{"id":"fb662531-382a-41f0-85c0-fa6c533d7625","arxiv_id":"2608.12982","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network can generate binary sensing matrices with lower mutual coherence than random matrices by optimizing coherence-based losses without any training data.","lead":"This paper trains a small neural network to generate binary sensing matrices for compressed sensing, using a loss function based on mutual coherence instead of training data. The resulting matrices show lower coherence than random Gaussian or Bernoulli matrices in simulations, but the paper does not test signal recovery.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LogSumExp surrogate in Section 5.3 omits the absolute value, so the training loss approximates the maximum signed Gram entry rather than the maximum absolute coherence reported in Table 1; this mismatch is load-bearing and should be tested directly.","rationale":"The reader's weakest assumption is exactly the missing absolute value in the LogSumExp loss, and I agree that it is the most load-bearing concern. The central claim is quantitative: the proposed binary sensing matrix achieves maximum mutual coherence about 0.281 versus 0.481 for Gaussian and 0.496 for Bernoulli (Table 1). That number is the direct output of an optimization whose loss is supposed to approximate max |Gij| but actually approximates max Gij. The asymmetry matters because binary columns produce both positive and negative Gram entries and the reported metric is the maximum of their absolute values. The Lp-norm term does contain absolute values, so this is not an immediate logical contradiction; however, with p=8 and ξ=30, the LSE term is the sharper surrogate and it selectively ignores large negative entries. A simple retraining experiment with an absolute-value LSE settles whether this actually changes the result. If the original worst-case entries are predominantly negative and the corrected loss reduces the maximum coherence, then the paper's headline comparison is not a fair test of its own objective. If the values are unchanged, the concern is a documentation issue rather than a correctness issue. I keep the reader's CONDITIONAL verdict because the concern is concrete and testable but does not, on its own, prove the empirical claim false; it identifies a specific condition that must be checked before the claim is relied upon. Missing recovery experiments and the weak baselines are additional concerns, but they are secondary because the paper's explicit contribution is coherence reduction, and the LSE mismatch directly threatens the reliability of the reported coherence values.","tokens_in":12826,"tokens_out":19471,"duration_ms":220871,"concrete_test":"Re-run the 20 training runs with identical latent vectors, architecture, and hyperparameters (α1=3, α2=1, α3=0.5, p=8, ξ=30), but replace LLogSum with (1/ξ) log Σ_{i≠j} exp(ξ|Gij|). Compare the final max |Gij|, the sign of the worst-case Gram entry, and the average/total coherence against Table 1. If the corrected loss lowers the max coherence by more than the run-to-run standard error, or if the original worst-case entries are predominantly negative, the misspecified surrogate is load-bearing and the reported claim requires revision; if the values are statistically unchanged, the concern is minor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 defines LLogSum(Aθ) = (1/ξ) log Σ_{i≠j} exp(ξ Gij) and says it converges to the maximum as ξ→∞. But the coherence metric used in Eq. (4) and Table 1 is max_{i≠j} |Gij|. The LSE surrogate converges to max_{i≠j} Gij. For a binary ±1/√M matrix, negative Gram entries are common; with ξ=30, a pair with Gij = -0.28 contributes e^{-8.4} ≈ 0.00022 to the sum, while the same positive value contributes e^{8.4} ≈ 4400. The loss is therefore nearly blind to large negative correlations, which are exactly the worst-case absolute coherences. The Lp-norm term (α1=3, p=8) does include absolute values and partially compensates, but the LSE term (α2=1, ξ=30) dominates the gradient for positive entries and creates an asymmetry that the Lp term cannot fully correct. Because the paper provides no sparse-recovery experiments or other independent validation, the only evidence for the central claim is the coherence of the produced matrix, and that matrix is produced by minimizing a surrogate that does not target the reported quantity. This concern is empirical rather than purely formal: if the final worst-case entries happen to be positive in all runs, the omission may be harmless, but the paper does not report the sign of the argmax entries.