{"id":"af03fcb0-f1fa-440e-b901-a8f2b47ef4fb","arxiv_id":"2607.18947","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For a rough-interface scattering layer, image resolution and signal-to-noise ratio are controlled by a single blur parameter βint = zint|1−c0/c1|√D, which produces the shower curtain effect.","lead":"This math paper treats a shower curtain as a random rough surface rather than a fog-like random volume, and derives formulas for how image blur and noise depend on where the curtain sits between object and viewer. It shows that a single parameter, built from distance, speed contrast, and surface roughness, controls both resolution and stability.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix D's fourth-order moment (D1) appears invalid: the local covariance expansion is applied globally, ignoring the zero manifold y1=0, y2=y3, so Proposition 4.3 and the optical SNR claim are not established.","rationale":"The reader's weakest assumption identifies the Gaussian and local-quadratic-covariance conditions as hidden assumptions. My stress-test goes further: even with V Gaussian, the way the local expansion is used in Appendix D is invalid. The exact fourth-order moment is concentrated near the zero manifold y1=0, y2=y3=R of the phase combination, not near the single point y1=y2=y3=0. The quadratic form near that manifold is D|y1|^2+a^2D|y2-y3|^2 for typical R, giving an integrand that factorizes; the paper instead uses D|y1+a(y2-y3)|^2, which is only valid when all points are within a correlation length of the origin. This changes the delta structure and the τ-scaling of (D1). Since (D1) is the sole input for Proposition 4.3, the universal Gaussian blur and the 'averaging is ineffective' claim are not supported. The proposed numerical/analytic check on (D2) would settle the matter. If the check confirms the missing-manifold contribution, Proposition 4.2 is false as stated, and the paper's central optical SNR result is not merely underproved; it requires a genuinely different fourth-order analysis. Hence I would move the conditional verdict to reject in its current form.","tokens_in":24640,"tokens_out":27514,"duration_ms":240963,"concrete_test":"Test the distributional identity (45) for Gaussian V with a finite-correlation smooth covariance, e.g., ρ(r)=exp(-|r|^2/2), in 2D (or a 1D analogue). Use the exact integral representation (D2) and evaluate it numerically at large τ (e.g., τσV≈20) against a smooth test function in k,k',q,q' chosen so that (k'-k)-(q'-q)=0 and (˜q'-˜q)-(˜k'-˜k)=0 but a(k'-k)-(˜k'-˜k)≠0. The paper's formula predicts zero; the corrected zero-manifold asymptotic predicts a nonzero contribution with an extra δ((˜q'-˜q)-(˜k'-˜k)) and no δ(a(k'-k)-(˜k'-˜k)). If nonzero, Proposition 4.2 fails.","verdict_should_be":"REJECT","load_bearing_attack":"Proposition 4.2 / Appendix D is the load-bearing step for the optical SNR results. Starting from (D2), the proof uses the local quadratic approximation C(r)=C(0)-D|r|^2/2 to replace the characteristic function by exp(-τ^2D|y1+a(y2-y3)|^2/2), as in (D3). It then changes variables z=y1+a(y2-y3), u=y2, v=y3 and integrates u,v as if the exponential factor were independent of them, obtaining the two extra delta functions in (45). This is not valid for a covariance of finite correlation length. The phase combination S=V(0)-V(y1)-aV(y2)+aV(y3) has variance exactly zero on the entire two-dimensional manifold y1=0, y2=y3=R, for every R, not only near the origin. Near that manifold, for R outside the correlation length, the leading variance is D|y1|^2 + a^2D|y2-y3|^2, not D|y1+a(y2-y3)|^2. The missing manifold contributes an additional delta δ((˜q'-˜q)-(˜k'-˜k)) and leaves the relative momentum (˜k'-˜k) unconstrained, with a 1/τ^4 rather than 1/τ^2 prefactor. Thus (45)/(D1) over-constrains the fourth-order statistics. Since Proposition 4.3's covariance formula and the claim that frequency/position averaging cannot reduce variance are direct consequences of (45), the central 'universal Gaussian covariance' claim is unsupported. The Gaussian assumption is not the only issue; the identical flaw persists for Gaussian V.