Algorithms explained

These pages explain how each algorithm in eigenkit works: what happens between the data that you put in and the result that it gives back. For the interface itself (the arguments of the constructors, the return types, saving a model), see the Estimators guide and the API Reference, which is generated from the code.

Page	What it computes	What it gives back
IQAE	An unknown probability amplitude `a`, to a target precision.	A number and a confidence interval.
HHL	The solution `x` of a linear system `Ax = b`, with a fast `"reference"` mode and an `"honest"` mode that is faithful to hardware.	A classical solution vector.
QFT	The Fourier transform of the amplitudes of a register.	A change of basis that other algorithms read from.
QAOA	A good bitstring for a combinatorial problem.	A bitstring and its cost.
VQE	The lowest energy that a molecule or a model Hamiltonian can have.	An energy, and the angles that reached it.
QSVC / QSVR	A similarity (kernel) matrix for a support vector model.	scikit-learn predictions.
HQNN	A trained hybrid quantum-classical neural network.	Labels, probabilities or values.
Reservoir computing	Features from a circuit that nobody trains.	A linear readout over a trajectory of states.

The names in the first column are acronyms, and each page defines its own: iterative quantum amplitude estimation (IQAE), the Harrow-Hassidim-Lloyd algorithm (HHL), the quantum Fourier transform (QFT), the Quantum Approximate Optimization Algorithm (QAOA), the Variational Quantum Eigensolver (VQE), the quantum support vector classifier and regressor (QSVC and QSVR), and hybrid quantum neural networks (HQNN).

Glossary

These terms come up on several pages. Each is also defined where it first appears.

Qubit: the quantum version of a bit. n qubits describe a state with 2ⁿ complex amplitudes.
Statevector: the full list of these 2ⁿ amplitudes.
Observable: a quantity that can be measured, here a Hermitian sum of the Pauli operators X, Y and Z. Its expectation value ⟨O⟩ is the average outcome over many measurements, a real number that is often in [-1, 1].
Ansatz: a circuit template with free (trainable) parameters, such as a Variational block. "Variational" means that these parameters are tuned by optimisation.
Amplitude amplification, or Grover operator: a unitary that repeatedly rotates probability towards a chosen "good" subspace, so that a property becomes easier to read. It is the engine of IQAE.
Quantum phase estimation (QPE): a part of a circuit that reads the eigenvalue of a unitary into a register of "clock" qubits. It is the engine of HHL.
Hamiltonian simulation, or Trotterisation: building the evolution e^{iAt} from elementary gates, by splitting A into a sum of simple terms and interleaving their evolutions over short times. The "honest" mode of HHL uses it. More trotter_steps lower the approximation error, and trotter_steps="auto" chooses a number of steps that keeps this error below the resolution of the clock register. The error is not only in the eigenvalues: a coarse grid mostly rotates the eigenvectors, which changes the direction of the answer while the spectrum still looks correct.
Gershgorin bound: a cheap upper bound on the largest magnitude of the eigenvalues of a matrix, which costs O(nnz): the largest absolute row sum. The honest mode of HHL uses it to choose the evolution time without diagonalising A.
Condition number κ: for a matrix, the ratio of its largest to its smallest eigenvalue, in absolute value. A large κ means that the system is hard to solve accurately.
Feature map: a circuit that embeds a classical sample x into a quantum state |φ(x)⟩. It is the basis of quantum kernels.
Angle encoding: the usual feature map here. One input value becomes one rotation angle, so n features need n qubits. This is what the Encoding block does.
Amplitude encoding: the other way to load data. A vector of 2ⁿ numbers is written directly into the amplitudes of a state of n qubits, so n qubits hold exponentially more values than with angle encoding. The vector is padded to the next power of two and scaled to a norm of 1, because a statevector must have a norm of 1. It is cheap in qubits and expensive in gates: preparing an arbitrary state is not an operation of fixed depth, which is why the AmplitudeEncoding block adds one gate to the intermediate representation and leaves each backend to use its own primitive for state preparation.
Kernel, or Gram matrix: a table of the similarities K(x_i, x_j) between pairs of samples. Classical support vector machines train directly on this table.
Barren plateau: a failure where the training gradients of a circuit shrink exponentially as the circuit gets wider or deeper, so it stops learning. It is a risk for the HQNN family.