{"id":"8fd3dcc0-13a4-4088-9ec1-c3793a7cfc31","arxiv_id":"2507.07667","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-phase VQE that penalizes iteration-to-iteration changes in the 1-RDM improves molecular properties and, in a small active space, energy accuracy relative to CISD.","lead":"This paper adds a penalty term to the VQE cost function that suppresses changes in the one-particle reduced density matrix between optimization steps. The modified optimizer improves dipole moments, electron densities, and charges for CH5+, and in one small active space it also fixes large energy errors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains are not causally tied to the 1-RDM penalty: Phase 2 lacks a control that continues energy-only optimization for the same number of extra steps, and the penalty term is numerically tiny relative to the energy term.","rationale":"The paper is a clear, honest benchmark study, and the reported numbers appear internally consistent. However, the advertised causal claim requires that the ΔRDM penalty, not merely extra optimization, drives the improved energies and properties. The reader's weakest assumption correctly identifies the missing control and the untested relation between ΔRDM and 1-RDM accuracy. My stress-test sharpens this by noting the large scale mismatch in Eq. 19, which makes it especially important to verify that the penalty actually changes the optimization trajectory. The proposed control with wRDM=0 would settle the causal question for the most dramatic result, Table 5. Since this concern is the same as the reader's and the appropriate verdict is already CONDITIONAL pending such a control, no verdict adjustment is needed.","tokens_in":23423,"tokens_out":5262,"duration_ms":62283,"concrete_test":"Rerun the GateFabric (2,2) case for R=1.3 Å with Phase 2 cost L = E only (wRDM=0), keeping the same SGD learning rate, the same Phase-1 stopping point, the same energy rejection threshold Elimit, and the same number of Phase-2 steps as in Table 5. Compare final energy, dipole moment, and 1-RDM error relative to CISD. If these match the reported VQE* values, the improvements come from continued optimization and the rejection rule, not from the 1-RDM penalty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that simultaneous optimization of energy and 1-RDM improves both energies and molecular properties. The mechanism is implemented through Eq. 19, L = wE E + wRDM ΔRDM, where ΔRDM (Eq. 18) is the RMSD between consecutive 1-RDMs. This is a convergence metric, not an error metric, and the paper never verifies that reducing ΔRDM reduces the distance to the CISD 1-RDM. More importantly, no control is reported in which Phase 1 energy minimization is continued for the same number of additional steps with wRDM=0. The GateFabric result in Table 5 is especially concerning: Phase 1 stops with energy -39.6578 and ΔRDM ≈ 1.5e-2, meaning the optimization was still moving along near-energy-null directions. Phase 2 then reaches -39.9176 in only 11-12 steps. This could simply be continued SGD escaping a plateau, aided by the rejection rule that discards energy-increasing updates, rather than an effect of the 1-RDM penalty. Numerically, the penalty is very small: for R=1.3, ΔRDM ≈ 1.5e-2 while E ≈ -39.7, so the penalty contributes about 4e-4 of the objective's magnitude. Its gradient could still matter if the energy gradient vanishes, but the paper does not demonstrate this. Similarly, for k-UpCCGSD the energy changes by only 10^-5 to 10^-7 Hartree, yet dipole-moment errors improve by orders of magnitude (Table 2); this is consistent with extra optimization moving parameters along flat energy directions, which would also happen without the penalty. Thus, the causal role of the ΔRDM term is the least secure part of the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-phase VQE protocol in which, after energy minimization (Phase 1), a penalty term is added to the cost function in Phase 2. The penalty is the RMSD between the 1-RDMs of consecutive iterations (ΔRDM, Eq. 18), yielding L = w_E E + w_RDM ΔRDM (Eq. 19). The method is tested on the CH5+ dissociation profile using the k-UpCCGSD ansatz with active space (4,4) and the GateFabric ansatz with active space (2,2), comparing to CISD references in the same active spaces. The authors report that for k-UpCCGSD the energies barely change (10^-5–10^-7 Hartree) while dipole moments, electron densities, electrostatic potentials, and Mulliken charges improve substantially, and for GateFabric the energy error drops from ~0.14–0.26 Hartree to ~10^-8 Hartree. The central claim is that simultaneous optimization of energy and 1-RDM is an effective strategy