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops an asymptotic theory of passive imaging through a single random rough interface in the paraxial, high-frequency regime. The source is a localized beam; the interface is modeled as a stationary random surface with amplitude σ and correlation length l_c, parameterized by a scaling exponent γ. The authors derive the transmitted field in terms of a random scattering operator, then analyze matched-field and optical imaging. In the critical scaling γ=1/2 and the strong-scattering regime, they claim that the optical image is the ideal image blurred by a Gaussian kernel of width β_int = z_int |1−c0/c1| √D, and that the image covariance is universal and driven by a single effective Gaussian displacement, so that frequency/position averaging is ineffective. A comparison with a published USAF-chart experiment shows approximately linear growth of the blur with source–interface separation.","tokens_in":24991,"tokens_out":14938,"duration_ms":129741,"significance":"If the fourth-order claims are correct, the paper would provide a rare tractable rough-interface model for the shower-curtain effect, with explicit formulas for resolution and SNR and a single dimensionless parameter β_int/ℓ_Ψ. The second-order matched-field and optical mean-image derivations are self-contained and have useful internal checks (flat-interface limit, the three γ-regimes). The manuscript also makes a concrete, falsifiable prediction—linear scaling of the blur with z_int—and connects it to data. However, the fourth-order statistical foundation is not sound in its current form, and the central 'universal Gaussian covariance' claim depends on that foundation.","major_comments":[{"comment":"The proof of the fourth-order moment is invalid. For S = V(0) − V(y1) − aV(y2) + aV(y3), on the noncompact manifold y1 = 0, y2 = y3 = R, S ≡ 0 for every R, so the characteristic function is identically 1 there. Near this manifold with |R| large, Var(S) ≈ D|y1|^2 + a^2 D|y2−y3|^2, not D|y1 + a(y2−y3)|^2; the cross term in the latter is absent when y2,y3 are outside the correlation length of the origin. Replacing the covariance by its local quadratic expansion for all y1,y2,y3 in (D3), then integrating over u = y2, v = y3, produces the extra delta δ(a(k′−k) − (k̃′−k̃)) in (45) and suppresses a separate Gaussian factor in the second momentum transfer. This is not a Gaussianity issue: the same failure occurs for Gaussian V with finite correlation length. Since Proposition 4.3’s covariance formula, Eqs. (47)–(48), and the SNR conclusions in Eq. (49) and Discussion 4.2 are direct consequences","section":"Appendix D, eqs. (D1)–(D3) and Proposition 4.2"},{"comment":"Proposition D.1 is stated for V satisfying only Assumption 1, but the proof uses Gaussianity at the step leading to (D3) ('as the interface fluctuations V are assumed to be Gaussian'). The Gaussian assumption is introduced informally in the text of §4.1, and the introduction claims that fourth-order statistics are computed 'without requiring a Gaussian approximation.' This is internally inconsistent. If the fourth-moment result is intended only for Gaussian V, the proposition statement, the text, and the abstract-level claims of universality must be revised accordingly.","section":"Appendix D, Proposition D.1 vs §4.1"}],"minor_comments":[{"comment":"The displayed identity T = 2s0/(s0+s1) = 2c0/(c0+c1) is algebraically false: with s_j = 1/c_j, one has 2s0/(s0+s1) = 2c1/(c0+c1), as correctly used in Appendix B. Please correct the displayed equality.","section":"Eq. (14)"},{"comment":"The linear fit α = 0.1 is used without error bars or a goodness-of-fit measure. The comparison is acceptable as an illustration, but the text should not describe it as 'direct evidence' without at least a rough quantitative assessment of the fit uncertainty.","section":"Example 1.1 Revisited and Figure 5"},{"comment":"The proof is only an outline; after applying Proposition 4.2, the text says 'then we can conclude in a similar fashion' and does not display the intermediate reduction. Given that the result is central, a fuller derivation (or a reference to a complete appendix) is needed.","section":"Proof of Proposition 4.3, §4.1"},{"comment":"The curvature parameter D is defined in Remark 3.1 as −C″(0) and in Appendix D as −ρ″(0) σ_V^2/ℓ_V^2. These are consistent under the Gaussian model, but the notation should be harmonized to avoid confusion.","section":"Appendix D, notation"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid second-order core and a clear physical story, but the fourth-order calculation is the linchpin of the novelty (universal Gaussian covariance and SNR). If Appendix D can be replaced by a correct derivation, the paper may be suitable for publication; if not, the claims in §4.1 need to be substantially scaled back. The reported experiment is only weakly quantitative, but the linear-scaling prediction is a useful benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this for the βint result and the matched-field analysis; ignore the optical SNR claims until the fourth-order moment is redone.\n\nWhat's new: the rough-interface formulation of the shower curtain effect is a genuine departure from the volumetric random-medium treatments, and the explicit formulas for resolution and variance in the critical scaling γ=1/2 are useful. I particularly like the clean statement that blur and matched-field fluctuation support are controlled by βint=zint|1−c0/c1|√D, and the flat-interface and fast-fluctuation limits behave correctly. The derivations of Proposition 3.2 and the mean optical image (Proposition 4.1) are careful and self-contained; Appendix C gives a reasonable Laplace argument for the second moment.