Measurement-efficient energy estimator for quantum chemistry of photodynamic materials

AI summary

Overview

Research area: Quantum computing for electronic structure — specifically measurement/readout strategies for extracting correlated energies from quantum-prepared states, applied to a transition-metal molecular model relevant to photodynamic and radiodynamic therapy.

Technical level: Advanced. The paper assumes familiarity with active-space quantum chemistry (CASCI, SHCISCF, NEVPT2), qubit Hamiltonians, and quantum measurement protocols (Hadamard/swap tests, classical shadows, Pauli grouping).

Scope (one sentence): The paper introduces a measurement-efficient "mixed estimator" workflow — pairing a quantum-prepared target state with sparse multiconfiguration classical references and magnitude-only overlap measurements — and validates it on an H₂DBB–Ru model across (6,6), (8,8) and (10,10) active spaces using emulator and trapped-ion hardware.

What This Paper Is About

Preparing a correlated electronic state on a quantum computer is only half the problem: reading its energy out with enough precision to distinguish chemically relevant states can require prohibitively many repeated measurements. This paper tackles that readout bottleneck by extending the "mixed estimator" — an energy functional borrowed from Quantum Monte Carlo — to use sparse multiconfiguration references and cheap overlap measurements, so the quantum device never has to sample individual Hamiltonian terms. The authors test it on a Ru-containing fragment of a hafnium metal–organic framework whose excited-state energetics govern singlet-oxygen generation, a system where different density-functional approximations disagree so strongly that predicted rates span more than five orders of magnitude.

Key Contributions

1. A mixed estimator built on sparse multiconfiguration references. The authors generalize the earlier single-determinant, classical-shadow-based mixed estimator to references |Φ⟩ = Σ_{x∈S_Φ} a_x|x⟩, splitting H|Φ⟩ = E_Φ|Φ⟩ + r_Φ|Λ⟩ into a classically known reference part and a residual state, so the quantum device only measures the overlaps h = ⟨Ψ|Λ⟩ and s = ⟨Ψ|Φ⟩.

2. A sign argument that eliminates phase estimation. Because the correction must be real and the variational principle fixes its sign for the lowest state in a symmetry sector, the energy correction is provably ≤ 0, so magnitude-only overlap measurements suffice — no Hadamard-test interferometry needed.

3. A new measurement protocol (flagged compute–uncompute, FCU) plus a systematic comparison of seven readout strategies — CU, FCU, real Hadamard test (RHT), real interferometric-swap test (RIST), standard and destructive swap tests (SST/DST), and classical shadows (CS) — with analytical shot-count formulas and bootstrap validation.

4. End-to-end hardware demonstration on the Quantinuum Helios-1E emulator and Helios-1 trapped-ion device up to a (10,10) active space, including a 50-qubit circuit that combines AI-assisted state preparation with leakage and spin-parity error detection.

Main Findings

• Shot counts collapse by orders of magnitude. Analytical variance estimates predict 1.6 mHa precision with a few hundred shots — about 100 times fewer than the single-determinant formulation and three orders of magnitude fewer than optimized Hamiltonian-term sampling. Concretely, at fixed total budget the bootstrap analysis shows CU/FCU give the narrowest energy distributions (standard deviations of 0.95–1.01 mHa), RHT/RIST close to 2 mHa, and SST/DST approximately 4.2–4.4 mHa at equal sample count.

• Analytical scaling laws for each protocol. For target standard error ε, the requirements are N_FCU ≤ N_CU ≈ (1/4)(r_Φ²/ε²), N_RHT = N_RIST ≈ r_Φ²/ε², N_CS ≈ 3 r_Φ²/ε², and N_SST = N_DST ≈ (1/4)(r_Φ²/ε²)(1/h²). Swap-test variants therefore become statistically unfavorable as the residual overlap shrinks.

• Reference quality and protocol efficiency compound. Moving from a one- to an eight-determinant reference cuts the classical-shadow requirement from roughly 40,000 to 5,000 shots; using CU with the eight-determinant reference cuts it further to roughly 400 shots — about 100 times fewer than single-determinant CS and approximately 1,700 times fewer than the grouped Pauli-averaging baseline.

