{"id":"3a967d4c-47aa-4389-9b27-6ad054f629b2","arxiv_id":"2608.04465","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A smoothed mixed ℓ2,1 objective on the sphere yields a provable Riemannian method for jointly sparse blind deconvolution, with sample complexity roughly K^6N^2 in the unit-impulse case and κ^8K^6N^3 in the general case, better than prior ℓ1-based analysis for K^2≤N.","lead":"This paper develops an optimization algorithm with proven guarantees for recovering an unknown shared filter and multiple sparse signals from convolutive observations when all signals share the same sparse support. It claims that exploiting this joint sparsity substantially reduces the number of measurement channels needed, with sample complexity improving on prior work when K^2 is at most N.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed sample-complexity improvement over [15] rests on an unproven hyperparameter restriction (μ≤θ); the paper's own [15] implementation violates it for small K, and if μ=Ω(1) the comparison reverses for part of the stated regime.","rationale":"I read the paper in good faith: it provides a real theorem (Theorem 5) with a detailed proof structure, and the real-valued landscape analysis appears internally coherent. The Gaussian assumption flagged by the reader is an explicit modeling assumption common in this literature; it limits applicability but does not by itself undermine the theorem. The more load-bearing issue is the headline comparison to [15]. The asserted μ≤θ restriction is unverified and contradicted by the paper's own numerical setup for small K. If the restriction fails, the claimed sample-complexity reduction is not uniform over the stated regime K^2≤N, and the central message of the paper weakens. This does not require rejecting the paper; it requires either a citation or proof of the μ condition, or a revised comparison. The reader's CONDITIONAL verdict remains appropriate, but for this additional reason.","tokens_in":66448,"tokens_out":30402,"duration_ms":233637,"concrete_test":"Check [15, Theorem 3] (and its proof) for the allowable range of μ. If μ=Ω(1) is allowed, recompute the Section 3.3 comparison: proposed O(κ^8 K^6 N^3) versus [15] O(κ^8 K^{-4} N^7), and test whether the proposed bound is smaller for all K^2≤N. Additionally, rerun Experiment 2 with [15]'s μ tuned per K (e.g., μ=cK/N for c in {0.1, 1, 10}) as well as μ=0.05, and compare success regions to see if the proposed method's advantage persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the general-filter sample-complexity comparison in Section 3.3. The paper writes [15, Theorem 3] as O(κ^8 θ^{-4} μ^{-2} N^3) and asserts that μ is an algorithmic hyperparameter whose order does not exceed that of θ, with θ≈K/N. This yields [15]'s rate as O(κ^8 K^{-6} N^9) versus the proposed O(κ^8 K^6 N^3) when K^2≤N. This assertion is load-bearing: if [15] permits μ=Ω(1), its rate becomes O(κ^8 K^{-4} N^7), which is smaller than the proposed rate for K^2≤N<K^{2.5}. The paper gives no proof or citation for μ≤θ. Moreover, Section 6.1 implements [15] with μ=0.05 for all K; for N=64 and K=1,2,3 this gives μ>K/N, so the numerical comparison uses [15] outside the regime assumed in the theoretical comparison. Thus the claimed significant reduction in sample complexity versus [15] is not established as stated; either the μ≤θ condition must be justified from [15], or the comparison must use the actual allowable μ.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies blind deconvolution of L channels that share a common filter and a common sparse support. It proposes a smooth approximation of the ℓ2,1-norm based objective on the unit sphere, after reparameterizing the problem through the inverse filter and applying a preconditioner built from the observations. Under a Gaussian generative model for the nonzeros and a non-uniform support assumption, it characterizes the landscape of the objective, proving that approximate second-order stationary points lie close to the target solution when the sample size satisfies (85). It then proposes RGD-NCS, proves that it reaches an approximate second-order stationary point in finitely many iterations, and combines the two results into a recovery guarantee in Theorem 5. A complex-valued extension and numerical experiments, including an array-processing application, are presented.","tokens_in":66721,"tokens_out":8348,"duration_ms":80530,"significance":"If the core comparison with Shi and Chi [15] is correct, the main theorem is a significant advance: it yields a finite-sample recovery guarantee whose dominant sample-complexity term is O(κ^8 K^6 N^3) in the regime K^2 ≤ N, versus the claimed O(κ^8 K^{-6} N^9) for [15]. The real-valued analysis is substantive: the landscape lemmas, the concentration bounds for EΩ and Δ, and the convergence proof for RGD-NCS are developed in detail, with the auxiliary technical estimates supplied in the supplement. The paper is also honest about the Gaussian and noiseless assumptions. However, the headline comparative claim rests on an unverified restriction on a parameter μ of [15], and the numerical experiments are run outside the theorem's stated regime; the complex-valued