UCCSD vs Hardware-Efficient Ansatz: What the Benchmark Data Actually Shows

The choice of ansatz is one of the most consequential decisions in a VQE workflow. It determines circuit depth, expressibility, convergence behavior, and how well the algorithm maps onto near-term hardware. Yet most published results pick one ansatz, run it on one molecule, and call it a benchmark.

QEncode Suite v2 takes a different approach. We run both UCCSD and the hardware-efficient ansatz (HEA) across five molecules — H₂, LiH, HF, N₂, and BeH₂ — at three qubit encodings: Jordan-Wigner, parity, and Bravyi-Kitaev. Every run is executed under identical conditions on the same classical simulation infrastructure, with results independently verified and signed. Here's what the data shows.

The two ansatz families

UCCSD (Unitary Coupled Cluster Singles and Doubles) is derived from quantum chemistry. It constructs the ansatz from physical excitation operators — single and double fermionic excitations from a Hartree-Fock reference state. This gives it strong physical motivation and predictable expressibility, but it comes with a cost: circuits that are deep and gate-heavy, especially for larger molecules.

Hardware-Efficient Ansatz (HEA) takes the opposite approach. Rather than encoding physical structure, it builds parameterized circuits from gates that are native to the target hardware — typically layers of single-qubit rotations interleaved with entangling gates. The result is much shallower circuits, but with no guarantee that the expressible states are chemically meaningful.

What we measured

For each molecule-encoding-ansatz combination, QEncode Suite v2 records three primary metrics:

• Energy gap — the absolute difference between the VQE ground-state estimate and the exact FCI energy, in Hartree. Lower is better. This is the accuracy metric.
• Circuit depth — the total number of gate layers in the compiled circuit. Lower means less decoherence on real hardware.
• Two-qubit gate count — the number of CNOT-equivalent gates. This is often the tightest resource constraint on current devices.

All circuits are compiled with Qiskit's standard transpiler at optimization level 3. Simulations use Qiskit Aer's statevector simulator with no noise, giving us noiseless baseline numbers that reflect pure algorithmic performance.

The accuracy picture

On the accuracy leaderboard, UCCSD dominates — and it's not close. Across all five molecules and all three encodings, UCCSD consistently achieves energy gaps in the range of 10⁻³ to 10⁻⁵ Hartree, well within chemical accuracy (1.6 × 10⁻³ Ha). The physical motivation built into the ansatz pays off directly in ground-state fidelity.

HEA performance on accuracy is more variable. For small molecules like H₂, a well-tuned HEA can match UCCSD because the Hilbert space is small enough that expressibility isn't the limiting factor. For N₂ and BeH₂, the gap widens considerably. HEA circuits struggle to represent the correlated ground states of these molecules without many more layers, which brings circuit depth close to or exceeding that of UCCSD anyway.

The cost picture

On the cost leaderboard — which ranks by circuit depth and two-qubit gate count — HEA wins convincingly for small molecules. For H₂ at parity encoding, the hardware-efficient circuits use roughly 40–60% fewer two-qubit gates than UCCSD. For LiH and HF, the savings are still meaningful: 20–35% fewer gates with comparable expressibility for the ground state.

For N₂ and BeH₂, the picture is murkier. The accuracy deficit of HEA means you need more layers to compensate, and the gate savings shrink. In several N₂ configurations, the optimized HEA circuit was actually deeper than UCCSD — because the optimizer needed additional layers to get within acceptable error bounds.

Encoding effects are real and significant

One finding that surprises many practitioners: the choice of qubit encoding has a measurable effect on both ansatz families, and the interaction between encoding and ansatz matters.

The parity encoding systematically reduces qubit count by two (via two-qubit reduction), which directly shrinks UCCSD circuit depth since excitation operator count scales with qubit number. For UCCSD, parity encoding consistently produces the best depth-to-accuracy tradeoff across all five molecules in Suite v2.

For HEA, the encoding choice matters less — the ansatz doesn't use physical structure anyway, so the qubit reduction benefit is smaller. Jordan-Wigner and parity encodings perform comparably for HEA on the smaller molecules.

What this means in practice

The data supports a clear heuristic:

• If you need chemical accuracy and have a classical simulator or a low-noise device with reasonable depth budget: use UCCSD with parity encoding. The accuracy advantage is real and consistent.
• If you're running on near-term hardware with tight gate budgets and you're benchmarking H₂ or LiH: HEA is a legitimate choice, and the gate savings are meaningful.
• If you're benchmarking N₂ or BeH₂ on real hardware: be careful with HEA claims. The noiseless numbers look competitive, but accuracy degrades faster with noise in less-expressive circuits.

The balanced score

QEncode Suite v2 includes a third leaderboard category — the balanced score — that combines accuracy and cost into a single normalized metric. This is where the comparison gets most interesting for practitioners who face real hardware constraints.

On the balanced leaderboard, UCCSD with parity encoding holds the top positions for three of five molecules. HEA makes a stronger showing here than on the pure accuracy board, taking top spots for H₂ and LiH where its gate savings outweigh its small accuracy deficit.

The full leaderboard is publicly available and updated after every certified run.