for improving both energies and molecular properties in VQE.","tokens_in":23753,"tokens_out":6665,"duration_ms":70784,"significance":"If the causal role of the 1-RDM penalty were established, the approach could be a practical, simple addition to VQE for improving density-derived properties, which are relevant for molecular simulations on near-term quantum hardware. The paper addresses a real limitation of energy-only VQE and the idea of borrowing density convergence criteria from classical electronic structure is intuitive. However, the current evidence does not support the central claim: the penalty term is not validated against any external error measure, no control experiment isolates the effect of the penalty from extra optimization steps and the rejection rule, and the hyperparameters are chosen without sensitivity analysis. Because the central mechanism is plausible but unproven, the paper's significance is conditional on substantial additional numerical experiments.","major_comments":[{"comment":"The penalty term ΔRDM is the RMSD between consecutive 1-RDMs, not the distance to a reference density. Reducing ΔRDM only indicates that successive iterates become similar; it does not imply that the 1-RDM is closer to the exact or CISD 1-RDM. In fact, an optimization that converges to a wrong local minimum will have small ΔRDM but large error. The manuscript nowhere compares ΔRDM to an external error measure, such as ||D - D_CISD||, either during training or at the end of optimization. The statement in Sec. 3.1 that 'the RMSD of 1-RDM was reduced, improving the quality of the description of this matrix' is therefore a logical non sequitur and undermines the central claim that the penalty improves 1-RDM fidelity.","section":"Sec. 2.4, Eq. (18)"},{"comment":"The reported GateFabric improvement (from -39.6578 to -39.9176 in 11–12 Phase 2 steps) is not causally tied to the 1-RDM penalty. Phase 2 differs from Phase 1 in two ways simultaneously: it adds the penalty term to the cost function and it introduces an energy-rejection rule (updates that increase energy beyond E_limit are rejected). No control is reported in which Phase 1 energy minimization is continued for the same number of extra steps with w_RDM = 0, with and without the rejection rule. Given that the penalty term is numerically tiny (ΔRDM ≈ 1.5e-2 versus E ≈ -39.7, so the penalty contributes about 4e-4 of the objective's magnitude), the observed 0.26 Hartree improvement could simply be continued SGD escaping a plateau, aided by the rejection filter, rather than an effect of the density penalty. The authors need to demonstrate that the gradient of the penalty term, not the energy gradient, is responsible for the escape.","section":"Sec. 3.2, Table 5"},{"comment":"For k-UpCCGSD, the energy changes by only 10^-5–10^-7 Hartree between VQE and VQE*, yet property errors improve by orders of magnitude (e.g., the dipole moment error at R = 1.8 Å changes from -4.65e-1 to -7.00e-5 Debye). This pattern is exactly what one would expect if the optimization is moving along near-null directions of the energy landscape, but such motion is not specific to the 1-RDM penalty; additional energy-only steps could produce the same effect. Without a control in which Phase 1 is extended for the same number of iterations, the improvements in molecular properties cannot be attributed to the penalty term. This is a load-bearing gap in the argument.","section":"Sec. 3.1, Tables 1 and 2"},{"comment":"The cost function L = w_E E + w_RDM ΔRDM is not a fixed objective because ΔRDM depends on the previous iteration's parameters; the reference density changes at every step. The optimization is therefore a temporal-difference scheme rather than a standard minimization of a static function. The manuscript does not discuss this aspect or its implications for convergence and for the meaning of 'simultaneous optimization of energy and 1-RDM'. Additionally, the hyperparameters (w_E, w_RDM, E_tol, RDM_tol, n_r, E_limit, and the SGD learning rate) are all fixed to hand-picked values with no sensitivity analysis, so the robustness of the reported improvements is unknown.","section":"Sec. 2.4, Eq. (19)"}],"minor_comments":[{"comment":"There are many typographical issues, including 'ans\\\"atzes' instead of 'ansätze' and inconsistent spacing in 'VQE *'. The manuscript would benefit from a careful proofread.","section":"Abstract and throughout"},{"comment":"The text refers to 'Figure 5' twice and also to 'Figure 4' with repetitive phrasing; the figure captions and cross-references should be reconciled.","section":"Sec. 3.1.2"},{"comment":"The 'Steps' column lists two numbers (e.g., '7 (13)') and the caption says