section also contains an apparent error in the complex gradient formula and stops short of proving the claimed complex guarantees. These issues do not obviously invalidate the real-valued theorem, but they do block acceptance in the current form.","major_comments":[{"comment":"The claimed sample-complexity improvement over [15, Theorem 3] is load-bearing and is not established. The paper states that μ is an algorithmic hyperparameter 'whose order does not exceed that of θ' and then converts O(κ^8 θ^{-4} μ^{-2} N^3) into O(κ^8 K^{-6} N^9). No proof or citation is given for the restriction μ ≤ θ. If [15] allows μ = Ω(1), the same formula gives O(κ^8 K^{-4} N^7), which is better than the proposed O(κ^8 K^6 N^3) for part of the stated regime. Moreover, the paper's own implementation of [15] in Section 6.1 uses μ = min{10 N^{-5/4}, 0.05}, so for N = 64 and K = 1,2,3 one has μ > K/N, i.e., the numerical comparison with [15] is itself conducted outside the regime assumed in the theoretical comparison. The authors should either justify μ ≤ θ from the assumptions of [15], or recompute the comparison using the actually allowable range of μ and update the conclusions accordingly.","section":"Section 3.3, paragraph after Eq. (101)"},{"comment":"The numerical experiments do not operate in the regime of Theorem 5, so the claim that the experiments validate the theorem's quantitative sample-complexity prediction is too strong. With N = 64 and K = 3, Theorem 5 requires ε ≤ c2 K^{-7} N^{-2} ≈ 1.9 c2 and ϵg ≤ c3 K^{-7/2} N^{-1/2} ≈ 8 × 10^{-4} c3 (for O(1) constants), while the experiments set ε = min{10^5 N^{-5}, 10^{-4}} ≈ 9.3 × 10^{-5} and ϵg = 10^{-2}. Both values exceed the required ranges unless the unspecified constants are extraordinarily large. Thus Figure 2 demonstrates heuristic success outside the proven range; it does not quantitatively confirm Theorem 5. The authors should either rerun within the theorem's parameter regime or explicitly state that the experiments are robustness checks beyond the guarantee.","section":"Section 6.1, Experiment 2"},{"comment":"The complex-valued gradient formula appears to be incorrect as written. For the term |y_l^T Γ_n R_C h|^2, the derivative with respect to the complex variable h involves the conjugate of y_l, i.e., the quadratic form should contain y_l^* y_l^T (equivalently y_l conjugate times y_l transpose), not y_l y_l^T. The printed expression R_C^H Γ_n^T (1/L Σ_l y_l y_l^T) Γ_n R_C h is not Hermitian and is not the Riemannian gradient of ρ_C under the metric in (117). Since Algorithm 1's complex variant and the Section 6.2 experiments rely on this formula, this needs to be corrected and verified. In addition, the section asserts that theoretical guarantees 'analogous to those established in Theorem 4 can be derived' but provides no proof; as written, the complex extension has no proven recovery guarantee and should be labeled accordingly.","section":"Section 5, Eq. (118)"}],"minor_comments":[{"comment":"The terms 'sample size' and 'number of channels' are both used for L; this is standard in the area, but a sentence explicitly fixing the notation would help avoid confusion with N, the filter length.","section":"Abstract and Section 1"},{"comment":"The ambiguity model states that the filter is recovered up to a global scaling and a circulant shift. It would be helpful to state explicitly that the success criterion in Section 6.1 corresponds to this ambiguity model, since the criterion uses ∥g ⊛ ĝ_inv∥∞ / ∥g ⊛ ĝ_inv∥2.","section":"Section 2.2, Eq. (13)"},{"comment":"There are minor language issues, for example 'to a array signal processing application' in the opening sentence; these should be corrected in copyediting.","section":"Section 6.2"},{"comment":"The conclusion correctly notes that noisy and off-grid settings are outside the proven model, but this limitation should also be stated when the array-processing experiment is described, so that the experimental section does not give the impression that the theorem covers the application setting.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The central comparison issue in Section 3.3 is likely fixable in revision, but it is the kind of point that determines whether the paper's headline claim survives. The apparent error in the complex gradient formula in Eq. (118) is concrete and needs a careful check, as it affects the complex experiments. There is no indication of questionable citation practice or novelty problems; the real-valued proof appears to be a serious and largely self-contained contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know two things about arXiv:2608.04465. First, it delivers the first non-asymptotic landscape analysis for jointly sparse blind deconvolution, using a smooth ℓ2,1 objective that couples channels. That part looks substantial and, from what I can verify, the real-valued proofs hold together. Second, the paper's headline comparison with Shi and Chi [15] is not established as stated. They need μ≤θ to get their improved rate, and they give no proof or citation for that. If μ can be Ω(1), the comparison flips in part of the regime.