{"id":"e7455d3f-c3c2-4612-b2ef-ee0c26ea43e7","arxiv_id":"1908.01996","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An adaptive two-stage receiver that first estimates the object centroid with direct detection and then applies a misalignment-corrected mode-sorting measurement can approach quantum-limited sub-Rayleigh imaging without prior knowledge of the object location, in Monte Carlo simulations.","lead":"A two-stage imaging receiver first uses ordinary direct detection to find an object's centroid, then switches to a quantum-inspired mode-sorting measurement aligned to that estimate, splitting the available photons between the two stages. Monte Carlo simulations suggest this adaptive scheme can beat direct detection by one to two orders of magnitude in mean squared error for sub-Rayleigh separation and length estimation when the centroid is unknown.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 13's integrated likelihood is not a valid likelihood: p(ξ) depends on the unknown θ and on the first-stage data, so the reported simulation gains may rely on oracle information.","rationale":"The reader correctly flagged the conditional-likelihood treatment in Eq. 13 as the weakest assumption: the first-stage data determine the BSPADE axis and hence the misalignment ξ, while the same data are treated as i.i.d. observations conditioned on ξ. This stress-test goes one step further: the prior p(ξ) used in Eq. 13 is itself θ-dependent via Eq. 16, so the maximum likelihood procedure in Eq. 14 is either circular (using the true θ to build the prior) or relies on an undocumented plug-in estimate. That is a sharper and more concrete version of the reader's concern. The paper's headline quantitative claim, the one-to-two-order MSE improvement over direct detection, rests entirely on the Monte Carlo implementation of this estimator; if the likelihood is misspecified or oracle, the simulated gains do not demonstrate an achievable receiver performance. The concern is not about disagreement with the broader super-resolution consensus, but about the internal statistical validity of the paper's own estimator. The proposed test would distinguish the two possibilities: a corrected non-oracle likelihood that retains the gain would validate the scheme, while a large degradation would invalidate the central claim. Because the paper contains useful derivations and a plausible two-stage architecture, rejection is too strong without the test, but the current evidence is not sufficient for acceptance; hence the verdict remains conditional pending the likelihood correction and the re-simulation.","tokens_in":17338,"tokens_out":10331,"duration_ms":159048,"concrete_test":"Re-run the Fig. 3A Monte Carlo at N=100,000 and θ/σ ∈ {0.1, 0.5, 1.0} with two estimators: (i) the paper's Eq. 14 using p(ξ) with the true θ in σP^2, and (ii) a corrected likelihood that treats φ_hat as observed and marginalizes over φ with a broad prior (or, equivalently, integrates over the estimator error conditional on φ_hat and over ξS), with no θ appearing in the prior. If estimator (ii) no longer beats direct detection by one to two orders of magnitude in the sub-Rayleigh regime, the central claim is unsupported; if the corrected estimator retains the gain, the concern is settled.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equation 13 is not a valid likelihood for the realized two-stage experiment. The nuisance prior p(ξ) in Eq. 7 is built from σP^2 = σ^2/n1(1 + θ^2/4σ^2) (Eq. 16), so it depends on the unknown separation θ that Eq. 14 is supposed to estimate. Evaluating Eq. 14 therefore requires knowing θ to construct the prior (oracle information) or an unstated estimate of θ; either way, the reported MSEs are not those of a well-defined, feasible estimator. In addition, ξP = φ - φ_hat is a function of the first-stage data through φ_hat = mean(x_i), so the same data D1 determine ξ and appear in PD(D1|ξ,θ). Marginalizing ξ over its across-trial sampling distribution N(0,σP^2) without conditioning on D1 double-counts D1: the first-stage data are used both to select the BSPADE axis and as independent evidence about θ conditional on ξ. The Monte Carlo gains in Figs. 3-4 may therefore be artifacts of this misspecified integrated likelihood rather than properties of the receiver. The central claim of one to two orders of magnitude improvement is not established until Eq. 13 is replaced by a coherent marginal likelihood, e.g., by integrating over the unknown φ with a stated prior or by conditioning on the observed φ_hat and integrating only over the systematic misalignment ξS.