the values in parentheses refer to Phase 2, but it is unclear whether the first number is Phase 1 steps and the second is total steps or Phase 2 steps. Please clarify the notation.","section":"Table 1 and Table 5"},{"comment":"The sentence 'In the code developed to optimize 1-RDM in the VQE flow, the following threshold condition:' is a fragment; it should be rewritten to introduce the list of parameters properly.","section":"Sec. 2.5"},{"comment":"Reference [31] lists 'MJ Frish' but the correct author is 'M. J. Frisch'.","section":"References"},{"comment":"The flowchart is difficult to follow, particularly the decision branches for Phase 1 and Phase 2. A cleaner diagram with labeled paths would improve readability.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The central idea is potentially interesting, but the current manuscript does not provide the control experiments needed to support the causal claim. The absence of a Phase 1 extension control, the lack of any validation of ΔRDM against an external error metric, and the confounding rejection rule are serious experimental-design issues that cannot be fixed by rewriting alone. I would encourage the editor to request a re-analysis with proper controls before considering publication. Also, the paper does not make any code or data publicly available, which limits reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful content is the two-phase VQE that appends a penalty on the RMSD between successive 1-RDMs (Eqs. 18–19) and reports better molecular properties for CH5+ with k-UpCCGSD and better energies for GateFabric. The property benchmarking is the strongest part: dipole moments, density maps, QTAIM critical points, MESP, and Mulliken charges, all against CISD, with consistent improvements. That is real work and useful for people who care about NISQ-era property calculations. The idea is simple, clearly explained, and honestly motivated by classical SCF convergence practice.\n\nThe soft spot is the causal claim. ΔRDM penalizes iteration-to-iteration change, not error relative to a reference. The paper never shows that reducing ΔRDM brings the 1-RDM closer to CISD. And there is no control where Phase 1 energy minimization continues for the same number of extra steps with wRDM=0. For GateFabric, Phase 1 stops at 23–41 steps while ΔRDM is still ~1e-2, so the optimizer is on a plateau; Phase 2 then reaches the CISD energy in 11–12 steps. That could just be continued SGD escaping the plateau, with the energy-rejection rule (which discards energy-increasing updates) doing the work. The penalty term is numerically tiny—about 4e-4 of the objective at R=1.3—so without evidence that the energy gradient vanishes, the penalty's causal role is speculative. For k-UpCCGSD, the energy changes by 1e-5 to 1e-7 while properties change by orders of magnitude; that is consistent with parameter movement along flat energy directions, which would also happen without the penalty.\n\nOther weaknesses are proportionate: one molecule, STO-3G, active spaces (4,4) and (2,2), single stochastic runs with no error bars, and hand-set weights without sensitivity analysis. No code or data release, which limits reproducibility.\n\nThat said, the central observation—energy convergence does not guarantee 1-RDM accuracy, and property estimation can be improved—is sound and worth testing properly. The paper deserves a serious referee, but it needs a control experiment, error bars, and a code release before the claim is accepted.","headline":"A plausible but unproven 1-RDM penalty for VQE: the property gains look real, yet without a Phase 1-extension control the causal claim is not established.","tokens_in":24360,"tokens_out":2137,"would_cite":false,"duration_ms":24664,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a penalty for 1-RDM drift between iterations to the VQE cost function improves both molecular-property accuracy and, for poorly expressive ansätze, the energy itself, reducing GateFabric errors from 0.14–0.26…","keywords":["variational quantum eigensolver","one-particle reduced density matrix","molecular properties","quantum chemistry simulation","CH5+ dissociation","k-UpCCGSD","GateFabric ansatz","density matrix optimization"],"falsifier":"Run the same two-phase protocol with the $\\Delta_{\\mathrm{RDM}}$ penalty weight set to zero, or with a penalty on an unrelated observable, keeping the Phase 2 step count, learning rate, and energy-rejection threshold identical; if the reported improvements in dipole moments and GateFabric energies persist, the 1-RDM penalty itself is not what causes them.","tokens_in":23207,"feed_emoji":"⚛️","tokens_out":11402,"duration_ms":107137,"temperature":0.7,"pith_summary":"Variational