{"id":"04f7bb33-801d-4421-87b1-0a5fb9fefc74","arxiv_id":"2502.03790","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An algorithm called eigenCWD uses a low-rank decomposition of spatially varying point spread functions to deblur simulated single-metalens images faster and with better quality than Wiener filtering.","lead":"Metalenses are thin flat lenses that blur images differently across the frame, so ordinary deblurring fails. This paper presents a fast algorithm, eigenCWD, that corrects spatially varying blur and distortion in simulated metalens images, outperforming the standard Wiener filter.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central derivation in Sec. 3 conflates a dense low-rank factorization with a sum of convolutions; Eq. (11) H ≈ QA and the algorithm's convolution steps describe different operators, so the scaling claim is not supported as written.","rationale":"The reader's verdict is CONDITIONAL primarily because of limited experimental validation and missing comparisons, while the weakest assumption is about PSF-model fidelity and finite-support limitations. The concern identified here is more fundamental: the algebraic derivation of the forward-model decomposition in Sec. 3 is inconsistent with the algorithm described in the same section. If H ≈ QA were taken literally as a rank-K matrix factorization, it could not represent even a shift-invariant blur faithfully, since a convolution operator is full rank. The paper's scaling argument therefore needs to be rewritten in terms of convolution operators and coefficient images, with the true memory and time complexity stated explicitly. This does not necessarily invalidate the empirical deconvolution results, because the implementation likely uses the convolution form, but it removes the current theoretical support for the headline scalability claim. As a result, the appropriate verdict is CONDITIONAL: the paper should be accepted only after the operator algebra is corrected, the exact storage complexity is stated, and the scaling experiments are re-run under the corrected formulation. This differs from the reader's stated weakest assumption, which focuses on the physical PSF model and experimental robustness rather than on internal mathematical consistency.","tokens_in":10271,"tokens_out":12157,"duration_ms":130935,"concrete_test":"Use the provided code (Zenodo 13329612) or a minimal reproduction with M=32 and a single shift-invariant PSF p. Compute (i) the forward output via the algorithm's convolution step with K=1 and a_1 = 1, and (ii) the literal QAf with Q = [vec(p)] and A = [1 ... 1]. Also form the true block-circulant H for p and compute ||H - QA||_F and the numerical rank of H. If the two outputs differ or the Frobenius error is large, Eq. (11) and the convolution algorithm are different operators, confirming the concern.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that eigenCWD scales because the forward model can be written as g ≈ QAf with Q ∈ R^{M^2×K}, A ∈ R^{K×M^2}, and K ≪ M^2 (Sec. 3, Eqs. 11-12). But the three-step efficient computation described immediately after Eq. (12) is not the literal matrix product QAf. That algorithm computes g = Σ_k q_k * (a_k ⊙ f), i.e., a sum of K convolutions of the coefficient-weighted object with the eigenPSFs q_k. The linear operator for this sum is Σ_k C_k diag(a_k), where C_k is the M^2 × M^2 convolution matrix for q_k. This operator is generally full rank: even a single shift-invariant PSF gives H = C_1, a block-circulant matrix of rank M^2, not the rank-1 outer product q_1 1^T that Q ∈ R^{M^2×K}, A ∈ R^{K×M^2} would imply for K=1. Thus Eq. (11) is not a valid decomposition of the transfer matrix H, and the statement that 'Q and A only have sizes M^2 × K' does not establish the claimed memory scaling for the operator actually used. The empirical results in Sec. 4 may still be reproducible if the supplied code implements the convolution form, but the mathematical foundation and complexity argument for the central claim are internally inconsistent as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes eigenCWD, a spatially-varying deconvolution algorithm for single-metalens imaging. The forward model is approximated by decomposing a set of sampled, space-variant PSFs into K eigenPSFs with interpolated eigencoefficients, yielding a forward operator expressed as a sum of K convolutions. Deconvolution is cast as an ADMM optimization with