{"id":"15898c52-191e-42bc-a33e-4152ded80a16","arxiv_id":"1908.09133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Fourier-mode reconstruction algorithm recovers radiative sources in 2D transport media with non-small anisotropic scattering, validated on numerical simulations for optical molecular imaging parameters.","lead":"This paper presents a numerical algorithm that reconstructs an unknown light source inside a two-dimensional absorbing and scattering medium from boundary measurements. It matters because it works for anisotropic scattering that is too strong for the diffusion approximation, which is the regime relevant to optical molecular imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No noise or regularization is tested; the algorithm's stability under measurement noise is unproven, and Proposition 1 does not bound noise in the data.","rationale":"The paper's central claim is a numerical reconstruction method. The theoretical exactness is cited to [14], but the numerical part is what this paper adds. Given that the method is intended for imaging, the lack of any noise test or regularization argument leaves the most important practical claim unsecured. The manuscript itself flags the inverse problem as ill-posed, and its own experiments are all noiseless. I agree with the reader's weakest assumption. I do not see a more internal inconsistency: the recursion and quadrature are clearly described, the forward data avoids inverse crime, and the reconstructions look quantitatively plausible. The truncation criterion is heuristic and can select suboptimal M (Experiment 1), but that is secondary to the missing noise analysis, because even a perfect M does not guarantee stability in an ill-posed problem. The verdict CONDITIONAL is appropriate: the method may work with added regularization, but as stated it is not ready for noisy data.","tokens_in":16202,"tokens_out":7500,"duration_ms":79025,"concrete_test":"Run Experiment 1 with additive Gaussian noise of relative amplitudes 1e-3, 1e-2, 1e-1 added to Imeasure, keeping M=6, S=128 and all other parameters unchanged. Compute the L2 reconstruction error (20) for each noise level. If the error does not scale mildly with noise (e.g., grows faster than the noise amplitude, or fails to decrease when noise is reduced), the algorithm lacks the stability required for the claimed reconstruction. A second check: repeat with M chosen by the proposed criterion (18) and with the true best M to see whether criterion-induced instability amplifies the noise effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the complete absence of noise analysis. The inverse source problem is explicitly stated to be ill-posed (Introduction, end of paragraph 3), yet the described algorithm contains no regularization step. All numerical experiments use noiseless, numerically generated data (Section 5), and the only error estimate provided, Proposition 1, bounds the difference between solutions for two scattering kernels; it does not address measurement noise. Steps 5 and 9 involve a Cauchy-type boundary integral and recursive Poisson solves; the latter includes differentiation of a finite-element solution to recover q (Step 10), which is a known noise amplifier. The claim that the truncation order M can be chosen so that data errors fall below the noise level is used to justify 'minimum residual' behavior, but the algorithm never actually minimizes a residual and the imaginary-part criterion (18) is shown in Experiment 1 to select M=6 when the true pseudo-error is smallest at M=3. Thus, for any realistic noisy measurement, the algorithm's output is unverified and may diverge.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a numerical algorithm for reconstructing a compactly supported source q in the stationary linear transport equation in a bounded convex two-dimensional domain, from knowledge of the angularly resolved outgoing boundary radiation. The medium is allowed to have anisotropic scattering that is neither negligible nor strong enough for the diffusion approximation. The method expands the intensity and the scattering kernel in angular Fourier modes; for kernels with finitely many angular modes it reduces the problem to a non-scattering Cauchy-type problem for the high modes plus recursive Poisson solves for the low modes, after which q is recovered from the zeroth and first Fourier modes. For the Henyey-Greenstein kernel, an M-th order truncation of the kernel is used, and a heuristic criterion based on the imaginary part of the reconstructed I0 is proposed to choose M. Two numerical experiments, using boundary data generated by an independent high-resolution forward solver and a coarser reconstruction mesh, produce quantitatively accurate reconstructions of piecewise constant sources. The exactness and stability for polynomial-type scattering kernels are stated as consequences of the authors' companion manuscript [14], which is listed as under review.","tokens_in":16346,"tokens_out":5824,"duration_ms":64336,"significance":"If the claims hold, this is a useful first numerical realization of the