{"id":"3783736d-a9d4-46d9-a4ce-8496befed07a","arxiv_id":"2506.15784","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"An optimized Trotter-based quantum algorithm reduces the estimated cost of simulating X-ray absorption spectra for a Li4Mn2O cathode cluster to 100 logical qubits and 3.1e8 Toffoli gates per circuit.","lead":"Researchers present an optimized quantum algorithm for simulating X-ray absorption spectra of battery materials, cutting the estimated gate count by five orders of magnitude. A daytime runtime on future fault-tolerant hardware could make quantum-based spectral fingerprinting of lithium-excess cathodes practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Resource estimates hinge on unvalidated extrapolation of |<Y3>| ≤ 1 Ha for CAS(22e,18o); if the true Trotter-error matrix element is larger, Δ shrinks and the 3.11e8 Toffoli count grows.","rationale":"The reader's weakest_assumption is exactly the load-bearing point. The algorithm itself is well structured, the N2 benchmark against FCI is genuine support, and the perturbative eigenvalue-error argument is plausible; however, the quantitative resource claims for the target Li4Mn2O system depend on a Trotter-error bound that the authors have not computed at the target size and explicitly defer to future work. Since the bound is a multiplicative input to the gate count, an order-of-magnitude error in B translates directly into a factor of about 3 in Δ and in all Toffoli counts, enough to move the system from 'under 4e8' to a materially larger resource envelope. I do not see a more fundamental flaw: the sampling analysis, the shot budget, the CDF rank choice and the active-volume compilation are all internally consistent given the assumed inputs. The correct response is to keep the paper's conditional status and require either the DMRG/MPO validation or a rigorous norm-based upper bound before the headline resource numbers are taken at face value. Hence no verdict change from the reader's CONDITIONAL.","tokens_in":38818,"tokens_out":16237,"duration_ms":191938,"concrete_test":"Compute |<E_l|Y3|E_l>| for the CAS(22e,18o) Li4Mn2O Hamiltonian using a DMRG/MPS approximation to the ground and low core-excited eigenstates (bond dimension 1000-2000), with Y3 built as an MPO from the CDF fragments via Eq. (25) and contracted, cross-checked by the finite-step estimator of Eq. (50) at t = 1. If the resulting value exceeds 1 Ha, recompute Table I with Δ = sqrt(η/B); if it is ≤ 1 Ha, the listed resource numbers stand. This is the same methodology used for Fig. 13, applied at the target active-space size.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central resource estimate (Table I, Sec. V.C) fixes the Trotter time step through Eq. (42) as Δ = sqrt(η/|<Y3>|) with η = 0.05 Ha, relying on the assumed bound |<E_l|Y3|E_l>| ≤ 1 Ha for the CAS(22e,18o) Li4Mn2O cluster. This bound is not evaluated on the target system; it is extrapolated from the small-system data of Fig. 13, whose largest active space is N = 14 and whose plotted values are for the first five eigenvalues, not specifically the core-excited states that dominate the XAS response. The text itself flags this in Sec. V.C.6: 'Future work should validate this thesis more robustly.' Since Δ enters NTrot linearly in Eq. (33) and enters the total sampling cost in Eq. (32), a true bound B > 1 Ha increases the longest-circuit Toffoli count by about sqrt(B) for the second-order formula, and could substantially erode the claimed five-orders-of-magnitude reduction. This is an input to the resource estimate, not a verified output, so the headline '100 qubits and 3.11e8 Toffoli gates' is conditional on an unvalidated number. The same Y3 also controls first-order eigenvector corrections whose size can be amplified by small energy denominators in the core-excited manifold, so a larger bound or near-degeneracies would affect peak intensities as well as positions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a heavily optimized quantum algorithm for time-domain X-ray absorption spectroscopy (XAS), building on earlier work by Fomichev et al. The main algorithmic contributions are: (i) use