{"id":"1a1e0e98-53f1-45ac-8274-97095592692b","arxiv_id":"2608.04829","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Joint 3D structured-Hankel completion (ALOHA) enriches sub-sampled multi-frequency far-field data, after which Fourier inversion reconstructs sparse electromagnetic sources with PSNR gains of 8-18 dB over a DCT-based l1-CS baseline.","lead":"This paper proposes a two-stage method to reconstruct electromagnetic current sources from sparse far-field measurements: first fill in missing frequency data using low-rank Hankel matrix completion, then apply Fourier inversion. If it works, it would make source imaging feasible when only 30-50% of the Nyquist-rate measurement grid is available.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth leak: the DC Fourier coefficient f0 is taken from the exact source (Sec. 4.1), so the experiments test ALOHA plus an oracle, not resolution of non-uniqueness from sparse far-field data.","rationale":"I focused on the exact-f0 leak rather than the reader's stated weakest assumption (manual rank selection) because the abstract's strongest claim is about resolving non-uniqueness, and the oracle DC value directly bypasses a null-space component that the method claims to recover. This is a controllable experimental choice, not an intrinsic algorithmic limitation, so a conditional verdict remains appropriate; if the proposed check shows a large performance drop, the headline must be revised or the experiments rerun. The same-model forward data and source-specific ranks noted in Section 5.6 are additional external-validity concerns, but they do not independently falsify the algorithm. The reader's rationale explicitly flags the f0 leak even though the structured weakest-assumption field does not, hence partial agreement. My concern reinforces the conditional verdict rather than moving it.","tokens_in":16701,"tokens_out":8803,"duration_ms":106947,"concrete_test":"Recompute the J2 Monte Carlo experiments (30%, 40%, 50% sampling; noise-free and 10 dB AWGN) with f0 not taken from the true source: either set f0 = 0 or estimate it from the sampled modes using the same low-rank Hankel model, keeping all ALOHA and l1-CS parameters fixed. If the paired PSNR advantage of Joint 3D ALOHA over l1-CS falls by more than roughly 3 dB (or SSIM by about 0.05) at 30% sampling, the reported gains depend on the ground-truth leak and the non-uniqueness claim fails as stated; if performance is essentially unchanged, the leak is not the load-bearing issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that the magnetic far-field data comprise 9260 Fourier modes and that 'f0 is recovered from the mean projection of the exact current onto the polarization direction.' In the reconstruction formula (7), f0 is the DC coefficient of the transverse-electric component; it is not part of the random mask (Sec. 4.2 says the origin is retained), so ALOHA never completes it from sparse measurements. The abstract's second claim, that the framework 'resolves non-uniqueness challenges', is therefore not tested by Tables 1-3: the missing modes that create non-radiating null-space components are exactly the coefficients the completion must recover, and the most global of them, the spatial mean, is inserted from ground truth. For J1 the leak is trivial because f=0, but the large-gain J2 case is where f0 is nontrivial. Section 5.6 candidly lists same-model data generation and manual rank selection as limitations, but it does not list this oracle. The central claim would require rerunning with f0 estimated from the observed data or omitted entirely.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage strategy for reconstructing a compactly supported electromagnetic current source from sparse multi-frequency magnetic far-field data. In stage one, the missing Fourier coefficients of the transverse electric and magnetic components are completed by solving a low-rank Hankel matrix completion problem using the ALOHA algorithm; stage two applies a Fourier inversion formula to synthesize the current density. The method is tested on two synthetic 3D source models (J1 and J2) at 30%, 40%, and 50% sampling rates, in noise-free and 10 dB AWGN conditions, and compared with zero-filled reconstruction and a complex DCT-based l1-CS baseline. The reported results show large paired PSNR and SSIM gains for Joint 3D ALOHA, with all paired comparisons significant after Holm correction. The paper also includes runtime measurements and a limitations discussion.","tokens_in":1685,"tokens_out":1856,"duration_ms":57757,"significance":"If the claims were fully supported, the paper would provide a useful computational tool for inverse source problems with sparse far-field data, extending low-rank Hankel completion to vector