{"id":"e3f33919-4c39-4bfa-bdfc-e171e7bf91f5","arxiv_id":"2505.15777","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A projection step that forces a deep network's reconstruction to satisfy y = Ax gives small PSNR gains in low-noise imaging tests, but the supporting theory is a restatement of the definition of a well-trained network.","lead":"This paper adds a simple final step to deep learning image-reconstruction networks: project the network's output onto the set of images that exactly match the measured data. The step improves PSNR in low-noise tests, but the theoretical framing restates known projection math and gains vanish when noise is high.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The noiseless projection step is mathematically sound, but the noisy-regime gains rely on λ chosen on the test set; when λ=0 the method is no correction, so the central improvement claim is not established.","rationale":"The reader's verdict is CONDITIONAL, and my concern does not move that verdict; it sharpens the condition. The reader's weakest_assumption focused on the null-space error assumption failing at high noise, while I identify the more directly actionable problem: the noisy-regime gains are obtained by choosing λ on the test set, and the paper reports rows where the best λ is 0, meaning no correction. This is a load-bearing issue for the practical claim because Eq. (17) has no principled selection rule and Proposition 2 does not characterize the corrected estimator. I also agree with the reader that Definition 1 is essentially tautological: condition (ii) already forces f(y)=x in the noiseless case, making Proposition 1 and Theorem 2 true by definition rather than by training dynamics. That is a theoretical weakness, but it does not invalidate the noiseless projection formula, which is a standard orthogonal projection and is provably non-increasing in L2 error when the true signal satisfies the constraint. The experimental concerns (duplicated tables, mismatched figure captions, no accessible code link) are real but secondary to the correctness question. A revision that adds a validation-based λ-selection protocol, a sensitivity analysis, and a corrected table could reasonably move the paper to acceptance; without those, the noisy-regime claim remains unsupported.","tokens_in":11355,"tokens_out":6791,"duration_ms":66168,"concrete_test":"Split each test set (DIV2K test, Set3C, CBSD68, Urban100) into a validation split and a held-out split. For each noise level σ ∈ {0.01, 0.05, 0.1, 0.2, 0.3}, select λ by maximizing PSNR on the validation split only, then report PSNR/SSIM on the held-out split. Also report PSNR as a function of λ over at least three orders of magnitude. If held-out gains reproduce Table 2, the λ-tuning concern is resolved; if gains vanish or become negative, the noisy-regime improvement claim depends on test-set oracle selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim splits into noiseless and noisy cases. In the noiseless case, Eq. (8) is the orthogonal projection of f(y) onto the affine subspace {x: Ax=y}. Since the true x lies in that subspace, ||x*−x||₂ ≤ ||f(y)−x||₂ for every f, so the noiseless gains are robust and not the weak point. The weak point is Section 5: for y=Ax+n, the proposed correction in Eq. (17) is a Tikhonov-type estimator with a free hyperparameter λ that the paper selects by grid search against ground truth on the same test images (Table 2). This is test-set leakage. In the high-noise rows of Table 2, the selected λ is 0 for several datasets and the 'projected' result is identical to the raw network; in the remaining high-noise rows, λ is chosen on the very data being reported. No validation-based selection rule, parameter-free criterion, or sensitivity analysis is provided. Proposition 2 only computes the noise bias of a well-trained network, Tr(A†Σ(A†)⊤), and does not give the expected error of Eq. (17) or an optimal λ. Thus the paper's claim that the regularized projection improves reconstruction accuracy across diverse inverse problems is not established for noisy measurements, which is exactly the regime where a correction is most needed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-hoc projection step for deep inverse networks. For noiseless measurements y = Ax, the corrected estimate is x* = A†y + (I - A†A) f(y), the orthogonal projection of the network output onto the affine set {x : Ax = y}; the paper shows algebraically that this enforces measurement consistency and that for a 'well-trained' network the projection is the identity. For noisy measurements, a Tikhonov-style regularized variant (Eq. 17) with hyperparameter λ is proposed. Experiments on inpainting, deblurring, single-pixel imaging, and