of compressed double factorization (CDF) for product-formula time evolution; (ii) a perturbative analysis showing that, for spectroscopy, it is sufficient to control the eigenvalue error of the approximate Hamiltonian, allowing the Trotter step size to be fixed independently of total simulation time; and (iii) a Lorentzian-kernel-aware sampling distribution that drastically reduces the shot budget. Additional circuit-level optimizations include the double-phase trick, combining consecutive basis rotations, reduced rotation precision, BLISS symmetry shifts, and a double-measurement scheme. The paper reports constant-factor resource estimates for the Li4Mn2O cluster with CAS(22e,18o): 100 logical qubits and 3.11e8 Toffoli gates for the longest circuit, with 4.58e11 Toffoli gates for the full spectrum, a claimed five-orders-of-magnitude reduction compared to the unoptimized algorithm. The algorithm is benchmarked on a simulator: it reproduces the N2 valence spectrum against full configuration interaction, and shows convergence for a smaller LiMnO CAS(14e,11o) core-excited spectrum under the core-valence separation approximation.","tokens_in":39097,"tokens_out":10096,"duration_ms":111542,"significance":"If the resource estimates are taken at face value, this is a significant step toward making XAS simulation of industrially relevant battery cathode models practical on early fault-tolerant quantum computers. The paper is commendable for its transparency: the cost formulas in Sec. IV are explicit, all free parameters are enumerated, the error sources are discussed one by one, and the simulator benchmarks do validate the core time-domain algorithm against an independent classical reference for N2. The CDF-based Trotter implementation and the sampling-distribution analysis are careful and appear internally consistent. The main weakness is that the headline resource numbers for CAS(22e,18o) rest on the assumed bound |<Y3>| <= 1 Ha, which is extrapolated from small systems and is explicitly flagged by the authors as needing future validation. A second concern is that the 'eigenvalue-only' error criterion neglects eigenvector corrections that can distort peak intensities. Both issues are load-bearing for the central claims, so the paper needs revision before the resource estimates can be considered reliable.","major_comments":[{"comment":"The displayed inequality Δ ≤ (|⟨Y_{2k+1}⟩| / 1 eV)^{1/(2k)} is inverted with respect to Eq. (41). Given ϵ_trot = Δ^{2k} ⟨Y_{2k+1}⟩ and the requirement ϵ_trot ≤ 1 eV, the correct bound is Δ ≤ (1 eV / |⟨Y_{2k+1}⟩|)^{1/(2k)}. As written, Eq. (42) would force Δ to shrink as the error matrix element decreases and is dimensionally inconsistent. The actual choice in Sec. V.C.6, Δ^2 |⟨Y_3⟩| ≤ 0.05 Ha, follows the corrected form, so this is likely a typographical error; nevertheless Eq. (42) is the formal basis for the Trotter step and must be corrected.","section":"Sec. IV.B.2, Eq. (42)"},{"comment":"The resource estimates in Table I rely on the input |⟨E_l|Y_3|E_l⟩| ≤ 1 Ha for the CAS(22e,18o) Li4Mn2O cluster. This bound is not evaluated on the target system: Fig. 13 only covers active spaces up to N = 14, and the data are for the first five eigenvalues of small systems, not specifically for the core-excited states that dominate the XAS response. Since the Trotter step is set by Δ^2 |⟨Y_3⟩| ≤ 0.05 Ha and the number of Trotter steps scales as τ/Δ, a true bound B would multiply the longest-circuit Toffoli count (and the total sampling cost) by roughly sqrt(B / 1 Ha). The manuscript itself says 'Future work should validate this thesis more robustly.' I request either a more direct estimate of the bound—for example, an MPO/tensor-network evaluation of Y_3 on approximate core-excited eigenstates of the actual CAS(22e,18o) Hamiltonian—or a sensitivity table for B = 1, 10, 100 Ha, so the reader can assess how the headline 3.11e8 and the five-orders-of-magnitude claim degrade with the true value of the bound.","section":"Sec. V.C.6, Fig. 13, Table I"},{"comment":"The argument that controlling the eigenvalue error is sufficient for spectroscopy ignores the first-order eigenvector corrections in