Adaptive operator-generated subspaces for effective many-body Hamiltonians

AI summary

Overview

Research area: Quantum computing for electronic structure — specifically quantum subspace diagonalization (QSD) methods and their interface with classical electronic-structure/embedding workflows that produce effective many-body Hamiltonians.

Technical level: Advanced. The paper assumes familiarity with second quantization, Jordan–Wigner mapping, Pauli-word measurement, generalized eigenvalue problems, and quantum subspace expansion/Krylov methods.

Scope (one sentence): The paper introduces and benchmarks A-CASE (Adaptive Clifford-Algebra Subspace Eigensolver), a single-reference adaptive subspace method that reconstructs overlap, Hamiltonian, and observable matrices from Pauli-word expectations on one fixed reference state, and evaluates it on a 72-run ladder of molecular, Hubbard, and Kitaev systems against fixed Krylov, Pauli QSE, generator-coordinate, ADAPT-VQE, and an exact ADAPT-GCIM implementation.

What This Paper Is About

Classical electronic-structure pipelines (DFT, Wannierization, screening, embedding) end by producing an effective many-body Hamiltonian, but quantum algorithm demonstrations typically start from a hand-built qubit Hamiltonian and report only a ground-state energy. The paper builds the missing downstream connection: a solver that consumes a standard FCIDUMP integral file, respects an explicit sector and reference-state contract, and returns energies, correlations, and spectroscopic response from a single measured data set.

The core methodological problem is that a subspace basis can be small in vector count yet expensive to measure, and its generalized eigenproblem can be ill-conditioned so small measurement errors are amplified. A-CASE is designed to keep the basis virtual on one reference state — so only one state ever needs preparing — while explicitly tracking measurement width, overlap conditioning, and symmetry leakage alongside energy.

Key Contributions

1. A strict, dependency-light FCIDUMP-to-qubit boundary. A parser reconstructs the chemist-notation Hamiltonian directly, restores packed integral symmetries, accepts Fortran D exponents, validates NORB, NELEC, MS2, sentinels, duplicates, finiteness, and electron–spin parity, rejects unrestricted IUHF=1 records, and binds every result to a SHA-256 digest of the input. The convention is independently checked against an OpenFermion reconstruction agreeing over 185 Pauli terms to a maximum coefficient difference of 8.74 × 10⁻¹⁶.

2. Overlap-aware, resource-aware adaptive growth. Candidate generators are scored by the predicted energy lowering of a 2 × 2 generalized pencil that retains the overlap block (so a rescaled duplicate of the current Ritz state scores zero), with optional penalty for newly introduced Pauli words, and rejection of candidates that fall below an orthogonality floor, annihilate the reference, or exceed a declared sector-leakage tolerance. Conditioning, word-universe growth, and retained rank are recorded together in a resource ledger.

3. Response without materializing the Ritz state. The same single-reference Pauli-expectation bank that supplies (S, H) also supplies projected observables Q, from which positive-frequency Lehmann lines, static susceptibility, and Lorentzian-broadened spectra are computed. Finite-shot variability is propagated through the entire nonlinear pipeline (thresholding, diagonalization, root matching, weights, susceptibility, broadening) by grouped bootstrap resampling, with the explicit label certified=False.

4. A matched-contract comparison suite. A 72-run committed ladder plus one explicitly fixed contract on H₄ (same FCIDUMP digest, same (N = 4, S_z = 0) sector, same Hartree–Fock reference, same eight-addition budget, exact arithmetic) comparing A-CASE against Pauli QSE, Hamiltonian-power Krylov, generator-coordinate, exact ADAPT-VQE, and an exact implementation of ADAPT-GCIM's published selector and fixed-angle rule on the same 26-excitation pool.

Main Findings

• The interchange boundary is independently reproducible. The two-site Hubbard dimer (t = 1 eV, U = 4 eV) reaches the analytic sector ground energy to 6.7 × 10⁻¹⁶ eV, and all six nontrivial observables agree to nine printed digits. On linear H₄ in STO-3G with CAS(4e,4o), the mapped (N = 4, S_z = 0) sector energy agrees with an external determinant-FCI result to 3.1 × 10⁻¹⁵ Ha.

• Cold A-CASE misses chemical accuracy on H₄. Eight adaptive additions give a nine-dimensional subspace with a 3.019 mHa error; the complete 27-dimensional singles/doubles subspace reaches 0.766 mHa, and the full (N = 4, S_z = 0) sector has dimension 36. Chemical accuracy is defined in the paper as 1.6 mHa.

• Changing the reference crosses chemical accuracy at fixed A-CASE budget. Replacing the Hartree–Fock determinant by a two-operator ADAPT-VQE state — which by itself is 27.091 mHa, an order of magnitude worse than the cold A-CASE result it rescues — moves the nine-vector error from 3.019 mHa to 0.342 mHa, with κ_S staying near unity and W growing by under two percent. The paper labels this a fixed-A-CASE-budget statement, not a matched total-cost claim.