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage adaptive receiver for estimating the separation of two incoherent point sources or the length of an extended object in the sub-Rayleigh regime when the object centroid is unknown. The first stage uses direct detection to estimate the centroid; the second stage performs a 0-BSPADE measurement aligned to that estimate. The authors construct a marginal likelihood (Eq. 13), define an ML estimator (Eq. 14), propose an adaptive rule for allocating the integration-time ratio alpha, and report Monte Carlo MSEs showing up to two orders of magnitude improvement over direct detection and near-QCRB performance. The central claims are that mode-sorting receivers can be made viable without prior centroid knowledge and that dynamic time allocation adds further benefit.","tokens_in":17586,"tokens_out":12101,"duration_ms":124045,"significance":"If the statistical construction were valid, this would be a practically important step for SPADE-type super-resolution imaging: it directly addresses the known sensitivity of mode sorters to centroid misalignment and gives a concrete feed-forward protocol with no fitted parameters, benchmarked against the external QCRB and direct-detection CRB. The centroid-error variance formulas (Eqs. 6, 16, 22) and the BSPADE target-mode probabilities (Eqs. 18, 23) are clean and useful. However, the validity of the likelihood in Eq. 13 is the load-bearing element, and it is currently not justified; the quantitative claims therefore need to be re-established with a coherent statistical model.","major_comments":[{"comment":"Equation 13 is not a valid likelihood for the realized two-stage experiment, and this invalidates the Monte Carlo MSEs in Figs. 3 and 4. The prior p(xi) in Eq. 7 uses sigma_P^2 = (sigma^2/n1)(1 + theta^2/4sigma^2) from Eq. 16, so it depends on the unknown separation theta that Eq. 14 is supposed to estimate; evaluating Eq. 14 therefore requires oracle knowledge of theta or an unstated estimate of theta. In addition, xi = phi - phi_hat + xi_S is a function of the first-stage data through phi_hat = (1/n1) sum x_i, so the same data D1 determine xi and also appear in P_D({x'_i}|xi,theta) in Eq. 9; marginalizing xi over its across-trial sampling distribution N(0,sigma_P^2+sigma_S^2) without conditioning on D1 double-counts the first-stage data. A coherent likelihood would integrate over the unknown phi with an explicit prior, or condition on the observed phi_hat and integrate only over the systematic misalignment xi_S. The abstract's central quantitative claims therefore rest on a misspecified estimator.","section":"III B, Eq. 13 (with Eqs. 7 and 16)"},{"comment":"The variance formula used to optimize alpha, <Delta theta^2_2-Stage> = (<Delta theta^2_D>^{-1} + <Delta theta^2_B>^{-1})^{-1}, treats the direct-detection and BSPADE estimates as independent contributors to a combined estimator. This is not justified: the BSPADE stage is aligned using the first-stage centroid estimate, so the two data sets are statistically dependent, and the BSPADE variance <Delta theta^2_B> in Eq. A15 is itself computed by averaging over p(xi), inheriting the same theta-dependent-prior problem as Eq. 13. Consequently the 'optimal' allocation ratio alpha*(theta) in Fig. 2 and the adaptive switching rule in Fig. 2C are built on an unproven approximation. The final MSEs are simulated, so the adaptive gains are not automatically invalid, but the paper should either derive the combined variance from the actual joint likelihood or present the allocation rule as a heuristic rather than as variance-minimizing.","section":"III C and Appendix A, Eq. A15"}],"minor_comments":[{"comment":"The exponentials in Eq. 17 are missing the square and the minus sign; they should read exp[-(x_i - xi +/- theta/2)^2/(2sigma^2)].","section":"II, Eq. 17"},{"comment":"The notation <Delta phi> should be <Delta phi^2> or should be explicitly defined as the variance, since the mean centroid error is zero.","section":"Appendix A, Eqs. A4 and A7"},{"comment":"The integral over x is written with both limits as infinity rather than -infinity to infinity; this appears to be a typographical error.","section":"Appendix A, Eq. A9"},{"comment":"The text contains the typo '0-BPSADE'; it should read '0-BSPADE'.","section":"III D, paragraph following Fig. 3"},{"comment":"The paper should state explicitly which estimator is used to produce the final Monte Carlo MSEs: the numerical ML estimator of Eq. 14 or the approximate closed-form estimator theta_B of Eq. A12. The current text describes both without specifying which one is evaluated in Figs. 3 and 4.","section":"III D and Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised in the reader's report is confirmed by the manuscript text: Eq. 13 is not a valid likelihood for the two-stage experiment, and the dependence of p(xi) on theta is explicit in Eq. 16. This is not an artifact of the review pipeline. I see a feasible path to revision, namely replacing the construction with a proper hierarchical or Bayesian treatment, or conditioning on phi_hat and integrating only over xi_S, and then rerunning the simulations. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuinely useful adaptive receiver idea, but the likelihood in Eq. 13 has a real statistical bug, so I wouldn't trust the MSE numbers until that is reworked. The paper still deserves a serious referee.\n\nWhat is new: previous unknown-centroid SPADE work was mostly asymptotic. This paper actually simulates finite photon number N with a two-stage receiver that spends part of the integration time on direct detection to estimate the centroid, then switches to BSPADE, and it adapts the switching time. The centroid-variance derivation in Eq. 6 and Appendix A is clean, the BSPADE target-mode probabilities are correct, and the Monte Carlo comparison is honest in the sense that the direct-detection baseline is the same code with alpha=1. There are no fitted parameters; the benchmarks are the QCRB and the direct-detection CRB. That is real work and worth credit.\n\nThe soft spot is exactly where the stress-test points. Eq. 13 is not a valid likelihood as written. The prior p(xi) in Eq. 7 is built from sigma_P^2 in Eq. 16, which depends on the unknown separation theta. So evaluating Eq. 14 requires either the true theta to build the prior (oracle information) or an unstated estimate of theta. Either way, the objective function is not a likelihood in the usual sense, and the Monte Carlo numbers in Figs. 3-4 are not the numbers a real receiver would produce without additional machinery. Second, the first-stage data are re-referenced to phi_hat, which is computed from those same data. After centering, the sample mean of x'_i is zero by construction. Treating D1 as i.i.d. samples from Psi(x'|xi,theta) and then marginalizing xi double-counts the centering information. A coherent formulation would integrate over the true centroid phi with a stated prior, or condition on phi_hat and integrate only over the systematic misalignment xi_S. The current construction is an empirical-Bayes or self-consistent likelihood, and the paper does not say so.\n\nI do not think this kills the central idea. A BSPADE receiver aligned to a noisy centroid estimate should still beat direct detection in the sub-Rayleigh regime; the rough factor-of-two-to-QCRB and the order-of-magnitude improvements are plausible in spirit. But as written, the quantitative claims are not established. The lack of code or error bars also makes it harder to see whether the adaptive allocation is doing the work or the likelihood misspecification is. The harmonic-sum approximation for optimizing alpha is heuristic, though that is a secondary issue.\n\nWho this is for: people working on quantum-inspired imaging and adaptive measurements. It deserves a serious referee. I would send it to review, not desk reject, but I would ask for a rewritten likelihood and rerun simulations.","headline":"A sensible adaptive two-stage receiver idea whose reported gains rest on a misspecified likelihood in Eq. 13; worth refereeing but not ready as stated.","tokens_in":18144,"tokens_out":9010,"would_cite":true,"duration_ms":101273,"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":"A two-stage adaptive receiver—direct detection to find the object, then a binary mode sorter aligned to it—estimates sub-Rayleigh separations and lengths without prior knowledge of the centroid, approaching the quantum Cramér–Rao bound.","keywords":["superresolution imaging","spatial-mode demultiplexing","BSPADE receiver","quantum Cramér–Rao bound","adaptive measurements","nuisance parameters","point source separation estimation","extended source length estimation"],"falsifier":"Run the two-stage receiver in simulation or experiment with the BSPADE axis set exactly to the sample mean of the first-stage photons, and estimate the separation with a likelihood that accounts for this selection rule instead of the paper's Eq. (13). If the mean squared error no longer beats idealized direct detection by one to two orders of magnitude in the sub-Rayleigh regime, the reported gains are an artifact of the assumed conditional independence.","tokens_in":1772,"feed_emoji":"🔭","tokens_out":1990,"duration_ms":80672,"temperature":0.7,"pith_summary":"The paper asks whether mode-sorting receivers—optical measurements that project light onto spatial modes before detection—can beat ordinary imaging when the object's location is unknown. It proposes a two-stage receiver: spend part of the integration time on standard direct detection, use those photons to estimate the object's centroid, align a binary spatial-mode-demultiplexing (BSPADE) measurement to that estimate, and use the remaining time on the BSPADE measurement. Monte Carlo simulations show that this scheme estimates the separation of two point sources and the length of an extended source with one to two orders of magnitude lower mean squared error than idealized direct detection alone, with