{"id":"0ae105d9-f259-479a-9dc0-8e951cbdbcfb","arxiv_id":"1908.01642","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TV-minimization solvers, especially NESTA and TVAL3, are presented as faster and more noise-robust than L1 solvers for natural images, but the L1 comparison is biased by an oracle-based setup.","lead":"This paper is a tutorial review of image compressive sensing algorithms, comparing off-the-shelf L1 and total-variation solvers on benchmark images with LiDAR-like noise. It recommends NESTA and TVAL3 for practical use, but its L1 baseline comparison relies on unrealistic access to image transforms.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"L1 and TV are tested under different forward operators (L1: y=ΨΦx; TV: y=Ψx) and different measurement rates, so the headline conclusion that TV beats L1 is confounded.","rationale":"The reader is in the right neighborhood but the specific mechanism is not quite right. The L1 simulation y = ΨΦx is a legitimate linear measurement of x, so it is not true that L1 is untested; the real flaw is that the TV arm uses a different forward operator (y = Ψx), so the comparison does not hold the sensing matrix fixed across regularizers. The paper's own §2.4 disclaimer that the L1 setup is 'not applicable to a real-world system' strengthens this concern. In addition, the L1 and TV experiments use non-overlapping measurement-rate ranges (1.7–80% vs 0.17–30%), so the claim that TV needs fewer measurements is not supported by matched comparisons. These confounds undermine the headline conclusion that TV is more appropriate than L1 for natural images. The REJECT verdict therefore stands, though for a more precise experimental-design reason than the one stated by the reader. Secondary reproducibility issues (no code link, no error bars) reinforce but are not the main load-bearing problem.","tokens_in":7821,"tokens_out":13167,"duration_ms":123802,"concrete_test":"Run a controlled comparison at matched measurement rates (e.g., m/n = 1.7%, 10%, 30%) on the same 35 images with a fixed measurement matrix A for both regularizers: (i) A = Ψ, with L1 as min ||Φx||_1 s.t. y = Ax and TV as min ||x||_TV s.t. y = Ax; (ii) A = ΨΦ, with both regularizers using the same A. Report PSNR/SSIM and L1 error. If TV does not consistently outperform L1 in both settings, the original comparison was confounded and the conclusion is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 defines the L1 model as y = Ψz with z = Φx, i.e. y = ΨΦx, and §2.4 says the simulation computes z = Φx as a pre-step. Section 3 defines TV as min ||x||_TV s.t. y = Ψx. Thus the L1 and TV algorithms are not fed the same kind of measurements: their effective sensing matrices differ (A_L1 = ΨΦ vs A_TV = Ψ). The author explicitly concedes in §2.4 that the L1 procedure 'is not applicable to a real-world system.' A second confound is the measurement schedule: L1 results are shown at m/n = 1.7–80% while TV results are shown at m/n = 0.17–30%, so the claim that TV needs fewer measurements is not based on matched rates. Since the paper's central conclusion is a practical 'which algorithm should I use' recommendation, the experiments must isolate the regularizer. As designed, the observed TV advantage could be an artifact of the forward operator or of the measurement rates, not of TV regularization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a tutorial-style review of classical compressive sensing algorithms for images, with emphasis on total variation (TV) methods and a target application to LiDAR. It introduces basic CS theory, reviews L1-based solvers (OMP, LASSO, GPSR, SPGL1, L1 Magic, NESTA) and TV-based solvers (L1 Magic TV, NESTA, TVAL3), and reports simulation comparisons of runtime and reconstruction error across measurement ratios and noise levels. Its headline conclusions are that TV minimization is more appropriate than L1 minimization for natural images and that NESTA and TVAL3 are the most practical algorithms.","tokens_in":8064,"tokens_out":9195,"duration_ms":87692,"significance":"A reliable practical comparison of off-the-shelf CS solvers would be valuable to practitioners, and the paper makes a reasonable effort to treat algorithms as black boxes, to include noisy measurements, and to provide its code. However, the experimental design confounds the regularizer with the forward operator and the measurement schedule, so the main comparative claim is not established. Because the paper's primary contribution is the practical recommendation, its significance depends entirely on correcting this flaw.","major_comments":[{"comment":"The comparison that drives the paper's central conclusion is confounded. The L1 experiments reconstruct z = Φx from measurements y = Ψz, i.e., y = ΨΦx, with z computed by a pre-step that the author states 'is not applicable to a real-world system' (§2.4). The TV experiments instead reconstruct x from y = Ψx (§3). Thus the L1 and TV solvers are given different forward operators, and the observed advantage of TV could be an artifact