{"id":"e27f4505-f979-4986-a938-d484ff709314","arxiv_id":"2412.05245","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For half-Gaussian priors on two-point separation, SPADE is close to optimal only for small separations; direct imaging can outperform SPADE for intermediate prior means.","lead":"This paper studies how to best measure the separation between two very close light sources when you already have a prior guess about that separation. It computes the fundamental Bayesian estimation error and compares two measurement schemes, SPADE and direct imaging, finding that the superior scheme depends on the prior.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (29) is invalid as printed (negative MSE for σ≳0.7); the analytic anchor of the non-displaced comparison is wrong, and the displaced-prior numerics behind the central DI-over-SPADE claim are not reproducible.","rationale":"The reader's weakest assumption (mode-map exactness) is not where I would put the weight: the Gaussian-PSF/HG expansion makes PNR↔SPADE and homodyne↔DI exact by a unitary relabeling of the complete HG basis, so that assumption is standard and internally consistent. The real problem is verification. Eq. (29) is presented as an analytic result for the non-displaced case, but as written it is dimensionally and algebraically impossible: for σ=1 it yields a negative MSE. I rederived the PNR MSE from Eqs. (26)–(28) (Poisson conditional likelihood and half-Gaussian prior); the correct posterior-mean second moment is 2σ²/(π√(σ²+1))[1+σ arcsin(σ/√(σ²+1))], and the printed expression differs by moving √(σ²+1) to the numerator. Since Fig. 1 is the only closed-form validation of the numerical/analytic pipeline, this error means the manuscript as submitted contains a false equation and cannot be considered internally verified. The displaced-prior comparison in Figs. 2–3, which supports the headline claim, is entirely numerical, with no code, no convergence criterion, and no discussion of the (μ,σ)→(μ_t,σ_t²) inversion. A single independent re-implementation is the cheapest decisive check. If the crossover persists after correction and reproduction, the central claim stands; if not, it falls. I therefore keep the reader's CONDITIONAL verdict, pending that check.","tokens_in":11900,"tokens_out":29042,"duration_ms":293698,"concrete_test":"Recompute the PNR/SPADE MSE for the half-Gaussian prior by direct numerical quadrature of Eqs. (26)–(28) for σ = 0.5, 1, 2, 3 and compare with the corrected closed form; the printed Eq. (29) will fail for σ>0.7. Then, in a fresh implementation of the displaced-prior Eqs. (26), (30), and (37), reproduce the DI/SPADE MSE ratio in Fig. 3(c)–(d) at σ_t²=1 with Fock cutoffs 35, 50, and 100; if the crossover where the ratio drops below 1 shifts by more than a few tenths in μ_t or disappears, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DI can beat SPADE in the middle range of μ_t rests on the numerical evaluation of Eqs. (26), (30), and (37), for which no code or convergence data is provided. The one closed-form check available, Eq. (29), is incorrect as typeset: substituting σ=1 gives σ² − (2/π)√2[(1)arcsin(1/√2)+1] ≈ 1 − 1.607 = −0.607, a negative MSE, and it is negative for all σ≳0.76. Repeating the PNR calculation from Eqs. (26)–(28) for the half-Gaussian prior gives E[Π_k²] = 2σ²/(π√(σ²+1))[1+σ arcsin(σ/√(σ²+1))], so the correct MMSE for PNR/SPADE is σ² minus that positive term; the printed formula has √(σ²+1) in the numerator instead of the denominator. Because Fig. 1 is supposedly generated from this formula, the analytic validation of the pipeline is invalid. The mapping PNR↔SPADE and homodyne↔DI is a standard unitary equivalence under the Gaussian-PSF assumption and is not the main risk. The load-bearing risk is that the numerical implementation used for the displaced-prior crossover is unverified; the demonstrable error in Eq. (29) makes it unsafe to accept the central quantitative claim without an independent reproduction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses Bayesian estimation of the separation of two incoherent point sources with a Gaussian PSF. For a half-Gaussian prior and a displaced half-Gaussian prior on the separation parameter, it compares the Bayesian MMSE with the MSEs of SPADE (modeled as photon-number-resolving detection in a single-mode coherent-state space) and direct imaging (modeled as homodyne detection). The non-displaced case is treated analytically through a squeezed-thermal-state representation of the operator Γ0 and a closed-form PNR MSE, while the displaced case is treated numerically by truncating a Fock expansion at up to 35 photons. The main conclusion is that SPADE is not universally optimal under a Bayesian prior: for intermediate prior means, direct imaging can outperform SPADE, in contrast to the Fisherian result.","tokens_in":12253,"tokens_out":24187,"duration_ms":214923,"significance":"If the numerical results are correct, the paper provides a useful Bayesian extension of superresolution imaging, showing that the choice between SPADE and