Eq. (26) and Eq. (B6). The spectral intensity depends on |⟨mρ|E_l⟩|^2, and the eigenvector correction is proportional to Δ^2 ⟨E_k|Y_3|E_l⟩ / (E_k - E_l). In the core-excited manifold of a transition-metal cluster, energy denominators can be much smaller than the 1 eV eigenvalue tolerance; near-degeneracies would amplify the eigenvector error and distort peak intensities even when the eigenvalue shifts are within tolerance. The manuscript should either bound the eigenvector correction and its effect on the spectral function, or provide numerical evidence (on the smaller N2 and LiMnO benchmarks where exact or converged spectra are available) that peak-height errors remain below the target under the same Δ selection.","section":"Sec. III.B and Appendix B"},{"comment":"The perturbative analysis in Sec. III.B, Eqs. (22)-(27), is developed for the deterministic second-order Trotter formula, but the benchmarks and resource estimates use a 'randomized second-order product formula' (Sec. V.B and Table I caption). The manuscript does not explain how randomization is incorporated into the effective-Hamiltonian analysis, nor whether Eq. (41) remains a valid spectral-error bound for the randomized circuit. Please clarify the randomized construction and either extend the analysis or provide numerical verification that the randomized formula obeys the same (or better) eigenvalue-error scaling.","section":"Sec. III.B vs. Sec. V.B and Table I"}],"minor_comments":[{"comment":"The abstract states 'less than 4×10^8 T gates per circuit', but Table I and the body report Toffoli gates (3.11e8). Since a Toffoli gate is not the same as a T gate, please harmonize the terminology throughout.","section":"Abstract and Table I"},{"comment":"The target error is given as '1 eV' in Sec. IV.B.2, but Sec. V.C.6 sets η = 0.05 Ha (approximately 1.36 eV) and the text elsewhere quotes 1 eV ≈ 0.039 Ha. Please clarify whether the target eigenvalue tolerance and the line-broadening parameter are meant to be identical, and use a consistent conversion between atomic units and eV.","section":"Sec. IV.B and Sec. V.C"},{"comment":"The expression jmax = π/(2ητ) log(1/ϵ_trunc) appears to have an extra factor of π and a missing factor of π inside the logarithm relative to the bound derived in Appendix B.3, ϵ_trunc ≤ e^{-2jmax τη}/π. Since the discrepancy is conservative, it does not invalidate the resource numbers, but the equations should be reconciled.","section":"Sec. IV.B.3, Eq. (45)"},{"comment":"The sentence 'For τ = π/4 Ha, this is satisfied for τ = 4Δ' is confusing because τ and Δ both have units of inverse energy; it would be clearer to write Δ = τ/4 and to give the value of Δ explicitly (Δ ≈ 0.196 Ha^{-1}).","section":"Sec. V.C.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the algorithmic work is careful, but the headline resource estimate is sensitive to an unvalidated Trotter-error bound and to the eigenvalue-only error criterion. The authors are transparent about the first issue, which is to their credit, but the manuscript should either validate the bound more directly or provide a sensitivity analysis. I would not reject the paper: the core algorithm and benchmarks are valuable, and the load-bearing concerns are addressable in revision. I also caution that the classical runtime comparison in Sec. V.D is extremely rough and should be framed strictly as an order-of-magnitude illustration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one is worth reading. The central idea is that for spectroscopy you don't need to control the error in the time-evolution operator; controlling the eigenvalue error of the effective Hamiltonian is enough. The perturbative analysis that fixes the Trotter step from the diagonal matrix element of the leading error term is genuinely new, and it correctly makes the cost linear in time. The kernel-weighted sampling and the double-measurement trick are also real, useful improvements, and the cost formula derivation is careful and transparent. The N2 benchmark against FCI gives the algorithm a credible sanity check, and the LiMnO convergence study is honest about not having a classical reference.