of the sensing matrix or of oracle access to the sparse coefficients rather than of the regularizer. Since §5's first conclusion is that TV is more appropriate than L1 for natural images, this load-bearing comparison must be redone with identical sensing matrices and without oracle access to the sparse coefficients.","section":"2.4, 3, 5"},{"comment":"The measurement-ratio schedules differ between the two algorithm classes. L1 results are shown at m/n = 1.7%, 10%, 30%, 50%, 80% (Fig. 2), while TV results are shown at m/n = 0.17%, 1%, 1.7%, 10%, 30% (Fig. 5). The claim that TV needs a smaller percentage of measurements is not established by this comparison, because the two classes are not evaluated on the same ratios; a matched comparison at common ratios (e.g., 1.7%, 10%, 30%) with the same sensing operator is required before drawing that conclusion.","section":"3.4.1, Figs. 2 and 5"},{"comment":"The experimental protocol omits key details needed to judge the comparison. The 'L1 error' metric is not defined; the tuning parameters (τ, σ, λ, μ) for each algorithm are not reported; no error bars or multiple noise realizations are shown in Figure 7; and the image dimensions are not stated. Because the paper's value is a practical 'which algorithm should I use' guide, these omissions make the runtime and quality comparisons irreproducible and potentially non-representative. The author should provide a complete protocol, including parameter values and a direct link to the code.","section":"2.4, 3.4, 4"}],"minor_comments":[{"comment":"In Eq. (1), the approximation −(k/n) log(k/n) − (1−k/n) log(1−k/n) ≈ (k/n) log(k/n) is incorrect in sign; the entropy of the sparse Bernoulli source is positive and behaves as (k/n) log(n/k) for small k/n. The error propagates into the intuitive bound that follows.","section":"Eq. (1)"},{"comment":"In Eq. (4), the constraint y_k = (φ x')_k is inconsistent with the model in §2, where measurements are taken in the sampling basis ψ (i.e., y = ψ x); the sparsifying basis φ should appear in the objective, not in the measurement constraint. This is a significant clarity issue in a review intended for beginners.","section":"Eq. (4)"},{"comment":"The heading 'Nestorov's algorithm' should read 'Nesterov's algorithm' to match the reference to Nesterov in the bibliography.","section":"3.2"},{"comment":"The phrase 'compression ration' should be 'compression ratio'.","section":"2.4"},{"comment":"The phrase 'the algorithms inability' should be 'the algorithm's inability'.","section":"3.3"},{"comment":"Several figure cross-references appear as 'Figure ??', and the image dimensions are not given, which prevents quantitative interpretation of the reported m/n ratios and runtimes.","section":"Appendix A"},{"comment":"The code availability statement says the code is on the author's website but provides no URL; a persistent repository link is needed for reproducibility.","section":"A"},{"comment":"The sentence 'the measurement basis in incoherent with the basis' should read 'the measurement basis is incoherent with the basis'.","section":"1.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a tutorial with a flawed central comparison. The authors could repair the main claim by re-running the L1 experiments with a non-oracle formulation (e.g., minimizing ||Φx||_1 subject to y = Ψx) and using matched measurement ratios and identical sensing matrices for all algorithms. This is a substantial rewrite of the experimental section, but it is in principle feasible; hence I recommend major revision rather than rejection. If the editor judges that the tutorial content alone is insufficient for the journal's scope, rejection is also defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Yoni's paper is a readable walk through classical compressive sensing solvers, and the noise-robustness comparison is a genuinely useful addition. But the headline claim that TV beats L1 for natural images is not supported by the experiments as designed.\n\nThe tutorial part is solid. The explanations of OMP, GPSR, SPGL1, NESTA, and TVAL3 are clear and appropriately caveated. The author is honest that L1 under a sparsifying basis is not directly implementable in hardware. The TV-only benchmark (L1 Magic vs NESTA vs TVAL3) on a 35-image set is a reasonable quick-start guide, and the low-SNR comparison is the most interesting piece: NESTA holds structure where TVAL3 falls apart, which matters for LiDAR-style noise. That alone is worth a cite for people choosing between these two.\n\nThe soft spot is exactly where the reader's report points. In Section 2.4, the L1 solvers are fed y = Ψz with z = Φx computed as a pre-step. The author admits this is 'not applicable to a real-world system.' So the L1 simulations are not testing a realizable CS pipeline. Meanwhile the TV solvers in Section 3 use y = Ψx on the pixels directly. Different forward operators, different measurement ratios (L1 from 1.7% to 80%, TV from 0.17% to 30%). The conclusion in Section 5 that TV is more appropriate than L1 is therefore confounded twice over. The intuition about local versus global zeroing is interesting, but it is not established by these experiments.