direct imaging depends on the prior. The mapping between the hypothetical single-mode coherent-state space and the imaging space (Table I) is clearly laid out, and the squeezed-thermal-state treatment of the non-displaced prior is elegant. The paper also gives a closed-form expression for the photon-number-resolving MSE, which can serve as a benchmark. However, the paper does not ship code or convergence data for its central numerical claims, and there are demonstrable errors in two analytic statements that support those claims. The significance is therefore conditional on correction and independent verification.","major_comments":[{"comment":"The formula as typeset is incorrect: the second term has √(σ²+1) in the numerator rather than the denominator. For σ=1, the printed expression gives 1 − (2/π)√2(π/4+1) ≈ −0.607, and it is negative for all σ ≳ 0.76, which is impossible for an MSE. Repeating the Gaussian integral from Eqs. (26)–(28) gives σ² − (2σ²/[π√(σ²+1)])[σ arcsin(σ/√(σ²+1)) + 1]. Because this equation is the analytic anchor of the SPADE curve in Fig. 1, the equation must be corrected and Fig. 1 regenerated or explicitly confirmed to have used the corrected formula.","section":"Sec. V A, Eq. (29)"},{"comment":"The statement \"We can easily simplify the equation and find that trΓ₂=σ²\" is incorrect for the displaced half-Gaussian prior. The trace of Γ₂ equals the second moment of the prior, ∫ q² P(q)dq = μ_t² + σ_t² (with μ_t and σ_t² as defined in Eqs. (35)–(36)), not the scale parameter σ². For example, with μ=1 and σ=1, the second moment is approximately 2.86, not 1. This error is load-bearing: δ = trΓ₂ − tr(BΓ₁), so if the numerical MMSE calculations use trΓ₂=σ², the MMSE curves in Figs. 2 and 3 are shifted. Please correct this claim and specify how trΓ₂ was actually evaluated numerically.","section":"Sec. VI, after Eq. (37)"},{"comment":"The central claim that DI outperforms SPADE in the middle range of μ_t rests entirely on numerical evaluations of Eqs. (26), (30), and (37), but the paper provides no code, no pseudocode, no quadrature details for the homodyne integrals, and no convergence study beyond the statement that the maximum truncated number is 35 for Fig. 2. The demonstrable typo in Eq. (29) weakens confidence that the numerics use the correct analytic anchors. The authors should provide the code or a step-by-step numerical recipe with tolerances and convergence data for both Figs. 2 and 3.","section":"Sec. VI, Figs. 2 and 3"}],"minor_comments":[{"comment":"There are typos: \"non-dispalced\" in the abstract and \"assumtions\" in Sec. II; also \"conjuction\" appears in Sec. II.","section":"Abstract and Sec. II"},{"comment":"The second integral in Eq. (17) is over qα from −∞ to 0, while the text states \"we always set α = qα/√2 ≥ 0\". Please clarify the notation by defining α=|qα|/√2 after the change of variables.","section":"Eq. (17)"},{"comment":"The derivation of Eq. (29) from Eqs. (26)–(28) is not shown; please include the intermediate integral steps in an appendix, especially since this closed form is used to validate Fig. 1.","section":"Sec. V A"},{"comment":"The paper states the maximum truncated number is 35 for Fig. 2 but gives no cutoff or convergence information for Fig. 3 or for the homodyne quadrature integration in Eq. (30). Please state the numerical parameters used.","section":"Sec. VI"},{"comment":"The sentence \"SPADE has advantage when the variance is larger than 1.1\" is difficult to reconcile with Figs. 2(c)–(d) and with the following sentence about increasing the variance; please rephrase the non-monotonic behavior more precisely.","section":"Sec. VII"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound conceptual core and an elegant non-displaced analysis, but the published Eq. (29) is impossible as typeset and the trΓ₂=σ² claim in Sec. VI is false for displaced priors. Because the central claim is numerical and no code is provided, I would ask for code, convergence data, and a clear statement of how trΓ₂ was computed before accepting. This seems fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this paper adds a genuinely new Bayesian layer to the SPADE-versus-direct-imaging question. Personick's MMSE framework applied to two-point separation with half-Gaussian priors hasn't been done before, and the observation that the optimal measurement is prior-dependent is worth taking seriously. The single-source appendix is clean and the imaging-space mapping in Sec. V C is clearly explained.\n\nThe problem is Eq. (29). As typeset it gives a negative MSE for σ beyond about 0.76. Substituting σ=1 gives roughly -0.61, which is not an estimator error. The correct expression, from a direct PNR calculation using Eqs. (26)-(28), is σ² - (2σ²)/(π√(σ²+1))[1 + σ arcsin(σ/√(σ²+1))]. The printed version has √(σ²+1) in the numerator instead of the denominator. That matters because Eq. (29) is the only closed-form check in the paper; Fig. 1 is built on it.