Bukhgeim-type Fourier reconstruction method for strongly anisotropic scattering, and the experiments are well designed: they avoid inverse crime by using a piecewise-constant upwind forward solver on a very fine mesh and a separate, coarser reconstruction mesh, and they report quantitative errors as well as computational cost. The paper also contains a clean treatment of the Hilbert transform in Lemma 1 and a detailed step-by-step algorithm that should be reproducible. However, the paper's central exactness/stability claim is not self-contained, and the absence of any noise-contaminated experiment or regularization analysis is a substantial gap for an inverse problem that the paper itself identifies as ill-posed. The reconstruction's behavior under measurement noise is the weakest part of an otherwise competent numerical study.","major_comments":[{"comment":"The exactness and stability assertion for scattering kernels of finite Fourier content is the load-bearing theoretical foundation of the paper, but it is not proved here: the manuscript cites only [14, Corollary 6.1], a companion paper listed as under review. The present paper cannot be fully evaluated without knowing the regularity conditions, the precise sense of convergence, and the stability estimate behind that corollary. Please state the relevant theorem with its hypotheses, include a proof or a detailed proof sketch, and explain how the discretized Steps 5-10 are intended to converge to that infinite-dimensional result.","section":"Abstract; Section 3; [14, Corollary 6.1]"},{"comment":"The inverse source problem is explicitly described as ill-posed in the Introduction, yet the algorithm contains no regularization step and all experiments in Section 5 use noiseless, numerically generated data. Proposition 1 bounds the data error caused by kernel truncation, not the effect of measurement noise, and the 'minimum residual' interpretation in Section 3 is not realized as an actual minimization. Steps 5, 9, and 10 involve a Cauchy-type boundary integral, recursive Poisson solves, and differentiation of a finite-element function to recover q, all of which can amplify high-frequency noise. Please add a stability estimate with respect to data noise, or at minimum numerical experiments with realistic noise levels, and state clearly whether and how the choices of M, S, and the mesh act as regularization.","section":"Section 3; Section 5; Steps 5, 9, 10"},{"comment":"The proposed optimality criterion (18) is not validated by the paper's own experiment. In Experiment 1, Figure 6 shows that the pseudo-error (20) is minimized at M = 3 while the imaginary-part criterion (18) is minimized at M = 6; the text even says 'Both errors take minimum around M = 5' before giving those conflicting values. Since the criterion is presented as a practical, source-independent substitute for the true error, this mismatch needs explanation and further validation, including under noise, before the claim that the algorithm includes a feasible optimality criterion is supported.","section":"Section 4.4; Experiment 1; Figure 6"},{"comment":"The continuous derivation argues by induction that I_m^{(M)} in H^1(D) because the right-hand side of (13a) lies in H^{-1}(D). In the discrete realization, however, each I_{m+1} and I_{m+2} is a P1 finite-element function whose derivative is piecewise constant and discontinuous across triangle edges, and the variational form in Step 9 evaluates the right-hand side cellwise. No convergence or stability analysis is given for this recursive finite-element scheme as m decreases, nor for the boundary interpolation in Step 8. This is directly relevant to whether the exactness claim for polynomial-type kernels applies to the algorithm as implemented; please add an error analysis or at least a systematic mesh-refinement study.","section":"Section 3; Eq. (13); Step 9"}],"minor_comments":[{"comment":"There is a typo in Lemma 1: 'defied' should be 'defined'.","section":"Section 4.1"},{"comment":"The sentence 'Both errors take minimum around M = 5' contradicts the following sentence and Figure 6, which state minima at M = 6 for the imaginary part and M = 3 for the pseudo-error; please correct the text.","section":"Section 5, Experiment 1"},{"comment":"Please state explicitly the number K of boundary measurement points in Experiment 1 (Experiment 2 gives K = 3000 and N = 360, but Experiment 1 only mentions 360 velocity intervals and 100 sampling points for the integral transforms).","section":"Section 5, Experiment 1"},{"comment":"Reference [13] is described only as 'accepted' with no year or venue, and [14] is under review; please update these entries or provide fuller bibliographic information so that the numerical method and the theoretical foundation can be located.","section":"References [13], [14]"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent numerical companion to an under-review theory paper, and the noise-free experiments are convincing as feasibility tests. The main risks are that the exactness/stability claim is not self-contained and that no noisy-data experiment or regularization