{"id":"276a33db-0205-4929-b6f6-96ccc8c5fb2a","arxiv_id":"2412.17434","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"Maximum likelihood estimation on model fluorescence distributions that include light-assisted-collision loss recovers atom-number proportions in a tight optical tweezer with few-percent accuracy from about 600 test runs.","lead":"This paper shows how to recover the number of atoms in a tiny optical trap from fluorescence photon counts, even when atom loss during detection makes the counts overlap and non-Poissonian. The method uses maximum likelihood estimation on a physical model of atom loss, and could improve atom counting in quantum computing and many-body physics experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy claim is validated in-sample only; a misspecified constant-rate loss model could bias fitted proportions without being detected by the same-pool resampling tests.","rationale":"The reader's conditional verdict is appropriate, and the weakest assumption is in the right neighborhood: the Sec. III model must be correct for the method to transfer beyond the calibration pool. My concern sharpens this into a methodological point: the same-pool resampling design cannot detect a misspecified model that affects control and test data in the same way. The paper partially addresses transferability with the temperature test in Sec. VIII B, but that test is explicitly aided by the cooling action of the detection light and does not cover other plausible model violations. I do not regard this as a fatal flaw; the model is plausible and the closed form Eq. (3) is algebraically consistent with the stated assumptions. Rather, it is the main unvalidated precondition for the central few-percent accuracy claim, and it justifies the reader's call for conditionality. Because the reader already assigned a conditional verdict and my concern does not overturn the central claim, the verdict should remain unchanged. The proposed simulation would settle whether the Markov assumption is truly load-bearing or whether the method is robust to this class of misspecification.","tokens_in":14018,"tokens_out":18415,"duration_ms":172600,"concrete_test":"Run a simulation study in which synthetic photon-count data are generated from a physically plausible non-Markovian loss process, e.g., Gamma_2(t) = Gamma_0 + Gamma_1 exp(-t/tau) with tau comparable to T_exp, while all other model parameters match the reported regime. Apply the full calibration-and-fit pipeline of Secs. IV-V to these synthetic test and control pools, and check whether the fitted atom-number proportions (a0, a1, a2) show biases beyond the claimed few-percent level. If such biases appear, the central claim is conditional on the Markov assumption; if not, the concern is settled in the method's favor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. V's performance test draws both the control and test histograms from the same experimental pool (blocks of 200 randomly assigned to test or control). This makes the reported few-percent deviation a statement about sampling variance under the pool's true per-atom distributions, not about the correctness of the Sec. III model. If Eq. (2) is misspecified—for example, if the light-assisted-collision loss rate is not constant during the exposure but has an initial transient, or depends on the instantaneous two-atom separation distribution—the calibration fits in Sec. IV B will absorb the misspecification into effective values of eta_i and gamma_ij. Because the same effective b_i then also describe the test pool, the fitted atom-number proportions still appear unbiased, and the claimed accuracy does not expose the error. The only external-validity test (Sec. VIII B) varies the initial atom temperature and finds insensitivity only because the detection light rethermalizes the atoms; the authors explicitly caution that this is design-dependent. Thus the load-bearing assumption that the constant-rate Markov model is the correct functional form for the two-atom distribution remains untested by the reported validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a maximum-likelihood method to infer the atom-number distribution of zero, one, or two atoms in a tightly focused optical tweezer from the photon-number distribution measured during fluorescence imaging. The method models the photon-count distribution for a fixed atom number as Poisson, and accounts for light-assisted-collision loss during the exposure via a Markov recursion (Eq. 2), yielding a closed form for up to two atoms (Eq. 3). Model parameters are calibrated on control histograms of known atom numbers, and the unknown proportions are found by maximizing a Poisson log-likelihood over a grid. The paper also studies an event-based variant that uses recorded histograms and