{"id":"7d2c04fe-922e-469c-90fd-59e8716969d6","arxiv_id":"2607.20020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Shearlets' superior sparsity does not yield a proportional reduction in Fourier sample complexity; the required measurements scale comparably to wavelets, improving at best logarithmically.","lead":"Shearlets represent cartoon-like images with far fewer coefficients than wavelets, but this paper finds that in Fourier compressed sensing this sparsity advantage does not reduce the number of measurements needed—either in theory or in numerical experiments. The result matters because it separates approximation efficiency from sampling efficiency, warning against assuming that better sparsity automatically means faster acquisition in imaging methods like MRI.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quadratic m_shear ~ s_shear^2 scaling in §3.4 depends on the local coherence lower bound of Remark 4, which is a heuristic; if a feasible shearlet frame achieved wavelet-like O(||n||^{-1}) coherence, the conclusion would flip.","rationale":"The reader's weakest_assumption identifies the same load-bearing step: the 'essentially optimal' claim in Remark 4 is a heuristic, not a theorem, and the §3.4 comparison requires that the local coherence upper bound from Proposition 2 be tight. I agree that this is the most vulnerable point in the theoretical argument. If the true local coherence for some feasible cone-adapted shearlet frame decayed faster than ||n||_8^{-3/4}, the computed ||ω_shear||_2^2 would shrink, potentially changing m_shear from s_shear^2 to s_shear and invalidating the paper's central negative theoretical conclusion. The numerical phase diagrams provide independent empirical support for the qualitative conclusion, but they are not a substitute for the missing optimality proof. Given this, the reader's CONDITIONAL verdict is appropriate: the paper is well-argued and the numerics are suggestive, but the central theoretical claim should not be taken as definitive until Remark 4 is made rigorous (or replaced with an explicit counterexample). No change to the reader's verdict is needed.","tokens_in":30086,"tokens_out":20640,"duration_ms":227029,"concrete_test":"For each of the two concrete systems used in the paper — the bandlimited Parseval shearlets of Theorem 4 and the ShearLab compactly supported shearlets — choose a fixed direction ρ∈(-1,1) and define n_j=(round(c2^j), round(ρ2^j)) with c such that n_j lies in the horizontal cone. Compute μ_j = sup_i |ψhat_i(n_j)| for j=1,...,20. If liminf_j μ_j / 2^{-3j/4} = 0, the local coherence decays faster than the rate assumed in §3.4 and the quadratic m_shear conclusion is not valid for that system; if the ratio is bounded below, verify whether the LP sum lower bound can be converted rigorously into a per-element lower bound. Ideally, prove a lemma: for bandlimited tight shearlets, sup_i |ψhat_i(n)| ≥ c 2^{-3j/4} for all n in the cone, using the fact that only O(1) scales and shear parameters contribute to the LP sum at a fixed frequency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2 proves an upper bound µloc_k ≲ (1+||n_k||_8)^{-3/4}. The §3.4 computation turns this into m_shear ~ s_shear^2 only if the bound is tight. The matching lower bound is Remark 4, which argues that the Littlewood-Paley lower bound on Σ_i |ψhat_i(n)|^2 implies a single shearlet at the dominant scale has |ψhat_i(n)| ≳ 2^{-3j/4}. This implication is not automatic: the LP sum runs over all scales and shear parameters, and for compactly supported (non-bandlimited) generators contributions from several scales/shears can accumulate, so the supremum over i could in principle decay faster than 2^{-3j/4} while the sum remains bounded below. In that case ||ω_shear||_2^2 would be smaller than the 2^{j0/2} used in §3.4; if the true decay were O(||n||_8^{-1}), the §3.4 estimate would become m_shear ~ s_shear, overturning the central theoretical claim. Thus the central negative scaling rests on an unproven optimality assertion, not on a theorem with hypotheses covering all feasible cone-adapted shearlet frames in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether the superior sparsity of shearlets for cartoon-like images translates into fewer Fourier measurements in compressed sensing. The theoretical part proves an upper bound for the local coherence of cone-adapted shearlets with Fourier measurements (Proposition 2), yielding decay O(||n||_∞^{-3/4}), which is slower than the wavelet decay O(||n||^{-1}) recalled in Remark 3; it discusses the localization factor and shows that controlling it requires lower frame bounds that may be exponentially small; and, assuming the coherence bound is optimal and ignoring localization/balancing, it derives m_shear ≈ 2^{j0} versus m_wave ≈ j0 2^{j0}, i.e. only a logarithmic improvement. The numerical part compares stationary-wavelet and ShearLab shearlet sparsification on synthetic disk images; it