analysis is provided for an ill-posed problem. If the companion paper [14] is not available to the reader, the central claim cannot be checked; I recommend coordinating the review or requiring the authors to include a self-contained statement of the relevant theorem. The noise issue should be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is apparently the first working numerical realization of the Bukhgeim/attenuated-Radon-transform style Fourier-mode reconstruction for inverse source problems in a scattering, non-diffusive medium. The numerical algorithm is new at the level of discrete details, and the two experiments are carefully set up to avoid inverse crime. The main caveats are that the exactness/stability theory is borrowed from an unpublished companion paper [14], and measurement noise is never introduced or analyzed. Both caveats are real, but they don't erase the contribution.\n\nWhat is actually new: the paper shows how to compute the integrating factor's Fourier modes via the Hilbert transform with the singularity removed, how to discretize the Cauchy-type boundary integral (17) with a stable quadrature, how to recursively solve the Poisson problems with P1 finite elements, and how to choose the truncation order M via the imaginary-part residual (18). These choices are non-obvious and are described concretely. The numerical experiments generate forward data with a different, much finer upwind scheme and a mesh about three orders of magnitude finer than the reconstruction mesh, with no knowledge of the source support or the absorption profile. That is a clean setup. The reconstructed images in Figures 4 and 9 are quantitatively accurate in regions where the source is constant, with expected edge artifacts at discontinuities. This is real evidence that the pipeline works as intended.\n\nSoft spots, in proportion. (1) The central claim—exact and stable reconstruction for scattering kernels of finite Fourier content—is cited to [14, Corollary 6.1], an unreviewed manuscript by the same authors. Section 3 outlines the recursion, but the functional-analytic convergence and stability are not proved here. A referee cannot verify the foundation without going to that manuscript. (2) Measurement noise is absent. The inverse problem is ill-posed, and Step 10 differentiates the P1 solution to recover q, so noise will be amplified. Proposition 1 bounds the data error caused by truncating the scattering kernel, not the noise sensitivity of the reconstruction. The phrase \"minimum residual\" in the introduction overstates things: the algorithm never actually minimizes a residual defined in terms of noise; (18) is a heuristic that is not guaranteed to pick the best M, and in Experiment 1 it picks M=6 while the true pseudo-error is smallest at M=3. (3) There is no convergence or error analysis for the discretized algorithm itself—how quadrature and FE errors propagate. That is common in numerical papers, but worth flagging.\n\nWho this is for: researchers in inverse transport who want a concrete starting point for the Fourier-mode method, and a benchmark for future regularized or noise-robust versions. I would send it to peer review, with the request that the companion theory in [14] be made accessible and that at least one noisy-data experiment or a stability argument be added. The paper is not complete as numerical analysis, but it is a solid and honest engineering contribution. I would engage with it and would cite it in related work once the companion theory is public.","headline":"A concrete numerical realization of the Fourier-mode inverse transport method, worth refereeing despite missing noise analysis and a load-bearing citation to an unpublished companion paper.","tokens_in":16875,"tokens_out":5046,"would_cite":true,"duration_ms":54777,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","30E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Fourier-mode splitting reduces source reconstruction in an absorbing, anisotropic scattering medium to a non-scattering inversion plus Poisson solves, and is exact when the scattering kernel has finite angular Fourier content.","keywords":["transport equation","inverse problems","numerical source reconstruction","attenuated X-ray transform","attenuated Radon transform","A-analytic maps","Hilbert transform","optical molecular imaging"],"falsifier":"Run the algorithm on boundary data generated from a known source with controlled Gaussian noise added at levels $\\delta = 10^{-2}, 10^{-3}, 10^{-4}$; if the reconstruction error does not stay bounded as the mesh is refined for a fixed $\\delta$, or if no truncation order $M$ yields an acceptable source, the exact-data premise is falsified. A simpler check is to compare the predicted kernel-truncation data error $g^{2M+2}/(1-g^2)$ with the actual measured boundary error for a range of $g$ and $M$ in the paper's settings.","tokens_in":15983,"feed_emoji":"💡","tokens_out":8571,"duration_ms":79297,"temperature":0.7,"pith_summary":"The paper presents a numerical algorithm that recovers the internal radiative source $q$ from outgoing boundary measurements in