electromagnetic sources. The manuscript has genuine strengths: the statistical protocol is unusually careful (30 Monte Carlo trials, nested masks, paired Wilcoxon tests with Holm correction, bootstrap confidence intervals, and effect sizes), the runtime benchmark is clearly delimited, the limitations section is candid, and the proposed joint completion of the three vector-current components is a sensible use of shared support. However, the experimental validation has a load-bearing oracle: the zero-frequency Fourier coefficient f0 is taken from the exact source (Section 4.1), and the measurement procedure never requires the completion method to recover it. Since f0 is precisely the type of global non-radiating null-space component the abstract claims to resolve, the headline 'resolves non-uniqueness challenges' is not tested by Tables 1-3. The same-model data generation and per-source rank selection further weaken the external validity of the reported gains.","major_comments":[{"comment":"The zero-frequency coefficient f0 is taken from the exact source, not reconstructed from the sparse measurements. Section 4.1 states that 'f0 is recovered from the mean projection of the exact current onto the polarization direction,' and Section 4.2 says 'the origin is retained in every reconstruction' so the random mask never removes f0. Consequently, the completion algorithm is never asked to recover the most global non-radiating component, and the abstract's claim that the framework 'resolves non-uniqueness challenges' is not supported by Tables 1-3. Please rerun the experiments with f0 estimated from the observed far-field data (e.g., via Eq. (9) or a data-driven estimate) or with f0 set to zero, and report both variants; alternatively, explicitly weaken the non-uniqueness claim to 'non-zero-frequency missing modes'.","section":"Section 4.1 and Section 4.2"},{"comment":"The same Fourier-Maxwell forward model is used both to generate the synthetic far-field data and to define the reconstruction formula, which is an inverse-crime setup. The paper acknowledges this in Section 5.6 ('both use the same Fourier-Maxwell forward model employed to construct the reconstruction data'), but the quantitative PSNR/SSIM gains are still likely to overstate performance under model mismatch. Please add at least one validation with a different forward solver (e.g., a boundary-element or finite-difference Maxwell solve) or with perturbed model parameters (wavenumber, polarization, domain size) to demonstrate that the enrichment is not an artifact of using the same spectral model for both data generation and inversion.","section":"Sections 4.1 and 5.6"},{"comment":"The ALOHA rank is fixed manually and source-specifically: ranks 22 for J1 and 30 for J2. Although the authors correctly note that the ranks are not changed across sampling rates, noise conditions, or trials, choosing a different rank for each source encodes prior knowledge of that source's complexity. Since the FRI rank is not estimated from the data, the claim that the method automatically exploits a finite rate of innovations is overstated. Please add a rank-sensitivity study (e.g., PSNR/SSIM versus rank for both sources) or an automatic rank-selection procedure, and discuss how the results would change if the rank were misspecified.","section":"Section 4.3.3"},{"comment":"The statistical comparison is internally sound, but the reference for PSNR/SSIM is the fully sampled Fourier reconstruction, which is generated with the same forward model used for data synthesis. This is appropriate for isolating completion error, but it does not measure fidelity to an independent ground truth. Please clarify in the text that the reported absolute PSNR/SSIM values are relative to a self-consistent numerical reference and cannot be directly compared with experimental imaging benchmarks.","section":"Section 5.3 and Tables 2-3"}],"minor_comments":[{"comment":"The notation in Eq. (12) is inconsistent: the vector es is declared as belonging to real 2M-dimensional space, but the Fourier coefficients f_m and g_m are complex-valued; the correct space is complex 2M-dimensional space.","section":"Section 3.3, Eq. (12)"},{"comment":"There is a missing parenthesis in the sentence about the rank of H_p(es), and the phrase 'if p(< M) is chosen larger' is awkward; please rewrite as 'if p < M is chosen larger than the minimum filter size'.","section":"Section 3.3, Eq. (13) area"},{"comment":"The phrase 'The magnetic far-field data comprise of 9260 Fourier modes' should read 'comprise 9260 Fourier modes'.","section":"Section 4.1"},{"comment":"The limitations paragraph would be strengthened by an explicit statement that no experimental or independently simulated far-field data were used, and by a note on whether the code and data will be made available for reproducibility.","section":"Section 5.6"},{"comment":"Some figure captions contain hyphenated phrases such as 'electromagnetic-sources' and 'Sub-sampled' in tables; please standardize the hyphenation and use 'subsampled' consistently.