{"id":"1f819f73-a526-4c52-a981-aa9396d91801","arxiv_id":"2507.05599","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A likelihood decomposition suggests the camera and optics explain about 99% of the signal in widefield single-particle tracking, implying post-processed trajectories carry little motion-model information.","lead":"This paper derives a likelihood for single-molecule widefield fluorescence tracking and claims that the imaging (emission) model, not the particle's motion model, dominates the likelihood by about 99 to 1. The authors argue this means motion models cannot be reliably learned from post-processed particle tracks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99/1 emission/motion split is a magnitude comparison, not an information measure; the emission term is independent of D, so the central conclusion does not follow from the equations.","rationale":"The reader's weakest assumption identifies precisely the central flaw: relative log-likelihood magnitude is not a measure of information about the motion model, because the emission term is independent of D. My independent reading of Eqs. (3), (6), and (8) confirms this. All likelihood-ratio information about diffusivity D is contained in the motion term; the emission term is a large but parameter-independent additive contribution. The paper's own admitted dependence on the intraframe interpolation number K further shows that the 99/1 split is not a robust property of the physics but a consequence of the chosen likelihood decomposition. I therefore agree with the reader that the headline conclusion does not follow from the derivation, and the appropriate disposition of the paper's central claim remains rejection. The tracking results and the demonstration that CONDOR and AnomDiffDB misclassify BM trajectories may support a weaker, more modest claim about specific post-processing pipelines, but they do not rescue the main information-theoretic argument. No additional independent concern is needed; the identified logical gap is sufficient and load-bearing.","tokens_in":28513,"tokens_out":3308,"duration_ms":42905,"concrete_test":"Generate one synthetic image stack from a BM trajectory at the reference parameters of Table 2, then compute the normalized likelihood L(D)/L(D_ML) over a grid of D values using the full likelihood Eq. (7). Recompute the same normalized curve using only the motion term Eq. (3), omitting the emission term. If the two normalized curves agree to numerical precision, the emission term contributes zero Fisher information about D, and the 99% emission share does not limit motion-model learning. As a secondary cross-check, repeat the likelihood-split calculation at K=1 versus K=100 interpolations; if the reported roughly 99% emission contribution reverses, the headline split is a modeling artifact rather than a property of the data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that the emission model 'contributes approximately 99% to the likelihood,' implying that little motion-model information reaches post-processed trajectories. This inference is unsupported. In Eq. (6), the emission term contains the observed pixel counts, the PSF, the photon rate, and the particle positions, but not the diffusivity D or any motion-model parameter. The motion term in Eq. (3) is the only part of the full likelihood in Eq. (7) that depends on D. Consequently, every likelihood ratio, posterior comparison, or model-classification statistic that varies D cancels the emission term exactly. The large numerical magnitude of the emission term comes from summing over P×N pixels and from the noise model, but a large constant in log-likelihood does not carry information about motion parameters; information is carried by the dependence on the parameter, i.e., by the motion term. The authors' 'Logic dictates' argument in the Introduction and the use of relative log-likelihood magnitudes in Figure 3 therefore conflate absolute likelihood weight with inferential information. This is not just a rhetorical issue: the ratio is also modeling-dependent, since the authors concede in the Results that an arbitrarily large number K of intraframe interpolations would reverse the split. The tracking demonstration itself is plausible and supported, but the central quantitative conclusion that motion models explain 'meager 1%' and that this 'casts doubt' on learned motion models does not follow from Eq. (7).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives a likelihood for widefield fluorescence single-particle tracking that explicitly separates an emission model (pixel intensities, PSF, camera noise) from a motion model (transition probabilities). It then compares the numerical magnitudes of the two contributions to the log-likelihood, reports that the emission term contributes roughly 99% and the motion term about 1%, and concludes that post-processed trajectories carry almost no information about the particle's motion model, thereby casting doubt on the widespread practice of learning motion models from trajectories. The paper also demonstrates that a Brownian-motion likelihood can track particles generated from several anomalous diffusion models with high coverage, and it shows that two existing classification tools (CONDOR and AnomDiffDB) frequently misclassify Brownian trajectories after trajectory extraction.","tokens_in":28709,"tokens_out":2547,"duration_ms":35149,"significance":"If the central claim were valid, it would have broad implications for the single-particle-tracking literature on anomalous diffusion. The paper is also commendable for shipping code and data, for formulating a reasonably explicit likelihood, and