{"id":"e7fa4324-6c9c-4319-97df-ae0d676682f1","arxiv_id":"2501.06154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A batched randomized-Halton variant of the iterative quasi-Monte Carlo method removes fixed-seed ray-effect bias in k-eigenvalue neutron transport while preserving O(1/N) convergence.","lead":"This paper applies randomized quasi-Monte Carlo sampling to a hybrid neutron transport solver, replacing fixed particle trajectories with fresh randomized sets each iteration. The batch scheme reduces the required particle histories by about two orders of magnitude while keeping the fast O(1/N) convergence of quasi-Monte Carlo.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(N^{-1}) convergence claim is asserted from one benchmark without a fitted slope and is partly masked by the 25^3 mesh plateau; the rate is not quantitatively established.","rationale":"The paper is a plausible and clearly written methods contribution. Owen's randomization of the Halton sequence is implemented in MC/DC, and the qualitative improvement over fixed-seed iQMC is demonstrated. The algorithm is described in sufficient detail for reproduction. However, the strongest claim—that the batch RQMC approach maintains the theoretical O(N^{-1}) QMC convergence rate—is not quantitatively established. The reader's weakest_assumption identifies exactly this gap: the measured error is not shown to be dominated by RQMC sampling error, the spatial mesh plateau limits the observable asymptotic range, and only one benchmark is used. I agree with that assessment. My independent reading found no deeper internal inconsistency or obvious algorithmic flaw: the iterative power iteration with randomized batches is coherent, and the sources of error are acknowledged in the text. The missing piece is a rigorous convergence-rate estimate, a mesh-convergence study, and at least one additional benchmark with batch sizes selected a priori. Those additions would turn the conditional acceptance into a solid verification of the central claim. Since the reader's verdict already reflects this conditionality, no change in verdict is needed.","tokens_in":7097,"tokens_out":4064,"duration_ms":45746,"concrete_test":"Use the Fig. 5 data to fit log10(error) = a + b*log10(N) over only the pre-plateau N points, reporting b and its 95% confidence interval for both k-effective and mean scalar flux. Then re-run Takeda-1 on 50^3 and 75^3 meshes with the same batch scheme, confirm the plateau shifts to higher N, and check that the fitted b remains consistent with -1 over an extended range. Finally, run a second 3D k-eigenvalue benchmark (e.g., C5G7 or Kobayashi) to show the rate is not problem-specific. If b differs significantly from -1 or the plateau cannot be separated from the sampling trend, the O(N^{-1}) claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the batch RQMC iQMC method converges both k-effective and mean scalar flux error at O(N^{-1}). The supporting evidence is Figure 5, from a single benchmark (Takeda-1) on a fixed 25x25x25 mesh. The paper reports no fitted slope, no confidence interval, and no quantitative comparison to the theoretical rate; 'O(N^{-1})' is inferred visually from a limited N-range. That range is further restricted because the scalar-flux error plateaus at large N, which Section 3 explicitly attributes to the coarse spatial mesh. Because the analog Monte Carlo reference is also tallied on the same 25^3 mesh, the comparison does not cancel iQMC's piecewise-linear source discretization error; the measured error mixes RQMC quadrature error with mesh-induced error. The k-effective curve is cleaner, but it is still one problem and one figure. Additionally, the inactive/active batch counts (200/100) are selected after inspecting convergence for this same problem, so the experiment is not an a priori test. These are methodological gaps rather than contradictions: the method may well achieve O(N^{-1}), but the paper as written does not provide the quantitative evidence needed to support that claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a batched power-iteration variant of the iterative Quasi-Monte Carlo (iQMC) method for multigroup neutron transport. Instead of resetting particles to the same low-discrepancy sequence every transport sweep (the fixed-seed approach), the new method uses Owen's randomization of the Halton sequence to generate fresh RQMC samples per sweep, with user-specified inactive and active batches to manage the stochastic noise. The method is implemented in MC/DC and tested on the Takeda-1 k-eigenvalue benchmark. The reported results show that the batched RQMC approach gives more stable and more accurate k-effective and scalar-flux estimates than fixed-seed iQMC, and the authors claim that the method converges at the theoretical QMC rate of O(N^-1) while needing roughly two orders of magnitude fewer particles per batch. However, the O(N^-1) claim rests on a single benchmark and is inferred visually without a fitted slope, and the paper itself notes that the scalar-flux error plateaus at high particle counts due to the 25x25x25 mesh.","tokens_in":7270,"tokens_out":3918,"duration_ms":38586,"significance":"The work addresses a genuine and well-known limitation of the iQMC method: the fixed-seed approach leaves parts of phase space permanently undersampled, producing