What is VQE and why does it excite chemistry?

TL;DR (status as of August 2026). VQE — the Variational Quantum Eigensolver — is a hybrid quantum–classical algorithm for estimating the ground-state energy of a molecule: a parametrized quantum circuit prepares a trial state, the processor measures its energy, and a classical optimizer adjusts the circuit parameters in a loop. Chemistry is the excited customer because ground-state energy is the genuinely quantum bottleneck of molecular simulation. The measured record is smaller than the excitement: the largest chemistry VQE ever run on hardware used 12 qubits — and computed Hartree–Fock, a mean-field quantity classical machines get trivially (Science 369, 1084, 2020) — no published VQE run has beaten a strong classical method (CCSD(T), DMRG) on the same molecule, and the frontier teams have started replacing VQE's own variational loop. Three brakes are measured and named: noise against a 1.6-millihartree accuracy target, barren plateaus in training, and a measurement bill that reaches 71 days for one energy evaluation of ethanol.

Status as of: August 2026.

What is VQE, exactly?

Four moving parts, one loop. (1) The molecule's electronic Hamiltonian H is encoded onto qubits. (2) A parametrized circuit — the ansatz U(θ) — prepares a trial state |ψ(θ)⟩. (3) The processor estimates the energy ⟨ψ(θ)|H|ψ(θ)⟩ term by term, averaging over many shots. (4) A classical optimizer proposes new angles θ, and the loop repeats until the energy stops dropping. That is the whole algorithm, as introduced on 2 photonic qubits for HeH⁺ in Peruzzo et al., Nature Communications 5, 4213 (2014).

One honest property is built into the math: by the variational principle, the measured energy is always an upper bound on the true ground-state energy. VQE can fail high; it cannot lie low.

The anatomy should look familiar if you read our QAOA post: parametrized circuit plus classical optimizer in a loop. QAOA points that loop at combinatorial cost functions; VQE points it at molecular eigenvalues. Both are heuristics — no proven runtime guarantee — and neither appears on the short list of proven quantum advantages.

Why chemistry is the structurally strong case

Electrons are quantum objects; simulating them exactly (full configuration interaction) scales exponentially on classical machines. Using amplitudes and interference to represent electronic states natively is the one place the hardware speaks the problem's language — the same structural argument behind quantum drug discovery.

The target that matters has a number: "chemical accuracy" is ~1 kcal/mol, about 1.6 millihartree relative to the exact answer — the precision at which computed energy differences begin to predict real reaction behavior (VQE survey, arXiv:2602.11384, 2026).

But the incumbents are not weak — and a weak baseline ruins any benchmark. CCSD(T), the classical "gold standard", and DMRG for strongly correlated systems run routinely on molecules far beyond anything a quantum processor has touched.

What has actually run on hardware?

Read the table coldly: in chemistry, a decade of VQE moved the hardware record from 2 to 12 qubits, and the 12-qubit run computed a mean-field quantity that a laptop evaluates in seconds — it was a hardware stress test, not a chemistry result classical machines lacked. The two large-qubit rows carry asterisks: 100 qubits was lattice field theory, and 77 qubits was not VQE at all.

What slows it down: three measured brakes

Brake 1 — noise against a 1.6 mHa target. Chemical accuracy is a microscopic error budget, and every gate error, decoherence event and readout fault on today's processors (what a qubit actually is) spends it. Hardware VQE results at chemical accuracy exist only for very small systems and lean heavily on error mitigation — which itself multiplies the number of shots you pay for (what a run costs).

Brake 2 — barren plateaus. For expressive ansätze with generic initialization, the training landscape flattens exponentially as qubits are added: gradients concentrate toward zero and the classical optimizer loses its signal (McClean et al., Nat. Commun. 9, 4812, 2018). The 2025 field review names barren plateaus "one of the main barriers to the trainability" of variational algorithms (Larocca et al., Nat. Rev. Phys. 7, 174–189, 2025). The recent twist cuts deeper: for broad circuit families, provable absence of barren plateaus comes bundled with classical simulability — if the landscape is provably trainable, a classical computer can often imitate the computation (Cerezo et al., Nat. Commun., 2025). The corridor between "trainable" and "still quantum" is narrow, and nobody has proven it is inhabited.

