Glossary of Terms
A
- Adjoint differentiation
- Simulator-only method that computes all parameter gradients of a circuit by running it backwards, costing about one extra pass.
- Amplitude encoding
- Storing 2^n normalised numbers in the amplitudes of n qubits; compact, but preparing the state usually needs deep circuits.
- Amplitude estimation
- Fault-tolerant algorithm estimating probabilities or averages with quadratically fewer samples than classical Monte Carlo.
- Angle encoding
- Encoding each feature as a rotation angle on a qubit; simple and shallow but uses one qubit per feature.
- Ansatz
- The chosen structure of a parameterised circuit, such as hardware-efficient layers or a problem-inspired design.
B
- Barren plateau
- A training landscape where gradients shrink exponentially with qubit count, making optimisation practically impossible.
- Basis encoding
- Writing a bit-string directly into qubit states |0> and |1>, one bit per qubit.
- BQP
- Bounded-error quantum polynomial time: the class of problems a quantum computer can solve efficiently with small error.
C
- Circuit Born machine (QCBM)
- A generative model that samples bit-strings from the measurement distribution of a trained quantum circuit.
- Classical shadows
- Compact classical descriptions of a quantum state from randomised measurements, used to predict many properties efficiently.
- Classical surrogate
- A classical model that reproduces a trained quantum model's predictions, testing whether the quantum part is needed.
- CUDA-Q
- NVIDIA's open-source programming model for hybrid quantum-classical programs, optimised for GPU simulation.
D
- Data re-uploading
- Encoding the same input several times between trainable layers to increase a quantum model's expressivity.
- Density matrix
- A matrix description of a quantum state that can represent noise and mixtures, needing 4^n entries for n qubits.
- Dequantisation
- Finding a classical algorithm that matches a claimed quantum speed-up, as Ewin Tang did for recommendation systems.
E
- Entanglement
- Correlation between qubits that cannot be described by treating each qubit separately.
- Error mitigation
- Techniques such as ZNE, PEC and readout correction that reduce noise bias in expectation values at a cost in shots.
- Expectation value
- The average outcome of measuring an observable, estimated on hardware from many shots.
- Exponential concentration
- The tendency of quantum kernel values or cost functions to cluster around a fixed value as qubits grow.
- Expressibility
- How well a parameterised circuit's states cover the space of all possible states.
F
- Fault-tolerant quantum computing
- Computing with error-corrected logical qubits so long algorithms can run reliably; not yet available at useful scale.
- Feature map
- A circuit that maps classical data into a quantum state; the quantum analogue of a kernel feature map.
- Fidelity kernel
- A quantum kernel defined as the squared overlap between the states encoding two data points.
G
- Generalisation
- How well a model trained on some data performs on new, unseen data.
H
- Hamiltonian
- The operator describing a quantum system's energy; used as observables, encodings and learning targets.
- Hardware-efficient ansatz
- A circuit built from gates and connections native to a device, minimising depth but prone to barren plateaus.
- HHL algorithm
- Harrow-Hassidim-Lloyd quantum algorithm for linear systems; exponential speed-up only under strict assumptions.
- Hilbert space
- The mathematical space of quantum states; its dimension doubles with every added qubit.
- Hybrid quantum-classical model
- A model where a classical computer optimises or post-processes the outputs of quantum circuits.
I
- Inductive bias
- The assumptions a model family builds in, which decide which patterns it learns easily.
- ISA circuit
- A circuit already transpiled to a specific device's native gates and qubit layout, as required by IBM's V2 primitives.
K
- Kernel method
- A learning approach, such as the support vector machine, that compares data points through a similarity function.
L
- Logical qubit
- An error-corrected qubit encoded across many physical qubits.
M
- Measurement
- Reading out qubits, which collapses the state and returns random outcomes with probabilities set by the state.
N
- NISQ
- Noisy intermediate-scale quantum: today's era of tens to hundreds of imperfect qubits without full error correction.
- Noise model
- A description of gate, readout and decoherence errors used to simulate realistic hardware.
O
- Observable
- A measurable quantity represented by a Hermitian operator, such as Pauli-Z on one qubit.
P
- Parameterised quantum circuit (PQC)
- A circuit with adjustable gate angles that are trained like the weights of a neural network.
- Parameter-shift rule
- Exact gradient method evaluating a circuit at shifted parameter values, usually two runs per parameter.
- PennyLane
- Xanadu's open-source framework for differentiable quantum programming, integrated with PyTorch and JAX.
- Primitive (Qiskit)
- Standard Qiskit interface for running circuits: the Estimator returns expectation values, the Sampler returns counts.
- Projected quantum kernel
- A kernel built from reduced (local) properties of encoded states, designed to avoid exponential concentration.
Q
- QAOA
- Quantum approximate optimisation algorithm: alternating problem and mixer layers trained to find good solutions.
- QRAM
- Quantum random access memory: a proposed device for loading classical data in superposition; not yet practical.
- QSVT
- Quantum singular value transformation, a framework unifying many fault-tolerant quantum algorithms.
- Quantum advantage
- A demonstrated benefit over the best classical methods: in speed, data needed, or practical outcome.
- Quantum annealing
- Analog quantum optimisation that slowly evolves a system towards low-energy solutions of an Ising problem.
- Quantum convolutional neural network (QCNN)
- A layered quantum model with convolution and pooling steps, used for phase recognition.
- Quantum GAN (QGAN)
- A generative adversarial network where the generator, discriminator or both are quantum circuits.
- Quantum-inspired algorithm
- A classical algorithm that borrows ideas from quantum computing, such as tensor networks or sampling tricks.
- Quantum kernel
- A similarity measure between data points computed from their encoded quantum states.
- Quantum natural gradient
- An optimiser that rescales gradients by the geometry of quantum states (the Fubini-Study metric).
- Quantum neural network (QNN)
- A parameterised quantum circuit used as a trainable model, often inside a hybrid network.
- Quantum reservoir computing
- Using a fixed quantum system as a feature generator and training only a classical readout layer.
- Qubit
- The basic unit of quantum information, able to be in a superposition of 0 and 1.
- QUBO
- Quadratic unconstrained binary optimisation, the problem format used by quantum annealers and QAOA.
S
- Sample complexity
- The number of training examples or experiments needed to learn a task to a given accuracy.
- Shot
- One run of a circuit followed by measurement; many shots are averaged, with error shrinking as 1/sqrt(shots).
- Simulability
- Whether a quantum computation can be reproduced efficiently on a classical computer.
- SPSA
- Simultaneous perturbation stochastic approximation: a gradient estimate from two evaluations, popular on noisy hardware.
- Superposition
- A quantum state that is a weighted combination of basis states, such as |0> and |1> together.
T
- Tensor network
- A way of representing large states or matrices as connected small tensors, used for simulation and classical ML.
- Transpilation
- Rewriting a circuit into a device's native gates and connectivity, adding SWAPs where needed.
U
- Unitary
- A reversible, norm-preserving matrix describing an ideal quantum gate or circuit.
V
- Variational quantum classifier (VQC)
- A classifier made of a feature map, a trainable ansatz and a measurement, trained in a hybrid loop.
- VQE
- Variational quantum eigensolver: a hybrid algorithm that minimises a Hamiltonian's energy with a parameterised circuit.
Z
- Zero-noise extrapolation (ZNE)
- Error mitigation that runs circuits at amplified noise and extrapolates results back to zero noise.