{"id":"fad97879-bf32-4f72-8c49-a83262ec75a3","arxiv_id":"2608.08948","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A real-time time-dependent Hartree method computes the dynamic screening term of the Bethe-Salpeter equation as a product in time, validated on SiH4.","lead":"The authors show how to include electron screening in molecular excited-state calculations by replacing a slow frequency-by-frequency computation with a single time evolution. The method is tested on silane, where it matches an existing static calculation and shifts the two lowest excitation energies downward, and it could eventually scale to larger molecules.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full-frequency claim rests on the unbenchmarked Shindo factorization (Eq. A15), not on Eq. (17); the paper validates only that Eq. (17) reproduces Eq. (3).","rationale":"The reader identified the Shindo factorization in Eq. (A15) as the weakest assumption, and I agree. Eq. (17) itself is not the weak point: it is a straightforward time-domain recasting of the frequency convolution in Eq. (3), and the paper's internal check of time-product vs direct convolution supports that equivalence. The static real-time vs static dielectric-inversion validation is also a meaningful consistency test for the TDH response machinery. What is not tested is the Shindo factorization that converts the true dynamical BSE into the approximate kernel of Eq. (3). The paper is honest about this, stating in the text that the factorization is adopted because the integration does not close, and in Appendix A that the derivation is within the Shindo approximation. However, no numerical evidence is provided that this approximation is accurate for molecular screening, and the reported dynamical corrections are of the same order as the likely factorization error. This is exactly the gap that makes the present evidence credible but insufficient, so the reader's CONDITIONAL verdict is appropriate; no verdict change is needed from this stress-test. A non-Shindo benchmark for the same test system would settle whether the concern lands.","tokens_in":11398,"tokens_out":9254,"duration_ms":93444,"concrete_test":"Implement a non-Shindo TDA-BSE solver for SiH4 using the same LDA orbitals, basis, and scissor correction: either expand W(omega) as a sum over poles (auxiliary-mode style) and diagonalize the enlarged frequency-dependent BSE, or evaluate the non-factorized two-particle equation (A8) by dense frequency-grid integration. Compare the lowest singlet and triplet roots with the real-time Shindo values. If the dynamical shifts differ by more than about 0.1 eV (the quoted dynamical correction scale), the central full-frequency claim needs qualification; if they agree, the Shindo factorization is validated for this system.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (17) is algebraically equivalent to the frequency convolution in Eq. (3), so the reported time-product vs frequency-convolution agreement validates the numerical transformation, not the underlying physics. The load-bearing physical assumption is the Shindo-type factorization in Eq. (A15), adopted explicitly because with frequency-dependent W the internal fermionic frequency integration does not close the BSE. Eq. (3) is therefore not the exact dynamical BSE kernel; it is a Shindo-approximated kernel. For a molecule like SiH4, whose discrete polarization modes may lie near the excitations, the factorization error is uncontrolled and is not benchmarked. The only external comparison (Ref. [13]) uses a plasmon-pole model and GW quasiparticle states, so it cannot validate the Shindo approximation either. Without a non-Shindo dynamical BSE reference using the same LDA + scissor input, the reported dynamical redshifts (0.12 eV singlet, 0.35 eV triplet) cannot be attributed unambiguously to full-frequency screening. The phrase 'full-frequency dependence' should be read as full-frequency W in a Shindo-approximated kernel.