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-free neural-network framework for designing binary sensing matrices (entries ±1/√M) with low mutual coherence. A small fully connected network maps M-dimensional latent vectors to columns of the sensing matrix; a sign activation with a straight-through estimator enforces the binary constraint. The training objective combines an Lp-norm penalty on off-diagonal Gram entries, a LogSumExp approximation to the maximum off-diagonal entry, and a tight-frame penalty. Experiments for (M,N) = (64,128), (64,256), (128,256) report lower maximum, average, and total mutual coherence than Gaussian and Bernoulli random matrices, and the paper argues this leads to better sparse-signal recovery via the Donoho–Elad bound.","tokens_in":13184,"tokens_out":4854,"duration_ms":47253,"significance":"If substantiated, the method offers a lightweight, storage-efficient alternative to random sensing matrices, with the notable feature of learning from a mathematical property rather than from data. The shared-column generator is simple and the coherence gains over random matrices are non-trivial. However, the current evaluation is limited to coherence values that are closely related to the training objective, with no sparse-recovery experiments and no comparison against established deterministic or optimization-based construction methods; hence the significance for compressive sensing practice remains to be established.","major_comments":[{"comment":"The LogSumExp loss is defined as LLogSum(Aθ) = (1/ξ) log Σ_{i≠j} exp(ξ Gij), which as ξ→∞ converges to max_{i≠j} Gij, not to the maximum absolute mutual coherence max_{i≠j} |Gij| used in Eq. (4) and Table 1. With ξ=30, a negative off-diagonal entry of magnitude 0.28 contributes e^{-8.4}≈2.2e-4 to the sum whereas a positive entry of the same magnitude contributes e^{8.4}≈4.4e3, so the gradient is overwhelmingly dominated by positive entries and the worst negative correlations are effectively invisible to this loss term. Since the reported maximum coherence is absolute, this is a load-bearing mismatch. The authors should either replace Gij by |Gij| inside the LSE (or use an equivalent sign-symmetric formulation), or report the sign of the largest off-diagonal entries in the Table 1 runs to demonstrate that the omission does not affect the final matrices.","section":"Section 5.3, LLogSum definition"},{"comment":"The evaluation is circular in the sense that the reported metric µ(A) is also the quantity that the combined loss is explicitly designed to reduce, and Section 6.2 states that hyperparameters were selected by monitoring µ(A). Consequently, Table 1 largely confirms that the optimization achieved its objective. To support the claim that the learned matrices 'consistently outperform' random matrices for compressive sensing, the paper needs independent validation: sparse-signal recovery experiments with Basis Pursuit or OMP on synthetic and real data, comparisons against deterministic or optimized constructions (e.g., Elad's coherence minimization, alternating projection methods, or ETF-based bounds), and error bars or standard deviations over the 20 independent runs. Without these, the lower coherence numbers cannot be interpreted as improved recovery performance.","section":"Section 6.4, Table 1"},{"comment":"The paper repeatedly claims that the low-coherence matrices enable 'perfect recovery' of sparse signals (Abstract, Section 7), but no recovery experiment is reported anywhere. The only indirect link is the Donoho–Elad bound in Eq. (5), but the paper never checks whether the actual sparsity level k for which recovery is guaranteed is relevant to any realistic signal model, nor whether the learned matrices actually recover signals in practice. A small set of recovery comparisons (e.g., phase transitions for Basis Pursuit) would make the claim testable.","section":"Abstract and Section 7"}],"minor_comments":[{"comment":"The statement 'Finally, the generated vectors are transposed and stacked column-wise to construct the sensing matrix of size 64×256' contradicts the experimental matrix dimensions (M,N) = (64,128) and (128,256) used in Section 6.4; please clarify whether this line refers only to one configuration.","section":"Section 5.2"},{"comment":"The sentence 'Instead of considering maximum absolute value of G, we considered smoother approximation to the maximum' is ambiguous; unless |Gij| is inserted, the LSE is not an approximation of the absolute maximum, so the text should be corrected to match the definition.","section":"Section 5.3, LogSumExp paragraph"},{"comment":"The definition of total mutual coherence, µtotal(A) = Σ_{i≠j} |gij|, is missing parentheses and does not clearly state that the diagonal is excluded; the same