{"id":"43b4e75d-7c9b-4b69-8a36-562c25a68649","arxiv_id":"2607.09593","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Multi-snapshot spike deconvolution via variable projection has an explicit basin of convexity whose radius is controlled by the PSF spectrum, spike separation, and amplitude dynamic range.","lead":"This paper derives explicit conditions under which the nonconvex spike-deconvolution objective is well-behaved near the true spike locations, and proves that gradient descent converges when started inside that region. The results give practitioners a principled, PSF-dependent rule for choosing sampling bandwidth and for knowing when iterative refinement is safe.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central basin-of-convexity proof depends on Lemmas 14–15 from unpublished preprint [15]; a flaw there would invalidate Theorem 1.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the proof of Lemma 11, and hence Theorem 1, imports Lemmas 14–15 and Lemma 10 from the unpublished preprint [15]. I agree this is the most serious risk. The manuscript is otherwise substantial: it states explicit basin, error, and convergence theorems; provides closed-form Hessian and cross-derivative computations; and includes reproducible experiments. The internal limitation noted in §3.2—that the theory does not explain behavior at large σ and large B—is honestly stated and supports a conditional rather than unconditional reading. A secondary concern is that the residual covariance condition (16) is not verified in the experiments, but that is an applicability gap, not a correctness flaw. Since the central mathematical claim is contingent on unproved external bounds, ACCEPT is premature, but REJECT is not warranted because no internal contradiction is evident and the missing proofs are plausibly valid. Thus the conditional verdict stands.","tokens_in":41676,"tokens_out":10850,"duration_ms":109351,"concrete_test":"Independently verify Lemma 14 computationally: for Gaussian PSFs with σ ∈ {0.05, 0.1, 0.3}, N = 51, T = 1, and random τ with minimum separation Δ ∈ {0.1, 0.2, 0.4}, compute σ_min(GΦτ) and σ_max(GΦτ) over 1000 trials and compare against the asserted bounds sqrt(T Eg (1 ± 0.5 ρg Δ^{-1})). Also count the nonzero diagonal entries of G for the Morlet PSF used in §3 to check whether the full-rank condition in Lemma 9 holds. If the bounds fail or the condition is violated, the basin radius in Theorem 1 must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1 and its key supporting Lemma 11 rest on conditioning bounds imported from the co-author's unpublished preprint [15]. Specifically, Lemma 11 (Appendix D) relies on Lemma 14 for σ_min/max(GΦτ) and Lemma 15 for ∥ΛGΦτ∥, ∥Λ²GΦτ∥; Lemma 9's proof also invokes Lemma 10, which is paraphrased from [15]. None of these is proved in the present manuscript; they are only stated as recalled results. If Lemma 14 is incorrect or missing an assumption, the Hessian lower bound σ_min(∇²ℓ) ≥ (1/3) Eg′ T rmin(X)² collapses, and with it the basin radius ϱ in (15) and the existence/uniqueness of the local minimizer. The concern is sharpened by an internal gap: Lemma 7's stated hypothesis requires BV and integrability only for Pg, but its proof applies Lemma 13 to Pg′ as well (to bound −Ĉ″_+(0) − Eg′), so an unstated assumption on Pg′/Pg″ is needed. Similarly, Lemma 9's full-rank argument requires G to have at least 2K nonzero diagonal entries; this is not listed in Theorem 1's assumptions and fails for PSFs with spectral nulls. The paper's 'arbitrary PSF' wording is therefore stronger than the actual hypotheses. These are correctness risks, not merely presentation issues, because the central claim is precisely a quantitative basin guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies multi-snapshot spike deconvolution with a known point-spread function (PSF). It adopts a variable-projection formulation (VarProSD) in which the amplitudes are eliminated, leaving a nonconvex least-squares problem over the spike locations. The main contribution is an explicit characterization of a basin of convexity around the ground truth: Theorem 1 states that, under a minimum-separation condition on the spikes, a spectral-regularity condition on the PSF, and a residual-covariance condition, the objective is strongly convex with Lipschitz gradient on a ball and admits a unique local minimizer. Theorem 2 gives an estimation-error bound for that minimizer under stochastic noise, Theorem 3 gives a deterministic adversarial-noise bound via the local Lipschitz property of the inverse map, and Theorem 6 establishes local linear convergence of gradient descent. The proofs rely on Beurling–Selberg extremal approximations to bound the conditioning of generalized Vandermonde matrices. Numerical experiments illustrate the theory and propose a PSF-driven sampling-bandwidth