one spectral imaging problem compare network outputs before and after projection across several architectures and noise levels.","tokens_in":11662,"tokens_out":9105,"duration_ms":75971,"significance":"The noiseless projection is a correct and potentially useful drop-in consistency step: for y = Ax it is non-expansive with respect to the true x, so it cannot increase reconstruction error in the noiseless case, and Table 1 shows consistent, sometimes substantial gains (e.g., deblurring Unrolled PSNR 23.77 to 25.98). The paper is also commendable for releasing code and reporting a null-space consistency diagnostic (Eq. 18). However, the theoretical 'well-trained' analysis is tautological, the noisy-regime evaluation uses λ selected on the test set, and parts of the appendix appear to duplicate Table 1. Thus the broader claims of guaranteed improvement and rigorous theoretical support are not established in the current version, though the core noiseless idea is sound and fixable.","major_comments":[{"comment":"The regularization parameter λ is chosen by a grid search against ground truth on the same test sets used for reporting ('we performed a grid search to select the optimal value for each noise level'; 'The best-performing λ was used to generate the final reconstruction results'). No validation split or repeated runs are described, so the noisy-regime results are affected by test-set leakage and lack error bars. In the σ = 0.10–0.30 rows the selected λ is 0 for several datasets, making Eq. (17) identical to the raw network output, and the remaining nonzero-λ rows show negligible or no improvement (e.g., σ = 0.20, Set3C PSNR 20.33 to 20.34; multiple SSIM values unchanged). Proposition 2 only computes the bias of a well-trained network and does not yield an optimal λ or an error expression for Eq. (17). The paper therefore does not establish the claimed improvement for noisy measurements.","section":"§6.3, Table 2"},{"comment":"A 'well-trained' network is defined by the conditions f(y) = A†y + N(y) and N(y) = (I - A†A)x for y = Ax, which immediately give f(y) = x. Consequently Proposition 1 (zero expected error) and Theorem 2 (projection reduces to identity) are true by construction. This is a tautology rather than a proof that trained networks satisfy the decomposition, and it does not explain the experimental behavior. The theoretical claim in the abstract ('we theoretically demonstrate...') should be reframed, or a nontrivial condition under which trained networks approximately satisfy Eqs. (4)–(5) should be provided.","section":"Definition 1, Proposition 1, Theorem 2"},{"comment":"The proof of Theorem 1 multiplies by (AA^T)^-1 and uses A† = A^T(AA^T)^-1, which requires A to have full row rank. This assumption is missing from the theorem statement. For rank-deficient A, which is common in inverse problems, the proof is invalid. The closed-form formula (8) is nevertheless correct for the Moore-Penrose pseudoinverse when y ∈ R(A); please state the needed assumption or replace the proof with one based on the Moore-Penrose equations.","section":"Theorem 1, Appendix A"},{"comment":"Several rows of Table 4 duplicate Table 1 exactly even though they are labeled for different datasets: Deblurring-Set3c and Inpainting-Set3c reproduce the DIV2K entries from Table 1. This indicates that the appendix does not actually contain the claimed evaluations on Set3c, CBSD68, and Urban100, so the cross-dataset generalization claim in Section 6.2 is unsupported. Moreover, some SPI SSIM values decrease after projection (e.g., Set3c DnCNN 0.544 to 0.540; Restormer 0.591 to 0.586), contradicting the blanket statement that the method improves all models. Please recompute these results and correct the tables.","section":"§6.2, Table 4"}],"minor_comments":[{"comment":"The text says 'four different datasets: DIV2K testing, Set3c, CBSD68, and Urban100, containing 3, 68, and 100 test images, respectively' but lists counts for only three datasets; please include the DIV2K test count.","section":"§6.1"},{"comment":"The proof of Theorem 1 contains a typo: 'sing the constraint Ax = y' should be 'Using the constraint Ax = y'; also, Theorem 2's proof refers to 'SM, Appendix A' inconsistently with the actual appendix location.","section":"Appendix A"},{"comment":"The caption reads 'he measurement is processed' and should read 'The measurement is processed'.","section":"Figure 1"},{"comment":"The conclusion says the method was validated on 'PnP models', but the experiments in Section 6.2 use Unrolled, DnCNN, Restormer, and DiffUNet; no PnP model appears in Tables 1 or 4.","section":"Conclusion"},{"comment":"When Σ is set to I, the meaning and scale of λ depend on the scaling of A and y; a sentence explaining the practical choice or range of λ would aid reproducibility.