no prior knowledge of the centroid. When the time split is adapted online, the receiver comes within about a factor of two of the quantum Cramér–Rao bound across most of the sub-Rayleigh regime. The central mechanism is the conversion of an unknown nuisance parameter—the centroid location—into a random misalignment with a known prior built from the first-stage data.","feed_headline":"Adaptive receiver beats direct detection with no location prior","feed_subtitle":"Splitting time between centroid estimation and a mode sorter cuts sub-Rayleigh error by 10 to 100 times.","key_machinery":"The central object is the adaptive two-stage receiver: first-stage direct detection produces photon arrival positions whose sample mean estimates the centroid, and the second stage is a 0-BSPADE measurement—a binary spatial-mode demultiplexer whose target mode matches the point spread function—aligned to that estimate. The argument is carried by the joint likelihood in Eq. (13), which multiplies the first-stage direct-detection likelihood by the binomial BSPADE likelihood and marginalizes over the residual misalignment $\\xi$ using a Gaussian prior whose variance comes from the centroid estimator, $\\sigma_P^2 = (\\sigma^2/n_1)(1+\\theta^2/4\\sigma^2)$ for two point sources. The adaptive loop then compares the elapsed-time fraction $\\alpha_t$ with the optimal fraction $\\alpha^*(\\hat{\\theta}_{D,t})$ computed from a coarse direct-detection estimate of the separation, switching to BSPADE only when the comparison says the mode sorter has become worthwhile. This machinery converts an unknown centroid into a known statistical misalignment model, which is what lets the mode sorter approach the quantum limit.","core_discovery":"The paper establishes that the known sensitivity of spatial-mode-sorting receivers to centroid misalignment can be overcome by a sequential passive measurement: a direct-detection stage estimates the centroid, and a 0-BSPADE stage, aligned to that estimate, provides near-quantum-optimal information about object scale. In simulations at $N=100{,}000$ photons, the two-stage receiver outperforms idealized direct detection by one to two orders of magnitude in mean squared error for sub-Rayleigh separation and extended-source length estimation, even with no prior on the centroid or the object scale. Adaptive choice of the allocation fraction $\\alpha$ improves on a fixed 50/50 split by up to a factor of two and automatically reverts to direct detection in the super-Rayleigh regime, where BSPADE becomes the inferior measurement. The same design works for both estimation tasks, supporting the claim that the strategy generalizes to more complex scenes with multiple nuisance parameters.","pith_inferences":["Editorial inference: a selection-aware likelihood that treats the first-stage centroid estimate as a deterministic function of the photon positions would likely raise the estimator variance, so the reported gains should be rechecked in a model that respects that selection rule.","Editorial inference: the same adaptive alignment principle could be applied to other nuisance parameters such as axial defocus, object rotation, or brightness imbalance, with preliminary measurements estimating each parameter before the mode-sorting stage; the paper sketches this extension but does not simulate it.","Editorial inference: as the total photon number $N$ decreases, the first-stage centroid estimate worsens, so the optimal first-stage time fraction should grow; a curve of optimal $\\alpha$ versus $N$ would be a direct, testable prediction of the framework.","Editorial inference: a Bayesian updating version, which forms a posterior over both centroid and separation from the direct-detection data before switching, could outperform the maximum-likelihood stopping rule, especially when partial prior information is available."],"forward_implications":["For sub-Rayleigh separation estimation with no centroid prior, the two-stage receiver beats idealized direct detection by one to two orders of magnitude in mean squared error at $N=100{,}000$ photons.","Adaptive allocation of integration time brings the receiver within about a factor of two of the quantum Cramér–Rao bound for most of the sub-Rayleigh regime, and it improves on a fixed 50/50 split by up to a factor of two.","In the super-Rayleigh regime, the adaptive receiver automatically spends the full integration time on direct detection, avoiding the degradation that a fixed two-stage receiver suffers when BSPADE becomes ineffective.","The same receiver design performs comparably for estimating the length of a uniform extended source, indicating the method is not specific to two-point-source separation.","The framework extends to additional nuisance parameters and multiple parameters of interest by using preliminary measurements to prepare one or more mode-sorting