{"id":"7cc0ef79-6cec-4a2a-9de2-506ffc36f883","arxiv_id":"2607.08723","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"IRC-safe optimal-transport metrics (EMD, sEMD) enable lower-bias cell resampling of negative-weight NLO Monte Carlo events without intermediate jet clustering.","lead":"This paper applies optimal-transport distance metrics to cell resampling algorithms that eliminate negative Monte Carlo event weights in collider simulations. It matters because negative weights waste computing resources at the LHC, and this method reduces bias compared to prior approaches while removing the need for jet clustering.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The ΣMD figure of merit uses EMD as its ground metric, structurally favoring EMD-based reweighting in the key comparison against Andersen et al.","rationale":"The reader correctly identified the sample-size limitation and the authors' own acknowledgment of it (Sec. 5.3). This is a legitimate concern about generalization to experimental-scale samples. However, I identify a more fundamental issue: the ΣMD, which provides the strongest evidence for the central comparative claim, uses EMD as its ground metric, creating a structural bias in favor of the EMD-based reweighting being evaluated. This is not addressed in the paper. The independent kinematic evidence partially mitigates this but is weaker than the ΣMD comparison — differences are mostly within statistical uncertainties at f_rw = 0.25, and the clear advantage at f_rw = 0.75 is concentrated in two of four observables. The reader's CONDITIONAL verdict is appropriate; this concern reinforces it without requiring a change. The concrete test (recomputing ΣMD with an alternative ground metric) would settle whether the comparative advantage is real or an artifact of metric choice. If the advantage persists under an independent ground metric, the claim is substantially strengthened; if not, the central comparative claim needs qualification.","tokens_in":21275,"tokens_out":3764,"duration_ms":231692,"concrete_test":"Recompute the ΣMD in Figure 14 using the Andersen et al. Euclidean jet-based metric as the ground metric instead of EMD. If the EMD-based reweighting still shows lower ΣMD under this alternative ground metric, the advantage is robust and metric-independent. If the ranking reverses or the gap closes substantially, the ΣMD comparison was an artifact of using the same metric for reweighting and evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OT-based metrics 'reduce the observed bias relative to other cell resampling techniques.' The clearest and most consistent evidence for this is the ΣMD comparison in Figure 14, where the Andersen et al. reweighting shows nearly double the ΣMD of the EMD reweighting at high f_rw. However, the ΣMD is defined with EMD as its ground metric (Eq. 4.2: Θ_ij = EMD_{β,R}(E_i, E'_j)). This creates a structural bias: the EMD-based reweighting defines event 'nearby' in the same metric space the ΣMD uses to measure distortion, so locally well-behaved reweightings in EMD space will naturally score better. The Andersen et al. metric defines locality differently (Euclidean distances over jet objects), so its reweightings may appear more distorting in EMD space even if they are locally well-behaved in their own metric. The paper does not discuss this potential circularity. The independent 1D kinematic comparisons (Figs. 12–13) provide weaker evidence: at f_rw = 0.25, differences between all methods are mostly within statistical uncertainties; at f_rw = 0.75, the EMD advantage is visible primarily in H_T and N_jets but not in ΔR(j1,j2) or pT,j1/pT,j2. The reader's sample-size concern (10^5 events vs. millions) is also valid but is about generalization; the ΣMD circularity questions whether the comparative advantage is real even at the studied sample sizes.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript proposes using optimal-transport (OT) based metrics — the Energy Mover's Distance (EMD) and the Spectral Energy Mover's Distance (sEMD) — as distance measures for cell resampling algorithms that eliminate negative Monte Carlo event weights in NLO QCD samples. Because these metrics are IRC-safe by construction, the resampling can be applied directly to particle-level information without intermediate jet clustering, and at any stage of event generation (hard scatter, parton shower, or hadronization). The authors also introduce the Cross-Section Mover's Distance (ΣMD) as a holistic, unbinned figure of merit for quantifying reweighting bias. The approach is validated on Z+jets and tt̄ samples generated with MadGraph5_aMC@NLO + Pythia8, and the OT-based metrics are shown to reduce bias relative to the Euclidean object-based metric of Andersen et al. [13].","tokens_in":22040,"tokens_out":1474,"duration_ms":264964,"significance":"The problem addressed — negative and pathological weights in NLO+shower MC samples — is timely and practically important for the HL-LHC computing program. The key methodological contribution is the use of IRC-safe OT-based metrics that eliminate the jet-definition dependence of prior cell resampling approaches. The introduction of the ΣMD as an observable-independent, unbinned figure of merit is a useful and broadly applicable tool. The study of reweighting at different generation stages (HS, PS, HAD) is a genuine advantage of IRC-safe metrics and is explored systematically. The constant-offset procedure for handling negative weights in the ΣMD (Sec. 