total-variation regularization. The authors simulate images from hyperbolic, parabolic, and spherical metalens phase profiles using an angular-spectrum PSF model, and demonstrate that eigenCWD removes coma and barrel distortion better than Wiener filtering. The paper includes a Zenodo code repository and explicitly acknowledges a finite-support limitation.","tokens_in":10554,"tokens_out":10894,"duration_ms":108694,"significance":"If the technical foundation is corrected, this is a useful contribution to computational imaging: it offers a space-variant deconvolution route that avoids the explicit M^2 by M^2 transfer matrix, with memory and computation scaling roughly as O(K M^2 log M^2) and O(K M^2) for K retained eigenPSFs. The paper's strengths are its concrete connection to the metalens aberration problem, systematic experiments over the PSF sampling density and component truncation, and the availability of runnable code. The claims are, however, demonstrated only on noiseless simulations that use the same nominal PSF model that generates the eigenPSF basis; practical superiority over Wiener filtering for real metalens images is not yet established.","major_comments":[{"comment":"The factorization H ≈ QA is not the operator computed by the three-step procedure described after Eq. (12). With Q in R^{M^2 x K} containing flattened eigenPSFs and A in R^{K x M^2} containing eigencoefficients, QAf equals sum_k q_k (a_k^T f), which for K=1 is a rank-one matrix q_1 a_1^T. The implemented procedure instead computes g = sum_k q_k * (a_k elementwise f), whose matrix is sum_k C_k diag(a_k), where C_k is the M^2 by M^2 convolution matrix of q_k; for K=1 this is a full-rank block-circulant matrix, not a rank-one outer product. The correct statement is H ≈ sum_k C_k diag(a_k). The memory and complexity argument should be rephrased in terms of storing K convolution kernels and K coefficient maps, rather than the sizes of Q and A.","section":"3, Eqs. (11)–(12)"},{"comment":"The notation in Eq. (3) writes p(u,v,x,y), but Eq. (4) substitutes q_k(x-u,y-v), indicating that x,y in Eq. (3) are offset coordinates relative to the source position. This ambiguity is the source of the algebraic inconsistency in Eq. (11). Please define explicitly whether the sampled PSFs and eigenPSFs are represented as functions of offsets or of absolute image coordinates, and keep that convention consistent throughout the matrix formulation. In Eq. (15), the expression inside the angular-spectrum propagator Pf appears to be a product of the phase factor and the lens phase phi(x,y); as printed, the notation is ambiguous and should be corrected.","section":"2, Eqs. (3)–(4) and Eq. (15)"},{"comment":"All reconstruction results are noiseless simulations generated from the same nominal PSF model used to construct the eigenPSF basis. No noise, background, or experimental data are demonstrated; the only experimental reference, [38], is mentioned in Sec. 5 as a source of ringing artifacts without showing the result. The central claim that eigenCWD 'surpasses Wiener filtering' for single-metalens imaging should therefore be qualified to the simulated, finitely supported, noiseless setting, or supported by experiments with realistic noise and background. As written, the practical relevance claim is stronger than the evidence.","section":"4 and 5"}],"minor_comments":[{"comment":"Please clarify whether C is the Gram matrix of mean-centered PSFs. If the mean is subtracted, Eq. (8) requires an additional mean term, because the reconstruction p_j = sum_i (w_i)_j q_i holds only when the decomposition is performed without centering or when the mean is explicitly included.","section":"2, Eq. (5)"},{"comment":"The simulation parameters contain apparent typographical errors: 'L = 1.25 mis placed s = 2 maway' should read 'L = 1.25 m is placed s = 2 m away' or similar, and the stated angular field of view of 46.9 deg does not follow from L=1.25 m and s=2 m. Please check the units and the calculation.","section":"4, first paragraph"},{"comment":"The explanation for the non-monotonic PSNR is that the interpolated eigencoefficients do not conserve total pixel energy. Please state explicitly whether the forward model in the reconstruction step normalizes the effective PSFs, and whether the same effect influences the PSNR improvements reported in Fig. 5.","section":"4.1, Figs. 3–4"},{"comment":"The ADMM updates are deferred to the paper by Sroubek [29]. Since the operator in Eq. (13) is a sum of convolutions rather than a standard matrix, a brief derivation of the proximal steps or a pseudocode listing would substantially improve reproducibility.","section":"3, Eq. (13)"},{"comment":"Reference [38] is cited as the source of the observed ringing artifacts, but [38] describes GaP quadratic metalenses; please clarify whether eigenCWD was actually applied to data from [38] or whether [38] is cited only as a related experimental imaging demonstration.","section":"5, reference [38]"}],"recommendation":"major_revision","confidential_remarks":"This is a technically competent algorithmic paper with reproducible code. The principal obstacle is the algebraic inconsistency between Eq. (11) and the implemented convolution-sum operator; this is fixable by rewriting the factorization and the complexity statement. The validation is narrow, but the underlying eigenPSF/convolution approach is sound and the experiments, once qualified, support the algorithmic claims. I recommend major revision rather than rejection, with the expectation that the formalism and validation framing can be corrected within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a method paper with real utility—space-variant deconvolution without constructing the full M^2 x M^2 transfer matrix—and the authors ship code on Zenodo. The new combination (eigenPSF basis from Lauer, CWD-SVD-style ADMM inversion from Sroubek, metalens application) is legitimate and the synthetic demonstrations are clean. But the central derivation in Sec. 3 is wrong as written. Eq. (11) states H ≈ QA with Q ∈ R^(M^2×K) and A ∈ R^(K×M^2). If you take that factorization literally, then QAf is a linear combination of the K eigenPSFs with scalar weights that are global dot products of f with each eigencoefficient row. That is not a space-variant blur at all. The actual computation the paper describes—and presumably implements—is g = Σ_k (f ⊙ a_k) * q_k, i.e., K spatially-weighted convolutions. The correct operator is Σ_k C_k diag(a_k), not a low-rank product of Q and A. So the mathematical foundation for the scaling claim is not established, and the claim that 'Q and A only have sizes M^2×K' does not do the work the paper says.\n\nThis matters because the whole pitch is scaling. The method may still deliver that scaling—using FFT-based convolutions avoids storing H—but the paper as written doesn't show it. A referee should ask for a corrected Sec. 3 that states the operator as a sum of convolutions, defines the memory and complexity in terms of K FFTs, and compares against the CWD-SVD baseline the paper modifies.\n\nOther soft spots, in proportion: all results are noiseless simulations; there's no noise analysis, which matters for deconvolution with TV; and the only experimental mention is a reference to the authors' own earlier work with ringing artifacts. The paper does flag that limitation (finite support, background breaks it), which I credit. The PSNR/SSIM inversion is explained but not deeply. Also, the running headers are contaminated from another submission—minor but sloppy.\n\nWhat's genuinely good: the eigenPSF forward-model approximation is tested against the explicit integral, the truncation behavior is characterized, and the code is public. The comparison to Wiener is a fair baseline even if not a strong one.\n\nFor a reader: this is worth engaging with as a practical method, not as a fundamental result. I'd send it to peer review with a request for major revision of Sec. 3 and the complexity analysis. If the authors fix the derivation, this could be a solid methods paper.","headline":"A practical and reproducible space-variant deconvolution method, but the paper's central matrix decomposition is misstated and needs a serious rewrite before the scaling claims can be trusted.","tokens_in":11152,"tokens_out":4193,"would_cite":false,"duration_ms":39137,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.30.