applies to the average coherence formula.","section":"Section 2.2"},{"comment":"All numerical entries are prefixed with '≈' and no standard deviations are reported despite 20 independent runs; please provide mean ± std and the number of seeds.","section":"Table 1"},{"comment":"The claim that the use of a mathematical property as a loss function is 'for the first time in literature' is an overstatement; coherence-based penalties appear in earlier dictionary-learning and sensing-matrix optimization literature (e.g., Elad 2007, Duarte-Carvajalino and Sapiro 2009), so the novelty should be more narrowly stated.","section":"Introduction and Conclusion"},{"comment":"The complexity comparison states that the network storage is 'almost equal to that of the random sensing matrix,' but this is only true for one specific N; for larger N the fixed-size network would be smaller, and for very small N it would be larger. Please state the storage cost as a function of M and N.","section":"Section 6.3"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and could be publishable after the load-bearing issues are addressed: the missing absolute value in the LogSumExp loss and the absence of sparse-recovery experiments. I would also recommend that the authors verify the novelty claim in the broader context of sensing-matrix optimization, since coherence-based objective functions have been used before."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely neat idea — a small neural net generates binary sensing matrices column-by-column from a shared rule, trained with no data, only a coherence-based loss. That combination is not in the paper's own cited literature. The experiments, taken at face value, show the learned matrices have lower max/average/total coherence than Gaussian and Bernoulli baselines, and the convergence plots look plausible. Credit is due: it is a simple, lightweight, hardware-friendly construction, and the storage argument is real.\n\nThe soft spots are real too, and one is load-bearing. The LogSumExp surrogate in Section 5.3 is written without absolute values: L = (1/ξ) log Σ exp(ξ Gij). It converges to max Gij, not max |Gij|. With ξ=30, a negative Gram entry of -0.28 contributes e^{-8.4} ≈ 0.0002 to the sum; the same positive entry contributes e^{8.4}. So the loss is nearly blind to large negative correlations — exactly the worst-case absolute values. The Lp-norm term does include absolute values and partially compensates, but the LSE term dominates the gradient for positive entries. The paper doesn't report the sign of the argmax entries, so we can't tell if this matters in their runs; it needs to be tested directly. This is not a purely formal nit, because the only claim supported by experiments is coherence, and the reported coherence is the training objective via surrogates.\n\nOther soft spots, in order: no sparse-recovery experiments despite claiming perfect recovery; baselines are only Gaussian and Bernoulli (no Elad-style optimized or deterministic matrices, no Welch-bound comparison); no error bars or standard deviations in Table 1 despite 20 runs; hyperparameter selection used the evaluation metric; and the number of network parameters is small but not actually reported, so the storage claim is asserted, not quantified. The novelty claim \"first time a mathematical property defines a loss\" is also overstated — coherence-based losses in sensing matrix learning are exactly that.\n\nNone of this kills the core idea. It's a modest but useful construction, and the missing recovery experiments and surrogate fix are addressable in revision. I'd send it to peer review; a knowledgeable referee can push for those. If the sign of the argmax is always positive in their runs, the abs-value issue is harmless in practice, but they should show that.","headline":"A data-free neural network for binary sensing matrices with lower coherence than random baselines is a solid, modest idea, held back by a surrogate loss that omits absolute values and by missing recovery experiments.","tokens_in":13655,"tokens_out":1971,"would_cite":false,"duration_ms":20019,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A lightweight neural network, trained only on the mutual-coherence property, generates binary sensing matrices with lower worst-case column correlation than Gaussian or Bernoulli random matrices.","keywords":["compressive sensing","sensing matrix design","mutual coherence","binary sensing matrix","property-driven learning","neural network","Gram matrix","random matrix"],"falsifier":"Retrain the same generator with the LogSumExp term changed to $\\frac{1}{\\xi}\\log\\sum_{i\\ne j} e^{\\xi |G_{ij}|}$; if the maximum absolute