\n\nWhat's genuinely new: the mixed ℓ2,1 objective is non-separable, which breaks earlier techniques, and the authors develop a new geometric analysis. The core result, Theorem 4, gives a non-asymptotic sample complexity of roughly κ^8 K^6 N^3 (up to logs) for approximate second-order stationary points to land near the target. The supplement has detailed proofs, and I didn't find an obvious gap. The non-uniform support condition is natural, and the probability bounds are standard.\n\nWhere I'd push back: the μ≤θ assumption. The stress-test note checks out. The paper writes [15]'s rate as O(κ^8 θ^{-4} μ^{-2} N^3) and then asserts μ is an algorithmic hyperparameter whose order does not exceed θ. That assertion is load-bearing. If μ=Ω(1), [15]'s rate is O(κ^8 K^{-4} N^7), which beats the proposed rate for part of K^2≤N. Moreover, their own implementation of [15] uses μ=0.05 for N=64, which exceeds θ for K=1,2,3, so the numerics compare against [15] outside the assumed regime. This doesn't sink the landscape theorem, but it means the central advantage claim is overstated.\n\nAlso, the experiments run with ε and ϵg far outside the theorem's tolerances, so they don't formally validate the bounds. That's common, but it should be said explicitly. The complex extension is asserted without the technical details—fine as a sketch, but not a proof. No code shipped.\n\nThe Gaussian assumption on nonzero entries is strong but standard in this literature. The comparison with [15] is the real issue.\n\nRecommendation: send it to review. The landscape result is important enough, and the proof appears careful. But the referee team should require the authors to either prove μ≤θ from [15] or revise the comparison to use the actual allowable μ. If the comparison doesn't survive, the paper is still a solid landscape contribution, just with a more modest headline.","headline":"A serious landscape theorem for jointly sparse blind deconvolution, but the claimed sample-complexity win over Shi-Chi rests on an unproven hyperparameter restriction.","tokens_in":67263,"tokens_out":3989,"would_cite":true,"duration_ms":33971,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C48","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that jointly sparse blind deconvolution can be solved by a Riemannian algorithm whose sample complexity grows like N^3 rather than N^9 when the sparsity level satisfies K^2 ≤ N.","keywords":["jointly sparse blind deconvolution","Riemannian optimization","sample complexity","nonconvex landscape","second-order stationary point","smooth ℓ2,1 objective","array signal processing"],"falsifier":"Generate inputs with a non-uniform common support but deterministic nonzero entries, such as all entries equal to one, run the proposed algorithm with a small number of channels below the theorem's bound, and check whether an approximate second-order stationary point exists far from every signed canonical basis vector.","tokens_in":66242,"feed_emoji":"📡","tokens_out":7670,"duration_ms":69776,"temperature":0.7,"pith_summary":"Jointly sparse blind deconvolution asks to recover an unknown filter and L input signals that share the same sparse support from their circulant convolutions, a model behind radar, passive localization, and fluorescence microscopy. The paper establishes that a smooth ℓ2,1-style objective on the unit sphere, rather than separate per-channel sparsity penalties, makes the optimization landscape benign: every sufficiently accurate second-order stationary point lies near the true inverse filter, provided enough channels are observed. The main quantitative claim is a sample-complexity bound of order $O(\\kappa^8 K^6 N^3)$ for $K^2 \\le N$, suppressing log factors, versus $O(\\kappa^8 K^{-6} N^9)$ for the strongest prior non-joint analysis. If correct, joint sparsity buys a substantial reduction in the number of channels needed, and the paper proposes a Riemannian gradient-descent method with negative-curvature search that provably reaches the required stationary point.","feed_headline":"Joint sparsity cuts blind-deconvolution sample needs","feed_subtitle":"A Riemannian algorithm with a joint-sparsity objective needs far fewer measurements than per-channel methods.","key_machinery":"The load-bearing object is a smooth surrogate for the $\\ell_{2,1}$ norm defined on the unit sphere: $\\rho(h)=\\sum_{n=1}^N (\\frac{1}{L}\\sum_{l=1}^L |y_l^\\top \\Gamma_n R h|^2+\\epsilon)^{1/2}$, where $R$ is the whitening preconditioner $(\\frac{1}{L}\\sum_l C(y_l)^\\top C(y_l))^{-1/2}$. The matrices $\\Gamma_n$ encode circulant shifts, so rows of the recovered signal matrix are penalized jointly, and the preconditioner partially inverts the unknown filter. The geometric argument shows that for non-uniform support the asymptotic objective has only signed canonical basis vectors as global minimizers, and then controls two perturbations, the sample covariance error $E_\\Omega$ and the preconditioning error $\\Delta$, so that every approximate second-order stationary point is forced into a small neighborhood of such a vector. A negative-curvature direction $d=e_o - h_o h$ is used to rule out distant saddle points.","core_discovery":"The paper's central claim is Theorem 5: under a non-uniform common