an effective likelihood, and it tests the method's accuracy by resampling from experimental test and control pools. The main claims are that the model-based fit achieves few-percent accuracy with about 600 test and 1400 control events, outperforms least-squares fitting, and can be extended to situations without a model.","tokens_in":14210,"tokens_out":17611,"duration_ms":159628,"significance":"If the accuracy claim can be fully supported, this is a useful and practical tool for the cold-atom and quantum-information community. The explicit modeling of light-assisted loss during detection, the comparison of Poisson and effective likelihoods, and the bootstrap error estimates are valuable contributions. The paper is honest about the assumptions made in the model, and the model fits the control histograms well. However, the validation strategy is entirely in-sample, and the treatment of pre-detection loss is inconsistent between the model-based and event-based analyses, so the quantitative few-percent accuracy claim is not yet fully established.","major_comments":[{"comment":"The accuracy claim is validated only by resampling from the same experimental pool that is used for calibration. Section V A randomly assigns blocks from a single pool to test and control, so both the fitted b_i and the test histograms are drawn from the same underlying conditional distributions. As a result, the reported few-percent RMSE in Figs. 4–6 measures sampling fluctuations under the empirically measured distributions, not the correctness of the model in Sec. III. If Eq. (2) is misspecified—for example, if the light-assisted-collision loss rate is not constant during the exposure or depends on the instantaneous two-atom separation—the calibration in Sec. IV B will absorb the misspecification into effective values of eta_i and gamma_ij, and the test will still appear unbiased. The only external check, Sec. VIII B, varies the initial temperature, but the authors state that the detection light rethermalizes the atoms, so it does not test a different physical regime. I recommend adding an out-of-sample validation, for example simulated data generated from a different (e.g., time-dependent-loss) model, or an independent atom-number measurement, and reporting how the method performs under those conditions.","section":"Sec. V A and Sec. VIII B"},{"comment":"The treatment of the pre-detection loss probability ell is inconsistent between the model-based and event-based fits. The model-based fit in Sec. IV C uses pure distributions b_i,k, i.e., the photon-count distributions conditional on i atoms being present at the start of the detection exposure. Consequently, the fitted proportions a_i are the actual atom-number distribution at that moment. In contrast, the validation ground truth in Sec. V A is the distribution of the preparation-stage labels (the number of atoms loaded into the tweezers), which is related to the actual distribution by binomial thinning with survival probability 1 - ell. The paper corrects the event-based fits for ell in Appendix A (Eqs. A1–A3) but applies no corresponding correction to the model-based fits before comparing them to the known input proportions. For ell ~ 0.03, this introduces a systematic offset of order 1–5 percentage points in the reported errors, which is not separated from the statistical scatter in Figs. 4–6 and could be mistaken for the distribution-overlap effects discussed in Fig. 5. The authors should state explicitly which quantity (pre-detection label distribution or post-detection actual-atom-number distribution) is being estimated, and apply the appropriate transformation consistently for both model-based and event-based fits in the performance comparison.","section":"Sec. IV C, Sec. V A, and Appendix A"},{"comment":"The bootstrap error estimation in Sec. VII resamples only the test data and keeps the model distributions b_i,k fixed. The uncertainty in the model parameters (eta_i and gamma_ij), which are themselves estimated from a finite control sample, is therefore not propagated into the reported confidence intervals. The sensitivity to N_c seen in Figs. 4 and 6 shows that this uncertainty is non-negligible for small control sets, and the effective likelihood method of Sec. VI A explicitly attempts to account for it only in the event-based case. To make the bootstrap CIs reliable in general, the authors should either resample the control histograms as well (e.g., a nested bootstrap) or provide an estimate of the additional uncertainty and state the range of N_c over which the fixed-b_i approximation is valid.","section":"Sec. VII"}],"minor_comments":[{"comment":"The inequality '1 < r < 100' in the definition of the averaging over resamples should read '1 <= r <= 100'.","section":"Eq. (6)"},{"comment":"The sentence 'This