coherence falls below 0.281 at $M=64$, $N=128$, the original loss was not minimizing mutual coherence. Separately, run OMP or Basis Pursuit on random $k$-sparse signals with the learned matrix; failure near the sparsity level allowed by the coherence bound would show the claimed link to perfect recovery is not demonstrated.","tokens_in":12658,"feed_emoji":"🧠","tokens_out":8487,"duration_ms":75993,"temperature":0.7,"pith_summary":"The paper sets out to show that a simple neural network can design a binary sensing matrix by learning from the mutual-coherence property alone, without any signal data or application-specific training. This matters because random sensing matrices are dense and costly to store, while data-driven learned matrices depend on large datasets and lose theoretical guarantees. The reported result is that the learned matrix achieves a maximum mutual coherence of about 0.281 for $M=64$, $N=128$, compared with about 0.481 for a Gaussian random matrix and 0.496 for a Bernoulli random matrix, with similar gains at other dimensions. The construction works through a shared rule: one small network maps different latent vectors to every column, and the sign of the output gives the binary entries. The paper claims this property-driven matrix design supports perfect recovery of sparse signals, although no recovery experiment is reported.","feed_headline":"Neural net cuts sensing-matrix coherence by ~40 percent","feed_subtitle":"Trained only on coherence, it beats Gaussian and Bernoulli random matrices on worst-case correlation.","key_machinery":"The central object is the shared-rule neural generator together with a coherence surrogate loss. A small feedforward network maps each latent vector $z_j$ to a real vector, and the sign function turns it into a $\\{-1/\\sqrt{M}, 1/\\sqrt{M}\\}$ column; because the same network weights produce all $N$ columns, the matrix is reproducible and can be stored as a few network parameters. The loss approximates the non-differentiable maximum absolute off-diagonal Gram entry with an $\\ell^p$ norm, a LogSumExp smooth maximum, and a tight-frame term that pushes $AA^T$ toward $\\frac{N}{M}I$, steering the binary matrix toward the incoherent structure that the classical coherence-to-sparsity bound connects to exact recovery.","core_discovery":"The central claim is that the mutual incoherence property can serve as the entire training signal for sensing-matrix design, replacing both random draws and data-driven optimization. The learned matrix $A_\\theta$ is generated column-wise by $a_j = \\frac{1}{\\sqrt{M}}\\operatorname{sgn}(\\mathcal{N}_\\theta(z_j))$, so the network acts as a compact, reproducible generator rather than a per-entry search. Training minimizes a combined Gram-matrix loss: an $\\ell^p$ norm over off-diagonal entries, a LogSumExp smooth-max term, and a tight-frame penalty $\\|AA^T - \\frac{N}{M}I\\|_F^2$. In the reported experiments, these three terms drive maximum mutual coherence down to about 0.281 from about 0.481 (Gaussian) and 0.496 (Bernoulli) at $M=64$, $N=128$, with average and total coherence also dropping across all tested sizes. The paper takes these results as evidence that a property-focused, data-free learning framework yields better incoherence guarantees than conventional random constructions.","pith_inferences":["The same template could be applied to other matrix properties, converting theoretical guarantees such as RIP or null-space property surrogates into trainable loss functions.","If the LogSumExp term were replaced with $\\frac{1}{\\xi}\\log\\sum_{i\\ne j} e^{\\xi |G_{ij}|}$, retraining might yield a lower maximum absolute coherence than the reported 0.281, which would indicate the current loss is not minimizing exactly what the tables report.","The paper's inference that lower coherence supports perfect recovery relies on the coherence-to-sparsity bound, but actual recovery was never measured; running OMP or Basis Pursuit on random sparse signals would settle whether the coherence gains translate into reconstruction gains.","Since the generator is continuous in the latent vectors, the same trained network could produce sensing matrices of other dimensions by changing the number or size of the latent inputs, a scalability claim the paper leaves implicit."],"forward_implications":["If the coherence gains persist at other sizes, compressive sensing systems can store the network weights instead of a dense random matrix and regenerate the sensing matrix on demand.","Lower maximum coherence widens the sparsity range for which recovery is guaranteed by the classical bound $k < \\frac{1}{2}(1/\\mu(A)+1)$, so the reported values promise recovery of less sparse