{"id":"a704fb5d-d3eb-4fbb-b685-a5d6536c943f","arxiv_id":"2607.29504","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A causality pre-filter that excludes retarded Green-function interactions whose delay cannot overlap the temporal basis support reduces MOT-JVIE assembly cost by 40.6% and speeds assembly up to 109x on 64 cores.","lead":"By checking, before any computation, which source–receiver pairs in a time-domain electromagnetic simulation can arrive at the right time delay, this paper skips about 41% of the interaction work and reports up to a 109x faster assembly on 64 CPU cores. The method changes no physics and claims no approximation, making it attractive for large transient scattering simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exactness claim of Eqs. (5)-(7) rests on an unstated single-delay-per-voxel-pair model; finite voxels produce a delay interval per pair, so center-based pruning can omit interactions whose quadrature points still satisfy causality.","rationale":"The reader's weakest assumption is precisely the single-delay-per-voxel-pair model. My analysis agrees: Eq. (5) uses τ_mm' = R_mm'/c0, a single separation distance, while the actual retarded interaction over finite voxels involves a continuum of distances. The reported voxel size and time step make the delay spread across a pair non-negligible relative to Δt, so the 'identically inactive' statement is not justified by the text. This is load-bearing because the paper's strongest claim is that pruning is exact and introduces no approximation; if the interval test reveals discrepancies, the method is approximate and the central contribution must be reframed. The concrete test would settle this by comparing center-based and interval-based admissibility and checking whether omitted contributions are numerically nonzero. I would keep the reader's CONDITIONAL verdict: the concern is real but addressable, and the numerical agreement to ~1e−15 suggests the issue may be benign in the specific discretization, though it must be stated and tested explicitly.","tokens_in":7138,"tokens_out":7896,"duration_ms":84355,"concrete_test":"Recompute the large-slab admissibility using the exact interval test: (m,m′,k) is active iff [min_{r∈V_m,r′∈V_m′}|r−r′|/c0, max(...)/c0] ∩ [kΔt−tb, kΔt−ta] ≠ ∅. Compare this set with S_act from Eq. (7). For every triple that is active under the interval test but absent from S_act, evaluate the corresponding retarded Green-function contribution (or a Monte-Carlo estimate of its quadrature integral). If the total L2 contribution of the omitted triples changes the scattered field by more than 1e−12 relative to the conventional assembly, the 'identically inactive' assertion is falsified; if the sets coincide or the omitted contributions are at machine precision, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Section III-A is that Eq. (5)-(7) defines an exact pre-assembly filter: interactions with kΔt − τ_mm' outside [ta,tb] are 'identically inactive' and pruning them 'introduces no approximation.' The inference depends on τ_mm' being the only retarded delay present between observation voxel m and source voxel m'. In a volumetric MOT-JVIE, the retarded interaction at delay index k is an integral over r ∈ V_m and r′ ∈ V_m′ of quantities containing T(kΔt − |r−r′|/c0), so the pair actually contributes whenever the delay interval [R_min/c0, R_max/c0] intersects the shifted temporal-basis support [kΔt − tb, kΔt − ta], not merely when the center-to-center delay does. Because the voxel side is 10 nm and c0Δt ≈ 6 nm in both reported configurations, the delay spread inside a voxel pair is of order several Δt; center-based and interval-based admissibility differ materially. Equation (5) can therefore classify as inactive an interaction for which the temporal-basis support is non-empty over a subvolume of the source/observation voxels. The 'identically inactive' phrase is only valid if the discretization is point-collocation with one distance per pair, a modeling restriction the manuscript does not state. This is the weakest point in the strongest claim because the 40.6% reduction, the 2.6× algorithmic speedup, and the 109× total speedup all inherit the exactness assertion; if the filter is approximate, the savings are still real but the headline 'no approximation' is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a causality-aware pre-assembly strategy for the retarded Green-function interactions in a marching-on-in-time volume integral-equation (MOT-JVIE) solver. For each observation-source voxel pair, the admissible MOT delay indices are restricted to those for which the center-to-center retarded delay falls within the support of the temporal basis function; inadmissible interactions are discarded before numerical evaluation. The remaining interactions are grouped into homogeneous vectorized batches and assembled