{"id":"ec0fcd49-c98c-4a29-ace9-e5e615b05e5b","arxiv_id":"2506.20653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Nulling interferometer differential-output noise is non-Gaussian and approximately follows an iteratively convolved Bessel distribution, so Gaussian significance thresholds are overconfident and need a correction factor.","lead":"This paper shows that the differential output of a dual Bracewell nulling interferometer contains non-Gaussian noise that arises from the product of phase and amplitude perturbations, and it models this noise with a Bessel-function distribution. A generalist should read it because Gaussian assumptions in detection statistics can overstate the significance of exoplanet detections, which matters for mission design for LIFE and similar space interferometers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IMB derivation in §4.1 is only valid for uniform template weights; for the actual varying planet template the matched-filter noise is not proven to be IMB, so fitted ν=36 and tail extrapolation may bias the TN/Tα correction.","rationale":"The central qualitative claim—that the differential output of a dual Bracewell nuller is non-Gaussian because the second-order phase-amplitude term is a product of Gaussian perturbations—is robust and well supported. My concern targets the specific quantitative claim: that the test-statistic distribution is exactly IMB and that the fitted ν=36 plus the lookup table provide the correct significance correction. The derivation in §4.1 proves the IMB form only for a uniform template weight η; for the actual planet modulation template, η varies in time, and the characteristic function of Σ η_i x_i y_i is a product of different-scale Bessel kernels that is not IMB. The paper papers over this by fitting ν to simulated data, which is reasonable as an approximation but is not a derivation of the 'true' distribution. Because the lookup table is used at extreme quantiles (p ~ 10⁻¹⁵) and extrapolated beyond the sampled range, a small misspecification of the tail shape can translate into a sizable error in TN/Tα. The concrete test I propose would settle whether the IMB approximation is accurate enough in the tail. I agree with the reader that the input-Gaussianity and common-mode-rejection assumptions also limit the IMB form, but the unequal-weight issue is more load-bearing because it is an internal mathematical gap that applies even under the paper's own assumptions. Credit is due for the public code, the cross-validation against LIFEsim, and the honest limitation discussion in §5.4; those do not resolve the missing derivation for the weighted matched-filter case. The verdict remains conditional: the paper should either prove or empirically validate the IMB form at the tails before the correction factor is used for detection/characterization claims.","tokens_in":36896,"tokens_out":8414,"duration_ms":104830,"concrete_test":"Generate N=10^10 samples of the exact matched-filter second-order term w = Σ_t η_t Σ_{ij} (∂²n/∂A_i∂φ_j) δA_i(t) δφ_j(t) under the reference-case Gaussian pink noise and real template, without assuming IMB. Use importance sampling or a saddlepoint approximation based on the exact characteristic function to estimate tail probabilities at Gaussian-equivalent thresholds TN = 5, 7, and 9.6, and compare the implied TN/Tα to the lookup table in Fig. 4. If the difference exceeds 1% at TN = 7.9, the IMB-based correction is not validated. As an analytic check, compute the characteristic function for n=2, η=(1,2) and show it is not proportional to the IMB characteristic function, confirming the derivation gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 derives the IMB distribution for w = Σ_i x_i y_i η_i, claiming via characteristic functions (Eqs. 25–27) that PDF_w differs from PDF_w' only by a scale. This is only correct when all η_i are equal: the characteristic function φ_w(ω) = ∏_i [1 + (2π σ_x σ_y η_i ω)^2]^{-1/2} reduces to the IMB form only in that limit. For the actual matched filter, η is the planet modulation template, which varies strongly with time (see Fig. 5), so φ_w is a product of Bessel kernels with different scales and is not the characteristic function of a single IMB distribution. The paper's subsequent numerical fit of ν in §4.2 (ν=36) can mask this misspecification in the bulk—the QQ R²=0.9992 is dominated by central quantiles—but the lookup table is applied at extreme tails (TN≈7.9, α≈3.6×10⁻¹⁵), far beyond the 10⁸ samples used to fit ν. No uncertainty is propagated from the ν fit or from the logistic extrapolation in Appendix D. Thus the abstract's claim that the 'true noise distribution' is an iterative convolution of Bessel functions is not actually established for the template-weighted, correlated case; the quantitative TN/Tα correction rests on an empirical approximation whose tail accuracy is unvalidated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a dual Bracewell nulling interferometer subject to Gaussian pink-noise perturbations in phase and amplitude. Starting from the Lay (2004) Taylor-expanded instrument response, it derives the differential-output