\n\nThe soft spot is exactly where the reader and stress-test put it: the headline resource numbers for CAS(22e,18o) rest on the assertion |<Y3>| ≤ 1 Ha, extrapolated from small systems in Fig. 13. The paper says 'future work should validate this thesis more robustly,' and that is the right way to put it. If the true bound is larger, Δ shrinks and the Toffoli count grows by roughly the square root of the ratio. It doesn't kill the paper, because the assumption is explicit and testable, but it means '100 qubits and 3.11e8 Toffoli gates' is a conditional estimate, not a verified cost. The abstract overstates things twice: the classical verification is for N2, not for LiMnO, and the per-circuit count is Toffoli gates, not T gates. Those should be corrected.\n\nNo code or data is released, which is a pity, but the formulas are detailed enough that a competent group could reimplement the small-system tests and check the extrapolation. The classical-vs-quantum runtime comparison is clearly labeled rough and is not load-bearing.\n\nBottom line: this deserves a serious referee. The core algorithmic insight is sound, the paper is honest about its main caveat, and the resource framework will be useful to the community even if the specific Y3 bound later needs revision. My recommendation: send it to peer review, and ask the referees to push on the Y3 extrapolation and on the abstract wording.","headline":"The Trotter step-selection idea is a real advance, but the headline resource numbers are conditional on an unvalidated Y3 bound and the abstract overstates what was verified.","tokens_in":39745,"tokens_out":2286,"would_cite":true,"duration_ms":25565,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81V55"],"pacs":[],"model":"deepseek-v4-flash","headline":"An optimized time-domain algorithm simulates X-ray absorption spectra of a battery cathode cluster with 100 logical qubits and $3.11 \\times 10^8$ Toffoli gates per circuit.","keywords":["quantum chemistry","X-ray absorption spectroscopy","Trotter product formulas","compressed double factorization","fault-tolerant quantum computing","resource estimation","time-domain simulation","battery cathode materials"],"falsifier":"Compute the diagonal matrix elements $\\langle E_l|Y_3|E_l\\rangle$ for the Li4Mn2O CAS(22e,18o) Hamiltonian directly, for example with a matrix-product-state approximation at increasing bond dimension, and check whether the largest value stays at or below 1 Ha; if it exceeds 1 Ha, the quoted Trotter step $\\Delta = \\sqrt{\\eta/1\\ \\mathrm{Ha}}$ is too large and the reported Toffoli counts are underestimates.","tokens_in":38543,"feed_emoji":"🔋","tokens_out":8171,"duration_ms":86013,"temperature":0.7,"pith_summary":"This paper aims to show that X-ray absorption spectroscopy (XAS) of industrially relevant battery cathode materials can be simulated on early fault-tolerant quantum computers at a practical cost. It optimizes the time-domain XAS algorithm for a Li4Mn2O cluster in a CAS(22e,18o) active space, reporting resource estimates of 100 logical qubits and $3.11 \\times 10^8$ Toffoli gates for the longest circuit, a reduction of about five orders of magnitude compared with the unoptimized implementation. The central reason the cost is low is that spectroscopy only requires the eigenvalues of the simulated Hamiltonian to be accurate, not the full time-evolution operator, which allows the Trotter step size to be fixed independently of total evolution time. The paper backs the resource estimates with simulator reconstructions of the absorption spectrum against classical references.","feed_headline":"100 logical qubits suffice for battery cathode X-ray spectra","feed_subtitle":"Optimized Trotter steps and sampling cut the gate count five orders for the Li4Mn2O cathode model.","key_machinery":"The load-bearing objects are the leading Trotter-error operator $Y_3$, obtained from the Baker–Campbell–Hausdorff expansion of the second-order product formula, and the compressed double factorization (CDF) of the electronic Hamiltonian. $Y_3$ controls the eigenvalue error through its diagonal matrix elements: the paper fixes the Trotter step by $\\Delta^2|\\langle