\n\nI'd also flag the lack of error bars and the unreachable code (the appendix says 'on the author's website' without a URL). For a benchmark paper that is a real omission. The noise experiment is a single realization per condition, I think, and no variance is shown.\n\nOverall, this is a useful tutorial with one flawed quantitative argument. The flaw is load-bearing only for the L1-vs-TV conclusion; the TV-only results and the noise comparison stand alone. A serious referee would catch the confound and ask for a redesign or a more modest claim. If the venue is a journal, I would not accept as is—but I would send it back rather than desk reject, because the tutorial and noise data are worth preserving. If the choice is only accept/reject, it's a reject in current form.","headline":"Readable tutorial with a useful noise comparison, but the headline L1-vs-TV conclusion is confounded by different forward models and measurement rates.","tokens_in":8559,"tokens_out":3035,"would_cite":false,"duration_ms":29993,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that total variation minimization reconstructs natural images better than L1 minimization on a sparsifying basis, with NESTA and TVAL3 as the most practical algorithms.","keywords":["compressive sensing","total variation","image reconstruction","LiDAR","NESTA","TVAL3","sparse recovery","measurement noise"],"falsifier":"Run the same image and SNR comparison with physically realizable measurements, such as random pixel masks on a digital micromirror device, instead of pre-computed transform coefficients; if L1-on-wavelets then matches or beats total variation, the paper's central conclusion is unsupported.","tokens_in":7606,"feed_emoji":"🖼️","tokens_out":7146,"duration_ms":65848,"temperature":0.7,"pith_summary":"This review argues that, for natural images, total variation (TV) minimization is the right reconstruction principle, outperforming L1 minimization under a sparsifying basis. The case rests on standardized simulations of off-the-shelf algorithms, including measurement noise matched to real LiDAR systems. The paper concludes that two solvers stand out: NESTA, which keeps the main image features at low signal-to-noise ratios, and TVAL3, which is the fastest for well-conditioned problems. The review is written as a quick-start guide, so a reader can pick a method without building custom software.","feed_headline":"Total variation beats L1 for compressed-sensing images","feed_subtitle":"Off-the-shelf solver comparison points to NESTA for noisy LiDAR-style measurements and TVAL3 when speed matters.","key_machinery":"The load-bearing mechanism is the total variation pseudo-norm, the L1 norm of the discrete gradient, which is nearly sparse for natural images and keeps reconstruction errors local instead of producing global interference patterns. NESTA works by applying Nesterov's accelerated gradient method with smoothing to reach a $\\mathcal{O}(1/k^2)$ convergence rate, which is what preserves image features under noise. TVAL3 works by minimizing an augmented Lagrangian with an alternating-direction non-monotone line search, which is what makes it fast. Both solvers rely only on matrix-vector products, so they can exploit fast measurement bases such as the Dragon wavelet group, computed in $O(n\\log n)$ time.","core_discovery":"The central discovery is that TV-regularized reconstruction of natural images needs far fewer measurements than L1 reconstruction on sparsified coefficients and produces visually better results at comparable L1 error. This is supported by controlled runs of L1 Magic, NESTA, TVAL3, GPSR, and SPGL1 on 35 images, with measurement ratios from about 0.17% to 80% and signal-to-noise ratios from 3 dB to 20 dB. In the noisy regime relevant to LiDAR, NESTA preserves overall structure at low SNR while TVAL3 fails; at high SNR, TVAL3 reconstructs roughly three times faster than NESTA and about ten times faster than L1 Magic. The paper therefore concludes that TV is more appropriate than L1 under a sparsifying basis and that the practical choice is between NESTA for noise tolerance and TVAL3 for speed.","pith_inferences":["A physically realizable sensor cannot measure transform coefficients directly, so the L1 comparison in this paper is a simulation convenience; a hardware test may show an even larger gap in favor of TV because the L1 route may not be implementable at all.","NESTA's low-SNR tolerance makes it a natural starting point for photon-starved imaging where each measurement is expensive; the noise model was chosen to match such data, so testing on the author's LiDAR setup is a direct next step.","The relative ranking of NESTA and TVAL3 could shift under a different sampling basis, since the paper never quantifies the choice of Dragon wavelets against alternatives such as scrambled Hadamard vectors."],"forward_implications":["For natural-image