selection criterion.","tokens_in":41940,"tokens_out":9433,"duration_ms":89797,"significance":"If correct, this is the first explicit, interpretable basin-of-convexity guarantee for multi-snapshot spike deconvolution under a nontrivial PSF, with the radius expressed in terms of PSF spectral parameters, spike separation, amplitude dynamic range, and bandwidth. The paper also contributes a deterministic adversarial-noise analysis and a practical bandwidth-selection heuristic. The proof structure is coherent and detailed, with closed-form gradient/Hessian expressions and a reproducible code repository. However, the central theorem rests on conditioning lemmas imported from an unpublished preprint, and several stated hypotheses are weaker than those actually used in the proofs, so the correctness risk is concentrated in a few load-bearing points.","major_comments":[{"comment":"The proof of Lemma 9 asserts that U is full column rank using 'since G has at least 2K non-zero diagonal entries'. This condition on the PSF is not stated in Lemma 9 or in Theorem 1's hypotheses. If the PSF has spectral nulls on the sampled frequency grid, the bound on Sγ—and hence the curvature lower bound (48) and the basin radius (15)—is unsupported. Add the assumption to the theorems or modify the proof, and qualify the 'arbitrary PSF' claim in Sections 1.3 and 6.","section":"Section 4, Lemma 9; Theorem 1"},{"comment":"The stated hypotheses of Lemma 7 require only BV and f²-integrability for Pg, but the proof bounds −Ĉ″₊(0)−Eg′ by applying Lemma 13 to Pg′ (see inequality (38)). This requires the same regularity for Pg′; since the conclusion already contains Eg′ and ρg′, the hypothesis is incomplete. Update Lemma 7 and the dependent results (Lemma 8 and Lemma 11) to state the needed regularity for Pg′ and Pg″.","section":"Section 4, Lemma 7"},{"comment":"The key conditioning bounds are paraphrased from the unpublished preprint [15] and are not proved in this manuscript. Lemma 11's lower bound σmin(∇²ℓ) ≥ (1/3) Eg′ T rmin(X)², and therefore Theorem 1's basin radius (15) and the existence/uniqueness of the local minimizer, depend directly on these lemmas. Appendix A claims self-containedness, but the central proof is conditional on an unreviewed source. Include complete proofs of Lemmas 14 and 15 (or of Lemma 10) or clearly state that verification requires [15] and ensure it is publicly available by the time this paper appears.","section":"Appendix A, Lemmas 10, 14, 15; Lemma 11"},{"comment":"The residual-covariance condition (16) is a load-bearing hypothesis for Theorems 1, 2, and 6, but the numerical experiments in Section 3 do not verify whether this condition holds in the simulated parameter regimes. Since the experiments are presented as corroborating the theory, please report the relevant ratio (or an upper bound) for the settings in Figures 4, 6, 8, and 11, or state explicitly that the experiments only validate the qualitative behavior rather than the sufficient condition.","section":"Section 2.1, condition (16); Section 3"}],"minor_comments":[{"comment":"The quantity rmax(X) is used in Theorems 1, 2, and 6 and in condition (16), but only rmin(X) is defined in the notation section. Define rmax(X) = max_j ||e_j^H X||_2 alongside rmin(X).","section":"Section 1.4"},{"comment":"The text 'Using ∥Aγ∥≲√TEg∥X∥ (cf. [15, Theorem 1])' is dimensionally/mathematically imprecise: Aγ is N×K, so its norm is bounded by √(TEg) times a separation-dependent factor, not by a quantity proportional to ∥X∥. The intended statement is likely ∥Y0∥ = ∥Aτ X∥ ≲ √(TEg)∥X∥. Please correct.","section":"Section 2.2.3"},{"comment":"The bound (25) is referred to as the 'gradient descent expression', but it is derived from Theorem 2, not from the gradient-descent convergence theorem (Theorem 6). Clarify the attribution to avoid confusion.","section":"Section 2.2.3, Eq. (25)"},{"comment":"The phrase 'XX^T in (23a) converges to the complex conjugate of the autocorrelation matrix RX = 1/L XX^H' is confusing because XX^T and XX^H differ, and R_X is defined with a different normalization. Please state the convergence in terms of (1/L)XX^T → I or similar, with the precise assumptions on X.","section":"Section 2.2.2, Remark 2"},{"comment":"Typos and minor issues: 'Frobenious' should be 'Frobenius' (Section 2.2.1); MSC classification '9408' should be '94A08'; the arXiv header dates conflict (v2, 24 Jul 2026 vs. July 27, 2026). Also, Figure 11(b) includes a '√CRB' curve whose definition for the L-vs-error experiment is not specified.","section":"General editorial"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the heavy reliance of Theorem 1 on Lemmas 10, 14, and 15 from the co-author's unpublished preprint [15]. Since these are not proved in the manuscript, I recommend that the editors require either full proofs in the appendix or a