in parallel using private buffers. The paper reports a 40.6% reduction in evaluated interactions, a 2.6x single-worker algorithmic speedup, up to 109x total assembly speedup with 64 workers, and a transient-field error of about 1e-15 relative to the conventional assembly on a small dielectric slab, with independent FDTD agreement.","tokens_in":7520,"tokens_out":9015,"duration_ms":93365,"significance":"The proposed causality filter is an attractive and low-risk optimization: it is derived from the temporal-basis support and the retarded delay, so the admissibility criterion has no fitted parameters; the only tunable parameter is the batch size B_b, which is a parallel-scheduling choice and is studied in Table IV. The arithmetic in Table III is internally consistent, the 40.6% figure is correctly computed, and the validation against an independent FDTD solver (Meep) is a genuine strength. If the modeling assumptions behind the exactness claim are made explicit and the speedup attribution is corrected, the work would be a useful engineering contribution to TDIE assembly. At present, however, the central exactness claim and the performance attribution are not fully supported.","major_comments":[{"comment":"The statement that non-admissible interactions are 'identically inactive' and pruning 'introduces no approximation' is valid only under a point-collocation (single-distance-per-voxel-pair) discretization. In the volumetric MOT-JVIE of Eq. (1), the interaction for pair (m,m') is an integral over r in V_m, r' in V_m' of quantities containing T(k*dt - |r-r'|/c0); the delays form an interval [R_min/c0, R_max/c0], not the center value tau_mm'. With voxel edge 10 nm and c0*dt ~ 6 nm (Table II), the intra-pair delay spread is several dt. An interaction whose center delay is outside [k*dt - tb, k*dt - ta] can still receive nonzero contributions from subvolumes whose delays intersect the temporal support. Thus the exactness claim and the 40.6% reduction require either justifying a single-delay model or replacing Eq. (5) by an interval-intersection test. The small-slab epsilon_E ~ 1e-15 verifies o","section":"Section III-A, Eqs. (5)-(7)"},{"comment":"The reported '2.6 times algorithmic speedup from causal-support screening' is not isolated from the vectorization/batching of Section III-B. Removing 40.6% of the interactions would yield at most 1/(1-0.406) ~ 1.68x if assembly time were proportional to interaction count; the observed 1.08e5/4.1e4 ~ 2.6x therefore includes the branch-free vectorized evaluation. Please provide a P=1 ablation with the same vectorized assembly code with and without the causality filter, and report the speedup attributable to pruning alone. Without this, the algorithmic speedup claim is overstated.","section":"Section IV-B, Figs. 3-4"},{"comment":"The 'no approximation' claim is checked only for the small slab; no transient-field accuracy comparison is reported for the large slab, which is the case used for the interaction-reduction and speedup numbers. Since exactness of the pruning is the paper's central assertion, an accuracy metric analogous to Eq. (10) for the large slab, or a clear statement that the point-collocation model makes the check unnecessary, is needed.","section":"Section IV-A"}],"minor_comments":[{"comment":"The summation notation 'n - l sum n'=n-1' is unclear; use, for example, sum_{n'=n-l}^{n-1}.","section":"Section II, Eq. (3)"},{"comment":"Define R_mm' explicitly as the center-to-center distance if that is the intended model. Also, 'a observation-source interaction' should read 'an observation-source interaction'.","section":"Section III-A"},{"comment":"The sentence 'the skipped-interaction ratio increasing from 0% at l=1 to 41.3% at l=7' is followed by 33.5% at l=8; the explanation for the decrease is plausible, but the wording 'pruning becomes more effective at larger delay levels' should be qualified.","section":"Table III(b)"},{"comment":"Clarify whether the conventional assembly is also parallelized, and explicitly label the 'total speedup' as relative to conventional single-worker assembly.","section":"Figures 3-4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering contribution: a pre-assembly causality screen plus vectorized batching cuts evaluated interactions by 41% and gives up to 109x assembly speedup, with transient fields matching the conventional solver to ~1e-15. The interaction counts are internally consistent, the MEEP comparison is a nice independent check, and the batching study is honest about the trade-off. The genuinely new piece is the screening-and-batching pipeline; inequality (6) is a straightforward formalization of a point already made in [7], [8], [10], that each source–observation pair occupies only a subset of the MOT history.\n\nThe soft spots are real but localized. The strongest claim is that pruning is exact: interactions outside [ta, tb] are \"identically inactive\" and discardal \"introduces no approximation.