E_l|Y_3|E_l\\rangle| \\leq 0.05$ Ha, and the perturbative estimate of this quantity is what decouples the step size from the total simulation time. CDF writes the two-electron integrals as a sum of $L$ fragments, each diagonal in its own single-particle basis, reducing the number of fragments from $O(N^4)$ in a direct Jordan-Wigner picture to $O(N^3)$ with $L = N$; the basis rotations between fragments are implemented with Givens rotations. The third machinery element is the Lorentzian-kernel sampling distribution, which allocates shots to evolution times proportionally to $e^{-\\eta\\tau|j|}$, replacing uniform sampling and saving a factor of about 18 in total shots.","core_discovery":"On the paper's own terms, the discovery is that a time-domain XAS calculation can be reorganized so that its cost is governed by spectral accuracy rather than by generic Hamiltonian simulation error. Using perturbation theory on the effective Hamiltonian produced by a second-order Trotter product formula, the eigenvalue shift is $E'_l = E_l - \\Delta^2 \\langle E_l|Y_3|E_l\\rangle$, so choosing the step $\\Delta$ with $\\Delta^2|\\langle Y_3\\rangle| \\leq 0.05$ Ha keeps peak positions within the 1 eV target broadening. Because this choice depends only on the Hamiltonian and not on how long the evolution is run, each Hadamard-test circuit costs linearly in time instead of superlinearly. Combined with the compressed double factorization of the Hamiltonian, a Lorentzian-weighted sampling distribution, and circuit-level optimizations, this brings the Li4Mn2O CAS(22e,18o) system to 100 logical qubits and $3.11 \\times 10^8$ Toffoli gates per circuit ($4.58 \\times 10^{11}$ including sampling), which the authors estimate runs in under a day on a MHz-clock fault-tolerant machine.","pith_inferences":["As an extension beyond the paper, the eigenvalue-error criterion for Trotter steps should apply to any spectroscopy computed by time evolution under a product formula, not just XAS, provided the observable of interest is dominated by eigenenergy differences.","The $|\\langle Y_3\\rangle| \\leq 1$ Ha bound could be tested classically on intermediate-size clusters between the small systems in Fig. 13 and the full CAS(22e,18o) target; a confirmation would turn the resource estimate from an extrapolation into a computed input.","The kernel-weighted sampling idea is generic: any damped spectral kernel (Lorentzian, Gaussian, or lifetime-broadened line shapes with known decay) could be used to reallocate shots, potentially reducing costs in other time-domain spectroscopy algorithms.","If the method scales as $O(N^3)$ as argued, the crossover against classical state-by-state methods should only widen with active-space size, making K-edge XAS of larger transition-metal oxide clusters a natural next target."],"forward_implications":["For the Li4Mn2O CAS(22e,18o) target, the longest circuit needs only 100 logical qubits and $3.11 \\times 10^8$ Toffoli gates, and the full spectrum with sampling costs $4.58 \\times 10^{11}$ Toffoli gates.","Because the Trotter step no longer needs to shrink as evolution time grows, the cost of each XAS circuit scales linearly with time, which is what makes long-time, fine-resolution spectra affordable.","The simulator benchmarks reproduce the valence absorption spectrum of N2 against a classical full-configuration-interaction reference and show convergence of the core-excited LiMnO spectrum with maximal evolution time.","Combined with an active-volume compilation on 350 logical qubits, the algorithm runs in about $5.6 \\times 10^7$ logical cycles for the longest circuit, under a day at a 1 MHz logical clock rate, versus an estimated month-scale runtime for a restricted-active-space classical approach on the same target."],"supporting_citations":[{"why":"proposed the time-domain XAS algorithm that this paper optimizes and singled out the Li4Mn2O CAS(22e,18o) cluster as a target","marker":"[11]"},{"why":"introduced compressed double factorization, the Hamiltonian representation used to cut the number of fragments to O(N^3)","marker":"[41]"},{"why":"supplies the sum-of-Slaters initial state preparation whose