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent application of existing low-rank Hankel completion technology to a vector electromagnetic inverse source problem, with unusually thorough statistical reporting. The main obstacle is the oracle use of f0 in Section 4.1, which directly undermines the paper's most prominent claim ('resolves non-uniqueness challenges'). The authors should be asked to either remove the oracle and rerun the experiments or substantially weaken the claim. I would also encourage the editor to require a rank-sensitivity analysis, since the manual per-source rank selection is a second source of information leakage. If these experiments cannot be rerun, the paper would be better framed as a proof-of-concept with clearly stated limitations rather than as a validated resolution of non-uniqueness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible extension of ALOHA to 3D vector electromagnetic source reconstruction, with a careful Monte Carlo study, but the evaluation leaks the zero-frequency Fourier mode from ground truth, so the paper's claim to 'resolve non-uniqueness' is not actually tested. The method itself may be fine; the experiments just don't support that particular conclusion.\n\nWhat's new: the authors take the annihilating-filter low-rank Hankel completion framework, previously applied to acoustic sources and image inpainting, and adapt it to jointly complete three vector components of a compact electromagnetic current source using the Helmholtz-Hodge decomposition. That joint 3D structured-Hankel completion is a genuine new application. The numerical study is thorough: 30 Monte Carlo trials, paired statistics, Holm-corrected Wilcoxon tests, effect sizes, bootstrap CIs, full-volume visual comparisons. They report large PSNR gains over a DCT-based l1-CS baseline, and the qualitative figures show meaningful artifact reduction. The runtime comparison is also fair, with the caveat that it is implementation-specific.\n\nThe soft spots are real. Section 4.1 states that f0—the DC coefficient of the transverse electric component—is recovered from the mean projection of the exact current. Since the origin is retained in the mask, ALOHA never completes that mode; it is handed to the reconstruction. For J1 that is harmless because f=0, but for J2, where the big gains occur, f0 is non-trivial. So the abstract's claim that the framework 'resolves non-uniqueness challenges' is not supported: the null-space component that creates the non-uniqueness is the one the experiment supplies from ground truth. Section 5.6 candidly lists same-model data generation and manual rank choice as limitations, but it does not mention the f0 oracle. The rank is set per source (22 for J1, 30 for J2), which encodes some prior knowledge of the signal complexity. That makes the results a proof of concept, not a black-box method.\n\nI would not call this a fatal flaw in the underlying idea. The completion of the other modes is not circular, and the method may well work without the leak. But the experiments as designed do not deliver the abstract's promise. The paper deserves peer review, but the referee should insist on three things: estimate f0 from observed data or exclude it, test with a genuine Maxwell forward solver rather than the same Fourier model used for inversion, and add an automatic rank-selection baseline or sweep.\n\nThis is a paper for the inverse-problems community—readers interested in FRI-based data completion for EM source imaging will get value from it. It is a solid incremental contribution with an overclaimed headline. I'd recommend accepting it for review with major revision required.","headline":"Joint 3D ALOHA for EM source reconstruction is a legitimate new application, but the headline claim about resolving non-uniqueness is not supported by the current experiments because the DC mode is taken from ground truth.","tokens_in":17459,"tokens_out":2210,"would_cite":false,"duration_ms":22575,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","78A46"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that missing multi-frequency far-field data for compact, geometrically sparse electromagnetic sources can be completed by low-rank Hankel matrix completion before Fourier inversion, eliminating under-sampling artifacts…","keywords":["electromagnetic inverse source problem","finite rate of innovations","Hankel matrix completion","ALOHA","multi-frequency far-field data","low-rank matrix completion","Fourier