\n\nThe displaced-prior results in Figs. 2 and 3, which carry the central DI-over-SPADE claim, come from a truncated-Fock numerical calculation with no code and no convergence data. The stated cutoff (35) is not sufficient. I don't think the misplaced square root directly invalidates the displaced-prior crossover—those curves are computed independently—but it makes the pipeline look unsafe, and the headline claim is exactly the kind of quantitative statement that needs a reproducible numerics package or an independent check.\n\nThe paper is otherwise fairly careful: the Personick/Helstrom formalism is applied correctly, the squeezed-thermal structure of Γ0 is a nice touch, and the authors do not oversell their conclusion—they explicitly note SPADE is not optimal in general. The literature coverage is adequate.\n\nBottom line: this is a legitimate paper for the quantum-imaging audience, and it deserves a serious referee. But it should not be accepted until Eq. (29) is fixed and the displaced-prior numerics are released or independently reproduced. I would not cite it in its current form.","headline":"Plausible and novel Bayesian take on superresolution, but the only closed-form check is wrong as printed and the headline numerics are unreproducible; worthy of review but not citable as-is.","tokens_in":12724,"tokens_out":8911,"would_cite":false,"duration_ms":72842,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P50","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under a Bayesian prior, direct imaging can outperform SPADE for mid-range source separations, contrary to the Fisherian result.","keywords":["Bayesian quantum estimation","two incoherent point sources","separation estimation","SPADE","direct imaging","minimum mean square error","half-Gaussian prior","superresolution"],"falsifier":"A concrete experiment: prepare two incoherent point sources separated by a value drawn from a displaced half-Gaussian prior with mean $\\mu_t = 2$ and variance $\\sigma_t^2 \\approx 1$, then compare the mean squared error of SPADE and direct imaging over many trials; if direct imaging does not beat SPADE in the middle range of the prior, the paper's central regime claim is falsified.","tokens_in":11728,"feed_emoji":"🔭","tokens_out":6002,"duration_ms":56047,"temperature":0.7,"pith_summary":"This paper asks what happens when a prior probability distribution on the separation of two incoherent point sources is available, and evaluates measurements by the Bayesian minimum mean square error instead of the Fisherian Cramér–Rao bound. It finds that with a displaced half-Gaussian prior, SPADE remains nearly optimal for small likely separations, but in an intermediate range of the prior mean, direct imaging (DI) actually outperforms SPADE. The regime dependence matters because real telescopes and microscopes often have prior knowledge about the objects they image, so the best measurement strategy can depend on that prior.","feed_headline":"Direct imaging can beat SPADE for mid-range source separations","feed_subtitle":"Under a Bayesian prior, SPADE is not optimal; the best measurement depends on the separation's prior distribution.","key_machinery":"The argument rests on two pieces. First, the Bayesian MMSE formalism expresses the optimal single-shot mean squared error as $\\delta = \\mathrm{tr}(\\hat{\\Gamma}_2 - \\hat{B}\\hat{\\Gamma}_1)$ with $\\hat{\\Gamma}_k = \\int_0^\\infty P(\\alpha) \\alpha^k \\hat{\\rho}(\\alpha) \\, d\\alpha$ and $\\hat{B}$ the optimal projector. Second, through the Gaussian point-spread function the imaging problem is mapped to a single-mode coherent-state space: the Hermite–Gauss modes in the image plane become Fock states in the hypothetical space, so that photon-number-resolving detection corresponds to SPADE and homodyne detection to direct imaging. This mapping lets the authors compute the MMSE and compare it with the two measurement strategies. For the half-Gaussian prior, $\\hat{\\Gamma}_0$ becomes a squeezed thermal state, which enables an analytic treatment of the optimal measurement.","core_discovery":"The paper's central claim is that SPADE is not the optimal measurement in general for estimating the separation of two incoherent point sources when a prior on the separation is available. Working with the Bayesian minimum mean square error, the authors show that for a displaced half-Gaussian prior, SPADE attains near-optimal performance when the prior mean is small (sources expected close together) and when the prior mean is large both SPADE and direct imaging converge to the fundamental MMSE. In the middle range of the prior mean, however, direct imaging produces a smaller mean squared error than SPADE. This contrasts with the Fisherian approach, in which SPADE achieves the quantum Cramér–Rao bound regardless of separation.","pith_inferences":["The qualitative regime dependence (DI beating SPADE in a middle band) likely extends to non-Gaussian point-spread functions, but the boundaries of the band would shift; a robustness check with a different PSF would be a natural next step.","In adaptive settings, a receiver could first perform a coarse direct-imaging measurement to localize the prior, then