parameter gathers single atom losses...' is grammatically incomplete and would read better as 'This parameter accounts for single-atom losses proceeding from collisions with background gas, from the merge process, and from false 1-atom detection with the EMCCD camera.'","section":"Sec. IV B"},{"comment":"The labels in the inset of Fig. 4 are very small; please enlarge the inset or describe the color scale in the caption so the reader can follow the dependence of Delta on both N_d and N_c.","section":"Fig. 4"},{"comment":"The symbol T in Eq. (2) is used before it is defined; the general model would benefit from an explicit statement at the start of Sec. III A that T denotes the exposure duration, with T = T_exp later in Sec. III B.","section":"Sec. III A"},{"comment":"The phrase 'the F = 3 -> F' = 3 component was turned off' is ambiguous; more precisely, the frequency component resonant with that transition was turned off, not the transition itself.","section":"Sec. VIII A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid applications study with a useful model and extensive experimental testing, but the two major issues—in-sample validation and the inconsistent handling of the pre-detection loss ell—need to be resolved before publication. The ell issue appears straightforwardly fixable by applying the appropriate transformation consistently or by clearly redefining the ground truth. The validation issue is more structural and may require a simulation study or an independent measurement. The manuscript fits the journal's scope well, and the authors seem capable of addressing the concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. The model-based recursion in Eq. (2) is the genuinely new and useful piece: it lets you construct fluorescence distributions for i atoms from those for fewer atoms, with Poisson photon statistics and constant loss rates, which turns overlapping non-Poisson histograms into a clean MLE problem. The paper shows it beats threshold methods and event-based histogram fits, especially when control statistics are limited. Second, the accuracy claim—few-percent errors from ~600 test events—is validated only in-sample, on resamples from the same experimental pool that calibrates the model. That doesn't make it wrong, but it does make it pool-specific.\n\nCredit where it's due: the derivation of Eq. (3) is clear and the assumptions are explicit (Poisson per-atom counts, constant loss rates, negligible one-body loss during the exposure). The performance characterization is careful—RMSE over the whole parameter simplex and bootstrap confidence intervals. The comparison of model-based vs event-based fitting and the effective-likelihood correction for limited control histograms is a practical guide experimentalists will appreciate. The temperature-insensitivity test is honest about its design-dependent caveat.\n\nThe soft spot is exactly the one the stress-test note flags: if the constant-rate Markov loss model is misspecified—say a transient in the loss rate at the start of exposure, or intensity dependence—the control fits will absorb it, and same-pool resampling can't detect it. The authors do provide one diagnostic: an additional two-body loss channel would make the 2-atom histogram fit fail. But that doesn't cover all misspecifications. Also, no code or data are shipped, and the fitted eta_i and gamma_ij values aren't reported, which makes independent replication harder than necessary. These are addressable and don't break the central claim for the tested regime.\n\nBottom line: solid methods paper for experimenters in tight-tweezer few-atom work. Send it to a serious referee. I'd want an external validation—synthetic data from a known non-Markov loss process, or a different temperature/power set with independent calibration—before trusting the few-percent number elsewhere, but the method itself is sound.","headline":"Useful model-based MLE recipe for tight-tweezer atom counting, with an in-sample validation that deserves an external test before you trust the numbers elsewhere.","tokens_in":14746,"tokens_out":3182,"would_cite":true,"duration_ms":28794,"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":"Maximum likelihood fits of model fluorescence distributions that account for light-assisted collision loss recover atom-number proportions in a tight optical tweezer to few-percent accuracy from about 600 test events where thresholds fail.","keywords":["maximum likelihood estimation","optical tweezers","atom number distribution","light-assisted collisions","photon counting","fluorescence distribution","bootstrap error estimation","few-atom experiments"],"falsifier":"Take two-atom control histograms at two different detection-light detunings, fit the model parameters ($\\{\\eta_j\\}$, $\\{\\gamma_{2j}\\}$) to each, then