a two-dimensional absorbing and anisotropic scattering medium, in the regime where scattering is neither negligible nor strong enough for the diffusion approximation. The central idea is to split the scattering kernel at a finite angular Fourier order $M$, which separates the problem into a non-scattering tail inversion and a finite chain of elliptic boundary value problems. For scattering kernels with genuinely finite Fourier content, the paper argues the reconstruction is exact and stable. For general kernels such as the Henyey-Greenstein kernel, the truncation error in the boundary data can be made smaller than the noise level, and the numerical experiments reconstruct piecewise constant sources quantitatively. This matters for optical molecular imaging, where realistic scattering is significant but not diffusive.","feed_headline":"Reconstructs internal sources through scattering media","feed_subtitle":"For finite angular Fourier content, boundary outflow determines the internal source exactly and stably.","key_machinery":"The machinery is the Fourier decomposition of the transport solution and scattering kernel in the angular variable, which converts the transport equation into an infinite elliptic system. When the scattering kernel has only $M$ nonzero angular Fourier modes, modes above $M$ satisfy the scattering-free system, which is solved by the Cauchy-type integral formula that recovers interior values from boundary values after conjugation by the integrating factor $e^{-h[\\mu_t]}$, where $h[\\mu_t]$ is the attenuated X-ray phase defined in the paper and chosen so that all negative Fourier modes vanish. The lower modes are then recovered recursively by solving Dirichlet problems for the Poisson equation, and the source $q$ is read off from the zero-mode equation. The algorithm also includes a locally optimal truncation criterion: choose $M$ to minimize the imaginary part of the reconstructed $I_0$, an indicator computable without knowing $q$.","core_discovery":"The central claim is that any finite Fourier content in the angular variable splits the inverse transport problem into a non-scattering part and a boundary value problem for a finite elliptic system, and that this split can be implemented numerically. For scattering kernels of polynomial type in the angular variable, the algorithm is claimed to recover the source exactly and stably, with no smallness assumption on the scattering modes. For general kernels, the paper gives an error estimate showing that the mismatch between the measured outflow and the outflow produced by the $M$-truncated kernel is bounded by a constant times the $L^2$ difference of the kernels; choosing $M$ so that this falls below the noise level yields a minimum-residual reconstruction. The numerical experiments with strongly scattering parameters (mean free path $1/5$, anisotropy parameter $g=1/2$) reconstruct piecewise constant sources with quantitatively correct values and support, using a truncation order selected by minimizing the imaginary part of the zero Fourier mode, a criterion that does not need the unknown source.","pith_inferences":["Editorial inference: the same angular-Fourier splitting should extend to three-dimensional transport with axial symmetry, replacing the unit circle by the sphere and the Cauchy formula by a spherical harmonics analogue; this is a testable extension the paper does not address.","Editorial inference: the minimum-imaginary-part criterion is a general model-selection heuristic: whenever the true quantity is real, a reconstruction parameter can be tuned by minimizing the spurious imaginary component, independent of ground truth.","Editorial inference: because the experiments use noiseless simulated data and the error estimate covers only kernel truncation, a natural next test is to add measurement noise and see how the optimal $M$ and the reconstruction error degrade; the paper's method as stated does not include regularization for that case.","Editorial inference: the recursive Poisson solves propagate information from high to low Fourier modes, so errors in the tail inversion should appear most strongly in the finer spatial features of $q$; comparing reconstructions at different $S$ and mesh resolutions would reveal this error propagation."],"forward_implications":["For any scattering kernel with finitely many nonzero angular Fourier modes, the algorithm recovers the source exactly from noiseless boundary data, without any smallness assumption on the scattering strength.","For the two-dimensional Henyey-Greenstein kernel, the error estimate yields explicit decay $g^{2M+2}/(1-g^2)$; selecting $M$ so this is below the noise level makes the reconstructed source's boundary data indistinguishable from the exact data within that noise.","The truncation criterion based on the imaginary part of $I_0$ selects a reasonable $M$ in experiments and tracks the actual reconstruction error, and it requires no knowledge of the unknown source.","The method operates in a strongly scattering regime (average of ten scatterings across the unit disc) where the diffusion approximation fails, and