\n\nNow the soft spots, in order of importance.\n\nThe load-bearing problem is Appendix D. The fourth-order moment is computed by replacing Var(V(0)−V(y1)−aV(y2)+aV(y3)) with D|y1+a(y2−y3)|². That quadratic form is the correct local approximation at the origin, but it is applied as if it were global. In reality the phase combination is exactly zero on the manifold y1=0, y2=y3 for every R, not just near zero. Near that manifold, with R outside the correlation length, the leading variance is D|y1|²+a²D|y2−y3|², no cross term. The change of variables in the proof then misses that continuum and fabricates an additional delta. The upshot is that equation (45) is not established; Proposition 4.3 and the SNR formula (49) are unsupported. The 'single Gaussian displacement' interpretation and the claim that frequency/position averaging cannot reduce variance are direct consequences of that faulty step. The Gaussian assumption isn't the problem—the error survives for Gaussian V.\n\nMinor but real: eq. (14) states 2s0/(s0+s1)=2c0/(c0+c1), but Appendix B correctly gives 2c1/(c0+c1). Looks like a typo, but it should be fixed. Also, the experimental comparison fits α in βint=αzint rather than measuring D and contrast independently; calling that 'direct evidence' is generous.\n\nBottom line: the second-order theory is a solid contribution and the paper is worth a serious referee. The optical SNR section needs a corrected fourth-order computation or the claims should be retracted. I'd send it to review, with a referee who will actually check Appendix D.","headline":"The second-order imaging theory and the βint parameter are solid and worth reading; the optical SNR section rests on an invalid fourth-order moment in Appendix D and should not be trusted as-is.","tokens_in":25508,"tokens_out":10841,"would_cite":false,"duration_ms":94132,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q60","78A45"],"pacs":[],"model":"deepseek-v4-flash","headline":"Imaging through a rough interface is governed by a single blurring length, βint, that grows linearly with distance from the object to the interface.","keywords":["shower curtain effect","rough interface scattering","speckle statistics","paraxial wave propagation","matched field imaging","optical imaging resolution","signal-to-noise ratio","random phase screen"],"falsifier":"Take a point-like or chartered source, a rough interface with known D, and vary the object–interface distance zint while recording the image width and the covariance of images at two colors. If the blur width does not grow as zint|1−c0/c1|√D, or if the two-color covariance is not equivalent to a single Gaussian displacement (for instance, if frequency diversity decorrelates the fluctuations), the central claim fails. A targeted version uses a non-Gaussian rough surface with the same D: the mean image should still be Gaussian-blurred, but the SNR should deviate from the paper's formula.","tokens_in":24481,"feed_emoji":"🚿","tokens_out":6122,"duration_ms":56231,"temperature":0.7,"pith_summary":"This paper claims that imaging through a randomly rough interface — the setting behind the shower curtain effect — is controlled by one parameter: βint = zint |1 − c0/c1| √D. Here zint is the distance from the object to the interface, c0 and c1 are the wave speeds on either side, and D is the mean-square slope of the interface fluctuations. In the critical scaling, where the interface correlation length matches the beam width, and in the strong-scattering limit, the mean optical image becomes the ideal image blurred by a Gaussian of width βint, and the signal-to-noise ratio depends only on βint. The same parameter sets the width of the speckle fluctuations in matched-field imaging. If the paper is right, the shower curtain effect is fundamentally a propagation phenomenon: blur and image fluctuations grow linearly as the scattering layer is moved away from the object, and they vanish when the interface is flat.","feed_headline":"One blurring length governs the shower curtain effect","feed_subtitle":"The blur grows linearly with the object-to-curtain distance, so moving the curtain closer to the scene restores sharpness.","key_machinery":"The load-bearing object is the interface scattering operator Kε(τ,k,k′), which maps a transverse Fourier mode k′ to k with random phase τV(x/ε^{γ−1/2}); the value τ = ωo|s0−s1| fixes the medium contrast. The critical step is the evaluation of the fourth moment of Kε in the strong-scattering limit using the local quadratic expansion of the covariance, C(x) ≈ C(0) − D|x|²/2, for a Gaussian V. This yields a universal Gaussian