switch to SPADE; the paper's MMSE comparison suggests such hybrid strategies could beat either fixed measurement.","For astronomical applications, the prior could come from a source catalog or magnitude distribution; the paper's framework suggests that the optimal instrument design depends on that catalog."],"forward_implications":["For small likely separations, SPADE remains the measurement of choice, approaching the Bayesian MMSE.","For intermediate prior knowledge, direct imaging can give lower mean squared error than SPADE, so measurement choice should be adapted to the prior.","For large likely separations, both measurements approach the MMSE and the Rayleigh limit becomes irrelevant.","The Bayesian MMSE gives a guaranteed single-shot attainable bound, unlike the Fisherian Cramér–Rao bound which is asymptotic.","The result suggests that prior information can be exploited to relax the requirement of exotic measurements in some regimes."],"supporting_citations":[{"why":"Establishes the Fisherian baseline where SPADE attains the quantum limit for all separations, which this paper contrasts with the Bayesian regime-dependent result.","marker":"[3]"},{"why":"Provides the mapping between the imaging space (Hermite–Gauss modes) and the hypothetical coherent-state space, used to equate PNR detection with SPADE and homodyne detection with direct imaging.","marker":"[6]"},{"why":"Supplies the Bayesian MMSE formalism (Eqs. 12–14) and the guarantee that MMSE is attainable by a projective measurement.","marker":"[20]"},{"why":"Provides the squeezed thermal state representation and the optimal receiver derivation (Appendix A) used for the half-Gaussian prior.","marker":"[27]"},{"why":"Supplies the analytic overlap of squeezed coherent states with Fock states, used in computing the matrix element in Eq. (24).","marker":"[28]"}],"fun_headline_variants":["Bayesian prior flips optimality: DI beats SPADE mid-range","For mid separations, direct imaging outperforms SPADE","SPADE not always optimal: DI wins with Bayesian prior","Bayesian estimation: direct imaging beats SPADE for mid-range","When priors matter: DI edges out SPADE in estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison assumes that the Gaussian point-spread function and paraxial approximation make the mapping from imaging space to a single-mode coherent-state space exact, so that PNR detection in the hypothetical space truly equals SPADE and homodyne detection truly equals direct imaging; if the point-spread function is not Gaussian, the quantitative crossover between DI and SPADE may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian prior flips optimality: DI beats SPADE mid-range","For mid separations, direct imaging outperforms SPADE","SPADE not always optimal: DI wins with Bayesian prior","Bayesian estimation: direct imaging beats SPADE for mid-range","When priors matter: DI edges out SPADE in estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":1053,"prompt_tokens":755,"completion_tokens":298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":371,"completion_tokens_details":{"reasoning_tokens":212}},"tokens_in":371,"tokens_out":298,"duration_ms":3850,"temperature":1.0,"reasoning_tokens":212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:53.427478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete experiment: prepare two incoherent point sources separated by a value drawn from a displaced half-Gaussian prior with mean $\\mu_t = 2$ and variance $\\sigma_t^2 \\approx 1$, then compare the mean squared error of SPADE and direct imaging over many trials; if direct imaging does not beat SPADE in the middle range of the prior, the paper's central regime claim is falsified.","supporting_citations":[{"cited_title":"Xxxi. investigations in optics, with spe- cial reference to the spectroscope,","cited_arxiv_id":null,"evidence_quote":"Establishes the Fisherian baseline where SPADE attains the quantum limit for all separations, which this paper contrasts with the Bayesian regime-dependent result."},{"cited_title":"Subdiffraction incoherent optical imaging via spatial-mode demultiplexing: Semiclas- sical treatment,","cited_arxiv_id":null,"evidence_quote":"Provides the mapping between the imaging space (Hermite–Gauss modes) and the hypothetical coherent-state space, used to equate PNR detection with SPADE and homodyne detection with direct imaging."},{"cited_title":"Bayesian parameter estimation us- ing gaussian states and measurements,","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian MMSE formalism (Eqs. 12–14) and the guarantee that MMSE is attainable by a projective measurement."},{"cited_title":"Quantum nonlocality in weak- thermal-light interferometry,","cited_arxiv_id":null,"evidence_quote":"Provides the squeezed thermal state representation and the optimal receiver derivation (Appendix A) used for the half-Gaussian prior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the analytic overlap of squeezed coherent states with Fock states, used in computing the matrix element in Eq. (24)."}],"review_version":1}