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The duplication between Table 4 and Table 1 needs careful verification; if it is not a formatting error, it calls the experimental record into question. I recommend that the editor require the authors to report a validation-based λ selection procedure, error bars or repeated runs, and the raw result tables. The noiseless contribution is salvageable, and the paper should not be rejected outright if these issues are correctable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the noiseless projection step is mathematically correct, and the empirical gains in noiseless deblurring look real. But the novelty is thin, and the noisy-regime claims are not supported because λ is tuned on the exact test data being reported.\n\nWhat's actually new: not much. Equation (8) is the standard projection onto the affine set {x : Ax = y}; Theorem 1 is explicitly from ref [16]. The value is empirical: applying this fixed post-processing step to several inverse networks (DnCNN, Restormer, DiffUNet, unrolled) consistently improves PSNR/SSIM in noiseless settings, with deblurring gaining up to ~2 dB. Figure 2, showing the training gap closing as the network matures, is a nice sanity check. As a practical consistency layer, the idea has legs.\n\nSoft spots. The \"well-trained\" theory (Definition 1, Propositions 1 and 2, Theorem 2) is true by construction: if N(y) = (I − A†A)x, then f(y) = x, so zero error and identity projection follow immediately. It provides no insight into real networks and should not be presented as a theoretical contribution.\n\nMore serious is Section 5 and Table 2. The regularized projection in Eq. (17) is just Tikhonov regularization with a free parameter λ, and the paper selects λ by grid search against ground truth on the same test images used to report PSNR. That is test-set leakage. In the highest-noise rows, the selected λ is 0, meaning the \"projected\" result is identical to the raw network; in the remaining rows, λ is chosen on the reported data. No validation-based selection, no sensitivity analysis. So the noisy-regime improvement claim is not established. Also, despite the text saying code is publicly available, I did not find code in the submission. There are duplicated tables (Table 1 rows appear identical to Set3c rows in Table 4), and Figure 5's caption looks copy-pasted from Figure 3.\n\nThis is a decent empirical note on noiseless data consistency, not a new method. The math is correct where it isn't tautological. I'd send it to peer review because a serious referee could push for cleaner noisy experiments and an honest re-framing, but I would not accept it as-is.","headline":"The noiseless projection is correct and gives real but modest gains; the noisy experiments are compromised by test-set leakage on λ, and the 'well-trained' theory is tautological.","tokens_in":12191,"tokens_out":2487,"would_cite":false,"duration_ms":21340,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F22","68T07","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a fixed, non-iterative projection step applied after any trained deep inverse network enforces measurement consistency and improves reconstruction accuracy, and that it reduces to the identity for a well-trained…","keywords":["inverse problems","deep learning","measurement consistency","projection-based correction","null-space decomposition","Tikhonov regularization","image reconstruction","data consistency"],"falsifier":"On a low-noise benchmark, compute the expected range-space error $\\mathbb{E}\\|A(\\hat{f}(y)-x)\\|^2$ and the expected null-space error $\\mathbb{E}\\|(I-A^\\dagger A)(\\hat{f}(y)-x)\\|^2$ for a trained network: if the range-space error is not small relative to the null-space error, the projection cannot remove the dominant part of the error, and a test set where the projected output has lower PSNR than the raw network output would count directly against the paper's claim.","tokens_in":11153,"feed_emoji":"🖼️","tokens_out":10626,"duration_ms":86553,"temperature":0.7,"pith_summary":"The paper tries to establish that a fixed, non-iterative projection step, applied after any trained deep inverse network, can enforce exact consistency with the measurement model and improve reconstruction quality. For clean measurements the correction replaces the part of the network's estimate that disagrees with $y=Ax$ by the least-norm consistent solution, yielding $\\hat{x} = A^\\dagger y + (I - A^\\dagger A)\\hat{f}(y)$. For noisy measurements the paper derives a weighted regularized version with a tunable parameter $\\lambda$ that interpolates between strict consistency and trusting the