{"id":"1821abb4-2a15-438e-9385-dd6034c7ab35","arxiv_id":"1908.03674","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"OSV-MP2 analytical gradients with explicit OSV relaxation are derived, implemented, validated against numerical and RI-MP2 references, and applied in BOMD simulations.","lead":"This paper derives and implements analytic energy gradients for OSV-MP2, a local approximation to second-order Møller-Plesset theory, including the exact response of the orbital-specific virtual orbitals. The gradients reproduce canonical MP2 forces to about 1e-4 a.u. and are used to run Born-Oppenheimer molecular dynamics for protonated water and ethanol.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OSV relaxation via non-degenerate perturbation theory is uncontrolled when a retained and a discarded OSV eigenvalue nearly coincide; Eq. 25 can become ill-conditioned near the l_osv selection boundary, and the numerical compensation shown in Fig. 1 is not a general proof.","rationale":"The reader's weakest-assumption analysis already identifies the near-degenerate eigenvalue treatment as the weakest point, and I agree that this is where the central claim is least secure. My read goes slightly further: the problem is not only the absence of a general proof, but also that first-order non-degenerate perturbation theory has no uniform error control as a retained/discarded eigenvalue gap shrinks, and that the fixed-partition derivative in Eq. 25 does not account for selection changes across the threshold. The paper's numerical evidence in Fig. 1, the analytical-versus-numerical gradient comparisons in Sec. 3.1, and the BOMD tests in Sec. IV are real supporting evidence, and they make the method plausible for the tested molecules and thresholds. However, they do not establish that Eq. 25 is valid in the regime most relevant to dynamics, where OSV eigenvalues can approach the selection boundary as atoms move. The correct response is not rejection: the derivation is detailed, the reported gradient errors are small, and the MD results are internally consistent. The concern warrants a conditional acceptance with a specific numerical test near a threshold crossing and either a full proof in the main text or a rigorous bound on the Hadamard-product cancellation. Because this is the same conditional posture the reader already adopted, I keep the verdict unchanged.","tokens_in":25220,"tokens_out":8711,"duration_ms":101204,"concrete_test":"Run a geometry scan for a symmetric molecule such as N2 or C6H6 with l_osv chosen so that one OSV eigenvalue crosses the selection threshold at an intermediate geometry. At geometries on both sides of the crossing, compare the analytical gradient from Eq. 53 with central-difference numerical gradients obtained by re-diagonalizing Tkk (Eq. 2) at each displaced geometry and re-applying the same l_osv partition. If the RMSD grows as |Δω| decreases, or if the central-difference step must be reduced below roughly 10^-5 a.u. to converge, Eq. 25 is not controlled in the near-degenerate regime. A complementary check is to run a short NVE BOMD trajectory through the crossing and monitor total-energy conservation; a force discontinuity at the crossing will appear as a jump or systematic drift in the energy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction is Eq. 25 in Sec. 2.2.2, where the OSV relaxation matrix O^λ_k is obtained from first-order non-degenerate perturbation theory with denominators ω_ν − ω_μ. The invariance theorem Eq. 28 removes rotations within the retained OSV block, but it does not protect rotations between a retained OSV and a discarded OSV whose eigenvalues are close: such pairs straddle the l_osv selection threshold and can have arbitrarily small Δω. Non-degenerate perturbation theory is only asymptotic in 1/Δω; the assertion that the Hadamard product N_{ij}∘ΔG^T in Eq. 36 remains bounded is supported by numerical examples in Fig. 1, not by a mathematical bound. Sec. 1 claims that the degenerate-eigenvalue issue 'does not occur for reasons that will be described', and Sec. 2.2.2 defers the proof of the diagonal Ω^λ to SI S3; neither source provides a general estimate for near-degenerate retained/discarded pairs. Moreover, Eq. 25 differentiates a fixed partition of OSVs into retained and discarded. When a geometry change carries an eigenvalue across l_osv, the selected energy can develop a kink, and the present formalism contains no term representing the discrete change in the OSV selection. Since the BOMD applications in Sec. IV require smooth forces, this is a load-bearing limitation of the 