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a real-time linear-response formulation of the dynamical Bethe-Salpeter equation for finite systems. The screened interaction entering the BSE kernel is constructed by propagating time-dependent Hartree equations under impulsive pair-density perturbations, and the frequency convolution that defines the dynamical kernel is evaluated as a product in the time domain followed by Fourier transformation. This avoids explicit frequency grids and storage of the full screened Coulomb operator. The method is validated on SiH4 in two ways: the static BSE from real-time TDH response agrees with a frequency-domain dielectric-matrix inversion to below 1 meV, and the time-product evaluation of the dynamical interaction reproduces direct numerical frequency convolution. The dynamical BSE redshifts the lowest singlet by 0.12 eV and triplet by 0.35 eV, in qualitative agreement with earlier work by Rohlfing and Louie. The paper also sketches a stochastic extension and a memory-based time-evolution formulation.","tokens_in":11517,"tokens_out":4033,"duration_ms":42331,"significance":"If the central claims hold, the method gives a practical route to dynamical BSE calculations that avoids explicit frequency integration and storage of the screened interaction, and its compatibility with stochastic sampling is a genuine algorithmic advantage. The algebraic transformation from the frequency convolution to a time-domain product is exact and internally consistent, and the reproduction of an independent static implementation is a useful validation. However, the physical significance of the reported dynamical shifts is conditional on the Shindo-type factorization adopted in Appendix A, which is not benchmarked against a reference that avoids that approximation. The validation is limited to a single small molecule with no convergence tests, so the quantitative redshifts should be treated as preliminary.","major_comments":[{"comment":"The central physical claim—that the method goes beyond the static screening approximation with full-frequency dependence in the screened interaction—rests on the Shindo-type factorization in Eq. (A15), but this factorization is not validated. The agreement between the time-product and frequency-convolution routes reported in Section III only confirms that Eq. (17) correctly implements Eq. (3); it does not test the approximation in Eq. (3) itself. Since the dynamical shifts in Table I are presented as physical predictions, a comparison against a dynamical BSE solver that avoids the Shindo factorization, using the same LDA+scissor input, is needed to support the attribution of these shifts to full-frequency screening.","section":"II.A and Appendix A, Eq. (A15)"},{"comment":"The comparison with Ref. [13] cannot serve as a quantitative validation because the reference uses GW quasiparticle states and a plasmon-pole model, as the authors acknowledge in the text. The statement that dynamical screening redshifts the excitations is supported, but the magnitudes (0.12 eV singlet, 0.35 eV triplet) are not benchmarked. Please either provide a same-input reference calculation or explicitly label these magnitudes as model-dependent within the Shindo approximation.","section":"III, Table I"},{"comment":"The reported numerical results have no convergence analysis. The choices N_c=24, dt=0.05 au, T_max=20 fs, gamma~100 meV, and eta~50 meV are stated, but their effect on the dynamical shifts is not quantified. Because the shifts are of order 0.1-0.35 eV, the numerical error from Gaussian damping, finite propagation time, and the frequency grid used for secant interpolation could be a substantial fraction of the effect. Please report convergence with respect to T_max (or gamma), eta, basis size, and number of frequency samples.","section":"III, Eqs. (13)-(19)"}],"minor_comments":[{"comment":"The retarded response in Eq. (14) uses Gaussian damping e^{-gamma^2 t^2/2}, while the time-domain convolution in Eq. (17) introduces e^{-eta|t|}. The relationship between gamma and eta, and how both appear in the final time-ordered W(t) entering Eq. (17), should be stated explicitly.","section":"Eqs. (14) and (17)"},{"comment":"Eq. (20) writes a relation for the full W(t), whereas Eq. (17) uses the polarization part W_pol(t), with the bare Coulomb term added separately in Eq. (18). Please clarify whether Eq. (20) includes the instantaneous delta-function term and how the delta function is handled in the stochastic sampling context.","section":"Eq. (20)"},{"comment":"The secant root-finding procedure is described in words but the frequency grid spacing and the number of sampled frequencies are not given. This information is needed to assess the numerical stability of the reported roots.","section":"Section III"},{"comment":"References [52] and [53] are identical (Romaniello, Guyot, and Reining, J. Chem. Phys. 131, 154111 (2009)); the duplicate should be removed or replaced.","section":"References"},{"comment":"The column layout of Table I is confusing: the columns labeled 'Static', 