• The warm start is not monotonic in ADAPT depth. Deeper warm starts give better ADAPT states (13.779 mHa at k = 4, 6.791 mHa at k = 6) but worse A-CASE results (0.612 mHa and 0.768 mHa), suggesting the determinant-excitation candidate pool is coupled to the choice of reference.

• Preparation count and measurement width trade against each other. Under the matched contract, the operator route needs one state preparation against ADAPT-VQE's ninety, bought at roughly an order of magnitude more measured words: ADAPT-VQE reads 2,424 words to score its pool and 185 to measure its final energy, against A-CASE's 14,401–15,783 and 2,240–7,371. The paper states which side of the trade is favorable is a hardware question it does not answer.

• Measurement width is a property of the generator representation, not the subspace. The same retained subspace costs 7,371 words when generated by determinant excitations and 2,240 when generated by the 160 odd-Y Pauli words they decompose into — identical energy to 4 × 10⁻¹⁶ Ha, same M = 9, same κ_S = 1, all nine principal angles zero, at 3.3× different measurement cost. On the 2 × 2 Hubbard plaquette the same substitution takes 2,519 words to 413, a factor of six. A W quoted without its generator resolution is therefore not comparable across methods.

• Fixed Krylov is more accurate on the molecular rungs and usually narrower. On equilibrium H₄, Krylov at M = 9 reaches 1.05 × 10⁻⁸ Ha against A-CASE's 3.02 × 10⁻³ Ha, using 4,224 words against 7,371, but with an overlap condition number of 6.6 × 10¹⁰ versus 1 for A-CASE. Across the molecular rungs Krylov's condition numbers sit in the 10⁸–10¹⁰ range. The one rung where the operator basis is narrower at determinant resolution is stretched H₂O (4e,4o), at 7,783 words against 8,192.

• The Krylov width is made computable by an algebraic identity. Setting A_k = H^k gives A_i†A_j = H^{i+j} and A_i†HA_j = H^{i+j+1}, reducing the union from O(M²) element operators to 2m + 2 powers, linear in M and running in seconds. On rungs small enough for direct enumeration the two agree exactly, including the tracked counts of 24 and 64 words for H₂ and LiH.

• Round-off makes untruncated Krylov word counts irreproducible. After seventeen products, thousands of words carry coefficients near 10⁻¹¹ absent from the exact operator; the reported counts use a 10⁻⁸ threshold under which A-CASE's 7,371 H₄ words are unchanged. The untruncated count differs between runs because a Hamiltonian perturbed at 10⁻¹⁶ has different round-off-level support.

• Exact ADAPT-GCIM on the same local pool. At the nearest size match (k = 4, M = 8), effective rank 6, κ_S = 32.9, error 13.364 mHa with 36/28 Hamiltonian/overlap state pairs; at the iteration match (k = 8, M = 16), effective rank 12, κ_S = 1.09 × 10³, error 10.674 mHa with 136/120 state pairs. The selector uses 98 and 180 candidate gradients with the published fixed angle θ = π/4 and no parameter optimization.

• Hubbard closure requires an extension, and it does not generalize. A-CASE reaches the 2 × 2 Hubbard sector ground state at M = 26 with κ_S = 1 only after compound/configuration generators are admitted (error 1.07 × 10⁻¹⁴ t), but the same pool adds no selected level-4 direction on 2 × 3, leaving a 1.09 t error at the same M. On the 2 × 2 Kitaev cluster A-CASE selects six Hamiltonian-power directions, reaches 8.12 × 10⁻¹³ K, and shows κ_S = 3.8 × 10⁴ — because there it is a pruned Krylov construction.

• The leakage rule is stricter than the physics. The odd-Y word pool is rejected outright under the declared tolerance (operator leakage √2 out of particle number), leaving M = 1 and the Hartree–Fock energy (56.057 mHa), yet with the rule disabled the resulting Ritz vector has sector weight 1 to machine precision. The warm-started arm is the only one whose Ritz vector leaves the sector, at weight 0.999882 — the case the operator-level test exists to catch.

• Finite-shot acceptance degrades with conditioning. Holding system, observable, words, groups, shots, basis size, and seeds fixed and varying only the generator family, acceptance falls from 200/200 at κ_S ≈ 1 to 139/200 at κ_S = 9.3 × 10³, with every rejection a moved thresholded rank. The percentile intervals are explicitly heuristic, pointwise, and conditional on surviving replicas; they are not finite-sample confidence certificates, and exclusion of failed draws is not neutral because those are precisely the draws that can widen a spectrum.

• Finite-shot growth is expensive and unreliable. Eight-qubit certified-growth attempts spend 1.5 × 10⁸ to 4.6 × 10⁸ physical shots and frequently stop or abstain with inaccurate energies; a noisy thresholded overlap can even yield an energy below the exact reference, which the paper notes does not violate Rayleigh–Ritz because the measured pencil no longer represents an exact subspace projection.