ray-effect-like bias. Replacing fixed seeds with randomized low-discrepancy samples in a batch power iteration is a natural and promising remedy, and the authors implement it in MC/DC in a way that avoids costly exits from compiled code. The qualitative improvement over fixed-seed iQMC is clearly demonstrated in Figures 3 and 4, and the paper is generally clearly written. If the O(N^-1) convergence claim is confirmed with rigorous quantitative evidence, the contribution would be significant for the iQMC research community. As it stands, the central rate claim is not yet supported to the standard required by a journal publication, because the evidence is limited to one benchmark, the rate is not measured, and the observable convergence range is partly masked by the mesh-induced plateau.","major_comments":[{"comment":"The central claim that the RQMC batch method converges both k-effective and mean scalar flux error at the theoretical QMC rate O(N^-1) is not quantitatively established. The paper reports no fitted slope, no confidence interval, and no comparison with the expected rate; the rate is inferred visually from a limited range of N. That range is further restricted by the scalar-flux plateau at high particle counts, which the text attributes to the coarse 25x25x25 mesh. Because the analog Monte Carlo reference is tallied on the same mesh, the comparison cannot cancel iQMC's piecewise-linear source discretization error, so the measured error mixes RQMC sampling error with spatial discretization error. Please fit the log-log slope in the pre-plateau region, report standard errors, and ideally repeat on at least one additional benchmark or refine the mesh to confirm that the plateau shifts and the asymptotic rate is reached.","section":"Section 3, Figure 5"},{"comment":"The inactive and active batch counts (N_inactive = 200, N_active = 100) are user inputs selected after inspecting convergence on the same Takeda-1 problem, as the text acknowledges when it says these values 'can be approximated from studying the convergence of the solution in each problem.' This makes the experiment not an a priori test of the method, and the batch counts could be tuned to this particular benchmark. To support the general claim of iterative stability, either implement an adaptive stopping criterion based on the relative change of the source strength per batch, or provide a sensitivity study showing that the reported accuracy and convergence rate are robust to reasonable variations in N_inactive and N_active.","section":"Section 2, Algorithm 1 and Section 3"},{"comment":"The paper claims 'nearly two orders of magnitude reduction in the number of particle histories per transport sweep' for the batch method relative to fixed-seed iQMC, but the quantitative support is unclear. The direct comparison in the text is between a fixed-seed run with N = 1.5E6 particles per sweep (Figure 3) and a batch run with N = 1E5 particles per sweep (Figure 4), which is a factor of 15, not two orders of magnitude. If the claim instead refers to total particle histories including inactive and active batches, that comparison should be made explicit with the relevant totals; otherwise the two-orders-of-magnitude statement should be revised.","section":"Section 3, Figures 3-5"}],"minor_comments":[{"comment":"Line 14 of Algorithm 1 contains a typo: 'N_inactve' should be 'N_inactive'.","section":"Algorithm 1"},{"comment":"The caption states that solutions are 'shown with one standard deviation of the ensemble,' but the definition of the plotted 'mean scalar flux error' is never given. Please specify the normalization and the set of cells over which the error is averaged (e.g., relative L2 error over all mesh cells).","section":"Section 3, Figure 5"},{"comment":"The text spells the library name as 'Scipy' in Section 2; the reference list correctly uses 'SciPy'. Please use a consistent spelling.","section":"Section 2"},{"comment":"The caption uses 'randomized-halton' while the abstract and elsewhere use 'Randomized-Halton' or 'randomized-Halton'; please standardize the capitalization.","section":"Section 3, Figure 4"},{"comment":"The paper would benefit from a brief discussion of the trade-off that the batched RQMC approach destroys the linearity assumption used for Krylov solvers in prior iQMC work; this is mentioned in Section 1.2 but never revisited in the conclusions, where the authors state the method 'negate[s] the use of linear Krylov solvers' without commenting on the practical cost of that loss.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computational physics or nuclear engineering methods journal. The concern is evidentiary rather than novelty-based: the method clearly improves on fixed-seed iQMC, but the headline O(N^-1) claim needs quantitative support and a more careful treatment of the mesh-induced plateau. I saw no issues with citation practice or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The batched RQMC variant of iQMC is a genuine, small step forward: it replaces the fixed-seed sweep with Owen-scrambled Halton sequences per batch, removes the ray-effect bias the authors were chasing, and appears to stabilize k-effective with roughly two orders of magnitude fewer particles per batch. The algorithmic description is clear, the implementation in MC/DC is concrete, and the authors are honest that N_inactive and N_active are user inputs selected after inspecting the same problem, and that the scalar flux error plateaus on the 25^3 mesh.