'Time Product', and 'Frequency Convolution' under the present work, followed by 'Static' and 'Dynamic' under Ref. [13], are not clearly separated. The caption should explicitly indicate which columns refer to which implementation.","section":"Table I caption"},{"comment":"The stochastic notation, particularly the distinction between overlined and double-overlined random states, is introduced abruptly. A brief definition of the sampling distribution and the normalization of the random states would improve reproducibility.","section":"Eq. (21)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the algorithmic core is real and likely useful, but the dynamical shifts reported are not yet evidence for the full-frequency BSE; they are shifts of a Shindo-approximated kernel.\n\nThe genuinely new piece is Eq. (17) — evaluating the frequency convolution as a product in time using W(t) from TDH propagation. That replaces a grid over frequency and avoids storing the full screened Coulomb operator. The static BSE from the real-time response matches a standard dielectric-matrix inversion below 1 meV, which is a meaningful sanity check. The time-product vs direct frequency-convolution agreement confirms the transformation is implemented correctly.\n\nThe soft spot is that this transformation is algebraically equivalent to the frequency convolution in Eq. (3), and Eq. (3) is already a Shindo-like factorization. The load-bearing physical assumption — that the four-point correlation function factors as in Eq. (A15) — is not tested anywhere. The only external comparison, Ref. [13], uses a plasmon-pole model and GW states, so it does not benchmark the Shindo error. Calling the result \"full-frequency dependence\" is fair only if you add \"within a Shindo-approximated kernel.\" For SiH4, the discrete polarization modes may sit close to the excitations, so the factorization error is not obviously small.\n\nEverything else is in proportion. Validation is one molecule, no error bars, no convergence study, TDA only, scissor fitted to an external GW gap. No code or data release. These are typical for a methods paper at this stage, and none are disqualifying alone, but together they make the physical claims provisional.\n\nWho is this for? Anyone working on dynamical BSE in molecules. The real-time stochastic extension is plausible, though not demonstrated. It deserves a serious referee, and I'd support peer review with the expectation that the authors add at least one non-Shindo benchmark (e.g., a frequency-domain dynamical BSE without the factorization) and at least one more molecule before publication.","headline":"A real algorithmic step for dynamical BSE, but the reported shifts test a Shindo-approximated kernel, not the full-frequency BSE.","tokens_in":12148,"tokens_out":2245,"would_cite":true,"duration_ms":22253,"reading_group":"maybe","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 the frequency-dependent kernel of the dynamical Bethe-Salpeter equation can be built by real-time propagation and a time-domain product, avoiding frequency-grid integration and storage of the screened Coulomb…","keywords":["dynamical Bethe-Salpeter equation","real-time linear response","time-dependent Hartree","screened Coulomb interaction","frequency-dependent kernel","Tamm-Dancoff approximation","stochastic sampling","silane"],"falsifier":"Compute the dynamical excitation energies of a small molecule whose low-lying polarization mode lies near the optical excitation energy, using an unfactorized frequency-domain solution of the Bethe-Salpeter equation that retains full frequency dependence in W, and compare with the real-time time-product result. Agreement within the roughly 50 meV numerical broadening would support the Shindo factorization; a larger mismatch would show that Eq. (17) is not the full-frequency BSE kernel.","tokens_in":1735,"feed_emoji":"⚛️","tokens_out":2535,"duration_ms":89666,"temperature":0.7,"pith_summary":"The paper tries to establish that the frequency-dependent screened interaction entering the dynamical Bethe-Salpeter equation can be constructed by propagating the density matrix in real time and then evaluating the frequency convolution as a product in time. If correct, the dynamical BSE can be solved without constructing or storing the full frequency-dependent screened Coulomb operator, removing a major memory bottleneck and opening a path to stochastic sampling for large systems. The authors validate the formalism on silane: the static limit reproduces a standard