for providing a concrete, potentially reproducible comparison of trajectory-extraction methods and classifiers. However, the headline claim is not supported by the paper's own equations: the emission term is independent of the motion-model parameters, so its large numerical magnitude cannot measure how much information the data carry about motion. The tracking-invariance result and the classifier-bias observations are of some independent interest, but the paper's principal conclusion—that motion models explain only about 1% of the data—is an artifact of comparing likelihood magnitudes rather than an information-theoretic statement.","major_comments":[{"comment":"The central inference conflates the magnitude of a log-likelihood term with its information content. In Eq. (6), the emission term depends on the observed pixel counts, the PSF, the photon rate, the background, and the particle positions, but not on the diffusivity D or on any parameter of the motion model. The motion term in Eq. (3) is the only place where D enters. Consequently, every likelihood ratio between different values of D—equivalently, every posterior comparison or model-selection quantity involving motion parameters—has the emission term cancel exactly. The large numerical size of the emission contribution in Eq. (8b) comes from summing over P×N pixels and from the camera noise model; a large additive constant in log-likelihood carries no information about D. The sentence in the Introduction beginning 'Logic dictates' is therefore not a logical consequence of the likelihood decomposition, and the abstract's claim that the emission model 'contributes approximately 99% to the likelihood, leaving motion models to explain a meager 1% of the data' is not a valid basis for the conclusion that little motion-model information permeates into post-processed trajectories.","section":"Eqs. (3), (6), (7), and (8); Introduction, 'Logic dictates'"},{"comment":"The 99/1 split is not a robust experimental finding because it depends on tunable modeling choices. The authors themselves concede that an arbitrarily large number K of intraframe interpolations would reverse the split, since this shrinks the lag time Δt and inflates the motion term in Eq. (8c). The choice of ROI size (32×32 pixels, Table 2) and the pixel count P also directly scale the emission term in Eq. (8b), so the reported 'approximately 99%' is a consequence of the authors' parameter choices rather than a property of widefield fluorescence experiments in general. The manuscript should at minimum quantify the sensitivity of the 99/1 ratio to K, ROI size, and P; without such an analysis the headline statistic is not a defensible summary of 'typical' experiments.","section":"Results, 'Likelihood Contributors' (paragraph beginning 'In principle, there exists a single exception')"},{"comment":"The classification results in Table 1 are presented as evidence for the 99/1 claim, but they support only a weaker and different statement. The observation that CONDOR and AnomDiffDB classify inferred trajectories differently from ground-truth trajectories shows that trajectory extraction can bias downstream classification; it does not quantify how much information about the motion model is present in the post-processed trajectory. The small sample (six trials for each tool in Table 1) and the already-known sensitivity of feature-based classifiers to localization error make these results suggestive but not load-bearing for the paper's central conclusion. The claim in the Discussion that the findings 'confirm that motion model classifications are biased by trajectory inference' is fine, but it should not be equated with the claim that the emission term explains 99% of the data or that motion models are 'a meager 1%' of the likelihood.","section":"Results, 'Motion Model Classification'; Table 1"}],"minor_comments":[{"comment":"The phrase 'the motion-induced variance introduced when measuring its average position over a single, may generate' appears to be missing the word 'frame' after 'single'.","section":"Introduction, paragraph 3"},{"comment":"The caption reads 'despite it's peak QE'; this should be 'its peak QE'.","section":"Table 2 caption"},{"comment":"The text refers to 'Table 3in the main text' where the main text numbering gives Table 2 for the reference parameters; please correct the cross-reference.","section":"Supplementary Information, Part IV"},{"comment":"The notation 'w1:P N' and 'R1:K 2:N' is used without an explicit definition of the indexing convention in the main text; a short notation table or explicit sentence would improve readability.","section":"General notation"}],"recommendation":"reject","confidential_remarks":"The paper is clear and the authors have shared code and data, but the core result is a magnitude comparison dressed as an information-theoretic conclusion. Because the emission term is independent of D, the 99/1 statistic cannot support the paper's central claim, and the acknowledged dependence on K further undermines its robustness. I do not see a way to repair the central claim within the manuscript's current scope; a future submission focused on the tracking-invariance and classifier-bias demonstrations, without the 'motion models explain meager 1%' framing, could be a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the 99/1 split at the heart of this paper is a magnitude comparison, not an information measure, so the headline conclusion doesn't follow from their own equations. The emission term in Eq. (6) has no D dependence; every likelihood ratio between motion models cancels it. The large emission magnitude comes from summing over pixels. So