\n\nWhat is new: this is the first application of RQMC batching to iQMC, as far as I can tell from the cited literature. The observation that the QMC O(N^{-1}) rate survives the added stochastic noise is useful if true. What is not established: the rate claim itself. Figure 5 is one benchmark, no fitted slope, no confidence band around the slope, and the O(N^{-1}) is a visual assertion. The plateau at high N, which the paper attributes to the coarse mesh, cuts off exactly the range where you would want to confirm the asymptotic rate. The k-effective curve is cleaner, but still one problem. The batch counts are not a priori; they are tuned on the same case, so the experiment is not a prediction. No code or data are included, so independent verification means reimplementation.\n\nNone of this is disqualifying. A plausible method with honest caveats and a clear numerical demonstration on a standard benchmark deserves referee time. I would send it for peer review and ask for a second benchmark, a fitted slope or at least a comparison to the theoretical rate on a range unaffected by the mesh, and a sensitivity study for N_inactive/N_active. If the code were released, that would help a lot.\n\nBottom line: useful for the neutron transport / hybrid MC community, not a major advance. I would not cite it in my own work unless I were actively working on iQMC, but I would recommend it to a student working on QMC transport. Yes to peer review.","headline":"A plausible and clearly described batching variant of iQMC, but the O(N^{-1}) rate claim is under-evidenced: one benchmark, no fitted slope, hand-set batch counts, and a mesh plateau that truncates the range.","tokens_in":7848,"tokens_out":3446,"would_cite":false,"duration_ms":30838,"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":"By replacing fixed-seed sampling with randomized Halton batches, the iQMC method removes ray-effect bias and preserves the O(N^{-1}) quasi-Monte Carlo convergence rate while cutting the required particles per batch by nearly two orders of…","keywords":["Quasi-Monte Carlo","Neutron Transport","k-Eigenvalue","Iterative Method","Hybrid Method","Randomized Halton sequence","Batch power iteration","Owen's randomization"],"falsifier":"Run the batched RQMC iQMC method on the Takeda-1 problem with a substantially finer mesh (or adaptive refinement near flux gradients) and check whether the k-effective and mean scalar flux errors continue to follow O($N^{{-1}}$) across a wider range of particle counts than shown in Figure 5. If the error curve flattens at the same particle count regardless of mesh, the claimed rate is an artifact of the mesh; if the curve continues, the claim is supported.","tokens_in":75,"feed_emoji":"⚛️","tokens_out":8198,"duration_ms":130050,"temperature":0.7,"pith_summary":"This paper claims that switching the Iterative Quasi-Monte Carlo (iQMC) neutron transport method from a fixed-seed to a batch approach with randomized Halton sequences preserves the theoretical QMC convergence rate of O($N^{{-1}}$) for k-effective and mean scalar flux error, while reducing the number of particle histories per batch by nearly two orders of magnitude. The authors argue that the fixed-seed approach, which resets particles to the same positions and directions each sweep, undersamples phase space and introduces ray-effect-like bias. They implement a batched power iteration that randomizes the Halton sequence with Owen's randomization each sweep, runs inactive batches to converge the source, then active batches to record statistics. In the Takeda-1 benchmark, the batch approach produces stable, accurate solutions with far fewer particles per sweep than fixed-seed, and averaged over 20 simulations the error converges at the QMC rate.","feed_headline":"Batched Halton sampling keeps neutron solver at 1/N rate","feed_subtitle":"The batch scheme reaches QMC accuracy with about 100x fewer particles per sweep.","key_machinery":"The central object is the randomized Halton sequence produced by Owen's randomization, which constructs each sample point as individually uniform on [0,1]^D while collectively preserving low discrepancy, and can be called any number of times to generate unique sample sets. It is used inside a batch power iteration: each transport sweep draws a new RQMC sample set to initialize particle positions and directions, inactive batches converge the source strength, and active batches accumulate tally statistics. This mechanism removes the fixed-seed bias while retaining the O($N^{{-1}}$) convergence of QMC sampling.","core_discovery":"The central claim is that the batch RQMC iQMC method converges both the k-effective and the mean scalar flux error at the theoretical QMC convergence rate of O($N^{{-1}}$) despite the stochastic noise introduced by randomizing the low-discrepancy sequence each transport sweep. Using Owen's randomization of the Halton sequence to generate a fresh set of samples per batch, the method removes the under-sampling bias of the fixed-seed approach while the averaging over active batches reduces the introduced noise. On the Takeda-1 problem, the batch method produces stable and accurate solutions with nearly two