compressive sensing, a total-variation solver should be preferred over an L1 solver applied to sparsified coefficients.","In low-SNR measurement settings, NESTA is the safer default because its output retains major image structure that post-processing can improve.","In well-conditioned high-SNR settings, TVAL3 is the faster default, roughly three times faster than NESTA and ten times faster than L1 Magic.","Reconstruction succeeds with a small percentage of measurements, but quality degrades sharply near the information limit and at low SNR.","Image content matters as much as algorithm choice: noise-like images and images with sharp edges reconstruct with higher error under every method."],"supporting_citations":[{"why":"Supplies the NESTA algorithm and its accelerated first-order smoothing, one of the two solvers the paper recommends.","marker":"[2]"},{"why":"Gives the sparsity-incoherence theory that underlies the reconstruction guarantees the review builds on.","marker":"[3]"},{"why":"Provides the introductory compressive-sensing background and the incoherence condition for measurement bases.","marker":"[4]"},{"why":"Source of the Dragon wavelet measurement vectors used in all the paper's simulations.","marker":"[5]"},{"why":"Provides compressive-sensing theory and the TVAL3 augmented-Lagrangian formulation used for the TV comparisons.","marker":"[6]"},{"why":"Presents the TVAL3 algorithm itself, the fast solver recommended for well-conditioned problems.","marker":"[7]"},{"why":"Introduces Nesterov's optimal gradient method whose convergence rate makes NESTA's noise tolerance possible.","marker":"[8]"},{"why":"Supplies the LiDAR noise and low-SNR conditions the noisy reconstruction scenarios are designed to match.","marker":"[11]"}],"fun_headline_variants":["TV beats L1 for compressed sensing images","NESTA for noisy LiDAR, TVAL3 for speed","Compressed sensing review: TV wins","TV needs fewer measurements for images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The L1-versus-TV comparison assumes a sensing system can directly measure the sparse transform coefficients $z=\\varphi x$; if a real camera cannot, the L1 experiments do not test a usable reconstruction method.","fun_headline_variants_meta":{"raw":{"variants":["TV beats L1 for compressed sensing images","NESTA for noisy LiDAR, TVAL3 for speed","Compressed sensing review: TV wins","TV needs fewer measurements for images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000343,"raw_usage":{"total_tokens":1817,"prompt_tokens":805,"completion_tokens":1012,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":955}},"tokens_in":421,"tokens_out":1012,"duration_ms":9620,"temperature":1.0,"reasoning_tokens":955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:07:05.834156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same image and SNR comparison with physically realizable measurements, such as random pixel masks on a digital micromirror device, instead of pre-computed transform coefficients; if L1-on-wavelets then matches or beats total variation, the paper's central conclusion is unsupported.","supporting_citations":[{"cited_title":"Nesta: a fast and accurate first-order method for sparse recovery","cited_arxiv_id":null,"evidence_quote":"Supplies the NESTA algorithm and its accelerated first-order smoothing, one of the two solvers the paper recommends."},{"cited_title":"Sparsity and incoherence in compressive sampling","cited_arxiv_id":null,"evidence_quote":"Gives the sparsity-incoherence theory that underlies the reconstruction guarantees the review builds on."},{"cited_title":"An introduction to compressive sam- pling","cited_arxiv_id":null,"evidence_quote":"Provides the introductory compressive-sensing background and the incoherence condition for measurement bases."},{"cited_title":"Power-efficient cameras using natural image statistics","cited_arxiv_id":null,"evidence_quote":"Source of the Dragon wavelet measurement vectors used in all the paper's simulations."},{"cited_title":"Compressive sensing for quantum imaging","cited_arxiv_id":null,"evidence_quote":"Provides compressive-sensing theory and the TVAL3 augmented-Lagrangian formulation used for the TV comparisons."},{"cited_title":"An efficient augmented lagrangian method with applications to total variation minimization.Computational Optimization and Applications , 56(3):507–530, 2013","cited_arxiv_id":null,"evidence_quote":"Presents the TVAL3 algorithm itself, the fast solver recommended for well-conditioned problems."},{"cited_title":"A method for unconstrained convex minimization problem with the rate of convergence o (1/k2)","cited_arxiv_id":null,"evidence_quote":"Introduces Nesterov's optimal gradient method whose convergence rate makes NESTA's noise tolerance possible."},{"cited_title":"Low intensity lidar using compressed sensing and a photon number resolving detector","cited_arxiv_id":null,"evidence_quote":"Supplies the LiDAR noise and low-SNR conditions the noisy reconstruction scenarios are designed to match."}],"review_version":1}