-d","42.30.Va"],"model":"deepseek-v4-flash","headline":"A spatially varying metalens blur can be inverted without building the full transfer matrix, using an eigen-decomposition of sampled point spread functions.","keywords":["metalens","spatially varying deconvolution","eigenPSF decomposition","point spread function","computational imaging","total variation regularization","ADMM","aberration correction"],"falsifier":"Measure the actual point spread functions of a fabricated metalens on a fine grid across its field of view and compare them with the nominal simulation used in Eq. (15); if the real PSFs differ substantially from the model or vary non-smoothly, the reconstruction quality shown in Figs. 5 and 6 will not be reproduced on experimental data. A simpler experimental check is to image a resolution target with known background on a hyperbolic metalens and see whether eigenCWD introduces the ringing artifacts the paper attributes to finite-support violations.","tokens_in":10034,"feed_emoji":"🔬","tokens_out":6562,"duration_ms":57780,"temperature":0.7,"pith_summary":"Single metalens cameras form images with aberrations that change across the field of view: a hyperbolic metalens focuses efficiently but adds coma and astigmatism off-axis, and the common Wiener filter cannot correct such spatially varying blur. This paper introduces eigenCWD, a deconvolution algorithm that treats the blur as a space-variant linear operator but never builds the full transfer matrix. Instead, it samples point spread functions across the field, decomposes them into a small set of eigen-PSFs via eigendecomposition, and interpolates the coefficients to every pixel. The resulting approximate forward model makes the inverse problem tractable, and the paper shows that eigenCWD outperforms Wiener filtering in correcting blur and geometric distortion for hyperbolic, parabolic, and spherical metalens profiles in simulation.","feed_headline":"eigenCWD deblurs metalens images where Wiener filtering fails","feed_subtitle":"A low-rank PSF decomposition corrects coma and barrel distortion in single-metalens cameras.","key_machinery":"The central object is the eigenPSF decomposition of the space-variant point spread function. From N PSFs sampled at different field positions, the paper forms a covariance matrix and takes its eigenvectors; each eigenPSF is a weighted sum of the sampled PSFs, and the eigenvector entries assigned to the sample positions are interpolated to yield field-dependent eigencoefficients. This yields the factorization H approximately equal to QA in which the forward blur is evaluated as K element-wise multiplications followed by K convolutions. The inverse problem is then solved by ADMM with total variation regularization using that factorized forward model.","core_discovery":"The paper claims that a spatially varying blur in metalens imaging can be efficiently inverted by factoring the transfer matrix H into a product QA, where Q holds K eigen-PSFs and A holds their interpolated coefficients, with K much smaller than the number of image pixels. The factorization, derived from an eigendecomposition of sampled point spread functions, captures the space-variant blur well enough to allow high-quality deconvolution of images from a single metalens. Using an ADMM solver with total variation regularization, eigenCWD reconstructs objects from simulated blurred images of hyperbolic, parabolic, and spherical metalenses, removing both field-dependent coma and barrel distortion that remain in Wiener-filtered results. The paper also shows that truncating the number of eigen-PSFs to about 200 for a 19 by 19 sampling grid preserves reconstruction quality while cutting computation time, demonstrating that the approach scales to realistic image sizes.","pith_inferences":["The rank K needed to represent the blur could serve as a diagnostic of space-variant complexity: lenses with severe coma require more eigen-PSFs, and a lens whose PSFs barely change would need only a handful, in which case Wiener filtering is nearly optimal.","The finite-support limitation points to an obvious preprocessing step: subtract the background or apply an edge-tapering mask before deconvolution, which the paper does not test but mentions as a future remedy.","The same eigendecomposition-of-PSFs strategy could be applied to other space-variant imaging systems, such as multimode fiber imaging or atmospheric turbulence, where the forward model is also too large to store explicitly.","Because the eigencoefficients are computed by interpolation, the method would likely fail at field positions where the PSF changes abruptly, e.g., near the edges of the aperture or at phase discontinuities; sampling adaptively at such positions could be tested."],"forward_implications":["A metalens image can be deblurred at full field of view without training data, using only a small set of measured or simulated PSFs.","The cost of