photon rate, constructs a matched-filter test statistic, and provides a semi-analytical frequency-domain expression for its variance. The central statistical claim is that the second-order phase-amplitude term in Equation (3) is a product of Gaussian perturbations, so the systematic noise in the differential output cannot be Gaussian; the paper derives an 'iterative-convolved modified Bessel' (IMB) distribution for the white-noise case, fits its degrees-of-freedom parameter to correlated pink-noise simulations (finding ν=36), bootstraps the test statistic, and constructs a lookup table correcting Gaussian-significance thresholds (TN) to the true thresholds (Tα). A reference-case noise budget for an Earth-twin observation yields TN=7.90 versus Tα=8.42, i.e., a ~6% overestimate of significance if Gaussianity is assumed. The method is implemented in the public package InLIFEsim, and the analytical response is cross-validated against LIFEsim in Appendix B.3.","tokens_in":37189,"tokens_out":6509,"duration_ms":72065,"significance":"The question addressed is important: nulling interferometry performance predictions for mission concepts such as LIFE typically assume Gaussian noise, and if systematic instrumental noise is genuinely heavier-tailed, detection thresholds and yield estimates could be biased. The paper contributes a computationally efficient semi-analytical instrument model, a public codebase, and a concrete quantitative recipe (variance formula plus correction factor) that can be integrated into yield simulations. The core observation that a product of Gaussian perturbations is non-Gaussian is robust, and the historical context (Mennesson et al. 2010; Hanot et al. 2011; Bonse et al. 2023) supports the relevance of non-Gaussian noise in high-contrast instrumentation. The paper is also commendably explicit about many of its limitations in Section 5.4. However, the specific analytical identification of the matched-filter noise as IMB-distributed in the realistic template-weighted, correlated case is not established, and the quantitative tail correction rests on an unvalidated extrapolation; these issues are load-bearing for the paper's headline claims.","major_comments":[{"comment":"The characteristic-function argument proves that the unweighted sum w' = Σ_i x_i y_i is IMB-distributed, but the extension to the weighted sum w = Σ_i x_i y_i η_i is not valid for a time-varying template. The characteristic function is φ_w(ω) = ∏_i [1 + (2π σ_x σ_y η_i ω)^2]^{-1/2}, which reduces to a single IMB characteristic function only when all η_i are equal. The text after Eq. (27) states that 'as η is a constant vector, PDF_w' and PDF_w differ only up to a constant,' but the actual planet template η is strongly time-dependent (Fig. 5). The claimed identification of the matched-filter noise distribution as a single IMB distribution is therefore not proven for the realistic case.","section":"§4.1, Eqs. (25)–(27)"},{"comment":"The numerical calibration of ν=36 uses 10^8 simulated template-matched noise values, and the reported QQ R²=0.9992 is dominated by central quantiles. The resulting lookup table is applied in Table 2 at TN=7.90, corresponding to α=3.6×10^-15, far beyond the 10^8-sample range. The logistic extrapolation in Appendix D is used without any uncertainty propagation from the ν fit or from the tail extrapolation. Consequently, the quantitative TN/Tα correction factor is an empirical approximation whose tail accuracy is unvalidated, and the statement that the correction is ~6% at TN=7.9 is not supported with confidence bounds.","section":"§4.2, Fig. 3, Table 2"},{"comment":"The analytical derivation treats w = Σ_i x_i y_i η_i, a single-index sum, whereas the actual second-order systematic noise in Eq. (18) is a double sum Σ_{i,j} (∂²n/∂A_i∂φ_j) δA_i δφ_j, with the mixed Hessian in Eq. (B43) non-diagonal. The reduction of this quadratic form to the simple product-sum model of §4.1 is not demonstrated. Since the fitted ν in §4.2 is calibrated on the full correlated model, a good central fit can mask this misspecification; the tail behavior, where the lookup table is used, is precisely where such misspecification would be most severe.","section":"§4.1 and Eq. (18)"},{"comment":"The abstract claims that the paper derives 'the true noise distribution of the differential output' and that it 'follows iterative convolutions of Bessel functions.' Section 5.4 itself lists major caveats: the actual perturbation distribution is likely far more complicated than Gaussian pink noise, the lookup table is not universal and depends on the target and instrument configuration, spectral correlations are ignored, and the matched filter assumes a priori knowledge of the planet position. The conclusions should be reframed as applying to the specific Gaussian-pink-noise, single-wavelength, perfectly common-mode-rejected model, and the 'true distribution' language should be softened accordingly.","section":"Abstract and §5.4"}],"minor_comments":[{"comment":"The characteristic function of the normalized K0 density in Eq. (23) should satisfy φ_ui(0)=1, but the expression in Eq. (25) gives φ_ui(0)=π