fit a known two-atom mixture recorded at the first detuning using the parameters obtained from the second; a systematic offset beyond the bootstrap confidence intervals would falsify the constant-rate assumption and the claimed few-percent accuracy.","tokens_in":13815,"feed_emoji":"⚛️","tokens_out":17841,"duration_ms":141810,"temperature":0.7,"pith_summary":"When several atoms share a tight optical tweezer, the light used to detect them causes rapid atom loss through light-assisted collisions. That loss makes the recorded photon-count distributions overlap and become non-Poissonian, so the usual threshold or Poisson fits cannot tell how many atoms were present. This paper presents a maximum likelihood method that fits a weighted sum of model distributions explicitly including the loss process, recovering the atom-number proportions from a relatively small number of experimental runs: about 600 test events plus roughly 1400 control events give errors at the few-percent level. The same idea works without an analytic model if more data are available. Accurate atom-number readout is central to tweezer experiments, which count surviving atoms after interaction-loss events, for instance to read out internal states.","feed_headline":"Maximum likelihood fits pin atom-number proportions to a few percent","feed_subtitle":"From about 600 test runs plus calibrations, the method recovers atom-number shares where thresholds fail.","key_machinery":"The central object is the recursion relation for the model photon-count distributions $b_{i,k}(T)$: a piecewise-Poisson description in which counts at fixed atom number $i$ are Poisson with rate $\\eta_i$, and loss events occur as a memoryless Markov process with constant rates $\\gamma_{ij}$ from $i$ to $j$ atoms. Its work is to turn control histograms from known atom numbers into an analytic basis that can be evaluated at any exposure time and atom number, and to provide a closed-form two-atom distribution that separates the collision-loss channels. The fitting machinery is the Poisson log-likelihood maximized by a grid search over the proportions, and, for the event-based variant, an effective likelihood that averages over Poisson fluctuations in the control histograms. Bootstrap resampling supplies the confidence intervals on the fitted proportions. This combination is what allows accurate atom-number extraction from datasets as small as about 600 test events.","core_discovery":"The paper's central claim is that the photon-count distribution for a known initial number of atoms in a tight tweezer can be modeled sufficiently well for fitting by assuming Poisson emission between loss events and memoryless loss at constant rates $\\gamma_{ij}$. For an $i$-atom sample, the distribution $b_{i,k}(T)$ obeys a recursion relation whose first term is the no-loss Poisson contribution and whose second term integrates over the time of the first loss event; for up to two atoms this recursion closes into an explicit formula with detection rates $\\eta_0,\\eta_1,\\eta_2$ and loss rates $\\gamma_{20},\\gamma_{21}$. Fitting these parameters to control histograms from known atom numbers, then fitting a mixture $f_k(\\theta)=N_d\\sum_i a_i b_{i,k}$ to a test histogram with the Poisson log-likelihood, determines the atom-number proportions $(a_0,a_1,a_2)$ with errors at the few-percent level from about 600 test and 1400 control events. The same procedure using measured instead of modeled basis distributions works when no model is available, at the cost of requiring better statistics, and the effective likelihood improves event-based fits in that regime.","pith_inferences":["The fitted loss rates $\\gamma_{ij}$ are measurable outputs, so the method could double as a probe of light-assisted-collision dynamics rather than only a counting correction.","The Poisson assumption per atom-number segment may fail if saturation or high collection efficiency changes the count statistics; the event-based effective-likelihood route would be the fallback in that regime.","Splitting the exposure into time bins and fitting the recursion jointly would yield time-resolved loss curves, a natural extension that the paper does not discuss.","A direct stress test would be to vary the detection-beam power after calibration and check whether a single set of loss rates reproduces the observed histograms; if not, the fitted proportions will be biased."],"forward_implications":["With roughly 600 test and 1400 control events, atom-number proportions are recovered to within a few percent, a level that threshold and Poisson fits do not reach in the loss-dominated regime.","The recursion relation extends to three or more atoms, so the method promises to replace rudimentary counting in multi-atom tweezer experiments.","When no analytic model for the photon-count