Quantum Lab

Quantum Lab runs quantum simulations on QVArena compute nodes. Hand it an OpenQASM circuit, a Hamiltonian, a lattice, or a Python program; it is queued, picked up by a node sized to the job, and the result comes back with a certificate recording exactly how it was produced.

A submitted program runs inside a container with no network, a read-only filesystem, and a capped share of one node — it cannot reach the platform or another user's work.

Choose a method explicitly, or leave it on auto and the node picks one for you.

The service accepts OpenQASM 2 and OpenQASM 3, which every major framework can export. Write your circuit in the tool you already use, export QASM, and submit it.

The client accepts a Qiskit circuit or a raw QASM string. Create an API token on the API tokens page, then submit and poll for the result.

Beyond sampling, you can request expectation values of weighted Pauli observables for a circuit — the primitive variational workflows such as VQE and QAOA are built on. Each observable is a Pauli string over I, X, Y, Z (qubit 0 first, leftmost) whose length equals the circuit's qubit count, with a coefficient.

Without noise, estimation uses an exact statevector and carries no shot noise; under a noise model it is an exact density-matrix expectation, and with mitigation a zero-noise-extrapolated (approximate) estimate.

The estimator says what an observable is worth on a fixed circuit. The optimizer answers the question VQE and QAOA actually ask — which parameters minimize it — and runs the whole loop on a compute node, so nothing has to stay open on your side across hundreds of evaluations.

Write free parameters into the circuit as theta[0], theta[1], … with indices running consecutively from 0; the worker substitutes values before each evaluation.

Five optimizers are available. COBYLA and Nelder-Mead are derivative-free, SPSA takes two evaluations per step regardless of width, L-BFGS-B uses gradients, and rotosolve solves each coordinate in closed form.

Two of them can be exact rather than approximate, under the same condition: every parameter appearing exactly once as the whole argument of a rotation. Then L-BFGS-B uses the two-term parameter-shift rule, which is the same derivative real hardware would report, and each coordinate slice is a single sinusoid so rotosolve locates that coordinate's exact minimum in three evaluations. Without that condition L-BFGS-B falls back to central differences and rotosolve is refused rather than approximated.

None of the five wins everywhere: COBYLA is the most reliable default, and rotosolve needs no tuning and holds up better on wide ansaetze where the gradient method stalls.

Beyond the optimum and the full convergence trace: up to 10 qubits the Hamiltonian is diagonalized directly, so the exact minimum, its degeneracy, and the overlap with the ground eigenspace sit beside the value that was found - a variational number means little without the answer it was chasing. The gradient norm at the reported optimum says whether the run reached a genuine stationary point or simply ran out of budget, which the energy alone cannot distinguish.

Asking for plateau samples additionally measures the gradient at random points and reports its variance: the barren-plateau measurement, reported as a number rather than a verdict, since a plateau is a statement about how that variance scales with width.

Sampling asks what a fixed circuit outputs; estimation what an observable is worth on it; optimization which parameters minimize one. None answer the question most of physics and chemistry actually asks — given this Hamiltonian, what does the system do over time?

Read the accuracy block first. exp(-iHt) has no exact circuit for a general H, so the trajectory is an approximation whose error is invisible in the output. Where the system is small enough the exact trajectory is computed alongside and the difference reported; above that size the error is unknown rather than small.

Five modes that describe what a circuit is rather than what it emits: its matrix, its channel with noise composed in, how its qubits are correlated, the state after every gate, and Clifford+T sampling whose cost tracks non-Clifford count instead of qubit count.

The step trace is the one to reach for when a histogram has not made the circuit clear; the entanglement mode is the one that can tell GHZ from a Bell pair, which the single-qubit numbers cannot.

Every request is authenticated with your access token as a Bearer header. Responses use the standard { success, data } envelope.

Circuits may use up to 36 qubits and up to 100,000 shots. Each account may submit up to 50 jobs per rolling 24 hours. Circuits too large for statevector simulation must request the matrix product state method.

Every completed job carries a certificate: the SHA-256 of the submitted QASM, the method used, the random seed, the shot count, the exact Qiskit and Aer versions, the compute node, and the start and finish timestamps. Anyone with the same circuit and seed can reproduce and verify the result.