measurements adaptively."],"supporting_citations":[{"why":"Defines the BSPADE measurement and shows spatial-mode demultiplexing beats direct detection for sub-Rayleigh separation estimation; supplies the 0-BSPADE target-mode probability used in Eq. (11).","marker":"[17]"},{"why":"Establishes the quantum limit for estimating an object's length and identifies 0-BSPADE as the optimal measurement with perfect alignment; supplies the extended-source intensity model and target-mode probability in Eqs. (20)–(23).","marker":"[33]"},{"why":"Analyzes sequential measurement strategies for the unknown-centroid case in limiting regimes; the two-stage design builds on its result that the optimal allocation approaches one half at small separation.","marker":"[38]"},{"why":"Provides the quantum Cramér–Rao bound used as the ultimate precision benchmark for the estimation tasks.","marker":"[3]"},{"why":"Supplies the semiclassical treatment of spatial-mode demultiplexing, grounding the single-photon mode probabilities for BSPADE.","marker":"[43]"},{"why":"Provides the classical Cramér–Rao bound and maximum-likelihood estimation background used for the direct-detection variance approximation.","marker":"[44]"},{"why":"Supports the marginalization over the nuisance parameter in the two-stage likelihood of Eq. (13).","marker":"[47]"}],"fun_headline_variants":["Adaptive two-stage receiver: 10-100x better sub-Rayleigh imaging without prior","Without a location prior, adaptive receiver beats direct detection by 100x","Two-stage receiver: 10-100x error reduction without any location info","Adaptive mode-sorting receiver outperforms direct detection with no prior"],"cache_read_input_tokens":20224,"weakest_assumption_plain":"The main assumption is that the first-stage photons, after being re-centered on the estimated object position, can be treated as independent samples from an intensity pattern whose center is the residual alignment error, with that error drawn from a Gaussian prior; in reality those same photons were used to pick the center, so they are not independent of the error.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive two-stage receiver: 10-100x better sub-Rayleigh imaging without prior","Without a location prior, adaptive receiver beats direct detection by 100x","Two-stage receiver: 10-100x error reduction without any location info","Adaptive mode-sorting receiver outperforms direct detection with no prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001662,"raw_usage":{"total_tokens":6583,"prompt_tokens":918,"completion_tokens":5665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":5582}},"tokens_in":534,"tokens_out":5665,"duration_ms":37056,"temperature":1.0,"reasoning_tokens":5582,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:00:18.429085+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-stage receiver in simulation or experiment with the BSPADE axis set exactly to the sample mean of the first-stage photons, and estimate the separation with a likelihood that accounts for this selection rule instead of the paper's Eq. (13). If the mean squared error no longer beats idealized direct detection by one to two orders of magnitude in the sub-Rayleigh regime, the reported gains are an artifact of the assumed conditional independence.","supporting_citations":[{"cited_title":"Tsang, R","cited_arxiv_id":null,"evidence_quote":"Defines the BSPADE measurement and shows spatial-mode demultiplexing beats direct detection for sub-Rayleigh separation estimation; supplies the 0-BSPADE target-mode probability used in Eq. (11)."},{"cited_title":"Dutton, R","cited_arxiv_id":null,"evidence_quote":"Establishes the quantum limit for estimating an object's length and identifies 0-BSPADE as the optimal measurement with perfect alignment; supplies the extended-source intensity model and target-mode probability in Eqs. (20)–(23)."},{"cited_title":"ˇReh´ aˇ cek, Z","cited_arxiv_id":null,"evidence_quote":"Analyzes sequential measurement strategies for the unknown-centroid case in limiting regimes; the two-stage design builds on its result that the optimal allocation approaches one half at small separation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the quantum Cramér–Rao bound used as the ultimate precision benchmark for the estimation tasks."},{"cited_title":"Basano, P","cited_arxiv_id":null,"evidence_quote":"Supplies the semiclassical treatment of spatial-mode demultiplexing, grounding the single-photon mode probabilities for BSPADE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the classical Cramér–Rao bound and maximum-likelihood estimation background used for the direct-detection variance approximation."},{"cited_title":"ˇReh´ aˇ cek, M","cited_arxiv_id":null,"evidence_quote":"Supports the marginalization over the nuisance parameter in the two-stage likelihood of Eq. (13)."}],"review_version":1}