still produces quantitative reconstructions of piecewise constant sources.","Reconstructions in the experiments run in minutes on a multicore workstation, indicating practical feasibility for optical molecular imaging parameters."],"supporting_citations":[{"why":"It supplies the theoretical basis of the reconstruction: exact and stable recovery for scattering kernels of finite Fourier content.","marker":"[14]"},{"why":"It introduces the A-analytic function framework and the Cauchy-type integral formula for the tail system that the algorithm evaluates numerically.","marker":"[7]"},{"why":"It adapts the A-analytic method to the absorbing, non-scattering case, giving the pattern for reducing the scattering-free system to the non-scattering system.","marker":"[2]"},{"why":"It provides the explicit integrating factor $h[\\mu_t]$ with vanishing negative Fourier modes, which the algorithm uses to convert measured modes into tail modes.","marker":"[12]"},{"why":"It supplies the convolution transformation between the $I$ and $J$ mode sequences and the range characterization used in the reconstruction steps.","marker":"[32]"},{"why":"It extends the range and series analysis to strictly convex sets, supporting the convergence of the Fourier series the method relies on.","marker":"[33]"},{"why":"It gives the well-posedness and stability estimates for the forward transport problem that justify the $L^2$ error estimate for kernel truncation.","marker":"[35]"},{"why":"It provides the piecewise-constant upwind forward solver used to generate the boundary data in the numerical experiments.","marker":"[13]"}],"fun_headline_variants":["Exact source reconstruction in scattering media from boundary data","Finite angular modes make source reconstruction exact and stable","2D inverse transport: exact source recovery for finite angular modes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction is demonstrated only on noiseless, numerically generated boundary data, and the paper provides no regularization or analysis of how the discretized Cauchy integral and recursive Poisson solves behave under measurement noise, even though the inverse problem is ill-posed.","fun_headline_variants_meta":{"raw":{"variants":["Exact source reconstruction in scattering media from boundary data","Finite angular modes make source reconstruction exact and stable","2D inverse transport: exact source recovery for finite angular modes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001039,"raw_usage":{"total_tokens":4312,"prompt_tokens":823,"completion_tokens":3489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":3437}},"tokens_in":439,"tokens_out":3489,"duration_ms":22552,"temperature":1.0,"reasoning_tokens":3437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:20:04.889401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on boundary data generated from a known source with controlled Gaussian noise added at levels $\\delta = 10^{-2}, 10^{-3}, 10^{-4}$; if the reconstruction error does not stay bounded as the mesh is refined for a fixed $\\delta$, or if no truncation order $M$ yields an acceptable source, the exact-data premise is falsified. A simpler check is to compare the predicted kernel-truncation data error $g^{2M+2}/(1-g^2)$ with the actual measured boundary error for a range of $g$ and $M$ in the paper's settings.","supporting_citations":[{"cited_title":"Fujiwara, K","cited_arxiv_id":null,"evidence_quote":"It supplies the theoretical basis of the reconstruction: exact and stable recovery for scattering kernels of finite Fourier content."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces the A-analytic function framework and the Cauchy-type integral formula for the tail system that the algorithm evaluates numerically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It adapts the A-analytic method to the absorbing, non-scattering case, giving the pattern for reducing the scattering-free system to the non-scattering system."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the explicit integrating factor $h[\\mu_t]$ with vanishing negative Fourier modes, which the algorithm uses to convert measured modes into tail modes."},{"cited_title":"Sadiq and A","cited_arxiv_id":null,"evidence_quote":"It supplies the convolution transformation between the $I$ and $J$ mode sequences and the range characterization used in the reconstruction steps."},{"cited_title":"Sadiq and A","cited_arxiv_id":null,"evidence_quote":"It extends the range and series analysis to strictly convex sets, supporting the convergence of the Fourier series the method relies on."},{"cited_title":"Stefanov and G","cited_arxiv_id":null,"evidence_quote":"It gives the well-posedness and stability estimates for the forward transport problem that justify the $L^2$ error estimate for kernel truncation."},{"cited_title":"Fujiwara, Piecewise constant upwind approximations to the stationary radiative transport equation","cited_arxiv_id":null,"evidence_quote":"It provides the piecewise-constant upwind forward solver used to generate the boundary data in the numerical experiments."}],"review_version":1}