network, reducing to Tikhonov regularization when the noise covariance is unknown. The authors prove that if the network is 'well-trained' in the sense of decomposing its output into range-space and null-space parts, the projection is exactly the identity, so the method only changes inconsistent outputs. The payoff is a drop-in post-processing layer for imaging applications where reconstructions must respect the physics of the sensing process, with experimental gains largest at low noise and on out-of-distribution data.","feed_headline":"Projection step enforces measurement fit in deep inverse networks","feed_subtitle":"A single closed-form correction makes reconstructions match the sensing model and helps most at low noise","key_machinery":"The load-bearing object is the Euclidean projection onto the affine measurement-consistent set $\\{x : Ax=y\\}$, whose closed form is $\\hat{x} = A^\\dagger y + (I - A^\\dagger A)\\hat{f}(y)$, where $A^\\dagger$ is the Moore-Penrose pseudoinverse. This operator replaces the inconsistent range-space component of the network output with the least-norm solution that satisfies $y=Ax$, while leaving the null-space component $(I - A^\\dagger A)\\hat{f}(y)$ untouched. For noisy measurements the machinery becomes the weighted projection $\\hat{x} = (I + \\lambda A^\\top\\Sigma^{-1}A)^{-1}(\\hat{f}(y) + \\lambda A^\\top\\Sigma^{-1}y)$, a one-step Tikhonov-type correction that can be precomputed for fixed $A,\\Sigma$. The supporting theory is Definition 1 (well-trained network), Theorem 1 (the closed-form projection is the optimal solution of the constrained problem), and Theorem 2 (for well-trained networks the projection is the identity).","core_discovery":"The central claim is that the output of any trained deep inverse network can be written as a range-space component plus a null-space component, and that only the range-space component needs to be corrected for measurement consistency. The paper defines a well-trained network as one for which $\\hat{f}(y) = A^\\dagger y + (I - A^\\dagger A)x$ holds over the data, and proves (Proposition 1) that such a network achieves zero mean-squared reconstruction error, and (Theorem 2) that the projection leaves its output unchanged. For a general network, the projection replaces the range-space part with $A^\\dagger y$ and keeps the null-space part, so any inconsistency $A\\hat{f}(y)\\neq y$ is removed in one closed-form step. In the noisy case, the paper replaces the hard constraint by the penalty $(\\hat{x} - \\hat{f}(y))^\\top(\\hat{x}-\\hat{f}(y)) + \\lambda(A\\hat{x}-y)^\\top\\Sigma^{-1}(A\\hat{x}-y)$, with closed-form solution $\\hat{x} = (I + \\lambda A^\\top\\Sigma^{-1}A)^{-1}(\\hat{f}(y) + \\lambda A^\\top\\Sigma^{-1}y)$, and shows empirically that the optimal $\\lambda$ decreases to zero as noise increases. Across deblurring, inpainting, and single-pixel imaging, the projection consistently improves PSNR/SSIM for convolutional, transformer, diffusion, and unrolled architectures, with the largest gains at low noise.","pith_inferences":["A natural extension the paper does not develop is an adaptive noise-level rule for $\\lambda$ (for instance, making it decrease with the noise variance), which the reported grid-search results suggest should work.","Because the correction is a fixed linear map, it could be composed with other post-processing or generative refinement steps, or applied iteratively, without retraining; that may compound gains in settings with very strong priors.","The method's behavior at high noise suggests that measurement consistency is not always desirable; a hybrid rule that switches the projection off when the estimated residual $A\\hat{f}(y)-y$ is dominated by noise would address a limitation the paper explicitly flags.","One could test the projector as a generic data-consistency layer inside untrained or equivariant imaging pipelines, since the projector needs no data to be constructed; this would connect it to broader unsupervised inverse-problem methods."],"forward_implications":["Any existing inverse network can be made exactly measurement-consistent in the noiseless case by one application of a fixed linear operator, with no retraining and no extra parameters.","At low noise, the projection typically improves PSNR and SSIM, including on out-of-distribution test sets, acting as a generalization booster for data-driven solvers.","For noisy measurements the method has a tunable knob: small $\\lambda$ trusts the network, large $\\lambda$ enforces the measurement, and the reported optimal values fall to zero as noise rises.","Networks that already satisfy the range/null-space decomposition are untouched by the correction, so the method is a safe refinement rather than an