4.2.1) is a clean, well-justified practical contribution grounded in the generalized Wasserstein cancellation property. The paper provides falsifiable, quantitative comparisons across two distinct final-state topologies.","major_comments":[{"comment":"Sec. 4.2, Eq. (4.2) and Fig. 14: The ΣMD uses EMD as its ground metric (Θ_ij = EMD_{β,R}(E_i, E'_j)), which means the ΣMD measures sample-level distortion in the same metric space that the EMD-based reweighting uses to define event locality. This creates a structural preference for EMD-based reweighting over the Andersen et al. Euclidean metric: reweightings that are locally well-behaved in EMD space will naturally score better in the ΣMD, while reweightings that are locally well-behaved in a different metric space may appear more distorting. The authors do not discuss this potential circularity. The 1D kinematic comparisons (Figs. 12–13) are independent of this concern, but at f_rw = 0.25 the differences between all methods are mostly within statistical uncertainties, and at f_rw = 0.75 the EMD advantage is visible primarily in H_T and N_jets but not in ΔR(j1,j2) or pT,j1/pT,j2. The ΣMD","section":null},{"comment":"Sec. 3 and Sec. 5.3: All comparisons use samples of 10^5 events, far smaller than the millions of events used in experimental productions. The authors acknowledge that 'some aspects of [the Andersen et al. metric's] performance may be due to its application to a much sparser metric space' (Sec. 5.3). This caveat is important because the Euclidean metric of Ref. [13] was developed and validated on much larger samples. If the relative performance advantage of EMD diminishes at realistic sample sizes — where the Euclidean metric's N log(N) scaling gives it a practical computational edge — the central comparative claim would not hold in the regime where it matters most for LHC experiments. The authors should either (a) provide at least one comparison at a larger sample size to test whether the ranking is stable, or (b) more clearly scope the comparative claim to the sparse-sample regime and","section":null},{"comment":"Sec. 2.1: The EMD computation scales as O(N^3 log N) per event pair (or O(N^2) with approximate solvers), and the authors note that 'further improvements related to runtime efficiency have been left to future studies.' However, no timing benchmarks are provided, making it difficult to assess whether the method is practical for production-scale samples. A table of typical per-event-pair computation times for the Z+jets and tt̄ samples (which have different particle multiplicities) would substantially strengthen the paper's claim of practical applicability.","section":null}],"minor_comments":[{"comment":"Sec. 2.1: The statement that R must be at least half of the maximum possible ground space distance when β > 1 is mentioned but the specific value R = 11.64 used throughout is only introduced later in the section. Consider stating the chosen R value when it is first needed.","section":null},{"comment":"Sec. 5.2, Fig. 8: The caption states the two events had H_T values of 187 GeV and 141 GeV, but the figure y-axis label reads 'max(H_T, H_T')' and '|H_T - H_T'|' without clear indication of which curve corresponds to which limit. Please clarify in the figure legend or caption.","section":null},{"comment":"Sec. 5.1, Fig. 3b: The x-axis is labeled 'Radius [GeV]' but the EMD ground metric is defined in the (η, φ) plane, which is dimensionless. Please clarify the units.","section":null},{"comment":"Sec. 4.1: The observable m_{b,l} is described as having a kinematic endpoint at √(m_t² - m_W²) ≈ 153 GeV, but this endpoint applies only to the specific decay topology where the b-jet and lepton come from the same top. In dileptonic tt̄ events, the endpoint differs. Please specify the decay channel or clarify.","section":null},{"comment":"Sec. 5.3: The sEMD is described as showing 'similar but consistently worse performance' than the EMD, but no explanation is offered for why this is the case. A brief discussion of whether the spectral representation's loss of geometric information (mentioned in Sec. 2.2) is responsible would strengthen the analysis.","section":null},{"comment":"The reference list includes Refs. [62, 63] which appear to be from 2025–2026; please verify these citations are complete and correctly attributed.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The ΣMD circularity concern is the most substantive issue raised in review. It does not invalidate the paper's contributions — the IRC-safety advantage and the ΣMD as a figure of merit are both genuine — but it does mean the comparative claim against Andersen et al. rests primarily on the 1D kinematic comparisons, which are less decisive than the ΣMD. The authors should be asked to address this directly rather than relying on the ΣMD as the primary evidence for superiority. The sample-size limitation is also worth flagging to the editor as a scope issue, though it may be beyond what can be reasonably addressed in a revision."