quantum eigensolvers (VQE) usually minimize only the energy, but molecular properties such as dipole moments and electron densities are read off from the one-particle reduced density matrix (1-RDM), and an energy-minimal state need not have an accurate 1-RDM. This paper proposes a two-phase VQE in which the first phase minimizes energy and the second adds a penalty proportional to the root-mean-square change of the 1-RDM between iterations, so energy and density matrix are forced to converge together. On the CH5+ dissociation curve, the procedure leaves already-accurate k-UpCCGSD energies nearly unchanged while cutting errors in dipole moments, electron densities, electrostatic potentials, and Mulliken charges, often by one to four orders of magnitude. For a GateFabric ansatz whose energy error before the penalty is 0.14–0.26 Hartree, the same procedure brings the energy to within about 1e−8 Hartree of the CISD reference. The paper concludes that energy alone is not a sufficient convergence criterion for VQE and that simultaneous optimization of energy and 1-RDM is an effective strategy.","feed_headline":"Density-matrix penalty cuts VQE error from 0.26 to 1e-8 hartree","feed_subtitle":"Penalizing 1-RDM drift between iterations also sharpens dipole moments and electron densities for CH5+.","key_machinery":"The central object is the one-particle reduced density matrix, $D_{pq} = \\langle \\Psi(\\theta) | a_p^\\dagger a_q | \\Psi(\\theta) \\rangle$, symmetrized as $\\tfrac{1}{2}\\langle \\Psi | a_p^\\dagger a_q + a_q^\\dagger a_p | \\Psi \\rangle$ so that measurements give real values. The mechanism that carries the argument is the two-phase cost function $L = w_E E + w_{\\mathrm{RDM}} \\Delta_{\\mathrm{RDM}}$, with $\\Delta_{\\mathrm{RDM}}$ defined as the RMSD between consecutive 1-RDMs; this penalty converts the classical electronic-structure convergence criterion of a stable density matrix into a term the variational optimizer must minimize while the energy-rejection rule keeps the energy from degrading. The active-space 1-RDM from VQE is merged with the Hartree-Fock 1-RDM for frozen orbitals, which is what lets the method report full-molecule densities, dipole moments, and charges from a small quantum circuit.","core_discovery":"The paper's central claim is that optimizing the 1-RDM alongside the energy, rather than relying on energy minimization alone, improves the accuracy of both energies and density-derived molecular properties in VQE. The algorithm runs ordinary energy minimization in Phase 1, then switches to the cost function $L = w_E E + w_{\\mathrm{RDM}} \\Delta_{\\mathrm{RDM}}$, where $\\Delta_{\\mathrm{RDM}}$ is the RMSD between the 1-RDMs of consecutive iterations, with a rule that rejects parameter updates that worsen the energy beyond a set limit. Tests on CH5+ show that with k-UpCCGSD in an active space (4,4), Phase 2 shrinks the reported errors in dipole moments, electron density topology, electrostatic potential, and Mulliken charges even though the energy barely moves. With GateFabric in active space (2,2), where ordinary VQE misses the CISD energy by 0.14–0.26 Hartree, the two-phase procedure reduces the error to about $10^{-8}$ Hartree. The authors take this as evidence that energy is not a sufficient convergence criterion and that simultaneous 1-RDM optimization is an effective strategy for variational quantum algorithms.","pith_inferences":["A control that continues Phase 1 energy minimization for the same number of extra steps, with the same rejection rule but no $\\Delta_{\\mathrm{RDM}}$ penalty, would separate the penalty's effect from the effect of simply optimizing longer; the GateFabric rescue may depend on that distinction.","Because $\\Delta_{\\mathrm{RDM}}$ measures self-consistency rather than distance to the true density matrix, a natural strengthening is to penalize a direct deviation from a reference-density proxy, such as natural-orbital occupancies from a cheaper classical calculation.","On noisy hardware, $\\Delta_{\\mathrm{RDM}}$ is a difference of measured expectation values, so its sampling noise will exceed the energy's; noise-aware weight tuning and extra measurement shots for Phase 2 are a practical follow-up."],"forward_implications":["For ansätze whose energies are already close to a high-level reference, the 1-RDM penalty does not move the energy much but can substantially improve density-derived properties.","For ansätze with large energy errors, the same Phase 2 penalty can rescue the energy itself, reducing errors by several orders of magnitude.","Energy convergence alone is an unreliable criterion for wavefunction quality; monitoring the 1-RDM