Brake 3 — the measurement wall. A molecular Hamiltonian in standard encodings has O(N⁴) terms as molecule size grows, and each energy estimate is a shot-averaged statistic over those terms. Costed end to end for combustion energies of small organic molecules at chemical accuracy: about 3.2×10⁹ measurements ≈ 1.9 days for one energy evaluation of methane; about 4.8×10¹⁰ ≈ 71 days for ethanol (Gonthier et al., Phys. Rev. Research 4, 033154, 2022). That buys one iteration; a full VQE optimization needs dozens to hundreds of them.

Has a VQE run ever beaten the classical incumbent?

Not that we can find, as of August 2026. We know of no published VQE hardware run that beats CCSD(T) or DMRG on the same molecule at chemical accuracy. The direction of validation is the tell: hardware demos are checked against classical references, never the reverse, and the February-2026 survey of the field benchmarks VQE variants in simulation against classical answers throughout (arXiv:2602.11384).

The sharpest evidence is the frontier's own pivot. IBM's flagship 2025 chemistry result — iron–sulfur clusters on 77 qubits, circuits of 10,570 gates — dropped the variational loop entirely: the paper names "deep circuits and a large number of measurements" leading to "prohibitive runtimes" as the reason to offload optimization classically (the SQD approach), and positions its results as upper bounds benchmarked against DMRG and SHCI — not victories over them (Robledo-Moreno et al., Sci. Adv. 11, eadu9991, 2025). When the team at the frontier stops using the algorithm's defining loop, that is data about the algorithm.

None of this is an accusation. Every demo above is honest work published with its limits stated. The gap between the anatomy (elegant) and the record (12 qubits, mean-field, zero wins) is simply what the state of the art looks like when it is measured — the same gap we track for the crossover point in every class.

What we know / What we don't know

What we know (measured, sourced): VQE's mechanics and its one-way error bar (variational upper bound). The chemistry hardware record: 12 qubits, and it was mean-field Hartree–Fock (2020). A 100-qubit VQE exists — in lattice physics, not chemistry (2024). Barren plateaus flatten training exponentially (2018 result, 2025 review), and provable trainability tends to travel with classical simulability (2025). The measurement bill at chemical accuracy is measured in days per single energy evaluation (1.9 for methane, 71 for ethanol; 2022 costing). Published VQE wins over strong classical methods on the same molecule: zero found.

What we don't know: whether adaptive, chemically structured ansätze plus better mitigation reopen the corridor barren plateaus close — the direction the 2026 survey maps. Whether measurement-reduction schemes (operator grouping, shadow tomography) bend the O(N⁴) wall enough to matter at useful sizes. Where the crossover would sit for industrially relevant molecules — estimates only, no measured point. Whether SQD-class hybrids inherit VQE's promise or retire it, and whether fault-tolerant phase estimation, not any variational method, ends up being chemistry's real quantum tool. Rosetta's own data: none in this class. Our sealed molecular series (RQ-0007) measures quantum walks on protein graphs — a different problem class (details here); we have no sealed VQE or ground-state-energy runs and claim no measurement in this class.

Sources

• Peruzzo et al., Nature Communications 5, 4213 (2014) — the first VQE (HeH⁺, photonic)
• Kandala et al., Nature 549, 242–246 (2017) — BeH₂ on 6 superconducting qubits
• Google AI Quantum (Arute et al.), Science 369, 1084–1089 (2020) — Hartree–Fock on 12 qubits
• Farrell et al., PRX Quantum 5, 020315 (2024) — Schwinger model vacuum on 100 qubits
• Robledo-Moreno et al., Science Advances 11, eadu9991 (June 2025) — SQD on 77 qubits
• McClean et al., Nature Communications 9, 4812 (2018) — barren plateaus
• Larocca et al., Nature Reviews Physics 7, 174–189 (2025) — barren plateaus review
• Cerezo et al., Nature Communications (2025) — absence of barren plateaus vs. classical simulability
• Gonthier et al., Physical Review Research 4, 033154 (2022) — measurements as a roadblock
• VQE methods & benchmarking survey, arXiv:2602.11384 (Feb 2026)

Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.