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper replaces the Euclidean jet-based distance metric in cell resampling (Andersen et al.) with IRC-safe optimal transport metrics (EMD and sEMD), eliminating the need for intermediate jet clustering. That's a genuine practical improvement. The constant-offset trick for extending the ΣMD to signed measures is also a nice, clean methodological contribution — it's justified by the generalized Wasserstein cancellation property and requires no modification to standard OT solvers. The study across two processes (Z+jets and tt̄) and three generation stages is thorough, and the finding that hadronization-level reweighting performs best is well-supported by both kinematic distributions and the ΣMD. The β scan is useful and the choice of β=1 is well-motivated. Credit where it's due: this is careful, systematic work on a real computational problem for HL-LHC experiments. The IRC-safety argument for skipping jet clustering is the cleanest part of the paper and holds up completely. The stress-test concern about ΣMD circularity is the real soft spot, and it's more than minor. The ΣMD uses EMD as its ground metric (Eq. 4.2), so when the paper compares EMD-based reweighting against the Andersen et al. Euclidean metric using ΣMD (Figure 14), the deck is structurally stacked: EMD reweighting defines locality in the same space the ΣMD uses to measure distortion. The Andersen et al. metric defines locality differently, so its reweightings will naturally look more distorting in EMD space even if they're locally well-behaved in their own metric. The authors don't discuss this. The independent 1D kinematic comparisons (Figs. 12–13) partially compensate — at f_rw=0.75, EMD does show visibly better performance in H_T and N_jets — but at f_rw=0.25 differences are mostly within statistical uncertainties, and ΔR(j1,j2) and pT,j1/pT,j2 don't clearly favor EMD. So the comparative advantage claim rests more heavily on the ΣMD than it should. The sample-size concern (10^5 events vs. millions in real productions) is also legitimate. The authors acknowledge it themselves for the Andersen et al. metric, but it cuts both ways: the O(N^3) EMD scaling is a real practical barrier that's deferred to future work. The paper doesn't ship code or data, which limits verification. This is a solid paper that deserves a serious referee. The core methodological contribution — IRC-safe metrics for cell resampling without jet clustering — is sound and useful. The comparative claim against Andersen et al. needs either an independent figure of merit (not EMD-based) or an explicit discussion of the circularity concern. A referee should push hard on that point and on the scaling question. I'd lean toward accept if the authors add that discussion and soften the comparative claim accordingly.","headline":"OT-based cell resampling is a solid practical advance, but the key comparison against the prior Euclidean metric has a structural bias in its favor that the authors don't address.","tokens_in":22298,"tokens_out":696,"would_cite":true,"duration_ms":141769,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Optimal transport beats Euclidean metrics for cleaning negative MC weights","keywords":["optimal transport","cell resampling","negative event weights","Energy Mover's Distance","Monte Carlo event generation","infrared collinear safety","Cross-Section Mover's Distance","NLO QCD"],"falsifier":"If, when applied to million-event samples comparable to those used in actual LHC productions, the EMD-based cell resampling does not produce measurably lower bias than the Euclidean object-based metric (or if its computational cost makes it impractical at that scale), the central claim of superiority would not hold in the regime where it matters for experiments.","tokens_in":21526,"feed_emoji":"🚚","tokens_out":1309,"duration_ms":136595,"temperature":0.7,"pith_summary":"The paper claims that cell resampling algorithms — which locally redistribute negative and pathologically large event weights among nearby events in a metric space to make all weights positive — perform better when the notion of 'nearby' is defined by optimal-transport distances rather than by the Euclidean object-based metrics used in prior work. The specific metrics studied are the Energy Mover's Distance (EMD), which computes the minimum work needed to rearrange one event's radiation pattern into another's, and a spectral variant (sEMD) that operates on a one-dimensional energy-weighted angular representation. Because these metrics are infrared- and collinear-safe by construction, the resampling can be applied directly to particles at any stage of Monte Carlo event generation — after the hard-scattering matrix element, after the parton shower, or after hadronization — without first clustering particles into jets, which prior Euclidean approaches required and which introduced an unwanted dependence on jet algorithm parameters. Applied to next-to-leading-order Z+jets and top-antitop samples, the EMD with angular exponent β=1 applied to fully hadronized events yields the lowest bias across kinematic distributions and a holistic sample-level distance metric the authors call the Cross-Section Mover's Distance (ΣMD), outperforming the Euclidean metric particularly at high reweighting fractions.","feed_headline":"Optimal transport beats Euclidean metrics for cleaning negative MC weights","feed_subtitle":"An IRC-safe distance lets cell resampling skip jet clustering and cut bias in NLO samples, but the test is on 100k events, not millions.","key_machinery":"The EMD between two collider events is defined as the minimum transport cost to rearrange one event's energy flow into the other's, computed over the two-dimensional rapidity-azimuth plane with a ground metric of Euclidean angular distance, an angular exponent β controlling sensitivity to different angular scales, and a parameter R setting the relative importance of transport cost versus total-energy