confirms better shearlet N-term approximation but phase diagrams show roughly comparable sample complexity.","tokens_in":30489,"tokens_out":9871,"duration_ms":96103,"significance":"A fully established negative result would be valuable: it would show that asymptotic approximation-rate advantages do not automatically translate into sample-complexity gains under coherence-based compressed sensing with subsampled Fourier data, and it identifies concrete technical bottlenecks (localization factor, lower frame bounds). The paper is unusually transparent about its optimistic assumptions. However, the central theoretical scaling rests on the unproven optimality lower bound in Remark 4; without it, the comparison in Section 3.4 is only an upper-bound heuristic. The numerical experiments are suggestive but lack confidence intervals and a noise-level specification.","major_comments":[{"comment":"The conclusion m_shear ≈ s_shear^2 rests on treating Proposition 2's upper bound μloc_k ≲ (1+||n_k||_∞)^{-3/4} as tight. Remark 4 is only a heuristic: the Littlewood-Paley lower bound on Σ_i |ψhat_i(n)|^2 does not imply a single shearlet at the dominant scale satisfies |ψhat_i0(n)| ≳ 2^{-3j/4}; several scales/shears could contribute to the sum while each individual Fourier transform decays faster than 2^{-3j/4}. If the true worst-case decay were O(||n||_∞^{-1}) (compatible with Proposition 2), then ||ω_shear||_2^2 would be O(j0) and the Section 3.4 estimate would give m_shear ≈ j0 2^{j0/2}, a substantial improvement over m_wave ≈ j0 2^{j0}. A matching lower bound for the specific shearlet frames covered by Proposition 2, or for the bandlimited Parseval frame of Theorem 4, is needed before the quadratic-in-sparsity claim is stated.","section":"§3.4 and Remark 4"},{"comment":"The numerical experiments are not fully reproducible because the noise level σ is never specified. Section 4.2 defines Gaussian noise with variance σ² and sets ε_Ω = σ√|Ω|, but no value of σ is reported for the phase diagrams. Since the success/failure thresholds ε_a=25 and ε_r=0.15 are fixed, the phase transition locations depend on σ. In addition, T=6 trials per grid point makes each displayed probability a multiple of 1/6; no confidence intervals or error bars are given, so the 'roughly factor 2' advantage is not quantitatively supported. Reporting σ, the number of trials, and per-point standard errors (or at least a few bootstrap intervals) is necessary to substantiate the empirical claim.","section":"§4.2–4.3.2"},{"comment":"The comparison assumes η_s,D≈1 and that the balancing property/localization factor do not degrade the shearlet bound. These are explicitly acknowledged as optimistic, but the paper states the result as 'even under optimal theoretical conditions.' As Section 3.3 shows, no frame satisfying all the hypotheses of Theorem 1 (Parseval, good NLA rate, and η_s,D=O(1)) is identified. The statement should be phrased as conditional on these unverified assumptions, and the abstract's 'even under optimal theoretical conditions' should be qualified accordingly, unless such a frame is exhibited or a matching lower bound on η_s,D is provided.","section":"§3.4"}],"minor_comments":[{"comment":"The text says 'bounded from below by an expression of the form C j0' but Theorem 8 gives an exponential lower bound of the form C^{j0}; the notation should be corrected.","section":"Remark 7"},{"comment":"Figure 2's y-axis label contains a typo ('mininum'). Figures 2 and 3 would benefit from error bars or confidence bands, especially because the phase diagrams use only T=6 trials per grid point.","section":"§4.3"},{"comment":"The sentence 'with ρ% different Fourier measurements' is unclear; ρ is only defined later as 100m/K². Please define ρ before using it.","section":"§4.1.2"},{"comment":"The shearlet system uses scales up to 2 while the stationary wavelet transform uses 4 levels. Although the total coefficient counts are matched, the finest scale/resolution differs; a sentence justifying that this does not bias the comparison would be useful.","section":"§4.1.1"},{"comment":"Equation (15) omits the log factors that appear in Theorem 1; this is acceptable for a scaling comparison but should be stated explicitly near the equation.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the unproven optimality of the coherence bound in Remark 4. If the authors can prove a matching lower bound for a natural shearlet frame (e.g., the bandlimited Parseval shearlets of Theorem 4), the paper would be a solid negative result. Otherwise, they should substantially soften the theoretical conclusion and make the numerical section reproducible by reporting the noise level and adding uncertainty estimates. I lean major revision rather than reject because the framework and upper-bound results are valuable and the issue is fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first direct comparison I know of shearlets versus wavelets for Fourier compressed sensing sample complexity, with both a theoretical obstruction and systematic phase diagrams. Second, the headline theoretical scaling m_shear ~ s_shear^2 is not fully proven: it depends on Remark 4, a Littlewood-Paley heuristic asserting that the local coherence upper bound of Proposition 2 is tight. The upper bound is rigorous; the matching lower bound is not.