factor exp(−|Δ|²/(2τ²D))/(2πτ²D) with momentum-conserving delta functions, which in turn produces the Gaussian blur kernel of width βint and the representation of all image fluctuations by one effective Gaussian displacement X. The same kernel κ(Δk,k) already controls the","core_discovery":"For a monochromatic paraxial source imaged through a single rough interface separating two homogeneous media, the paper derives a reduced observation model in which all random scattering enters through the operator Kε(τ,k,k′) = (2π)^(−2)∫ e^{i(k′−k)·x} e^{iτV(x/ε^{γ−1/2})} dx. In the critical regime γ=1/2 with strong scattering, the fourth-order statistics of this operator converge to a universal form that depends on the interface only through D, the curvature of its covariance at zero. As a consequence, the mean optical image is the ideal image convolved with a Gaussian kernel of width βint, and images at different frequencies fluctuate as if a single Gaussian random displacement were appli","pith_inferences":["The universality suggests a simple calibration strategy: measure the blur at one object distance and one contrast, extract D, and predict resolution and SNR at any other geometry without knowing the full interface spectrum.","The same fourth-moment machinery should extend to stacked rough interfaces; if the βint contributions add in quadrature, a multilayer curtain would behave like one effective rough interface, which could be tested numerically.","Because the randomness is effectively one-dimensional (a common displacement X), methods that scan frequency or detector position will not decorrelate the noise; only changing the interface realization itself—moving the curtain—can average it out.","The mean Gaussian blur may survive for non-Gaussian interface fluctuations, but the SNR formula and the conclusion about averaging should not; an experiment comparing Gaussian and engineered non-Gaussian surfaces would separate the robust geometric part from the model-dependent part."],"forward_implications":["Image blur grows linearly with object–interface distance: βint = zint |1−c0/c1| √D, so placing the scattering layer close to the object restores resolution while leaving the layer itself unchanged.","In the strong-scattering critical regime the image statistics become frequency-independent and universal; neither multifrequency averaging nor local spatial smoothing reduces the fluctuations because a single random Gaussian displacement drives them.","Optical intensity-only imaging is far more robust than matched-field imaging: random phase distortions barely affect intensity, so the dominant effect is the βint blur rather than phase noise.","In the rapidly varying interface regime the image shape is preserved but contrast is reduced by |φV(Δk)|², while in the slowly varying regime the image approaches the flat-interface result; only the critical regime produces simultaneous blur and fluctuations."],"fun_headline_variants":["Rough interface model pins down shower curtain blur","Shower curtain blur scales with object-to-curtain gap","How rough interfaces blur images: a precise law","Shower curtain effect: blur grows with distance","Image blur through rough interfaces: a unified formula"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The Gaussian blur formula and especially the SNR and covariance results assume the interface fluctuations are Gaussian and that their covariance is locally quadratic at the scattering scale; that quadratic-Gaussian condition is stated in Appendix D and is not guaranteed by the paper's general Assumption 1.","fun_headline_variants_meta":{"raw":{"variants":["Rough interface model pins down shower curtain blur","Shower curtain blur scales with object-to-curtain gap","How rough interfaces blur images: a precise law","Shower curtain effect: blur grows with distance","Image blur through rough interfaces: a unified formula"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1032,"prompt_tokens":701,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":258}},"tokens_in":445,"tokens_out":331,"duration_ms":20029,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:48:21.120014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a point-like or chartered source, a rough interface with known D, and vary the object–interface distance zint while recording the image width and the covariance of images at two colors. If the blur width does not grow as zint|1−c0/c1|√D, or if the two-color covariance is not equivalent to a single Gaussian displacement (for instance, if frequency diversity decorrelates the fluctuations), the central claim fails. A targeted version uses a non-Gaussian rough surface with the same D: the mean image should still be Gaussian-blurred, but the SNR should deviate from the paper's formula.","supporting_citations":[],"review_version":1}