verifiable public version of [15] before acceptance. The missing hypotheses in Lemma 7 and Lemma 9 are fixable with statement changes, but they are load-bearing; the 'arbitrary PSF' framing should be adjusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"If you work on spike deconvolution or super-resolution, this paper is worth your time. It gives the first explicit basin radius for the variable-projection objective under a non-trivial PSF, with local linear convergence of gradient descent and an O(1/sqrt(L)) stochastic error bound. That is a genuine contribution, not a repackaging of known results. The Hessian decomposition and the Beurling–Selberg conditioning machinery are laid out carefully, and the experiments are honest, including the modified-ESPRIT comparison and the bandwidth-selection criterion.\n\nThe soft spots are real but manageable. The biggest one is structural: Theorem 1 and its key Lemma 11 rely on Lemmas 14–15, quoted from an unpublished preprint [15] by one of the authors. Those are load-bearing — if they are wrong, the basin radius collapses. The paper would be much stronger if the proofs were included or the preprint were published. I would not call this a fatal flaw, but it is a genuine referee concern.\n\nThere are also two concrete hypothesis gaps the authors should fix. Lemma 7 is stated with assumptions only on Pg (bounded variation plus f^2-weighted integrability), but its proof applies Lemma 13 to Pg′ to bound the second moment of the majorant; that needs an unstated regularity assumption on Pg′. Similarly, Lemma 9’s proof requires G to have at least 2K nonzero diagonal entries, but the lemma statement omits it. So the “arbitrary PSF” phrasing is too strong — the actual hypotheses are spectral regularity plus nonzero sampled Fourier weights. These are fixable, but they are correctness issues, not just presentation.\n\nOne more minor point: the residual-covariance condition (16) is a sufficient condition that the experiments never verify directly. That is not unusual, but it means the empirical success is not a direct test of the theory’s hypotheses. The authors also openly admit their theory does not explain the large-σ, large-B failures, which I appreciate — it is honest.\n\nBottom line: this is a serious paper by capable people. The central argument is coherent, the new result is significant within the subfield, and the gaps are addressable. I would send it to peer review with a request for proofs or published citations for the imported lemmas and tightened lemma statements. A good referee will need to check the imported conditioning bounds carefully, but the paper deserves that attention.","headline":"A serious, novel basin-of-convexity analysis for VarProSD with a real new result, but the central conditioning bounds ride on an unpublished preprint and a couple of lemmas have missing hypotheses — worth a careful referee, not a desk reject.","tokens_in":42470,"tokens_out":2627,"would_cite":true,"duration_ms":32174,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","65T99","94A12","94A20","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the nonconvex variable-projection objective in multi-snapshot spike deconvolution has an explicit, computable basin of convexity—a ball around the true spike locations whose radius is determined by PSF spectral flatnes","keywords":["spike deconvolution","variable projection","basin of convexity","Beurling-Selberg approximation","gradient descent","multi-snapshot","point spread function","local convergence"],"falsifier":"Evaluate the smallest eigenvalue of the Hessian of ℓ(γ) at a point γ on the sphere d₂(γ,τ)=ϱ for the Gaussian-PSF setup of Figure 2(b) (σ=0.3, K=2, Δ=0.4); if it is below the strong-convexity constant (1/3)Eg′Tr_min(X)², the claimed basin radius is too large.","tokens_in":41509,"feed_emoji":"📡","tokens_out":7403,"duration_ms":67328,"temperature":0.7,"pith_summary":"The paper tries to establish that the nonconvex least-squares problem over spike locations—obtained by eliminating amplitudes via variable projection—is strongly convex on a ball around the true locations whose radius is given in closed form. The radius is set by the flatness of the point spread function's power spectrum, the minimum spike separation, the amplitude dynamic range, and the sampling bandwidth. Within this ball, a unique local minimizer exists, gradient descent converges to it linearly, and the minimizer's error decays as one over the square root of the number of snapshots under stochastic noise; an adversarial-noise bound is also derived. If true, this is the first interpretable basin-of-convexity guarantee for arbitrary point spread functions under multiple snapshots, and it yields a practical way to select the sampling bandwidth.","feed_headline":"Explicit basin of convexity proved for spike deconvolution","feed_subtitle":"Radius