σx σy. The final density is renormalized in Eq. (28), but the notation is inconsistent and should be clarified.","section":"Eq. (25)"},{"comment":"The manuscript contains numerous unresolved cross-reference placeholders such as '?? section 2', '?? section 3', and '?? section 5'; these should be fixed before submission.","section":"Throughout"},{"comment":"The y-axis of the left panel appears garbled ('10□6' instead of a power-of-ten label), and the legend entries for the QQ plot do not clearly map to the plotted lines and R² values.","section":"Fig. 3"},{"comment":"The statement that 'a pure PDF-based comparison is not sufficient' is followed by a QQ analysis, but the QQ coefficient of determination is also dominated by central quantiles; specifying the quantile range used or adding a tail-weighted diagnostic would be more informative for the extreme-event behavior.","section":"§4.2"},{"comment":"The remainder term is denoted R_2 but the text describes it as 'the third order remainder'; the indexing should be made consistent.","section":"Eq. (2) and Appendix B.4"},{"comment":"The sum N(0,σ_N) + IMB(0,σ_IMB,ν) is not itself IMB-distributed; the text should state explicitly that the distribution of the test statistic T is obtained numerically by sampling from this sum, not by a closed-form convolution.","section":"§4.3, Eq. (30)"}],"recommendation":"major_revision","confidential_remarks":"The main reason for not recommending acceptance at this stage is the gap between the analytical IMB derivation (valid for unweighted sums of i.i.d. Gaussian products) and its application to the template-weighted, correlated matched-filter statistic. The direction of the paper is sound and the non-Gaussianity message is likely correct, but the 'true distribution' claim and the quantitative tail correction go beyond what the current derivation and the 10^8-sample fit support. I would advise the editor that the paper is promising and should be reconsidered after the authors either prove the template-weighted case or explicitly reframe the IMB fit as a heuristic approximation with uncertainty propagation. The paper's self-reported limitations in §5.4 are honest, but they should be reflected in the abstract and conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Felix – quick take: the central claim is solid. A product of Gaussian perturbations is heavy-tailed, so the differential output of a dual Bracewell cannot be Gaussian, and earlier S/N thresholds were overconfident. That point is worth taking seriously, and the paper backs it with a clean derivation for white noise, public code, and a validated instrument model.\n\nWhat's actually new: the analytic test statistic in Section 3 (frequency-domain expression for T), the Bessel-family IMB distribution as a candidate model for systematic noise, the semi-analytical correction factor TN/Tα, and the reference noise budget for LIFE. The validation against LIFEsim to <0.7% in the fundamental-noise-only case is a good check. The code and data are on Zenodo, which makes the work reproducible.\n\nWhere it's soft: the step from the white-noise IMB derivation to the correlated, template-weighted case is the load-bearing bridge, and it is asserted rather than proved. In §4.1, η is taken constant; when the planet template varies (as it does), the characteristic function is a product of Bessel kernels with different scales, not a single IMB. The numerical fit of ν=36 can absorb that misspecification in the bulk—the QQ R² is dominated by central quantiles—but the lookup table is then used at TN≈7.9, α≈3.6×10⁻¹⁵, far beyond the 10⁸ samples used to fit. No uncertainties are given for ν or for the logistic extrapolation in Appendix D. The common-mode rejection assumption (δβ₋=0) is also idealizing, as the paper's own limitations section notes.\n\nNone of this kills the core message. The non-Gaussianity is real, the direction of the correction is right, and the paper is honest about the limits. But the quantitative correction should be treated as an empirical approximation until the matched-filter case is either derived properly or validated at high quantiles (e.g., by importance sampling).\n\nWho benefits: mission yield modelers (LIFE especially), nulling interferometry designers, and anyone doing hypothesis tests on differential outputs of high-contrast instruments. It deserves a serious referee. I'd send it to review, but ask the referee to dig into §4.1 and Appendix D, and ask the authors for uncertainty on ν and a tail-validation check.","headline":"Non-Gaussian noise in dual Bracewell nullers is real; the IMB model is a useful semi-empirical fit, but the analytic proof doesn't cover the matched-filter case and the tail correction is extrapolated far beyond the fitted range.","tokens_in":37795,"tokens_out":2500,"would_cite":true,"duration_ms":28660,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual Bracewell nuller's differential output cannot have Gaussian noise: the phase–amplitude cross term is a product of Gaussians whose true distribution is an iterated Bessel convolution.","keywords":["nulling interferometry","non-Gaussian noise","instrumental