\" That rests on treating each voxel pair as having one retarded delay. In a volumetric Galerkin MOT-JVIE, the interaction integral sees a range of delays because the source and observation basis functions extend over finite voxels, and the correct admissibility test is whether the delay interval [R_min/c0, R_max/c0] intersects the temporal-basis support. The paper does not state that its solver uses point collocation or otherwise restricts to one distance per pair. With 10 nm voxels and c0Δt = 6 nm, the delay spread across a voxel pair is several Δt, so center-based pruning can, in principle, discard nonzero interactions. The 1e-15 agreement on the two test cases suggests this did not happen here, but it does not establish exactness in general. This is the weakest point in the strongest claim.\n\nA separate, less serious issue: the per-level explanation in Section IV-B contradicts Table III(b). The text says \"more candidate source–observer pairs lie inside the admissible causal support\" for larger ℓ, but the skip fraction increases with ℓ. The trend is plausible, but the explanation is backwards. Minor issues: no code or data released, runtimes lack error bars, and the batch size is tuned on the test case itself.\n\nWho this is for: people building MOT-JVIE or other TDIE solvers who care about assembly time. The engineering payoff is probably real even if the exactness claim is softened to \"negligible approximation\" or clarified to the point-collocation case. Deserves a serious referee, because the method is practical and the validation is well designed, but the exactness claim should be fixed before the paper is trusted as a general-purpose exact strategy. I would not cite it in my own work until that is resolved.","headline":"Useful engineering result, but the 'exact' pruning claim overstates what the paper actually shows.","tokens_in":8019,"tokens_out":3778,"would_cite":false,"duration_ms":38652,"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 retarded causality is an exact pre-assembly filter, pruning 40.6% of Green-function interactions with no approximation.","keywords":["causality","retarded Green function","marching-on-in-time","time-domain integral equations","volume integral equation","interaction pruning","parallel assembly","transient scattering"],"falsifier":"For a source-observation pair with center-to-center delay just outside the admissible interval, evaluate the full retarded Green-function interaction integral using the actual spatial basis functions and multiple quadrature points per voxel; if the resulting entry is not exactly zero within machine precision, the 'identically inactive' premise is violated. Alternatively, rerun the small-slab validation with voxel size increased to a substantial fraction of the wavelength and check whether the relative error ε_E remains ~1e-15.","tokens_in":7026,"feed_emoji":"🕒","tokens_out":6911,"duration_ms":62759,"temperature":0.7,"pith_summary":"In transient electromagnetic scattering simulations, the most expensive step is often assembling the interaction matrix: for every pair of small volume cells and every time-delay index, the solver must evaluate a retarded Green function. This paper proposes to use causality itself as a filter: for a given observation cell and source cell, only those delay indices whose retarded delay falls inside the temporal basis support can possibly contribute; every other candidate interaction is claimed to be identically zero. The paper shows that discarding those candidates before numerical evaluation does not change the simulated transient field (agreement to about 1e-15 relative error), while cutting the number of evaluated interactions by 40.6% and accelerating assembly by 2.6x on one CPU core and up to 109x with 64 cores. If this is right, a universal physical principle becomes a cheap, exact algorithmic optimization for a whole class of time-domain integral-equation solvers.","feed_headline":"Causality pruning cuts Green-function assembly by 41%","feed_subtitle":"Skipping causally inactive retarded interactions gives a 2.6x single-core speedup and up to 109x total.","key_machinery":"The operative mechanism is the causality-admissibility test in Eqs. (5)–(7): for each observation voxel m and source voxel m', the retarded delay τ_mm' = R_mm'/c0 maps the pair into a discrete delay shell; an interaction with MOT delay index k is retained only when kΔt − τ_mm' ∈ [ta,tb], the support of the temporal basis function. This converts causality into a pair-level integer set Q_mm' that depends only on separation distance and basis support, allowing all inactive triples to be pruned before quadrature. The remaining triples are grouped into homogeneous batches and evaluated with branch-free SIMD vectorization, with each worker accumulating into private delay-indexed