inversion"],"falsifier":"Apply the same Joint 3D ALOHA pipeline to a source deliberately containing more than 30 well-separated geometric innovations (so the true FRI rank exceeds the fixed rank 30) at 30% sampling in noise-free conditions; if the PSNR advantage over $\\ell_1$-CS shrinks toward zero or reverses, the claim that low-rank completion resolves non-uniqueness is limited to sources whose rank is known and small.","tokens_in":16528,"feed_emoji":"📡","tokens_out":7766,"duration_ms":73448,"temperature":0.7,"pith_summary":"The paper tries to establish that the non-uniqueness caused by under-sampled multi-frequency far-field measurements can be tamed by first completing the missing Fourier spectrum, rather than by regularizing the inversion directly. It claims that compactly supported, geometrically sparse current sources have a finite rate of innovations, which makes a wrap-around block Hankel matrix built from the measured spectral coefficients low-rank. Recovering the missing coefficients by low-rank matrix completion and then applying Fourier inversion yields artifact-free reconstructions at 30% to 50% sampling and 10 dB noise, outperforming a DCT-based $\\ell^1$-CS baseline by 8 to 18 dB PSNR in paired Monte Carlo trials. A sympathetic reader would care because it offers a practical route to sparse-sensor electromagnetic imaging in biomedical, non-destructive testing, and antenna synthesis applications.","feed_headline":"Missing data completed first: 8-18 dB gain over compressed sensing","feed_subtitle":"Filling missing spectrum with low-rank Hankel matrices suppresses under-sampling artifacts in source imaging.","key_machinery":"The load-bearing object is the block-diagonal, wrap-around Hankel matrix $\\mathbf{H}_p(\\mathbf{s})$ built from the concatenated transverse electric and magnetic Fourier coefficients $f_\\ell$ and $g_\\ell$, with $\\mathbf{s}=[f^T,g^T]^T$. Its low rank follows from the finite-rate-of-innovations property: a compactly supported sparse source admits an annihilating filter, and the rank-deficiency of the wrap-around Hankel matrix is what turns missing-data recovery into a nuclear-norm minimization problem. The minimization is solved by ALOHA, the annihilating-filter low-rank Hankel matrix completion algorithm, via ADMM on a factorized representation, and the completed spectrum is fed into the Fourier inversion formula to reconstruct the current density.","core_discovery":"The central claim is that the inverse problem of recovering a current source density from sparse multi-frequency far-field data becomes tractable if the missing spectrum is treated as a finite-rate-of-innovations signal and completed by low-rank Hankel matrix completion before Fourier inversion. The paper derives Fourier formulas linking electric and magnetic far-field data to the transverse electric and magnetic scalar components of the source, constructs a block-diagonal wrap-around Hankel matrix from those spectral components, and argues its low rank follows from the finite rate of innovations. Joint 3D ALOHA completes the missing coefficients, and the enriched Nyquist-sampled spectrum is then fed into the Fourier inversion formula. Numerical experiments on two 3D Maxwell source models with 30% to 50% conjugate-paired sampling and 10 dB AWGN show paired PSNR gains of 8.02 to 18.18 dB and SSIM gains of 0.1842 to 0.3451 over the $\\ell^1$-CS baseline, with every trial won by the proposed method.","pith_inferences":["Editorial inference: replacing the manually fixed ranks (22 for J1, 30 for J2) with an automatic rank estimator, such as singular-value thresholding or cross-validation on the observed coefficients, should extend the method's applicability to sources whose FRI order is not known in advance.","Editorial inference: because the enrichment is a preprocessing stage rather than a new inversion operator, it should compose with other reconstruction schemes; a natural test is to run the same completed spectrum through TV-regularized or iterative refinement solvers and check whether the 8-18 dB gain persists.","Editorial inference: the same wrap-around Hankel structure should transfer to other inverse problems in which compact sparse sources are observed through Fourier-type data, such as acoustic source reconstruction or MEG/EEG source localization; the paper itself only validates the Fourier-Maxwell forward model, so this remains an open test."],"forward_implications":["At 30% to 50% sampling, the completed spectrum yields stable reconstructions where direct zero filling and DCT-based $\\ell_1$-CS leave heavy ringing and speckle.","The paired gains over $\\ell_1$-CS range from 8.02 to 18.18 dB PSNR and 0.1842 to 0.3451 SSIM across all twelve tested source-noise-sampling conditions.","The reconstruction