the claim that post-processed trajectories carry almost no motion-model information is unsupported. That said, the paper does real work: the likelihood decomposition separating emission from motion is clean, the tracking demonstration (BM likelihood recovers anomalous trajectories with 99.9% CI coverage) looks solid, and the small classifier comparison showing CONDOR and AnomDiffDB misclassify BM is a useful caution. The problem is the interpretation. The authors themselves admit that an arbitrarily large number K of intraframe interpolations can reverse the split, which shows the 99/1 statistic is a modeling artifact, not a robust fact about information. The tracking result doesn't depend on the 99/1 claim, so it stands on its own. The classification experiments are small (18 trials) and tool-specific, so they don't support broad conclusions about the whole anomalous-diffusion literature. Bottom line: this is a paper with a useful toolkit and a flawed central argument. It deserves a serious referee, but the referee should push for a major reframing: report the likelihood decomposition and tracking robustness, drop the 'meager 1%' information claim, and present the classifier results as a specific caution about existing tools rather than a global invalidation of the field.","headline":"The 99/1 emission/motion split is a magnitude comparison, not an information measure; the paper's strongest claim fails, but the tracking demo and likelihood decomposition are worth a careful look.","tokens_in":29317,"tokens_out":2284,"would_cite":false,"duration_ms":25776,"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":"The paper argues that in typical widefield fluorescence single-particle tracking, the emission model contributes about 99% of the log-likelihood, so post-processed trajectories carry little information about the particle's motion model","keywords":["single-particle tracking","widefield fluorescence microscopy","motion model inference","anomalous diffusion","emission model","likelihood factorization","localization error","Bayesian inference"],"falsifier":"Run a controlled simulation with known ground-truth Brownian and anomalous trajectories, compute the Bayes factor between Brownian motion and each anomalous model from the raw image-stack likelihood and from the post-processed trajectory likelihood, and check whether the image-stack comparison is substantially more decisive; if it is not, the claim that post-processing discards most motion information would be refuted.","tokens_in":28212,"feed_emoji":"🔬","tokens_out":9413,"duration_ms":102980,"temperature":0.7,"pith_summary":"This paper asks whether motion models—the transition probabilities that describe how a particle moves—can actually be learned from the trajectories that single-particle tracking software outputs from widefield fluorescence movies. The authors write down a likelihood that separates the emission model (how photons hit pixels and camera noise corrupts them) from the motion model (how positions evolve), and compute how much each contributes in typical diffraction-limited experiments. They find that the emission term contributes roughly 99% of the log-likelihood, leaving the motion term with about 1%, across normal and anomalous diffusion models and over a wide range of parameters. If true, this means that the tracks analyzed in most single-particle-tracking studies carry very little motion-model information, that trajectory extraction is nearly insensitive to the assumed motion model, and that motion-model classification should be attempted from raw image stacks rather than post-processed tracks.","feed_headline":"Fluorescence noise holds 99% of the likelihood in tracking data","feed_subtitle":"If true, post-processed tracks carry little motion information, so anomalous-diffusion findings need rechecking.","key_machinery":"The load-bearing object is the factorized tracking likelihood $\\mathbb{L} = \\mathbb{P}(\\mathrm{Data}\\mid\\mathrm{Position}) \\times \\mathbb{P}(\\mathrm{Position}\\mid\\mathrm{Motion})$, which separates the emission model—a Gaussian point-spread function integrated over pixels, compounded over frames and pixels with a Gamma-distributed EMCCD readout—from the motion model, a Gaussian transition density with diffusivity $D$ and a tunable number $K$ of intraframe interpolations. The argument is carried by comparing the logarithms of the two factors: the emission term accumulates over the number of pixels and frames, giving it an enormous magnitude, while the motion term depends on displacement statistics and is comparatively small. The comparison is performed under Markov chain Monte Carlo sampling with a Brownian-motion prior.","core_discovery":"The central claim is that in widefield fluorescence single-particle tracking, the likelihood factorizes cleanly into an emission part and a motion part, and the emission part dominates. For a particle diffusing in three dimensions imaged on an EMCCD with a Gaussian point spread function, the emission log-likelihood—arising from pixel integration, photon shot noise, and detector gain—is typically about two orders of magnitude larger than the motion log-likelihood, with the motion share never exceeding 10% and often falling near 0.1%. The authors verify this by synthesizing image stacks from several motion models, tracking them with a Brownian-motion likelihood, and measuring the numerical contribution of each term. They find that a Brownian assumption recovers 99.9% of true positions even for anomalous motion models, confirming that tracking is robust to the assumed dynamics, yet two benchmark classifiers misclassify pure Brownian trajectories as anomalous