auxiliary qubit count enters the logical qubit formula","marker":"[46]"},{"why":"provides the linear-depth Givens rotation implementation of the basis-change unitaries U^{(l)}","marker":"[49]"},{"why":"supports the L ~ O(N) rank scaling for factorized Hamiltonians that justifies choosing L = N","marker":"[54]"},{"why":"introduced the Monte Carlo sampling of evolution times that the Lorentzian-weighted sampling distribution adapts","marker":"[80]"},{"why":"supplies the Baker–Campbell–Hausdorff expansion used to derive the perturbative eigenvalue error estimate","marker":"[83]"},{"why":"gives the phase-kickback adder used for the single-qubit rotations that dominate the Toffoli cost per Trotter step","marker":"[84]"},{"why":"provides the active volume architecture and cost model used to convert circuit estimates into runtime","marker":"[40]"},{"why":"supports the randomized second-order product formula choice as an optimal balance of precision and cost","marker":"[77]"}],"fun_headline_variants":["Quantum XAS: 100 qubits for battery cathodes","Spectral accuracy cuts Trotter steps for XAS","Battery spectra on a quantum computer: 100 qubits","Faster XAS simulation via eigenvalue-error control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The resource counts rest on the extrapolated bound $|\\langle E_l|Y_3|E_l\\rangle| \\leq 1$ Ha for the CAS(22e,18o) Li4Mn2O cluster, obtained from small-system data in Fig. 13; the authors explicitly call for future work to validate this bound more robustly, and if the true value is larger the Trotter step must shrink and the gate counts rise.","fun_headline_variants_meta":{"raw":{"variants":["Quantum XAS: 100 qubits for battery cathodes","Spectral accuracy cuts Trotter steps for XAS","Battery spectra on a quantum computer: 100 qubits","Faster XAS simulation via eigenvalue-error control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1480,"prompt_tokens":1098,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":714,"completion_tokens_details":{"reasoning_tokens":316}},"tokens_in":714,"tokens_out":382,"duration_ms":4539,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:51:20.660611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the diagonal matrix elements $\\langle E_l|Y_3|E_l\\rangle$ for the Li4Mn2O CAS(22e,18o) Hamiltonian directly, for example with a matrix-product-state approximation at increasing bond dimension, and check whether the largest value stays at or below 1 Ha; if it exceeds 1 Ha, the quoted Trotter step $\\Delta = \\sqrt{\\eta/1\\ \\mathrm{Ha}}$ is too large and the reported Toffoli counts are underestimates.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduced compressed double factorization, the Hamiltonian representation used to cut the number of fragments to O(N^3)"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the sum-of-Slaters initial state preparation whose auxiliary qubit count enters the logical qubit formula"},{"cited_title":"Maganas, J","cited_arxiv_id":null,"evidence_quote":"provides the linear-depth Givens rotation implementation of the basis-change unitaries U^{(l)}"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supports the L ~ O(N) rank scaling for factorized Hamiltonians that justifies choosing L = N"},{"cited_title":"Tensor Network enhanced Dynamic Multiproduct Formulas","cited_arxiv_id":"2407.17405","evidence_quote":"introduced the Monte Carlo sampling of evolution times that the Lorentzian-weighted sampling distribution adapts"},{"cited_title":"Hagan and N","cited_arxiv_id":null,"evidence_quote":"supplies the Baker–Campbell–Hausdorff expansion used to derive the perturbative eigenvalue error estimate"},{"cited_title":"Pocrnic, M","cited_arxiv_id":null,"evidence_quote":"gives the phase-kickback adder used for the single-qubit rotations that dominate the Toffoli cost per Trotter step"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the active volume architecture and cost model used to convert circuit estimates into runtime"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supports the randomized second-order product formula choice as an optimal balance of precision and cost"}],"review_version":1}