preserves the full three-dimensional source topology, not just central slices, as shown by consistent isosurface thresholds.","The method handles 10 dB complex Gaussian noise by relaxing the data-consistency weight, and it runs about 1.74 to 2.51 times faster than the implemented $\\ell_1$-CS solver.","Accurate recovery of the zero-frequency coefficient, the mode most distorted by subsampling, is a direct corollary of completing the spectrum before inversion."],"supporting_citations":[{"why":"establishes the finite rate of innovations concept that underlies the low-rank Hankel property.","marker":"[25]"},{"why":"supplies the annihilating-filter low-rank Hankel completion method (ALOHA) and the rank-deficiency theorem.","marker":"[28]"},{"why":"provides the structured Hankel matrix model for missing-data recovery that the wrap-around construction extends.","marker":"[29]"},{"why":"extends the low-rank Hankel framework to joint multidimensional completion, the basis of Joint 3D ALOHA.","marker":"[30]"},{"why":"gives the dense-data Fourier inversion formulas that map far-field data to the source coefficients.","marker":"[16]"},{"why":"documents non-uniqueness in electromagnetic inverse source problems, the failure mode the enrichment addresses.","marker":"[17]"},{"why":"provides the TE/TM decomposition of electromagnetic sources that separates the spectral components into the Hankel blocks.","marker":"[27]"}],"fun_headline_variants":["Hankel completion fixes under-sampled far-field data","Low-rank Hankel beats compressed sensing in source imaging","Enrich sparse far-field data with finite-rate innovation","Matrix completion lifts EM source reconstruction","8-18 dB gain: low-rank Hankel over l1-CS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the source is compactly supported and geometrically sparse enough that the wrap-around Hankel matrix has a known low rank; the paper fixes the ranks at 22 and 30 manually, so if the true FRI rank is unknown or larger, completion degrades.","fun_headline_variants_meta":{"raw":{"variants":["Hankel completion fixes under-sampled far-field data","Low-rank Hankel beats compressed sensing in source imaging","Enrich sparse far-field data with finite-rate innovation","Matrix completion lifts EM source reconstruction","8-18 dB gain: low-rank Hankel over l1-CS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1501,"prompt_tokens":1023,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":399}},"tokens_in":639,"tokens_out":478,"duration_ms":5198,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:31:54.085999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same Joint 3D ALOHA pipeline to a source deliberately containing more than 30 well-separated geometric innovations (so the true FRI rank exceeds the fixed rank 30) at 30% sampling in noise-free conditions; if the PSNR advantage over $\\ell_1$-CS shrinks toward zero or reverses, the claim that low-rank completion resolves non-uniqueness is limited to sources whose rank is known and small.","supporting_citations":[{"cited_title":"Sampling signals with finite rate of innovation,","cited_arxiv_id":null,"evidence_quote":"establishes the finite rate of innovations concept that underlies the low-rank Hankel property."},{"cited_title":"Compressive sampling using annihilating filter-based low-rank interpolation,","cited_arxiv_id":null,"evidence_quote":"supplies the annihilating-filter low-rank Hankel completion method (ALOHA) and the rank-deficiency theorem."},{"cited_title":"Annihilating filter-based low-rank Hankel matrix approach for image inpainting,","cited_arxiv_id":null,"evidence_quote":"provides the structured Hankel matrix model for missing-data recovery that the wrap-around construction extends."},{"cited_title":"A general framework for compressed sensing and par- allel MRI using annihilating filter based low-rank Hankel matrix,","cited_arxiv_id":null,"evidence_quote":"extends the low-rank Hankel framework to joint multidimensional completion, the basis of Joint 3D ALOHA."},{"cited_title":"Fourier method for identifying electromag- netic sources with multi-frequency far-field data,","cited_arxiv_id":null,"evidence_quote":"gives the dense-data Fourier inversion formulas that map far-field data to the source coefficients."},{"cited_title":"Nonuniqueness in the inverse source problem in acoustics and electromagnetics,","cited_arxiv_id":null,"evidence_quote":"documents non-uniqueness in electromagnetic inverse source problems, the failure mode the enrichment addresses."},{"cited_title":"TE/TM decomposition of electromagnetic sources,","cited_arxiv_id":null,"evidence_quote":"provides the TE/TM decomposition of electromagnetic sources that separates the spectral components into the Hankel blocks."}],"review_version":1}