convergence is needed when molecular properties are the goal.","The approach preserves the standard VQE structure, so it can be added without iterative Hamiltonian reconstruction or full 1-RDM diagonalization."],"supporting_citations":[{"why":"Introduces the variational quantum eigensolver framework that the two-phase algorithm modifies.","marker":"[17]"},{"why":"Defines the variational quantum algorithm framework and the role of the cost function in parameter optimization.","marker":"[18]"},{"why":"Supplies the 1-RDM expressions for dipole moments, electron density, and Mulliken populations used to judge accuracy.","marker":"[25]"},{"why":"Supplies Hartree-Fock orbitals and the atomic-orbital integrals needed to convert the active-space 1-RDM into full molecular densities and properties.","marker":"[30]"},{"why":"Supplies the CISD reference energies and optimized geometries that define the accuracy targets.","marker":"[31]"},{"why":"Defines the k-UpCCGSD ansatz used in the accurate-energy scenario.","marker":"[37]"},{"why":"Defines the GateFabric ansatz used in the poor-energy scenario.","marker":"[38]"},{"why":"Provides the quantum simulator and automatic-differentiation optimizer in which the two-phase VQE is implemented.","marker":"[66]"},{"why":"Performs the QTAIM topological analysis of electron density at nuclear and bond critical points.","marker":"[68]"}],"fun_headline_variants":["Penalizing 1-RDM drift cuts VQE energy error to 1e-8 hartree","Two-step VQE: energy first, then density matrix, boosts accuracy","Optimizing 1-RDM improves VQE molecular properties and energy","Adding 1-RDM penalty to VQE slashes error and sharpens dipoles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that stopping the 1-RDM from changing between iterations is the same as making it more accurate; the paper does not measure the 1-RDM against an external reference during training or compare Phase 2 with a control that keeps optimizing the energy alone for the same number of extra steps.","fun_headline_variants_meta":{"raw":{"variants":["Penalizing 1-RDM drift cuts VQE energy error to 1e-8 hartree","Two-step VQE: energy first, then density matrix, boosts accuracy","Optimizing 1-RDM improves VQE molecular properties and energy","Adding 1-RDM penalty to VQE slashes error and sharpens dipoles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3914,"prompt_tokens":1101,"completion_tokens":2813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":717,"completion_tokens_details":{"reasoning_tokens":2725}},"tokens_in":717,"tokens_out":2813,"duration_ms":21071,"temperature":1.0,"reasoning_tokens":2725,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:35:19.371624+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same two-phase protocol with the $\\Delta_{\\mathrm{RDM}}$ penalty weight set to zero, or with a penalty on an unrelated observable, keeping the Phase 2 step count, learning rate, and energy-rejection threshold identical; if the reported improvements in dipole moments and GateFabric energies persist, the 1-RDM penalty itself is not what causes them.","supporting_citations":[{"cited_title":"A variational eigenvalue solver on a photonic quantum processor","cited_arxiv_id":null,"evidence_quote":"Introduces the variational quantum eigensolver framework that the two-phase algorithm modifies."},{"cited_title":"Modern quantum chemistry: introduction to advanced electronic structure theory","cited_arxiv_id":null,"evidence_quote":"Supplies the 1-RDM expressions for dipole moments, electron density, and Mulliken populations used to judge accuracy."},{"cited_title":"Pyscf: the python- based simulations of chemistry framework","cited_arxiv_id":null,"evidence_quote":"Supplies Hartree-Fock orbitals and the atomic-orbital integrals needed to convert the active-space 1-RDM into full molecular densities and properties."},{"cited_title":"Gaussian 09, revision a","cited_arxiv_id":null,"evidence_quote":"Supplies the CISD reference energies and optimized geometries that define the accuracy targets."},{"cited_title":"Generalized unitary coupled cluster wave functions for quantum computation","cited_arxiv_id":null,"evidence_quote":"Defines the k-UpCCGSD ansatz used in the accurate-energy scenario."},{"cited_title":"Local, expressive, quantum-number-preserving vqe ansätze for fermionic systems","cited_arxiv_id":null,"evidence_quote":"Defines the GateFabric ansatz used in the poor-energy scenario."},{"cited_title":"Critic2: A program for real-space analysis of quantum chemical interactions in solids","cited_arxiv_id":null,"evidence_quote":"Performs the QTAIM topological analysis of electron density at nuclear and bond critical points."}],"review_version":1}