dielectric-matrix-inversion implementation to below 1 meV, and the time-product evaluation agrees with the direct frequency convolution. The resulting dynamical correction redshifts the lowest singlet and triplet excitations, consistent with earlier dynamical-BSE studies.","feed_headline":"Real-time route to dynamical Bethe-Salpeter kernel","feed_subtitle":"Propagate density matrix, multiply in time, transform: no frequency grid or screened-Coulomb storage; silane test.","key_machinery":"The load-bearing object is Eq. (17), the real-time expression for the polarization contribution to the dynamical kernel: it converts the convolution over frequency in Eq. (3) into multiplication in time by the phase factors $e^{-i(\\varepsilon_b-\\varepsilon_i+\\Delta)t}$ and $e^{i(\\varepsilon_a-\\varepsilon_j+\\Delta)t}$, followed by a Fourier transform. The retarded response $W^{\\mathrm{pol},ij}_{ab}(t)$ is generated by time-dependent Hartree propagation: an impulsive perturbation from each occupied pair $kl$ produces an induced density matrix, which is contracted with the Coulomb tensor to give the polarization part of the screened interaction. A causal-to-time-ordered transformation puts this response on the full time axis. The kernel itself rests on the Shindo-type factorization of the four-point correlation function, Eq. (A15), which allows the two internal fermionic frequency integrals to be evaluated by contour integration so that the BSE closes.","core_discovery":"On the paper's own terms, the central discovery is that Eq. (17) is an equivalent time-domain rendering of the frequency convolution defining the dynamical BSE kernel. The retarded screened interaction is obtained by propagating the time-dependent Hartree density matrix after impulsive perturbations built from occupied-orbital pair densities; the causal response is converted to a time-ordered interaction, multiplied by phase factors carrying the bare electron-hole transition energies, and Fourier transformed to give the full-frequency-dependent kernel without explicit frequency integration or storage of the full real-space screened Coulomb operator. The construction is validated on silane in two ways: static BSE from real-time response matches a standard dielectric-matrix-inversion implementation with mean absolute deviations below 1 meV, and the time-product evaluation of the dynamical interaction reproduces the direct frequency convolution. Dynamical screening shifts the lowest singlet excitation down by 0.12 eV and the triplet by 0.35 eV, a redshift in the same direction as earlier dynamical-BSE results.","pith_inferences":["A natural test of the Shindo factorization would be to compare the real-time kernel with an unfactorized dynamical BSE calculation on a small molecule whose low-lying polarization mode overlaps the optical excitation; the paper does not provide such a benchmark.","The real-time construction may be most beneficial where memory rather than floating-point operations is the bottleneck, since no screened-Coulomb operator is stored on a frequency grid; profiling memory against system size would identify the crossover.","If the stochastic extension succeeds, the dynamical BSE could become available precisely for the molecules where the static approximation is most questionable, such as weakly screened systems and triplet excitations, because the phase factors factorize onto occupied and unoccupied subspaces.","The Appendix B memory-based evolution suggests that the same time-dependent interaction could bypass the nonlinear root search entirely by propagating a density matrix; whether that evolution is numerically stable is left open."],"forward_implications":["For systems where polarization modes lie near optical excitation energies, the static approximation can misestimate singlet and triplet energies by roughly 0.1-0.3 eV; the real-time dynamical kernel corrects this at the cost of propagating TDH equations rather than storing a frequency grid.","The nonlinear BSE can be solved by diagonalizing the dynamical matrix on a sparse grid of trial frequencies and locating roots by secant interpolation; for silane the matrix varies smoothly, so few sampling frequencies suffice.","The same real-time response supplies both the static and the dynamical kernel, so a single code path covers both approximations and can be checked against independent static implementations.","The time-domain form is compatible with stochastic sampling, in which random occupied and unoccupied orbital combinations carry the orbital-energy