'complete' analytical gradient claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents an analytical gradient theory for RI-OSV-MP2 that explicitly includes the relaxation of the orbital-specific virtual orbitals. The OSV response is obtained by solving a perturbed non-degenerate eigenvalue problem for each diagonal pair-amplitude matrix T_kk, with the key relaxation matrix O^λ_k given by a Hadamard product involving 1/(ω_ν − ω_μ) (Eq. 25). The gradient is decomposed into OSV-specific, MO-specific, and AO-specific contributions, and the implementation uses a Z-vector approach together with coupled-perturbed localization for Pipek–Mezey/meta-Löwdin orbitals. The numerical section compares analytical gradients against numerical gradients (RMSD ~10^-6 to 10^-7 a.u.), against RI-MP2 reference gradients, and against DLPNO-MP2 results, and applies the gradients to geometry optimizations, Born–Oppenheimer molecular dynamics of protonated water clusters, and well-tempered metadynamics of ethanol.","tokens_in":1582,"tokens_out":1414,"duration_ms":92522,"significance":"If the central derivation is accepted, this is a significant contribution: it is, to the best of my knowledge, the first complete analytic OSV-MP2 gradient implementation that treats OSV relaxation explicitly, and it enables local-correlation-quality molecular dynamics. The derivation is largely self-contained, uses no fitted parameters (only the convergence thresholds l_osv and l_pair), and the implementation is validated against independent RI-MP2 references and numerical gradients. The reported BOMD and metadynamics applications demonstrate practical utility. The main weakness is the treatment of near-degenerate OSV eigenvalues and of changes in the OSV selection with geometry, which is addressed only by numerical evidence and not by a mathematical argument.","major_comments":[{"comment":"The claim that the degenerate-eigenvalue issue 'does not occur' is not substantiated in the main text. Equation (25) contains denominators ω_ν − ω_μ, and a retained OSV paired with a discarded OSV of nearly equal eigenvalue that straddles the l_osv threshold can make these denominators arbitrarily small. The invariance in Eq. (28) only removes rotations within the retained OSV block; it does not control the retained/discarded pairs that enter Eq. (25). The numerical compensation shown in Figure 1 for N2 and C6H6 is encouraging but is not a general mathematical argument. The authors should either provide a general bound on the Hadamard product N_ij ∘ ΔG^T (e.g., showing that the numerator vanishes at the same order as Δω when the eigenvalues approach each other) or clearly state the non-degeneracy conditions under which Eq. (25) is valid, and correspondingly qualify the 'complete'/'exact' claim.","section":"Sec. 2.2.2, Eq. (25); Sec. 1"},{"comment":"The gradient derivation differentiates a fixed partition of OSVs into retained and discarded sets. When a geometry change carries an eigenvalue across the l_osv selection threshold, the selected energy can develop a kink, and the working equations (30)–(53) contain no term representing this discrete change in the OSV selection. This is load-bearing for the BOMD applications, which require smooth forces. The paper itself provides related evidence: in Table 7, the H13O6+ simulation with l_pair = 0.001 and losv = 10^-3 exhibits a large energy drift (δE ≈ 51 kJ/mol), and Sec. 4.1 attributes similar pair-screening drift to temporal changes in the number of kept pairs. The same mechanism is expected for OSV selection changes. Please either demonstrate that no OSV eigenvalue crosses the threshold in the reported simulations or extend the theory to account for selection-boundary changes.","section":"Sec. 2.2.2 and Sec. IV (BOMD)"}],"minor_comments":[{"comment":"The invariance of the gradient with respect to rotations among retained OSVs is stated as a theorem, but the proof is deferred to SI S2. Since Eq. (28) is the key simplification that reduces the OSV response to the retained/discarded block, a short derivation or at least a clear statement of the underlying assumptions should appear in the main text.","section":"Sec. 2.2.2, Eq. (28)"},{"comment":"The caption does not state the geometry or the l_osv threshold used for the N2 and C6H6 examples. Please specify these details so that the reader can judge whether the plotted quantities are representative of the near-threshold regime.","section":"Figure 1 caption"},{"comment":"The central working equation (53) relies on intermediates A_ij,αβ, B_kl,ai, and C_kl,ij that are defined only by reference to Ref. 62. For a self-contained gradient theory, the definitions of these intermediates (or at least a brief statement of their structure) should be included or summarized.","section":"Eq. (53) and Sec. 2.4"},{"comment":"There are several typographical and formatting issues, including duplicated words, non-ASCII artifacts in the reference list (e.g., 'MÃÿller', '484âĂŞ–507'), and inconsistent capitalization in Figure 2 ('THE OSV-MP2 gradients'). These should be corrected during revision.