difference. Cell resampling uses this metric to define hyperspherical neighborhoods (cells) around negatively-weighted seed events; when the total weight in a cell is positive, weights are redistributed via the transformation w_i → (Σ w_j / Σ |w_j|) |w_i|, making all cell weights positive while保","core_discovery":"The central discovery is that optimal-transport-based distances — specifically the EMD with β=1 on hadronized events — provide an infrared-safe, jet-definition-independent metric for cell resampling that introduces less bias into kinematic distributions than the Euclidean object-based metric from prior literature, while also enabling reweighting at any stage of event generation. The authors further introduce the ΣMD, an unbinned optimal-transport distance between entire event samples, as a general-purpose figure of merit for quantifying reweighting bias that does not depend on choosing specific observables or binning.","pith_inferences":["The comparison against the Euclidean metric was conducted on samples of 10^5 events, far smaller than the millions used in experimental productions. The authors acknowledge that the Euclidean metric's N log(N) scaling may give it a practical advantage at realistic sample sizes, so the demonstrated superiority of EMD may not hold in the regime where LHC experiments actually operate.","The O(N^3 log N) computational cost of exact EMD computation for high-multiplicity events (e.g., tt̄ after hadronization) could become a bottleneck when applied to millions of events, even though the bias reduction is demonstrated convincingly at the smaller scale studied.","The constant-offset trick for negative weights in the ΣMD works because cell resampling preserves event support (no events added or removed), but extending it to samples with genuinely different supports — as in generative model evaluation — would require ghost events or a different approach."],"forward_implications":["If the EMD-based cell resampling scales to the millions-of-events samples used in real LHC productions, it could reduce the effective sample sizes needed for NLO simulations by factors of 6–16 (depending on the negative-weight fraction), directly alleviating the storage and computing bottleneck projected for the High-Luminosity LHC era.","The IRC-safety of the EMD means cell resampling can be applied as a post-hoc afterburner at any stage of the simulation pipeline without modifying the generator itself, making it compatible with existing production workflows.","The ΣMD provides a standardized, binning-independent benchmark that any group developing reweighting or sample-correction algorithms could adopt to report bias in a comparable way.","The constant-offset procedure for handling negative weights in optimal transport — adding a uniform shift to make all weights non-negative, exploiting the cancellation property of generalized Wasserstein distances — is a general technique applicable to any signed-measure optimal transport problem, not just cell resampling.","If the approach generalizes beyond Z+jets and tt̄ to more complex final states (e.g., multi-leg merged samples with higher negative-weight fractions), it could become a standard post-processing step in Monte Carlo production chains."],"fun_headline_variants":["IRC-safe optimal transport cuts bias in negative-weight cell resampling","Energy Mover's Distance enables jet-free cell resampling with less bias","Optimal transport distances let cell resampling skip jet clustering","EMD-based cell resampling reduces NLO weight bias without jet definitions","Cross-Section Mover's Distance quantifies reweighting bias without binning"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The performance advantage of the optimal-transport metrics over the Euclidean object-based metric is established on samples of only 100,000 events, whereas real LHC productions use millions. At larger sample sizes, the metric space becomes denser, cell radii shrink, and the Euclidean metric's superior computational scaling may erode or reverse the bias advantage demonstrated here. The authors themselves flag this concern.","fun_headline_variants_meta":{"raw":{"variants":["IRC-safe optimal transport cuts bias in negative-weight cell resampling","Energy Mover's Distance enables jet-free cell resampling with less bias","Optimal transport distances let cell resampling skip jet clustering","EMD-based cell resampling reduces NLO weight bias without jet definitions","Cross-Section Mover's Distance quantifies reweighting bias without binning","IRC-safe metrics beat Euclidean resampling for negative MC weights","Optimal transport resamples negative weights with less kinematic bias","ΣMD offers unbinned bias metric for full-phase-space reweighting"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1085,"prompt_tokens":455,"completion_tokens":630,"prompt_tokens_details":null},"tokens_in":455,"tokens_out":630,"duration_ms":37952,"temperature":1.0,"reasoning_tokens":530,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T02:14:21.523839+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If, when applied to million-event samples comparable to those used in actual LHC productions, the EMD-based cell resampling does not produce measurably lower bias than the Euclidean object-based metric (or if its computational cost makes it impractical at that scale), the central claim of superiority would not hold in the regime where it matters for experiments.","supporting_citations":[],"review_version":1}