noise","IMB distribution","matched filter","hypothesis testing","exoplanet detection","LIFE mission"],"falsifier":"Generate $10^8$ planet-free matched-filter outputs from the reference dual Bracewell model with Gaussian pink-noise perturbations and count realizations exceeding the Gaussian $5\\sigma$ threshold: the paper's IMB model predicts a false-positive rate above $2.9\\times10^{-7}$, so an observed rate indistinguishable from the Gaussian value would falsify the central claim.","tokens_in":36685,"feed_emoji":"🔭","tokens_out":14417,"duration_ms":129187,"temperature":0.7,"pith_summary":"This paper sets out to establish that the differential output of a dual Bracewell nulling interferometer cannot have Gaussian instrumental noise. The structural reason is that the perturbed photon rate contains a second-order phase–amplitude term, essentially $\\delta A \\cdot \\delta\\phi$, and the product of two Gaussian perturbation processes is not Gaussian. The paper derives the resulting distribution as an iterated convolution of modified Bessel functions of the second kind (the IMB distribution), fits its shape parameter for an Earth-twin reference case, and uses it to correct matched-filter hypothesis tests. If the derivation is right, Gaussian-based S/N thresholds used in nulling-interferometry mission studies set the false-positive rate too low and therefore overstate detection confidence. For the reference case the correction is small for broadband detection but reaches roughly 8% in significance for per-channel spectroscopic characterization, so ignoring it mainly distorts high-significance single-wavelength claims.","feed_headline":"Instrumental noise in planet nullers is not Gaussian","feed_subtitle":"Phase–amplitude cross terms give the noise Bessel tails, so Gaussian S/N thresholds understate false positives.","key_machinery":"The load-bearing objects are: (1) the van Cittert–Zernike instrument model of Equation (1), which writes the photon rate as Fourier transforms of the sky brightness evaluated at the array baselines; (2) the second-order Taylor expansion of Equation (3), in which identical perturbations in the two dark outputs cancel all symmetric terms and leave only the first-order phase term and the mixed phase–amplitude cross term; (3) the IMB probability density of Equation (28), obtained from the characteristic function of the $K_0$-Bessel product-of-Gaussians distribution via the convolution theorem; and (4) the matched-filter test statistic of Equation (16), evaluated in Fourier space using the pink-noise power spectral density of Equation (19). The mechanism is that a product of Gaussian perturbations is heavy-tailed, temporal correlations reduce the effective number of independent samples from $N$ to the fitted $\\nu=36$, and mixing this IMB component with the Gaussian approximation of photon noise shifts the critical values of the test statistic away from the Gaussian ones.","core_discovery":"The paper's central claim is that the null-hypothesis distribution of the matched-filter output of a dual Bracewell nulling interferometer is non-Gaussian and follows an iteratively convolved Bessel distribution. After subtraction of the two dark outputs, the Taylor-expanded differential rate of Equation (3) contains only a first-order phase term and a second-order mixed term $\\delta A^T (\\partial^2 n/\\partial A\\partial\\phi)\\delta\\phi$; this mixed term is a quadratic form in two Gaussian vectors, so the distribution of each frequency component is a modified Bessel function $K_0$, and the sum over many independent modes is the IMB distribution of Equation (28). The paper calibrates the shape parameter by drawing $10^8$ correlated time series from its instrument model and finds $\\nu=36$ for the reference setup, then bootstraps $10^{10}$ realizations of the test statistic to build a lookup table for the correction factor $T_N/T_\\alpha$. The consequence is that quoting a Gaussian S/N with $5\\sigma$ or $7\\sigma$ thresholds understates the true false-positive rate; for the reference setup a nominal $7.90\\sigma$ detection at $10\\,\\mu$m corresponds to about 6% lower significance, and the paper quantifies this with the $T_N/T_\\alpha$ correction.","pith_inferences":["Beyond the paper, if the input perturbations are non-Gaussian or common-mode rejection between the two dark outputs is imperfect, the specific Bessel form and the fitted $\\nu=36$ will not hold; the paper itself notes in Section 5.4 that the true perturbation distribution is likely more complicated, and a testbench measurement of the null distribution could act as a diagnostic of second-order pertu","The fitted degrees of freedom $\\nu$ probably equals the effective number of independent samples set by the overlap of the perturbation PSD with the planet-modulation template; if so, $\\nu$ could be predicted analytically from the quantities in Equation (15) rather than fitted from $10^8$ simulated observations.","Because spectral correlations can be used to calibrate systematic noise, the single-channel correction is likely an upper bound for broadband yield calculations; the main mission-design