Chapter 4).
SWAP test
Estimating an overlap with an ancilla controlling a swap of two registers between two Hadamards: \(P(0)=\tfrac12(1+\lvert\langle\psi\vert\phi\rangle\rvert^2)\) (Chapter 7).
Unitary
A matrix \(U\) with \(U^\dagger U=I\); it preserves inner products and is exactly invertible. Gates and circuits are unitary (Chapters 2 and 3).
Variational quantum classifier (VQC)
A model in which the circuit's gate angles are trained by a classical optimiser: encode, transform with \(U(\theta)\), measure (Chapter 9).
Zero-noise extrapolation (ZNE)
Error mitigation by amplifying noise, measuring at several noise levels and extrapolating the result to zero noise (Chapter 12).
ZZ feature map
An entangling feature map built from Hadamards and data-dependent two-qubit \(ZZ\) phases (Chapter 5).
2-design
A distribution of unitaries that reproduces the random (Haar) distribution up to second moments; deep random ansätze approximate one, with the consequence of concentration (Chapters 8 and 12).