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely publishable after a major revision. The numerical implementation is thorough and the gradient accuracy checks are convincing, but the 'exact'/'complete' claim and the BOMD smoothness requirement depend on a rigorous treatment of near-degenerate OSV eigenvalues and of OSV selection changes. The authors should be asked to either supply a mathematical bound for the Hadamard product in Eq. (36) or to soften the claim and discuss the practical safeguards. I would not reject the manuscript, as the outlined issues appear addressable within the scope of a revised paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is the first OSV-MP2 analytic gradient that includes OSV relaxation, and the derivation is mostly solid. The novelty is real—prior gradient work used PNOs or dropped PNO relaxation; here they solve the OSV response as a perturbed eigenvalue problem and prove invariance to rotations among kept OSVs. The result is validated against numerical gradients to ~1e-6/1e-7 a.u. and against RI-MP2 to ~1e-4 a.u. at normal thresholds, and the BOMD test on protonated water clusters reproduces RDF and VDOS features. Credit where due: this is careful, reproducible-by-reimplementation work, with the gradient equations spelled out clearly enough that a competent group could code it up.\n\nThe soft spots, in proportion. First, the near-degenerate OSV eigenvalue issue is real. Eq. (25) uses 1/(omega_nu - omega_mu) denominators, and when a retained and a discarded OSV eigenvalue nearly coincide the relaxation matrix is formally ill-conditioned. The paper argues that the Hadamard product with Nij stays bounded and shows numerical examples, but that is not a proof. The invariance theorem removes kept-kept rotations, not kept-discarded near-degeneracies. The claim in Sec. 1 that the degenerate issue 'does not occur' is too strong without a formal argument; the SI apparently gives a diagonal-relaxation proof, but the main text leans on numerical compensation. This matters for the 'complete' in the title: the gradient is derived at fixed OSV selection, and when a geometry change moves an eigenvalue across l_osv the selected energy can kink. The BOMD energy drift numbers in Table 7 are small, so the practical impact looks minor at normal thresholds, but it is a limitation that should be stated more honestly.\n\nSecond, no code is released. The implementation details are enough to reproduce in principle, but 'principle' is doing work. Third, the invariance proof is in the SI, which weakens the main text for a reader who wants to check the central step.\n\nWho this is for: people working on local correlation gradients, especially OSV/PNO reduced-scaling dynamics. It is an incremental but real capability, and the benchmark suite is sensible. I would send it to a serious referee—the derivation deserves scrutiny, most of it will pass, and the near-degenerate gap should be addressed in revision.\n\nWould I cite it? Yes, as the OSV gradient reference. Bring to reading group? Probably yes for the derivation.\n\nRecommendation: accept for peer review, with a request for a formal or at least numerical-bounded treatment of near-degenerate retained/discarded OSV pairs and a less absolute claim about degeneracy.","headline":"First OSV-MP2 analytic gradient with explicit OSV relaxation, carefully derived and tested; near-degenerate OSV pairs and missing code are the real soft spots, but this deserves refereeing.","tokens_in":26061,"tokens_out":2516,"would_cite":true,"duration_ms":25600,"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 establishes an exact analytic gradient for OSV-MP2 by solving the OSV response as a perturbed non-degenerate eigenvalue problem, with in-subspace rotations proven to vanish.","keywords":["OSV-MP2","analytical gradient","orbital-specific virtuals","local correlation","resolution-of-identity","Born-Oppenheimer molecular dynamics","metadynamics","coupled-perturbed localization"],"falsifier":"Compare OSV-MP2 analytical gradients with finite-difference gradients along a path that forces two eigenvalues of $T_{kk}$ to cross, for instance a symmetric stretch of a linear or high-symmetry molecule; if the analytical-minus-numerical RMSD jumps or the Hadamard product in Eq. (25) diverges, the non-degenerate relaxation treatment is not generally valid.","tokens_in":24975,"feed_emoji":"⚛️","tokens_out":7487,"duration_ms":73395,"temperature":0.7,"pith_summary":"OSV-MP2 is a local approximation to second-order Møller–Plesset theory that replaces canonical virtual orbitals with compact, orbital-specific virtuals, but analytic derivatives of its energy have been difficult because the OSVs themselves respond to nuclear motion. This paper establishes that the OSV response can be solved exactly enough by treating it as a perturbed non-degenerate eigenvalue problem for the diagonal pair-amplitude matrix, and that rotations among the OSVs kept in the calculation never contribute to the gradient, provided the unperturbed Hylleraas residual is well converged. The resulting gradient reproduces canonical RI-MP2 forces to within about $10^{-4}$ a.u. at the normal OSV threshold, matches finite-difference gradients to $10^{-6}$–$10^{-7}$ a.u., and is stable enough to drive Born–Oppenheimer molecular dynamics, including a 200 ps well-tempered metadynamics run for ethanol. A sympathetic reader would care because this removes the main obstacle to using OSV-MP2, not just for energies, as a black-box engine for geometry optimization and finite-temperature dynamics.","feed_headline":"OSV-MP2 forces become exact enough for dynamics","feed_subtitle":"OSV relaxation is solved as a perturbed eigenvalue problem, reproducing canonical RI-MP2 gradients to 1e-4 a.u.","key_machinery":"The carrying object is the semi-canonical MP2 diagonal pair-amplitude matrix $T_{kk}$, whose orthonormal eigenvectors define each occupied orbital's OSV set; the OSVs are selected by the size of its eigenvalues $\\omega_{\\bar\\mu k}$. Under a perturbation, the perturbed OSVs are written $Q_k(\\lambda) = Q_k^0 O_k(\\lambda)$, and the relaxation matrix $O_k^\\lambda$ is obtained from the first-order change of the eigenvalue equation. The load-bearing identity is Eq. (25), $O_k^\\lambda = \\Delta G_k \\circ [Q_k^\\dagger T_{kk}^\\lambda Q_k]$, where the Hadamard product with the inverse eigenvalue-difference matrix $\\Delta G_k$ encodes the non-degenerate perturbation solution, together with Eq. (28), which states $\\partial E_c^\\lambda / \\partial [O_k^\\lambda]_{\\bar\\mu\\bar\\nu} = 0$ for rotations inside the retained subspace; together they reduce the OSV response to a well-defined rotation between kept and discarded OSVs.","core_discovery":"The central claim, stated as the author would state it: the OSV-MP2 correlation-energy gradient can be written as a sum of AO-, MO-, and OSV-specific parts, and the OSV part is computable exactly from the first-order perturbation of the eigenvalue equation $T_{kk}(\\lambda) Q_k(\\lambda) = Q_k(\\lambda) \\Omega_k(\\lambda)$ that defines the OSVs. Differentiating this equation gives the OSV relaxation matrix $O_k^\\lambda = \\Delta G_k \\circ [Q_k^\\dagger T_{kk}^\\lambda Q_k]$ with $[\\Delta G_k]_{\\bar\\mu\\bar\\nu} = 1/(\\omega_{\\bar\\nu} - \\omega_{\\bar\\mu})$, and a proof in the paper's Supporting Information shows the gradient is invariant to rotations among retained OSVs, so only rotations between retained and discarded subspaces enter. Because those pairs carry different eigenvalues, the non-degenerate formula is the right tool, and the paper demonstrates numerically that near-degeneracies are smoothed by the Hadamard-product numerator. The working equations are implemented with resolution-of-identity integrals, a Z-vector equation for occupied-virtual response, and a coupled-perturbed localization treatment for the meta-Löwdin localization function.","pith_inferences":["The same relaxation ansatz should transfer to OSV-based coupled-cluster methods, because it decouples pair-specific rotations and only uses the diagonal amplitude matrix that defines the OSVs.","The invariance result suggests gradient domains can be chosen smaller than energy domains without changing the gradient, which would directly reduce the dominant $O(N^3)$–$O(N^4)$ cost of the working equations.","The near-degenerate safety of the Hadamard-product formula is only demonstrated numerically; a targeted test on a molecule with enforced exact eigenvalue degeneracy would show whether a degenerate perturbation treatment is ever needed.","The ethanol metadynamics comparison hints that MP2 and CCSD(T) rotational free-energy surfaces may differ by about 0.5 kcal/mol, within thermal