impact would then be on per-channel spectroscopy rather than on overall detection rates.","The same matched-filter IMB machinery may transfer to single-aperture high-contrast imaging, where speckle noise is also known to be non-Gaussian, potentially giving a common false-positive-control framework across direct-imaging architectures."],"forward_implications":["Standard Gaussian S/N thresholds in nulling interferometry set the false-positive rate too low; for the reference Earth-twin setup the nominal $7.90\\sigma$ detection in the $10\\,\\mu$m channel is about $6\\%$ less significant, equivalent to roughly $14\\%$ more integration time.","The correction factor $T_N/T_\\alpha$ collapses onto the ratio $\\sigma_{\\mathrm{IMB}}/\\sigma_N$, so a lookup table calibrated for one shape parameter $\\nu$ can be reused across different noise strengths without new Monte Carlo runs.","In the reference case the second-order phase–amplitude term dominates the first-order phase term, so the systematic noise in the differential output is set by the imaging baseline rather than the nulling baseline.","The non-Gaussian correction is negligible for broadband Earth-twin detection, where per-channel significances are low, but it matters for high-significance single-wavelength characterization, where $T_N/T_\\alpha\\approx 0.92$ at $10\\,\\mu$m.","Equation (16) evaluates a multi-day observation's significance in seconds from perturbation power spectra, making the correction cheap enough to embed in exoplanet yield simulations."],"supporting_citations":[{"why":"Supplies the perturbed instrument model, the van Cittert–Zernike formulation, and the differential-output Taylor expansion that produce Equation (3).","marker":"Lay (2004)"},{"why":"Establishes the modified Bessel $K_0$ distribution for the product of two independent Gaussian variables, the building block of the IMB derivation.","marker":"Nadarajah & Pogány (2015)"},{"why":"Provides the reference astrophysical noise models and the LIFEsim baseline used for validation and for the noise budget.","marker":"Dannert et al. (2022)"},{"why":"Supplies the hypothesis-testing framework, false-positive-rate notation, and the $T_N$-style Gaussian-equivalent significance used for the correction.","marker":"Bonse et al. (2023)"},{"why":"Explains the self-calibrating kernel-nulling property of the dual Bracewell output, which underlies the cancellation of symmetric first-order noise.","marker":"Martinache & Ireland (2018)"},{"why":"Provides empirical evidence from the Palomar fiber nuller that nulling outputs show non-Gaussian noise, motivating the distributional analysis.","marker":"Hanot et al. (2011)"},{"why":"Documents non-Gaussian null noise and self-calibration ideas that the paper builds on for interpreting the derived distribution.","marker":"Mennesson et al. (2010)"}],"fun_headline_variants":["Nulling interferometer noise has Bessel tails, not Gaussian","Non-Gaussian noise skews exoplanet detection stats","Gaussian S/N overstates exoplanet detection confidence","True noise distribution for nulling interferometers revealed","Bessel noise: Why Gaussian thresholds fail for planet hunters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the inputs being zero-mean Gaussian pink-noise perturbations with the power spectral density of Equation (19), and on the two dark outputs experiencing identical perturbations so common-mode terms cancel; Section 5.4 concedes that the real perturbation distribution is probably more complicated, and if either assumption fails the specific Bessel form with $\\nu=36$ does not apply, though heavy-tailed non-Gaussianity would likely survive.","fun_headline_variants_meta":{"raw":{"variants":["Nulling interferometer noise has Bessel tails, not Gaussian","Non-Gaussian noise skews exoplanet detection stats","Gaussian S/N overstates exoplanet detection confidence","True noise distribution for nulling interferometers revealed","Bessel noise: Why Gaussian thresholds fail for planet hunters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1500,"prompt_tokens":1058,"completion_tokens":442,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":674,"tokens_out":442,"duration_ms":4282,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:44:38.369715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate $10^8$ planet-free matched-filter outputs from the reference dual Bracewell model with Gaussian pink-noise perturbations and count realizations exceeding the Gaussian $5\\sigma$ threshold: the paper's IMB model predicts a false-positive rate above $2.9\\times10^{-7}$, so an observed rate indistinguishable from the Gaussian value would falsify the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Explains the self-calibrating kernel-nulling property of the dual Bracewell output, which underlies the cancellation of symmetric first-order noise."},{"cited_title":"2010, in Ground-based and Airborne Instrumentation for Astronomy III, ed","cited_arxiv_id":null,"evidence_quote":"Documents non-Gaussian null noise and self-calibration ideas that the paper builds on for interpreting the derived distribution."}],"review_version":1}