# Quantum chemistry study tests two UCJ initialization methods

Published 2026-11-19 · Quantum.world Team · Quantum.world

> A Quantum paper presents two classical preprocessing methods for initializing unitary cluster Jastrow circuits used in quantum chemistry algorithms. The researchers tested them in simulations and on superconducting quantum processors, with hardware results varying by molecule and method.

## Two classical routes to better starting parameters

Two classical preprocessing techniques could improve starting parameters for unitary cluster Jastrow quantum chemistry circuits, according to a paper published in *[Quantum](https://quantum-journal.org/papers/q-2026-11-19-2278/)* on November 19, 2026. The study by researchers affiliated with [IBM Quantum](https://quantum.world/companies/ibm-quantum) and other institutions evaluates the methods for the unitary cluster Jastrow (UCJ) ansatz and its local variant, LUCJ.

These circuits can start with parameters derived from a classical coupled cluster singles and doubles (CCSD) calculation. But shortening the circuit or removing interactions to fit a processor’s connectivity can make that starting approximation less accurate. The paper proposes two ways to prepare parameters for circuits constrained in this way.

The first method uses compressed double factorization to fit the circuit parameters to CCSD amplitudes after truncation or connectivity constraints. It applies to both algorithms that estimate expectation values and those based on samples. The second method uses approximate tensor network simulation to refine parameters for sample-based algorithms: a classical optimizer uses simulated samples to evaluate and lower an approximate quantum-selected configuration interaction (QSCI) energy.

## Simulation gains, with limits

The researchers tested the methods using exact state-vector simulations on systems of up to 52 qubits. In simulations of nitrogen, compressed factorization improved QSCI results over a simple truncation strategy; the paper says it also outperformed the full-layer ansatz in most tested cases. Tensor-network refinement added further improvements for nitrogen at a stretched bond length and for an iron-sulfur cluster.

Those results depended on the initial classical calculation. The team used CCSD amplitudes for the nitrogen systems, but CCSD did not converge for the iron-sulfur cluster, so they instead derived amplitudes from configuration interaction singles and doubles (CISD). The paper reports that errors for this system remained above 0.1 Hartrees, which the authors attributed to the CISD parameters.

## Hardware results vary by molecule

Hardware experiments ran LUCJ circuits on IBM’s ibm_kingston, a 156-qubit Heron r2 processor. The systems included nitrogen with 10 electrons and 26 orbitals, and a formamide dimer with 36 electrons and 30 orbitals. The paper says simulations of the study’s largest hardware case were not possible with exact state-vector methods.

In QSCI tests on the processor, truncated and compressed initializations produced errors an order of magnitude lower than random parameters for both molecules. The researchers did not find a significant difference between those two approaches in that comparison, which they attributed to hardware noise. A control with no LUCJ circuit yielded the lowest QSCI error, a result the paper links to the shallow state-preparation circuit and hardware errors that generated additional configurations.

With configuration recovery, a step used to obtain sample-based quantum diagonalization (SQD) results, compressed initialization achieved lower error than simple truncation for the formamide dimer and was the only tested method to beat the no-LUCJ control within 10 iterations. For nitrogen, the paper reports no advantage for compression over truncation. The researchers say their hardware results suggest the compressed initialization can outperform naive truncation, and that either method is likely to beat random parameters.

## Not a quantum optimization test

The reported tests assess the initialized states and parameters; the researchers did not follow the preprocessing with parameter optimization using quantum measurements or run a variational quantum eigensolver optimization loop. Whether the starting methods speed later quantum optimization or improve trainability remains open, the paper says. The authors contributed an implementation of compressed double factorization to the open-source ffsim library.

## Sources

1. Original source [Quantum · page 1](https://quantum-journal.org/papers/q-2026-11-19-2278/) quantum-journal.org

Other pages read

2. [Quantum · page 2](https://quantum-journal.org/papers/q-2026-11-19-2278/pdf/) quantum-journal.org

## Related stories

- [Quantum Motion adds prototype CUDA-Q backend to QUICHE chemistry framework](https://quantum.world/news/quantum-motion-quiche-cuda-q-chemistry-pipeline)
- [Researchers train chemistry circuit generator on 70-qubit hardware](https://quantum.world/news/hardware-trained-70-qubit-chemistry-circuit-generator)
- [IBM VQE Tests Find Error Bars Missed Energy Residuals by Up to 32×](https://quantum.world/news/ibm-vqe-error-bars-energy-residuals-32-times)
- [IBM, RIKEN and NVIDIA report iron-sulfur SQD result near DMRG](https://quantum.world/news/ibm-riken-nvidia-79-fold-iron-sulfur-simulation)
- [Qunova says HI-VQE reaches chemical accuracy in an 80-qubit calculation](https://quantum.world/news/qunova-hi-vqe-chemical-accuracy-80-qubit-simulation)

## Explore

[More articles](https://quantum.world/news)

---

Canonical: https://quantum.world/news/ucj-quantum-chemistry-initialization-methods