depends on PSF flatness, spike separation, and bandwidth; guarantees linear convergence and shrinking error.","key_machinery":"The argument rests on the variable-projection objective and the structured Vandermonde matrix A_γ = G Φ_γ, where G is the diagonal PSF sampling matrix and Φ_γ is the Fourier-Vandermonde matrix of spike locations. Its curvature is controlled through Beurling–Selberg extremal approximations: bandlimited majorants and minorants of the truncated power spectral density of the PSF yield sharp bounds on the singular values of A_γ, Λ A_γ, and the derived matrix S_γ = A_γ^H Λ^H P⊥_γ Λ A_γ (Lemmas 9–10). These bounds, independent of the number of spikes K, translate directly into the strong-convexity constant and the gradient Lipschitz constant of the objective on the ball, and hence into the contract","core_discovery":"The central claim is Theorem 1: when the spikes are separated by more than 2/(3 ρ κ²) and the residual covariance satisfies the smallness condition (16), the VarProSD objective ℓ(γ)= (1/2L)∥P⊥_γ Y∥_F² is strongly convex with Lipschitz gradient on the ball N(τ,ϱ) defined in (14)–(15), so a unique local minimizer γ⋆ lies in that ball. Theorem 6 adds that gradient descent from any initialization in the ball converges linearly to γ⋆, and Theorem 2 bounds the recovery error by c₂√K √(Eg/Eg′) ∥R∥/(T Eg r_min(X)²), which under i.i.d. Gaussian noise becomes O(1/√L). The paper also constructs a worst-case rank-one adversarial noise and shows the resulting deterministic bound is √K sharper in the high","pith_inferences":["Our inference: the same Beurling–Selberg conditioning machinery could extend to single-snapshot or unknown-PSF settings, where no explicit basin characterization currently exists.","Our inference: the residual-covariance condition (16), which the paper assumes but does not verify in experiments, is the most likely place for the theory to be conservative; a testable extension is to check empirically how tightly (16) holds at the bandwidths predicted optimal.","Our inference: the inverse-map Lipschitz analysis is a general template—the explicit Jacobian expression could be adapted to other separable nonlinear least-squares problems to derive adversarial stability bounds."],"forward_implications":["Any initialization inside the explicit ball N(τ,ϱ) is guaranteed to yield linear convergence to the unique local minimizer, so the radius quantifies exactly how accurate an initialization must be.","Under i.i.d. Gaussian noise, the local minimizer achieves consistency with O(1/√L) error decay in the number of snapshots, matching the rate of ESPRIT for trivial PSFs but now for an arbitrary PSF.","The spectral parameter ρ gives a principled, PSF-driven rule for selecting sampling bandwidth: the bandwidth minimizing ρ predicts the empirically optimal B for Gaussian and Morlet PSFs, and larger B is not always better because it shrinks the basin.","The adversarial-noise analysis shows that a rank-one perturbation aligned with the top singular vector of the inverse-map Jacobian is asymptotically worst case, and the resulting deterministic error bound is sharper by a factor of √K than the stochastic bound in the high-SNR large-L regime."],"fun_headline_variants":["Explicit convexity basin for spike deconvolution","Spike separation sets radius of convex basin","Gradient descent converges linearly within convex basin","Error bounds shrink with snapshots in spike recovery"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof's load-bearing premise is that the truncated power spectral density of the point spread function and its first two derivatives are of bounded variation and integrable with an f² weight, so that the Beurling–Selberg majorant/minorant bounds on the conditioning of the structured matrices carry through; if that spectral regularity fails, the basin radius and curvature bounds collapse.","fun_headline_variants_meta":{"raw":{"variants":["Explicit convexity basin for spike deconvolution","Spike separation sets radius of convex basin","Gradient descent converges linearly within convex basin","Error bounds shrink with snapshots in spike recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001232,"raw_usage":{"total_tokens":4943,"prompt_tokens":837,"completion_tokens":4106,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":4048}},"tokens_in":581,"tokens_out":4106,"duration_ms":29976,"temperature":1.0,"reasoning_tokens":4048,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:30:07.103573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the smallest eigenvalue of the Hessian of ℓ(γ) at a point γ on the sphere d₂(γ,τ)=ϱ for the Gaussian-PSF setup of Figure 2(b) (σ=0.3, K=2, Δ=0.4); if it is below the strong-convexity constant (1/3)Eg′Tr_min(X)², the claimed basin radius is too large.","supporting_citations":[],"review_version":2}