orders of magnitude fewer particle histories per batch than the fixed-seed method, and the ensemble-averaged convergence follows the O($N^{{-1}}$) rate until the spatial mesh discretization error becomes the limiting factor.","pith_inferences":["If the O(N^{-1}) rate reproduces on other benchmarks and meshes, batch RQMC iQMC could make quasi-Monte Carlo transport practical for deep-penetration or void-dominated problems where ray effects are severe.","Because the Halton sequence is less sensitive to sample count than Sobol, the batch scheme could naturally support varying particle numbers per batch, enabling adaptive refinement in the active phase.","The method effectively turns iQMC into a stochastic iterative map; existing theory of stochastic power iteration might help set inactive/active batch counts automatically rather than by user input.","One testable extension is to apply the same batch RQMC approach to the piecewise-linear source tilting from earlier iQMC work, which may push the convergence plateau to higher particle counts."],"forward_implications":["The batch method produces stable k-effective estimates in the Takeda-1 benchmark with roughly two orders of magnitude fewer particles per transport sweep than the fixed-seed method needs to avoid diverging.","Averaging the tally scores over active batches cancels most of the stochastic noise introduced by randomization, so the ensemble-mean error still follows the QMC O(N^{-1}) rate.","The scalar flux error plateaus at high particle counts because the 25x25x25 mesh limits accuracy; refining the mesh should extend the observable O(N^{-1}) regime.","By randomizing the sequence each batch, the method loses the linearity that fixed-seed iQMC used for Krylov solvers, so the batch approach trades solver acceleration for unbiased sampling."],"supporting_citations":[{"why":"Introduces the iterative Quasi-Monte Carlo method that this work modifies.","marker":"[1]"},{"why":"Supplies the Halton low-discrepancy sequence that is randomized here.","marker":"[3]"},{"why":"Provides Owen's randomization, the RQMC technique used to generate unique sample sets per batch.","marker":"[12]"},{"why":"Describes the Monte Carlo Dynamic Code (MC/DC) in which the batch RQMC method is implemented and tested.","marker":"[13]"},{"why":"Defines the Takeda-1 benchmark used to measure k-effective and scalar flux convergence.","marker":"[17]"},{"why":"Presents the iQMC flattened power iteration for k-eigenvalue problems that this batch algorithm adapts.","marker":"[6]"}],"fun_headline_variants":["Batch randomized-Halton hits O(N^-1) with 100x fewer particles","Randomized Halton per sweep removes ray effects in iQMC","iQMC batch scheme reaches QMC limit with ~100x less work","Owenized Halton batches: neutron transport converges at 1/N"],"cache_read_input_tokens":9984,"weakest_assumption_plain":"The O($N^{{-1}}$) convergence rate is measured on a single benchmark under the assumption that the scalar flux error is dominated by the RQMC sampling error, not by the coarse spatial mesh that visibly flattens the error at high particle counts.","fun_headline_variants_meta":{"raw":{"variants":["Batch randomized-Halton hits O(N^-1) with 100x fewer particles","Randomized Halton per sweep removes ray effects in iQMC","iQMC batch scheme reaches QMC limit with ~100x less work","Owenized Halton batches: neutron transport converges at 1/N"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1919,"prompt_tokens":1051,"completion_tokens":868,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":786}},"tokens_in":667,"tokens_out":868,"duration_ms":8311,"temperature":1.0,"reasoning_tokens":786,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:15.518983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the batched RQMC iQMC method on the Takeda-1 problem with a substantially finer mesh (or adaptive refinement near flux gradients) and check whether the k-effective and mean scalar flux errors continue to follow O($N^{{-1}}$) across a wider range of particle counts than shown in Figure 5. If the error curve flattens at the same particle count regardless of mesh, the claimed rate is an artifact of the mesh; if the curve continues, the claim is supported.","supporting_citations":[{"cited_title":"A Quasi–Monte Carlo Method With Krylov Linear Solvers for Multigroup Neutron Transport Simulations","cited_arxiv_id":null,"evidence_quote":"Introduces the iterative Quasi-Monte Carlo method that this work modifies."},{"cited_title":"On the efficiency of certain quasi-random sequences of points in evaluating multi- dimensional integrals","cited_arxiv_id":null,"evidence_quote":"Supplies the Halton low-discrepancy sequence that is randomized here."},{"cited_title":"A randomized Halton algorithm in R","cited_arxiv_id":null,"evidence_quote":"Provides Owen's randomization, the RQMC technique used to generate unique sample sets per batch."},{"cited_title":"3-D Neutron Transport Benchmarks","cited_arxiv_id":null,"evidence_quote":"Defines the Takeda-1 benchmark used to measure k-effective and scalar flux convergence."},{"cited_title":"iQMC: Iterative Quasi-Monte Carlo for k-Eigenvalue Neutron Transport Simulations","cited_arxiv_id":null,"evidence_quote":"Presents the iQMC flattened power iteration for k-eigenvalue problems that this batch algorithm adapts."}],"review_version":1}