• Approximate target states suffice. With pp-tUPS targets, two layers place both the direct expectation value and the mixed estimator within ±1.6 mHa of CASCI for both geometries. For sMCS targets, the mixed estimator generally reaches CASCI with fewer retained determinants than direct expectation-value estimation, with the clearest separation in the larger active spaces (the estimator is non-variational, so monotonic convergence is not guaranteed).

• Resource savings on the largest active space. In the (10,10) active space, direct expectation-value estimation needs 96 determinants for the R_x geometry and more than 100 for R_0, with circuits exceeding 1,600 two-qubit gates. The mixed estimator reaches the target accuracy with 27 and 23 determinants respectively — 224 and 159 two-qubit gates at depths of 211 and 134.

• AI-assisted truncation and circuit optimization help at (10,10). Constrained by energy accuracy and spin symmetry, the search left the R_x construction at (m,k,l) = (1,21,24) and selected (1,21,22) for R_0, reducing the combined ZZPhase-gate count from 857 to 755 (approximately 12%) and the R_0 two-qubit depth from 351 to 262.

• Classical benchmarks are not yet converged with active-space size. CASCI gives ΔE(R_x − R_0) ≈ −6.3 mHa in both (6,6) and (8,8), but +0.1 mHa in (10,10); with an NEVPT2 correction the value lies between +35 and +47 mHa. The authors therefore present these as controlled benchmarks for quantum energy estimation rather than converged physical energies. For the (6,6) space, SHCISCF energies lie about 40 mHa below Hartree–Fock, the dominant determinant has a squared coefficient of roughly 0.93, and natural-orbital occupations depart from two and zero by about 0.07.

• Compute–uncompute was chosen for experiments; FCU wins under a runtime condition. CU combines low sampling variance with an ancilla-free n-qubit circuit, so it was used for all scans and hardware runs. FCU becomes preferable when the flag is measured mid-circuit and flagged shots terminate early, with the high-quality-reference crossover reducing to τ_Φ + τ_flag < τ_Λ.

Methodology in Plain English

1. Pick a chemically hard but tractable model. From a Ru-functionalized Hf–DBB–Ru metal–organic framework, the authors extract an H₂DBB–Ru fragment that reproduces the parent framework's reported absorption and phosphorescence behavior. They study two geometries: the relaxed triplet geometry (R_0) and the geometry associated with singlet-oxygen activation (R_x). The target quantity is the energy difference between them, which dominates the Marcus activation energy that enters the rate exponentially.

2. Build classical benchmarks. Semistochastic heat-bath configuration interaction with self-consistent orbital optimization (SHCISCF) generates the orbitals; CASCI then gives exact energies within (6,6), (8,8) and (10,10) active spaces. The triplet sector is targeted with (N_α, N_β) = (4,2), (5,3) and (6,4) respectively.

3. Rewrite the energy so the quantum computer does less work. Instead of measuring every Hamiltonian term, the energy is expressed relative to a classically computable reference: E_ME = E_Φ + r_Φ(h/s). The reference energy, the residual norm, and the residual state's determinant expansion are all obtained classically by acting with H on |Φ⟩; only the two overlaps are measured.

4. Exploit the sign. For a ground state in the same symmetry sector as the reference, the correction is provably non-positive, so only squared overlap magnitudes are needed — measured as all-zero return probabilities of compute–uncompute circuits.

5. Compare readout protocols. Six circuits/protocols are benchmarked against classical shadows and a Pauli-averaging baseline using greedy qubit-wise-commuting grouping under the Jordan–Wigner mapping, with bootstrap resampling at a fixed 10,000-shot budget.

6. Validate on hardware. Finite-shot emulator runs on Helios-1E and device runs on Helios-1 test the workflow with leakage detection and spin-parity checks; the largest demonstration is a 50-qubit circuit.

Why This Matters

The paper reframes a problem that is often treated as an implementation detail — how to read an energy off a quantum state — as a first-order driver of quantum resource requirements. By showing that a carefully chosen classical reference plus a cheap overlap circuit can beat term-by-term Hamiltonian sampling by orders of magnitude, it makes near-term quantum chemistry workflows plausible for systems where the state-preparation step alone is already at the edge of what hardware can do.