fluctuations at 300 K; repeating the OSV-MP2 run with a smaller thermostat coupling constant would separate thermostat artifact from correlation error."],"forward_implications":["At the normal OSV threshold, OSV-MP2 analytical gradients reproduce canonical RI-MP2 forces to about $10^{-4}$ a.u., and analytic gradients agree with finite-difference gradients to $10^{-6}$–$10^{-7}$ a.u., so the method can be used as a reliable force engine.","Optimized geometries at def2-TZVPP recover RI-MP2 bond lengths within 0.017 pm, bond angles within 0.03 degrees, and dihedrals within 0.2 degrees, with errors decreasing as the OSV threshold is tightened.","Including OSV relaxation lowers gradient errors by roughly an order of magnitude, so omitting it, as some earlier local-MP2 gradient implementations did, is the main accuracy cost at loose thresholds.","The gradients conserve energy well enough for NVE Born–Oppenheimer dynamics: drifts below about 0.2 kJ/mol for protonated water cations with normal OSV selection, and the simulated O–H and O–O radial distribution functions and vibrational densities of states match canonical-MP2 dynamics.","The 200 ps NVT well-tempered metadynamics simulation of ethanol produces a hydroxyl-rotation free energy surface that distinguishes trans from gauche ethanol, providing OSV-MP2 finite-temperature barriers."],"supporting_citations":[{"why":"Defines the OSV ansatz, the diagonal pair-amplitude eigenvectors, and the OSV selection and pair-screening scheme reused throughout the gradient derivation.","marker":"[20]"},{"why":"Supplies the local-MP2 gradient working equations, the coupled-perturbed localization formalism, and the A, B, C intermediates reused in the final working equation.","marker":"[62]"},{"why":"Prior PNO-MP2 gradient computed without PNO relaxation; provides the direct comparison for geometry errors and the baseline that OSV relaxation improves upon.","marker":"[67]"},{"why":"Prior exact DLPNO-MP2 analytic derivatives that avoided PNO relaxation via a block-diagonal density; the paper contrasts its own retained-subspace invariance argument with that approach.","marker":"[69]"},{"why":"Provides DLPNO-MP2 gradient RMSDs and optimized structures used as the main numerical comparison for gradient accuracy and for bond and angle errors.","marker":"[70]"},{"why":"The Hylleraas residual equation whose convergence underpins the invariance proof and the Lagrangian form of the correlation energy.","marker":"[73]"},{"why":"Resolution-of-identity approximation used to evaluate exchange integrals and their derivatives in the OSV basis.","marker":"[78]"},{"why":"Defines the trajectory and RDF/VDOS protocol for the protonated water cation BOMD benchmarks.","marker":"[83]"},{"why":"sGDML CCSD(T) force-field dynamics for ethanol; provides the reference rotational barriers that OSV-MP2 metadynamics is compared against.","marker":"[86]"}],"fun_headline_variants":["Exact OSV-MP2 gradients for dynamics","OSV-MP2 forces: exact relaxation from eigenvalue perturbation","Analytic gradients in OSV-MP2: now exact and efficient","OSV-MP2 gradient theory enables Born-Oppenheimer dynamics","Perturbed OSV relaxation yields exact MP2 forces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction rests on treating the OSV response with ordinary non-degenerate perturbation theory, which requires that no two of an orbital's virtual-space eigenvalues be exactly equal; for near-equal eigenvalues the paper relies on numerical cancellation rather than a proof.","fun_headline_variants_meta":{"raw":{"variants":["Exact OSV-MP2 gradients for dynamics","OSV-MP2 forces: exact relaxation from eigenvalue perturbation","Analytic gradients in OSV-MP2: now exact and efficient","OSV-MP2 gradient theory enables Born-Oppenheimer dynamics","Perturbed OSV relaxation yields exact MP2 forces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1418,"prompt_tokens":1116,"completion_tokens":302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":216}},"tokens_in":732,"tokens_out":302,"duration_ms":3117,"temperature":1.0,"reasoning_tokens":216,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:05:55.477529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare OSV-MP2 analytical gradients with finite-difference gradients along a path that forces two eigenvalues of $T_{kk}$ to cross, for instance a symmetric stretch of a linear or high-symmetry molecule; if the analytical-minus-numerical RMSD jumps or the Hadamard product in Eq. (25) diverges, the non-degenerate relaxation treatment is not generally valid.","supporting_citations":[],"review_version":1}