alteration of a well-trained model.","The magnitude of the gain tracks how much information the forward operator keeps: large gains appear in deblurring and inpainting, while single-pixel imaging at very low compression shows only small gains."],"supporting_citations":[{"why":"Supplies the closed-form solution of the constrained least-squares projection stated as Theorem 1 and used throughout as the correction operator.","marker":"[16]"},{"why":"Inspires the range/null-space decomposition that defines a well-trained network, on which Theorem 2 and Proposition 1 rest.","marker":"[28]"},{"why":"Provides the implementation platform used to run the simulations and benchmarks reported in the paper.","marker":"[27]"},{"why":"One of the base deep inverse networks used in the benchmark and in the out-of-distribution noise experiments; its output is what the projection post-processes.","marker":"[4]"},{"why":"Serves as the transformer-based architecture in the benchmark across inverse problems.","marker":"[7]"},{"why":"Serves as the diffusion-based architecture in the benchmark across inverse problems.","marker":"[8]"},{"why":"Serves as the hybrid unrolled-optimization architecture in the benchmark across inverse problems.","marker":"[11]"},{"why":"Is the architecture used in the training-dynamics and well-trained-condition experiment.","marker":"[3]"}],"fun_headline_variants":["Projection fix aligns deep inverse nets with physical constraints","One-step projection boosts deep inverse reconstruction accuracy","Closed-form projection enforces forward-model fidelity in deep nets","Deep inverse networks get a projection boost for low-noise cases","Projection correction makes deep reconstructions match measurements"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the assumption that a trained network's reconstruction error lies almost entirely in the null space of the forward operator, so replacing the range-space component with the least-norm measurement-consistent value discards no meaningful signal information.","fun_headline_variants_meta":{"raw":{"variants":["Projection fix aligns deep inverse nets with physical constraints","One-step projection boosts deep inverse reconstruction accuracy","Closed-form projection enforces forward-model fidelity in deep nets","Deep inverse networks get a projection boost for low-noise cases","Projection correction makes deep reconstructions match measurements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00054,"raw_usage":{"total_tokens":2607,"prompt_tokens":979,"completion_tokens":1628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":1553}},"tokens_in":595,"tokens_out":1628,"duration_ms":10894,"temperature":1.0,"reasoning_tokens":1553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:11:26.204595+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a low-noise benchmark, compute the expected range-space error $\\mathbb{E}\\|A(\\hat{f}(y)-x)\\|^2$ and the expected null-space error $\\mathbb{E}\\|(I-A^\\dagger A)(\\hat{f}(y)-x)\\|^2$ for a trained network: if the range-space error is not small relative to the null-space error, the projection cannot remove the dominant part of the error, and a test set where the projected output has lower PSNR than the raw network output would count directly against the paper's claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form solution of the constrained least-squares projection stated as Theorem 1 and used throughout as the correction operator."},{"cited_title":"Data-proximal null-space networks for inverse problems","cited_arxiv_id":"2309.06573","evidence_quote":"Inspires the range/null-space decomposition that defines a well-trained network, on which Theorem 2 and Proposition 1 rest."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the implementation platform used to run the simulations and benchmarks reported in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the base deep inverse networks used in the benchmark and in the out-of-distribution noise experiments; its output is what the projection post-processes."},{"cited_title":"We conducted three main experiments to assess its impact","cited_arxiv_id":null,"evidence_quote":"Serves as the transformer-based architecture in the benchmark across inverse problems."},{"cited_title":"Visual results of the evaluated method","cited_arxiv_id":null,"evidence_quote":"Serves as the diffusion-based architecture in the benchmark across inverse problems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the hybrid unrolled-optimization architecture in the benchmark across inverse problems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the architecture used in the training-dynamics and well-trained-condition experiment."}],"review_version":1}