Real-world applications:

• Photodynamic and radiodynamic cancer therapy. Understanding and tuning how MOF-based sensitizers generate reactive oxygen species could support spatially controlled, minimally invasive treatment with less collateral damage to healthy tissue.
• Photosensitizer and materials design. The workflow is described as transferable to photochemical energetics and rate calculations for other RDT/PDT photosensitizers.
• Transition-metal excited-state chemistry. Systems with near-degenerate states, where density-functional approximations can change state ordering or the balance between metal-to-ligand charge-transfer (MLCT) and metal-centred (MC) character, are exactly where multireference methods are needed and classical costs explode.
• Reaction-rate prediction. Because the computed energy difference feeds exponentially into a Marcus rate, modest precision gains translate into large changes in predicted rates.

Industry relevance: The author list spans Quantinuum (trapped-ion quantum hardware and algorithms) and Johnson & Johnson (pharmaceutical R&D), signaling direct interest from both the quantum-computing and drug-discovery sectors in making correlated energy estimation practical. The work also provides a benchmarking template — an electronically challenging but classically tractable model — that hardware vendors and end users can share.

Future Directions

• Signed and complex readouts for excited states. Magnitude-only measurement relies on a known correction sign. Excited states without a rigorously known sign require signed protocols (RHT/RIST), and complex states require both overlap quadratures — at greater control overhead.
• Joint optimization of state sizes. The reference size m, residual truncation k and target size l are coupled; the authors note the required sizes vary nontrivially and that approximate target information could enable more selective discarding than the conservative condition in Eq. (5).
• Mid-circuit early termination for FCU. The predicted crossover τ_Φ + τ_flag < τ_Λ — plausible when the residual state's worst-case support scales as O(m n⁴) — remains to be demonstrated as a runtime advantage on hardware rather than an idealized circuit-duration estimate.
• Beyond a single energy difference. The complete kinetic model also depends on electronic couplings and reorganization energies, which this work does not compute; and the classical benchmarks themselves are not converged with active-space size, leaving open how these estimators behave as active spaces grow toward classically inaccessible regimes.

Target Audience

Quantum-computing researchers working on near-term algorithms for electronic structure; quantum-chemistry method developers interested in measurement and readout costs rather than state preparation; computational chemists and materials scientists working on transition-metal photosensitizers, MOFs, or photodynamic therapy agents; and industry R&D teams in quantum hardware or pharmaceutical discovery evaluating whether current devices can deliver chemically meaningful numbers. Readers seeking a first introduction to variational quantum eigensolvers or active-space methods will find the paper dense; those already familiar with the mixed estimator (e.g., from Lenihan et al.) will find it a directly useful extension.

Note: the supplied paper text is truncated partway through the hardware results section, so the final quantitative hardware comparisons on Helios-1 for the larger reference sizes and the (10,10) device results are not fully reported here.

Authors’ abstract

Accurate modelling of chemical reaction rates requires precise correlated energy differences. Quantum computing offers a potential route to their determination, but extracting energies from quantum-prepared states can demand prohibitive sampling. We demonstrate measurement-efficient correlated energy estimation using an electronically challenging H$_2$DBB--Ru model derived from a Ru-functionalized hafnium metal-organic framework relevant to radiodynamic and photodynamic therapies. The mixed estimator combines a quantum-prepared target with sparse multiconfiguration references and recovers ground-state corrections using magnitude-only compute-uncompute measurements. Analytical variance estimates predict 1.6 mHa precision with a few hundred shots, about 100 times fewer than the single-determinant formulation and three orders of magnitude fewer than optimized Hamiltonian-term sampling. With approximate targets, it reaches comparable accuracy using simpler state preparation than direct expectation-value estimation. Emulator and trapped-ion experiments on Quantinuum Helios validate the approach up to (10,10), including a 50-qubit circuit combining AI-assisted state preparation with error detection, providing a measurement-efficient route to correlated energy differences for challenging chemical problems.