buffers that are s","core_discovery":"The central claim is that the set of admissible interactions S_act = {(m,m',k): kΔt − τ_mm' ∈ [ta,tb]} is the exact set of non-vanishing retarded interactions in the discretized MOT-JVIE. Each observation-source voxel pair is assigned the single retarded delay τ_mm' = R_mm'/c0, and an interaction at delay index k contributes if and only if the delay offset overlaps the support of the temporal basis function T(t). All other candidate interactions are 'identically inactive' and can be discarded before evaluation without modifying the TDIE formulation or introducing approximations. Numerical validation on two dielectric slabs shows the pruned assembly reproduces the conventional transient field","pith_inferences":["The exactness of the pruning hinges on the one-delay-per-pair model; if voxel sizes are large relative to the wavelength or spatial basis functions extend over the voxel, quadrature points sample a range of delays, and a pair whose center-to-center delay falls outside [ta,tb] could still contribute through parts of its volume. This is not tested in the paper, which uses 10 nm voxels at 378 nm wave","A natural test of the single-delay assumption is to compute full quadrature interactions for pairs near the admissibility boundary and compare their magnitude to machine precision; if any are non-negligible, the no-approximation claim would need qualification for non-uniform or coarse grids.","The same delay-shell structure could be used to schedule data movement in distributed-memory TDIE solvers, because the admissible sets define which processes need which source voxels at which time steps.","For wider temporal basis functions, the admissible window [ta,tb] grows and the skip fraction will shrink; the method's benefit is largest when Δt is small relative to the basis support, which is typical for stable MOT schemes."],"forward_implications":["Because the admissibility test depends only on the retarded delay and temporal-basis support, the same screening applies to other marching-on-in-time TDIE formulations that use compactly supported temporal basis functions.","Interaction reduction grows with delay level: the skip ratio climbs from 0% at level 1 to 41.3% at level 7 in the large-slab test, so longer propagation histories benefit more from the pruning.","Pruning before quadrature removes the light-cone conditional checks from inner loops, which is what enables the branch-free vectorized batches and the additional parallel speedup.","The observed 109x total speedup combines the 2.6x algorithmic reduction with the 42x parallel speedup from the vectorized batching.","Preservation of the transient response to ~1e-15 means the pruned and conventional matrices are numerically indistinguishable in the tested configuration, so the filter can be treated as exact within the discrete model."],"fun_headline_variants":["Exact causality pruning cuts 41% of retarded interactions","Causality skips 41% of interactions, speeds assembly 109x","No approximation: causality prunes 41% interactions","Prune by causality: 41% less work, 109x speedup","Causality-aware assembly: 41% fewer interactions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Every observation-source voxel pair is assigned a single retarded delay based on center-to-center distance; if voxel volumes actually produce a range of delays, an interaction whose center delay falls just outside the temporal-basis window can still have non-negligible contributions, and the claim that excluded interactions are identically inactive would no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Exact causality pruning cuts 41% of retarded interactions","Causality skips 41% of interactions, speeds assembly 109x","No approximation: causality prunes 41% interactions","Prune by causality: 41% less work, 109x speedup","Causality-aware assembly: 41% fewer interactions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001186,"raw_usage":{"total_tokens":4702,"prompt_tokens":679,"completion_tokens":4023,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":3932}},"tokens_in":423,"tokens_out":4023,"duration_ms":24814,"temperature":1.0,"reasoning_tokens":3932,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:37:52.137582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a source-observation pair with center-to-center delay just outside the admissible interval, evaluate the full retarded Green-function interaction integral using the actual spatial basis functions and multiple quadrature points per voxel; if the resulting entry is not exactly zero within machine precision, the 'identically inactive' premise is violated. Alternatively, rerun the small-slab validation with voxel size increased to a substantial fraction of the wavelength and check whether the relative error ε_E remains ~1e-15.","supporting_citations":[],"review_version":1}