signals than random matrices of the same dimensions.","Binary entries turn acquisition and reconstruction operations into additions and sign flips, which simplifies hardware implementation and reduces storage costs.","Because training uses no signal data, the resulting matrix is not tied to any signal class, so the same generator could be reused across applications without retraining.","A deterministic shared rule makes the construction reproducible: the same network parameters always produce the same sensing matrix."],"supporting_citations":[{"why":"Supplies the bound connecting sparsity level to mutual coherence, which is the theoretical reason the loss targets coherence.","marker":"[18]"},{"why":"Provides the definition of mutual coherence that the Gram-matrix loss approximates.","marker":"[15]"},{"why":"Shows verifying RIP is NP-hard, justifying the paper's choice of coherence as a tractable training target.","marker":"[31]"},{"why":"Establishes the restricted-isometry guarantee for random Gaussian and Bernoulli matrices, the baselines the learned matrix is compared against.","marker":"[3]"},{"why":"Represents the data-driven sensing-learning approach the paper contrasts with its property-based, data-free training.","marker":"[26]"},{"why":"Represents a deep-unrolling framework whose dependence on large training data the method aims to avoid.","marker":"[36]"}],"fun_headline_variants":["Data-free neural net builds low-coherence binary matrices","Coherence property alone trains sensing-matrix generator","Neural net cuts sensing coherence without data or apps","Binary sensing matrices learned from mutual coherence loss","No data, just coherence: neural net designs sensing matrices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the training loss actually shrinks the worst column correlation; the LogSumExp term omits the absolute-value bars, so a strongly negative off-diagonal Gram entry would not be penalized, and the reported coherence gains may come from a loss that is not minimizing what the paper claims.","fun_headline_variants_meta":{"raw":{"variants":["Data-free neural net builds low-coherence binary matrices","Coherence property alone trains sensing-matrix generator","Neural net cuts sensing coherence without data or apps","Binary sensing matrices learned from mutual coherence loss","No data, just coherence: neural net designs sensing matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3542,"prompt_tokens":1068,"completion_tokens":2474,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":684,"tokens_out":2474,"duration_ms":17516,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:12:37.365070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same generator with the LogSumExp term changed to $\\frac{1}{\\xi}\\log\\sum_{i\\ne j} e^{\\xi |G_{ij}|}$; if the maximum absolute coherence falls below 0.281 at $M=64$, $N=128$, the original loss was not minimizing mutual coherence. Separately, run OMP or Basis Pursuit on random $k$-sparse signals with the learned matrix; failure near the sparsity level allowed by the coherence bound would show the claimed link to perfect recovery is not demonstrated.","supporting_citations":[{"cited_title":"Optimally sparse representation in general (nonorthogonal) dictionaries viaℓ 1 minimization.Proceedings of the National Academy of Sciences, 100(5):2197–2202, 2003","cited_arxiv_id":null,"evidence_quote":"Supplies the bound connecting sparsity level to mutual coherence, which is the theoretical reason the loss targets coherence."},{"cited_title":"Introduction to compressed sensing., 2012","cited_arxiv_id":null,"evidence_quote":"Provides the definition of mutual coherence that the Gram-matrix loss approximates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows verifying RIP is NP-hard, justifying the paper's choice of coherence as a tractable training target."},{"cited_title":"A simple proof of the restricted isometry property for random matrices.Constructive approxi- mation, 28:253–263, 2008","cited_arxiv_id":null,"evidence_quote":"Establishes the restricted-isometry guarantee for random Gaussian and Bernoulli matrices, the baselines the learned matrix is compared against."},{"cited_title":"A deep learning approach to structured signal recovery","cited_arxiv_id":null,"evidence_quote":"Represents the data-driven sensing-learning approach the paper contrasts with its property-based, data-free training."},{"cited_title":"Ista-net: Interpretable optimization-inspired deep network for image compressive sensing","cited_arxiv_id":null,"evidence_quote":"Represents a deep-unrolling framework whose dependence on large training data the method aims to avoid."}],"review_version":1}