in most trials. The authors conclude that post-processed trajectories are primarily informed by the emission model and that motion models should be learned from raw image stacks.","pith_inferences":["A complementary sensitivity analysis—computing the Fisher information of $D$ and the anomalous exponent rather than comparing log-likelihood magnitudes—would test whether the motion term's small numerical share also means low identifiability, since the emission term does not itself contain $D$.","The factorized likelihood suggests a practical diagnostic: before reporting a motion model from any widefield single-particle-tracking dataset, researchers could simulate under the fitted camera model and report the emission share of the log-likelihood.","The same emission/motion split could be applied to other detector architectures such as sCMOS, confocal, or MINFLUX to identify acquisition regimes where motion information survives post-processing.","Because the motion term grows with the number of intraframe interpolations $K$, the framework implies that adding unobserved interpolated positions does not recover lost motion information in practice, since interpolated positions are highly uncertain and inflate the motion term artificially."],"forward_implications":["Most anomalous-diffusion classifications obtained from post-processed widefield fluorescence trajectories may be artifacts of static and dynamic localization errors rather than genuine motion.","Learning motion models from raw image stacks is necessary, because post-processed trajectories discard most of the information the data contain about position and therefore about motion.","A Brownian-motion assumption in the tracking step is not the main source of bias: particle positions are recovered with 99.9% accuracy even when particles move anomalously.","Existing classifiers that do not include Brownian motion as a candidate model will systematically over-report anomalous diffusion, and the paper documents that 12 of 15 benchmark tools never test for Brownian motion.","The framework offers a quantitative diagnostic: before reporting a motion model, researchers can compute the emission share of the likelihood on simulated data and report it alongside the inferred model."],"supporting_citations":[{"why":"Establishes the hidden-Markov likelihood factorization into emission and motion terms that the paper adopts as its starting point.","marker":"[55]"},{"why":"Provides the statistics-optics perspective that justifies the photon-counting, point-spread-function-convolved emission model.","marker":"[68]"},{"why":"Supplies the Gamma-distributed EMCCD multiplication noise model used in the emission term.","marker":"[69]"},{"why":"Provides the EMCCD gain and conversion-factor calibration values used in the reference parameter set.","marker":"[70]"},{"why":"Defines the anomalous-diffusion benchmark, its trajectory generators, and the classifier competition that the paper re-examines.","marker":"[20]"},{"why":"A modular tracking platform whose extracted trajectories are compared against ground truth in the classification benchmark.","marker":"[63]"},{"why":"A Bayesian tracker built on the paper's likelihood formulation, used to show that a Brownian prior recovers anomalous trajectories with high accuracy.","marker":"[65]"},{"why":"One of the two classification tools benchmarked, evaluated on both ground truth and inferred trajectories.","marker":"[35]"},{"why":"The other classification tool benchmarked, which shows drift toward anomalous classifications when fed inferred trajectories.","marker":"[37]"}],"fun_headline_variants":["Emission model holds 99% of fluorescence tracking likelihood","Motion models explain only 1% of widefield tracking data","Post-processed single-particle tracks reveal little motion info","Tracking analysis dominated by camera noise, not motion","Anomalous diffusion results may hinge on tiny data share"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the relative numerical sizes of the emission and motion terms in the log-likelihood measure how much information each contributes to learning the motion model, with the split also depending on the chosen number of intraframe interpolations $K$.","fun_headline_variants_meta":{"raw":{"variants":["Emission model holds 99% of fluorescence tracking likelihood","Motion models explain only 1% of widefield tracking data","Post-processed single-particle tracks reveal little motion info","Tracking analysis dominated by camera noise, not motion","Anomalous diffusion results may hinge on tiny data share"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1643,"prompt_tokens":1052,"completion_tokens":591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":526}},"tokens_in":668,"tokens_out":591,"duration_ms":7016,"temperature":1.0,"reasoning_tokens":526,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:22:35.099531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled simulation with known ground-truth Brownian and anomalous trajectories, compute the Bayes factor between Brownian motion and each anomalous model from the raw image-stack likelihood and from the post-processed trajectory likelihood, and check whether the image-stack comparison is substantially more decisive; if it is not, the claim that post-processing discards most motion information would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the statistics-optics perspective that justifies the photon-counting, point-spread-function-convolved emission model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EMCCD gain and conversion-factor calibration values used in the reference parameter set."}],"review_version":1}