\n\nWhat the paper does well: Proposition 2 is a clean and careful coherence decay bound for cone-adapted shearlets. The localization-factor analysis in Section 3.3 is a genuine contribution, showing explicitly why standard frame-based recovery theorems stall on shearlet redundancy—lower frame bounds for subframes can be exponentially small. The numerical study is also honest and useful: it confirms the sparsity advantage of shearlets on cartoon-like images but shows only a factor-of-2 sample-complexity gain, consistent with the qualitative claim that better sparsity does not automatically buy proportional sampling savings.\n\nThe soft spots, in proportion: the stress-test concern is on point. The §3.4 quadratic scaling holds only if the coherence lower bound of Remark 4 is true. The Littlewood-Paley sum over scales and shear parameters does not automatically force a single shearlet Fourier transform to decay that slowly; in principle a feasible shearlet frame could have faster decay, which would flip the theoretical comparison. The authors are transparent that Remark 4 is a heuristic, but the abstract and conclusions state the quadratic scaling as the theoretical result, which overstates what is actually proven. The numerics also have limitations: T=6 trials per phase point, no confidence intervals, the noise level is not reported, and no code is supplied. These are not fatal, but they mean the theoretical claim should be read as conditional.\n\nWho this is for: anyone working on structured compressed sensing, shearlets, or sampling theory for imaging. It deserves a serious referee: the question is important enough and the partial results are solid enough to warrant careful checking. The referee should ask the authors to either prove the optimality lower bound or soften the theoretical conclusion and present the numerics as the primary evidence. I would send it to review.","headline":"The negative result on shearlet Fourier sample complexity is likely right but its theoretical core leans on a heuristic optimality claim; the numerics are the stronger evidence.","tokens_in":30868,"tokens_out":2457,"would_cite":true,"duration_ms":30264,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["42C15","94A20","65T60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Shearlets' superior image sparsity does not translate into fewer Fourier samples for compressed sensing.","keywords":["compressed sensing","shearlets","wavelets","sample complexity","local coherence","Fourier measurements","frame theory","cartoon-like images"],"falsifier":"Exhibit a cone-adapted shearlet frame whose elements satisfy |ψhat_i(n)| ≤ C (1+||n||)^{-1} on integer grids while retaining the frame and approximation properties; the paper's comparison then gives m_shear ~ s_shear, overturning the quadratic conclusion. Alternatively, produce a phase-diagram study on cartoon-like images in which shearlets require an order of magnitude fewer Fourier samples than wavelets at matched recovery quality.","tokens_in":30019,"feed_emoji":"📡","tokens_out":4879,"duration_ms":48557,"temperature":0.7,"pith_summary":"This paper asks whether the known advantage of shearlets in sparsely representing images—an N-term approximation error of order N^{-2} log^3 N versus N^{-1} for wavelets—also buys a proportional reduction in the number of Fourier measurements needed for compressed sensing. The answer it argues is no. The local coherence between shearlets and Fourier exponentials decays only as (1+||n||)^{-3/4}, slower than the wavelet decay (1+||n||)^{-1}; this slower decay puts a quadratic-in-sparsity term in the sample-complexity bound even under the most optimistic assumptions about frame redundancy. Numerically, phase diagrams on piecewise-smooth test images confirm that shearlets need only a minor (roughly factor-two) measurement advantage over wavelets, far below the sparsity gain. A sympathetic reader should care because it delimits when directional representation systems can actually accelerate Fourier-sampling modalities such as MRI.","feed_headline":"Shearlets' sparsity edge does not cut Fourier sample counts","feed_subtitle":"Phase diagrams show shearlets need about as many Fourier measurements as wavelets despite far sparser representations.","key_machinery":"The two quantities that carry the argument are the local coherence between the sparsifying frame and the Fourier measurement basis, and the localization factor η_{s,D} that measures the redundancy of the frame. For cone-adapted shearlets, the paper proves a decay bound |ψhat_i(n)| ≤ C ||n||^{-3/4} for