deconvolution scales with the number of retained eigen-PSFs K rather than the number of pixels, so megapixel metalens images remain feasible.","Truncating the eigen-PSF components gives a controllable speed-quality tradeoff, with roughly 200 components sufficient for the hyperbolic lens studied.","The same algorithm corrects coma and barrel distortion for hyperbolic, parabolic, and spherical metalens profiles, so one method covers common metalens designs."],"supporting_citations":[{"why":"Supplies the eigenPSF decomposition method that produces the basis PSFs and interpolated coefficients.","marker":"[28]"},{"why":"The CWD-SVD deconvolution algorithm that eigenCWD modifies; provides the ADMM-based inverse-problem strategy.","marker":"[29]"},{"why":"The Wiener filter baseline that eigenCWD is compared against and outperforms on space-variant blur.","marker":"[10]"},{"why":"The ADMM optimization framework used to solve the regularized inverse problem.","marker":"[32]"},{"why":"Total variation regularization used to stabilize the deconvolution.","marker":"[35]"},{"why":"The angular spectrum propagator used in Eq. (15) to compute the PSFs from the metalens phase profile.","marker":"[36]"},{"why":"The SSIM metric used to quantify reconstruction quality.","marker":"[37]"},{"why":"The experimental metalens data on which ringing artifacts were observed, motivating the noted finite-support limitation.","marker":"[38]"}],"fun_headline_variants":["eigenCWD fixes metalens blur that Wiener filter misses","Spatially-varying deconvolution for metalens imaging","New algorithm deblurs metalens images with low-rank PSF","Metalens deblurring beyond Wiener filtering","eigenCWD: fast deconvolution for single-metalens cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that the real point spread functions of the metalens match the nominal phase-profile simulations and vary smoothly enough that a modest number of sampled PSFs plus interpolation can represent every pixel's blur; fabrication errors, partial coherence, or a nonzero background around the object would break this assumption, and the paper itself notes that finite-support objects are required to avoid ringing artifacts.","fun_headline_variants_meta":{"raw":{"variants":["eigenCWD fixes metalens blur that Wiener filter misses","Spatially-varying deconvolution for metalens imaging","New algorithm deblurs metalens images with low-rank PSF","Metalens deblurring beyond Wiener filtering","eigenCWD: fast deconvolution for single-metalens cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1313,"prompt_tokens":954,"completion_tokens":359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":570,"tokens_out":359,"duration_ms":3488,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:43:24.909045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual point spread functions of a fabricated metalens on a fine grid across its field of view and compare them with the nominal simulation used in Eq. (15); if the real PSFs differ substantially from the model or vary non-smoothly, the reconstruction quality shown in Figs. 5 and 6 will not be reproduced on experimental data. A simpler experimental check is to image a resolution target with known background on a hyperbolic metalens and see whether eigenCWD introduces the ringing artifacts the paper attributes to finite-support violations.","supporting_citations":[{"cited_title":"Deconvolution with a spatially variant PSF","cited_arxiv_id":null,"evidence_quote":"Supplies the eigenPSF decomposition method that produces the basis PSFs and interpolated coefficients."},{"cited_title":"Decomposition of Space-Variant blur in image deconvolution","cited_arxiv_id":null,"evidence_quote":"The CWD-SVD deconvolution algorithm that eigenCWD modifies; provides the ADMM-based inverse-problem strategy."},{"cited_title":"Distributed optimization and statistical learning via the alternating direction method of multipliers","cited_arxiv_id":null,"evidence_quote":"The ADMM optimization framework used to solve the regularized inverse problem."},{"cited_title":"Nonlinear total variation based noise removal algorithms","cited_arxiv_id":null,"evidence_quote":"Total variation regularization used to stabilize the deconvolution."},{"cited_title":"Image quality assessment: from error visibility to structural similarity","cited_arxiv_id":null,"evidence_quote":"The SSIM metric used to quantify reconstruction quality."}],"review_version":1}