distributions exists, event-based fitting with the effective likelihood remains usable, at the price of requiring larger datasets.","The observed insensitivity of the photon-count distributions to atom temperature between 27 and 68 µK supports the use of control data taken under slightly different conditions, as long as the detection light rethermalizes the atoms.","Maximum likelihood outperforms least-squares for these count histograms, because the Poisson likelihood respects the discrete, non-negative nature of the data."],"supporting_citations":[{"why":"Supplies the assumption that photon counts at fixed atom number are Poisson distributed, the starting point of the model in Eq. (1).","marker":"[25]"},{"why":"Documents light-assisted-collision loss in a tight tweezer and the loading technique, defining the regime where threshold counting fails.","marker":"[1]"},{"why":"States that a collisional loss event can eject one or both atoms, justifying the two loss channels $\\gamma_{20}$ and $\\gamma_{21}$ in the closed-form two-atom distribution.","marker":"[26]"},{"why":"Demonstrates multi-atom tweezer experiments that rely on counting remaining atoms and where event-based fitting of measured distributions has been used.","marker":"[12]"},{"why":"Supplies measured photon-count distributions for known atom numbers, an event-based fitting basis that the paper compares against its model-based approach.","marker":"[20]"},{"why":"Introduces the effective likelihood used in Eq. (7) to account for statistical fluctuations in the control histograms during event-based fitting.","marker":"[39]"},{"why":"Presents three-atom tweezer experiments to which the recursion-based method is claimed to extend beyond the two-atom demonstration.","marker":"[3]"}],"fun_headline_variants":["Loss-aware likelihood fits few-atom distributions","Six hundred runs are enough for atom-number fits","Photon-count modeling beats thresholds for few atoms","Tweezer loss modeled for accurate atom counts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method relies on the model in which a fixed number of atoms produces photons at a steady average rate and loses atoms at constant, memory-free rates set by the control data; if the real loss rates depend on conditions the controls do not match, the resulting atom-number proportions will be biased.","fun_headline_variants_meta":{"raw":{"variants":["Loss-aware likelihood fits few-atom distributions","Six hundred runs are enough for atom-number fits","Photon-count modeling beats thresholds for few atoms","Tweezer loss modeled for accurate atom counts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":1902,"prompt_tokens":912,"completion_tokens":990,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":932}},"tokens_in":528,"tokens_out":990,"duration_ms":8275,"temperature":1.0,"reasoning_tokens":932,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:27:12.091987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two-atom control histograms at two different detection-light detunings, fit the model parameters ($\\{\\eta_j\\}$, $\\{\\gamma_{2j}\\}$) to each, then fit a known two-atom mixture recorded at the first detuning using the parameters obtained from the second; a systematic offset beyond the bootstrap confidence intervals would falsify the constant-rate assumption and the claimed few-percent accuracy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the assumption that photon counts at fixed atom number are Poisson distributed, the starting point of the model in Eq. (1)."},{"cited_title":"Grünzweig, A","cited_arxiv_id":null,"evidence_quote":"Documents light-assisted-collision loss in a tight tweezer and the loading technique, defining the regime where threshold counting fails."},{"cited_title":"Grünzweig, M","cited_arxiv_id":null,"evidence_quote":"States that a collisional loss event can eject one or both atoms, justifying the two loss channels $\\gamma_{20}$ and $\\gamma_{21}$ in the closed-form two-atom distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates multi-atom tweezer experiments that rely on counting remaining atoms and where event-based fitting of measured distributions has been used."},{"cited_title":"Sompet, S","cited_arxiv_id":null,"evidence_quote":"Supplies measured photon-count distributions for known atom numbers, an event-based fitting basis that the paper compares against its model-based approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the effective likelihood used in Eq. (7) to account for statistical fluctuations in the control histograms during event-based fitting."},{"cited_title":"Weyland, S","cited_arxiv_id":null,"evidence_quote":"Presents three-atom tweezer experiments to which the recursion-based method is claimed to extend beyond the two-atom demonstration."}],"review_version":1}