support, i.i.d. Gaussian nonzero entries, and an invertible filter with condition number $\\kappa$, once $L \\ge c_1 \\kappa^8 \\log^2 N (\\sqrt{K} + \\sqrt{\\log L})^2 \\max\\{K^7 N^2, K^5 N^3, \\epsilon^{-1}, \\epsilon^{-1} K^{-2} N\\}$, the estimate produced by RGD-NCS satisfies the error bound in Eq. (102) with probability at least $1 - 4L^{-8}$. Ignoring logarithmic factors and choosing $\\epsilon \\sim K^{-7}N^{-2}$, this is roughly $O(\\kappa^8 K^6 N^3)$ samples when $K^2 \\le N$, compared with $O(\\kappa^8 K^{-6} N^9)$ for the per-channel analysis it compares against. The improvement comes from aggregating information across channels through the joint-sparsity objective, at the cost of a landscape with many saddle points that must be excluded by second-order arguments.","pith_inferences":["The i.i.d. Gaussian generative model is likely stronger than needed; a concentration analysis for sub-Gaussian or bounded nonzero entries would probably preserve the landscape conclusion with modified constants, but the paper does not prove this.","The numerical experiments show graceful behavior under noise and off-grid delays even though the theorems are noiseless and grid-aligned; a natural next step is a noise-aware objective with error bounds that depend explicitly on SNR and grid mismatch.","The same group-sparsity mechanism could be ported to other bilinear inverse problems, such as dictionary learning with common support across samples, where the circulant-structure arguments would need to be replaced by incoherence assumptions.","A direct testable prediction is that the success phase transition should shift left as the number of channels grows roughly like $N^3$ for fixed sparsity, which could be checked by Monte Carlo trials across $N$ and $L$."],"forward_implications":["With $L$ satisfying (101), the recovered inverse filter obeys the explicit error bound (102), so both filter and input signals can be reconstructed to arbitrary accuracy by shrinking $\\epsilon$ and the stationarity tolerances.","In the common regime $K^2 \\le N$, joint sparsity lowers the sample requirement from order $N^9$ to order $N^3$ in the filter length, meaning far fewer antennas or snapshots are needed in array-processing applications.","The non-uniformity condition on the support is both necessary and sufficient in the asymptotic problem, so supports with cyclic symmetry are the only algebraic obstruction to unique recovery.","The same framework extends to complex-valued inputs, covering radar and array-signal models directly."],"supporting_citations":[{"why":"Provides the per-channel baseline whose sample complexity is improved, and supplies the concentration argument adapted in Lemma 7 for the preconditioning error $\\Delta$.","marker":"[15]"},{"why":"Supplies the non-asymptotic tail bound for the sample covariance matrix $E_\\Omega$ used in Lemma 3 and Theorem 2.","marker":"[58]"},{"why":"Supplies the Riemannian geometry tools, including gradient projection, retractions, and Lipschitz properties, used in the landscape analysis and in Proposition 8's convergence proof.","marker":"[59]"},{"why":"Introduces the negative-$\\ell_4$ objective with preconditioning on the sphere that the proposed framework extends to joint sparsity.","marker":"[13]"},{"why":"Shows that an $\\ell_1$-based objective improves sample complexity and provides a prior sample-complexity comparison point.","marker":"[14]"},{"why":"Establishes the unit-sphere reparameterization and preconditioning framework for multichannel sparse blind deconvolution that this work builds on.","marker":"[12]"}],"fun_headline_variants":["Joint sparsity reduces sample needs in blind deconvolution","Riemannian method leverages joint sparsity for blind deconvolution","Fewer measurements for blind deconvolution via joint sparsity","Joint sparsity lowers sample complexity in blind deconvolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis assumes the nonzero entries of the input signals are independent standard Gaussian random variables across all channels; if those entries are deterministic, bounded, or heavy-tailed, the concentration bounds behind the landscape theorem have no guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Joint sparsity reduces sample needs in blind deconvolution","Riemannian method leverages joint sparsity for blind deconvolution","Fewer measurements for blind deconvolution via joint sparsity","Joint sparsity lowers sample complexity in blind deconvolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3029,"prompt_tokens":867,"completion_tokens":2162,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2092}},"tokens_in":483,"tokens_out":2162,"duration_ms":14283,"temperature":1.0,"reasoning_tokens":2092,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:39:38.054690+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate inputs with a non-uniform common support but deterministic nonzero entries, such as all entries equal to one, run the proposed algorithm with a small number of channels below the theorem's bound, and check whether an approximate second-order stationary point exists far from every signed canonical basis vector.","supporting_citations":[],"review_version":2}