Two ways of setting circuit parameters improved simulated results and, in some tests, outputs from quantum processors before any later optimization.
The study focuses on unitary cluster Jastrow circuits, which are designed to represent chemical systems while limiting the number of operations needed on a quantum processor. The researchers used a classical coupled-cluster calculation as a starting point, then introduced compressed representations of its interaction data and, for sampling-based calculations, an approximate tensor-network simulation to refine the circuit parameters.
They tested the approaches with exact state-vector simulations of systems reaching 52 qubits and with experiments on superconducting quantum processors using up to 65 qubits. The work evaluated the circuits as initialized; the researchers did not run a later parameter-optimization loop using quantum measurements.
What the circuits produced
In simulations, compressed double factorization improved on simply cutting away circuit interactions or repetitions. For the sampling-based quantum-selected configuration interaction method, it could sometimes do better than the untruncated circuit because it produced a more diverse range of sampled configurations. Approximate tensor-network refinement provided further energy improvements, particularly when the initial classical approximation was less accurate.
In hardware experiments on molecular systems, the compressed and simply truncated settings produced quantum-selected configuration interaction errors about an order of magnitude lower than random parameters. The researchers did not find a significant difference between those two settings in the hardware results, which they attributed to noise. A circuit with no local unitary cluster Jastrow operations produced the lowest error in those particular tests, which the researchers linked to its shallower state-preparation circuit and the effects of hardware errors. Overall, the results indicate that the proposed initialization methods can improve expectation-value and sampling-based calculations before parameter optimization.
Why better starts matter
Quantum chemistry circuits often have to be shortened or adapted to the limited connections between qubits, which can weaken their approximation to a classical chemistry calculation. These methods use classical preprocessing to compensate for some of that loss while preserving a circuit that can be run on near-term quantum processors.
The results also show that the best initialization can depend on the calculation and the hardware. For sampling-based methods, improving the variety of configurations produced by a circuit can be useful, while hardware noise can obscure the difference between more carefully constructed starting parameters.
Evidence and caveats
The evidence combines noiseless exact state-vector simulations with experiments on superconducting quantum processors. Simulations covered systems of up to 52 qubits, and hardware experiments used up to 65 qubits. For larger simulated systems, the researchers evaluated quantum-selected configuration interaction energies and used semistochastic heat-bath configuration interaction as the reference because exact full-configuration-interaction calculations were too costly. The largest formamide simulation was beyond exact state-vector capacity.
The study assessed the quality of the initial parameters and did not perform the later parameter optimization that these methods are intended to support. Hardware noise limited the distinction between truncated and compressed parameters, and the sampling-based energy depends on measurement outcomes. The results were demonstrated on the tested molecular systems and devices, so they do not establish that the same ranking will hold for all molecules or quantum processors.