• Scope limits are stated explicitly. The work does not establish materials accuracy, favorable scaling, or a quantum advantage, and the synthetic dimer is not represented as the output of any DFT, Wannier, cRPA, or DMFT calculation.

Methodology in Plain English

Start from a small set of electrons in a small set of orbitals — here, four electrons in four orbitals extracted from a standard FCIDUMP integral file. Rather than optimizing a deep quantum circuit, prepare just one reference state (typically Hartree–Fock) and consider a family of "generator" operators applied to it. Each generator produces a candidate direction in a small subspace; the matrices describing that subspace are obtained not by preparing new states but by measuring Pauli-word expectation values on the single reference.

Growing the subspace is a greedy loop. For each candidate, build the tiny 2 × 2 problem of "current best state plus this one candidate," including the overlap between them, and ask how much the lowest eigenvalue would drop. Pick the best-scoring candidate, provided it is not nearly a duplicate of what is already there and does not break particle number or spin symmetry. Append one row and column to the cached matrices. A shared matrix-element bank records which Pauli words were needed, so the same measured data serves energy, correlations, and response.

For the fair comparison, every method is run inside one fixed contract: same input file and digest, same symmetry sector, same reference, same eight-addition budget, exact arithmetic, every arm runs its budget out. Because A-CASE's natural candidates are whole excitations while ADAPT-VQE consumes single Pauli words, A-CASE is run at both resolutions rather than one being declared canonical. Costs are reported as separate currencies — state preparations, Pauli words, state-pair measurements — rather than summed, since converting between them would be an assumption.

For the noise study, one qubit-wise-commuting partition is built for the union of words needed by S, H, and Q_sub, each group yields a joint outcome histogram, and a grouped bootstrap redraws those histograms and reruns the entire downstream chain. Replicas where the retained rank changes, roots collide, or the solver fails are counted and excluded, and the acceptance rate is reported beside every interval.

Why This Matters

The paper argues that a downstream quantum solver is only scientifically useful if its input contract — orbital order, electron number, spin sector, energy units, one- and two-body conventions, provenance, and reference determinant — is as explicit as its eigensolver. It supplies that contract and an executable, independently checked path from an interchange Hamiltonian to energies, correlations, and response, plus a diagnostic ledger that prevents a small basis size from being mistaken for a cheap calculation.

Potential real-world applications:

• Catalysis and reaction-path screening, where active-space Hamiltonians are repeatedly re-derived along a coordinate and a solver must accept them without hand-built qubit models.
• Spectroscopy and response properties — the projected Lehmann formulation targets excitation frequencies, spectral weights, and static susceptibility, which are the quantities of interest in optical and magnetic measurements.
• Embedding workflows (DFT, cRPA, DMFT fronts), where the active-space Hamiltonian arrives with a known sector and provenance that the solver must preserve and validate.
• Benchmarking and validation infrastructure — the FCIDUMP digest binding, external determinant-FCI cross-checks, and the evidence-label taxonomy (reference, exact, finite_sample, heuristic) give auditors a way to distinguish certified from diagnostic results.

Industry relevance: The finding that a W quoted without its generator resolution is not comparable across methods matters for hardware resource estimation — the same nine-dimensional subspace is priced at 7,371 or 2,240

Authors’ abstract

Electronic-structure and embedding workflows terminate in effective many-body Hamiltonians, whereas quantum eigensolver studies often start from hand-built qubit models. We present the Adaptive Clifford-Algebra Subspace Eigensolver (A-CASE), a single-reference, operator-generated Rayleigh--Ritz method. Overlap, Hamiltonian, observable, and response matrices are reconstructed from one shared Pauli-expectation bank, while adaptive growth scores overlap-aware local pencils and rejects symmetry leakage or near-linear dependence. A strict FCIDUMP adapter supplies the active-space boundary. For linear H$_4$ in STO-3G with CAS(4e,4o), the mapped sector agrees with independent determinant FCI to $3.1\times10^{-15}$ Ha. At a nine-vector budget A-CASE has a $3.019$ mHa error; replacing the determinant reference by a two-operator ADAPT-VQE state reduces it to $0.342$ mHa, without implying a matched total-cost advantage. Under a matched contract, the fixed-reference route uses one state preparation versus ADAPT-VQE's ninety but measures roughly an order of magnitude more Pauli words. Exact fixed-angle ADAPT-GCIM gives $13.364$ mHa at the nearest size match and $10.674$ mHa at the iteration match, with its transition-pair burden reported separately. Across a broader benchmark ladder, fixed Krylov bases are generally more accurate and often narrower but substantially less well conditioned. A grouped bootstrap propagates finite-shot variability through thresholding, diagonalization, root matching, spectral weights, susceptibility, and broadening; its bands are explicitly heuristic and conditional, not finite-sample confidence certificates. The work establishes an executable path from an interchange Hamiltonian to energies, correlations, and response, without claiming materials accuracy, favorable scaling, or quantum advantage.