the Fourier transform on integer grids, and argues this rate is essentially optimal via a standard Littlewood-Paley heuristic; for wavelets the corresponding decay is ||n||^{-1}. The difference in decay exponents is what turns the ideal linear sample complexity m ~ s into m ~ s^2 for shearlets. The localization factor, which would require lower frame bounds for arbitrary finite shearlet subfram","core_discovery":"The central claim is that, for subsampled Fourier compressed sensing, shearlets do not deliver the sample-complexity reduction their approximation rates suggest. For cartoon-like images, shearlets have effective sparsity s_shear while wavelets have s_wave approximately s_shear^2; however the general frame recovery bound m ≳ η^2 ||ω||_2^2 s, combined with the essentially optimal local-coherence bound |ψhat_i(n)| ≲ ||n||^{-3/4} for cone-adapted shearlets, yields m_shear ≳ s_shear^2 even when the localization factor η is taken to be of order one. Comparing this with the wavelet bound m_wave ≳ s_wave yields only a logarithmic (resolution-dependent) improvement for shearlets; the paper's phase-di","pith_inferences":["One testable extension: if a cone-adapted shearlet family can be engineered so that its generators decay like wavelets (exponent -1) at integer frequencies, the paper's own comparison would predict m_shear ~ s_shear, restoring a proportional advantage; the remaining obstacle would then be the localization factor.","The paper's analysis suggests a broader design principle: for structured Fourier measurements, the sampling distribution should be matched to the coherence profile of the dictionary, and dictionaries with excellent approximation rates but slow Fourier decay will be throttled by the coherence term. Other directional systems may face a similar trade-off.","The roughly factor-two numerical advantage might partly reflect the redundancy and finite-scale implementation of digital shearlets rather than the asymptotic coherence bound; experiments at higher resolutions and with different wavelet families would separate finite-size effects from the asymptotic story."],"forward_implications":["If the paper's bounds are tight, standard shearlet–Fourier compressed sensing will require m growing as s_shear^2, not as s_shear, so the theoretical regime where shearlets beat wavelets in sample count is at best logarithmic in resolution.","The practical message for imaging: switching from a wavelet to a shearlet sparsifying dictionary in an ℓ1-analysis recovery pipeline is unlikely to let one drop the sampling rate proportionally to the reduction in coefficients; observed gains are of order a factor of two.","The bottleneck is incoherence, not sparsity: the slower Fourier decay of shearlets offsets their better N-term approximation, and the unresolved localization factor only adds uncertainty on the pessimistic side.","Future approaches seeking to make directional systems pay off in Fourier sampling must either alter the sampling distribution, use multilevel sparsity-aware schemes, or find shearlet generators with faster Fourier decay while retaining frame bounds."],"fun_headline_variants":["Shearlets' sparsity doesn't shrink Fourier samples","No sample win from shearlets' sparsity in Fourier CS","Why shearlets' efficiency fails to cut measurements","Shearlets vs wavelets: no Fourier sample savings","Sparsity advantage lost in Fourier compressed sensing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The negative conclusion depends on the claim that the decay |ψhat_i(n)| ≲ ||n||^{-3/4} is essentially optimal for cone-adapted shearlets; that optimality is argued from a Littlewood-Paley heuristic, and if some shearlet generator achieved the wavelet-like decay ||n||^{-1}, the sample-complexity comparison would flip.","fun_headline_variants_meta":{"raw":{"variants":["Shearlets' sparsity doesn't shrink Fourier samples","No sample win from shearlets' sparsity in Fourier CS","Why shearlets' efficiency fails to cut measurements","Shearlets vs wavelets: no Fourier sample savings","Sparsity advantage lost in Fourier compressed sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1158,"prompt_tokens":828,"completion_tokens":330,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":251}},"tokens_in":572,"tokens_out":330,"duration_ms":3318,"temperature":1.0,"reasoning_tokens":251,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:01:19.384215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhibit a cone-adapted shearlet frame whose elements satisfy |ψhat_i(n)| ≤ C (1+||n||)^{-1} on integer grids while retaining the frame and approximation properties; the paper's comparison then gives m_shear ~ s_shear, overturning the quadratic conclusion. Alternatively, produce a phase-diagram study on cartoon-like images in which shearlets require an order of magnitude fewer Fourier samples than wavelets at matched recovery quality.","supporting_citations":[],"review_version":1}