phase factors, offering a route to dynamical BSE for systems with hundreds to thousands of valence electrons.","The time-ordered interaction can also drive a memory-based density-matrix evolution, outlined in Appendix B, potentially yielding spectra or steady states without a frequency-domain diagonalization."],"supporting_citations":[{"why":"Supplies the GW quasiparticle gap and the static/dynamical BSE reference values used for the scissor correction and for comparing the direction and size of the dynamical redshift.","marker":"[13]"},{"why":"Provides the frequency formalism of the dynamical BSE and the Shindo-like frequency-dependent kernel from which Eq. (3) is taken.","marker":"[4]"},{"why":"Gives the Schwinger functional-derivative derivation of the BSE and the Shindo approximation underlying Eq. (A15).","marker":"[11]"},{"why":"Supplies the secant interpolation procedure used to locate roots of the nonlinear BSE eigenvalue problem.","marker":"[25]"},{"why":"Establishes the real-time stochastic GW prescription for converting retarded responses to time-ordered interactions, which Eq. (15) follows.","marker":"[38]"},{"why":"Documents numerical divergences from resonant-antiresonant coupling of the dynamical kernel, motivating the Tamm-Dancoff approximation used here.","marker":"[24]"},{"why":"Provides the LDA density functional used for the plane-wave DFT ground state on which the BSE is built.","marker":"[48]"},{"why":"Supplies the Martyna-Tuckerman approach used to reduce periodic-image effects in the finite-grid DFT calculation.","marker":"[51]"}],"fun_headline_variants":["Real-time BSE without frequency grids or screened-Coulomb storage","Dynamical Bethe-Salpeter kernel via time-domain multiplication","Time-propagation trick for full-frequency BSE kernels","Silane test: real-time BSE matches static, captures dynamical shifts","Full-frequency BSE from real-time Hartree propagation"],"cache_read_input_tokens":14208,"weakest_assumption_plain":"The load-bearing premise is the Shindo-type factorization of the four-point correlation function in Eq. (A15): if that factorization is not quantitatively accurate for molecular screening, the computed dynamical shifts are not the true full-frequency BSE shifts, and the paper gives no benchmark against a dynamical BSE solver that avoids it.","fun_headline_variants_meta":{"raw":{"variants":["Real-time BSE without frequency grids or screened-Coulomb storage","Dynamical Bethe-Salpeter kernel via time-domain multiplication","Time-propagation trick for full-frequency BSE kernels","Silane test: real-time BSE matches static, captures dynamical shifts","Full-frequency BSE from real-time Hartree propagation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1593,"prompt_tokens":863,"completion_tokens":730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":645}},"tokens_in":479,"tokens_out":730,"duration_ms":6960,"temperature":1.0,"reasoning_tokens":645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:20:26.962213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the dynamical excitation energies of a small molecule whose low-lying polarization mode lies near the optical excitation energy, using an unfactorized frequency-domain solution of the Bethe-Salpeter equation that retains full frequency dependence in W, and compare with the real-time time-product result. Agreement within the roughly 50 meV numerical broadening would support the Shindo factorization; a larger mismatch would show that Eq. (17) is not the full-frequency BSE kernel.","supporting_citations":[{"cited_title":"Romaniello, D","cited_arxiv_id":null,"evidence_quote":"Provides the frequency formalism of the dynamical BSE and the Shindo-like frequency-dependent kernel from which Eq. (3) is taken."},{"cited_title":"Zhang, J","cited_arxiv_id":null,"evidence_quote":"Supplies the secant interpolation procedure used to locate roots of the nonlinear BSE eigenvalue problem."},{"cited_title":"Neuhauser, Y","cited_arxiv_id":null,"evidence_quote":"Establishes the real-time stochastic GW prescription for converting retarded responses to time-ordered interactions, which Eq. (15) follows."},{"cited_title":"Loos and X","cited_arxiv_id":null,"evidence_quote":"Documents numerical divergences from resonant-antiresonant coupling of the dynamical kernel, motivating the Tamm-Dancoff approximation used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Martyna-Tuckerman approach used to reduce periodic-image effects in the finite-grid DFT calculation."}],"review_version":1}