{"id":"74ae09d0-235b-4294-a30a-658fb144d018","arxiv_id":"2412.11313","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A signal is stably recoverable from noisy linear measurements if and only if the kernel of the measurement operator intersects the tangent cone of the conjugate-subdifferential image only at zero.","lead":"This mathematics paper gives a complete geometric condition that decides when a regularized inverse problem recovers the true signal at the same rate as the noise shrinks. The condition uses second-order curvature information of the regularizer, which prior work missed, and yields testable sufficient conditions for group-sparse and total-variation problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3's 'if' direction rests on an unverified convergence lemma: [21, Prop. 3.1] is cited to get xk→x0 from uniqueness alone, so the iff characterization inherits any hidden assumptions of that lemma.","rationale":"The central claim of the paper is Theorem 3.3, an iff characterization of stable recovery. I read the proof carefully and found no internal inconsistency: the 'only if' direction's construction of admissible y_k and μ_k is sound, and the 'if' direction's contradiction argument is valid once convergence x_k→x0 is granted. The one place where the argument relies on an external result without re-derivation is the assertion that uniqueness of x0 implies convergence of arbitrary Tikhonov minimizers, citing [21, Proposition 3.1 and Remark 3.2]. That assertion is load-bearing because it is what turns a purely algebraic failure of stable recovery into a unit vector in the forbidden intersection. The paper's examples and corollaries are consistent with the theorem, and the numerical section, while methodologically weaker, does not affect the proof of the main theorem. I agree with the reader that this is the weakest assumption. However, convergence of regularized minimizers to a unique constrained minimizer is a standard and plausible property for continuous convex functions in finite dimensions, and the subsequent argument does not use any rate; so I do not regard the citation as an observed error, only as an unverified dependency. The proposed test settles it: if the cited proposition's hypotheses match, the theorem stands; if they do not, the authors must supply the missing lemma or restrict Theorem 3.3. Since the reader's conditional verdict already captures this and other addressable concerns, I leave the verdict unchanged.","tokens_in":27861,"tokens_out":21218,"duration_ms":201120,"concrete_test":"Obtain [21, Proposition 3.1 and Remark 3.2] and verify their assumptions against Theorem 3.3's: R continuous convex, x0 unique solution of (3.1), y_k→y0, μ_k→0, and x_k any solution of P(y_k, μ_k). If the cited statement assumes strictly more (e.g., coercivity, bounded sublevel sets, sharp minimum), then prove or disprove the needed convergence lemma directly from the optimality condition 0∈Φ*(Φx_k−y_k)+μ_k∂R(x_k). Specifically, show that any sequence of Tikhonov minimizers is bounded and that every cluster point is feasible with R ≤ R(x0), hence equals x0 by uniqueness. If the direct proof requires an extra assumption not present in Theorem 3.3, the theorem must be amended; a concrete non-coercive convex regularizer (e.g., R(x,y)=|x|+log(1+y) on y≥0) can serve as a counterexample test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"On page 9, in the proof of Theorem 3.3's sufficiency, after (3.8) is used to establish uniqueness via [20, Thm 4.5], the proof asserts 'Since x0 is the unique solution of problem (3.1), it follows from [21, Proposition 3.1 and Remark 3.2] that xk → x0.' This convergence is the only mechanism that produces tk = ||xk−x0|| ↓ 0 and hence the unit vector w in the tangent-cone intersection that contradicts (3.8). The paper supplies no proof and no statement of the cited proposition's assumptions. If [21, Prop. 3.1] requires, say, R to be coercive, x0 to be a sharp/strong minimizer, or the feasible set to be bounded, then Theorem 3.3 does not hold for all continuous convex R as stated, and the claimed generality of the characterization is unsupported. This is a genuine dependency risk, not an observed contradiction; all internal steps from the citation onward are consistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies stable recovery, meaning the linear (order δ) convergence of Tikhonov-regularized solutions to the true signal, for convex regularized linear inverse problems in finite dimensions. The central result, Theorem 3.3, characterizes stable recovery at an optimal solution x0 of the constrained problem (3.1) by the tangent-cone condition Ker Φ ∩ T_{∂R*(Im Φ*)}(x0) = {0}. The authors then derive consequences for smooth regularizers, for convex piecewise linear-quadratic regularizers, and for analysis group sparsity (ℓ1/ℓ2) and isotropic total variation regularizers, giving explicit sufficient conditions (Corollaries 4.8 and 4.9) and reporting numerical experiments on those two problem classes.","tokens_in":28038,"tokens_out":16739,"duration_ms":151382,"significance":"If Theorem 3.3 holds as stated, it is a substantial contribution: it gives a complete geometric characterization of stable recovery that exposes its second-order nature, and it shows that stable recovery can occur at non-sharp minimizers where the classical sufficient conditions (null space property, nondegeneracy source condition with restricted injectivity, minimum gain property) all fail. The examples in Section 3 are well constructed and illustrate the new phenomena convincingly. The extension to convex piecewise linear-quadratic regularizers (Corollary 3.8) and the group-sparsity sufficient conditions are useful and appear to be new. The numerical experiments, while not rigorous, provide encouraging evidence that the conditions are practical. The main theorem, however, has a load-bearing dependency on an external convergence result that is neither stated nor proved in the paper, and this gap must be addressed before the central characterization can be considered fully justified.","major_comments":[{"comment":"The assertion immediately after (3.10) that \"Since x0 is the unique solution of problem (3.1), it follows from [21, Proposition 3.1 and Remark 3.2] that xk → x0\" is load-bearing for the 'if' direction: it is the only mechanism that yields tk ↓ 0 and hence the unit vector w in the tangent-cone intersection that contradicts (3.8). The paper neither states the cited proposition nor verifies its hypotheses under the assumptions of Theorem 3.3 (R continuous convex, x0 optimal). If [21, Proposition 3.1] requires, for instance, coercivity of R, boundedness of the feasible set, or some form of strong/sharp minimum, then the theorem as stated for all continuous convex regularizers is not justified. Please provide a self-contained proof of the convergence of the Tikhonov minimizers to the unique solution of (3.1), or state the proposition explicitly and confirm that its assumptions are satisfied under the paper's hypotheses.","section":"Section 3, proof of Theorem 3.3 (sufficiency direction)"}],"minor_comments":[{"comment":"The definition of stable recovery quantifies over \"any optimal solution x(y,µ) of problem (3.3)\" but the paper does not discuss existence of these minimizers for arbitrary continuous convex R. If the Tikhonov problem has no minimizer for some small δ, the property is vacuous; please add a short existence argument or an explicit standing assumption that minimizers exist for the relevant parameters.","section":"Definition 3.1 and Section 3"},{"comment":"The optimization problem is stated as min over x ∈ R^6, but the regularizer uses four groups of two variables and the vector x0 has eight components; the correct setting is x ∈ R^8.","section":"Example 4.10, equation (4.51)"},{"comment":"The verification of conditions (4.45) and (4.50) is performed by solving nonconvex quadratic problems with Gurobi's NonConvex parameter and a 5-second time limit; Gurobi's nonconvex solver is a heuristic and does not provide global optimality certificates for general nonconvex QPs. Please state that the numerical conclusions are indicative rather than certified, or use a method that guarantees global optimality.","section":"Section 5, problems (5.3)–(5.5)"},{"comment":"The name \"Crome\" should be \"Cromme\" to match reference [14], and the external results [20, Theorem 4.5] and [21, Proposition 3.1 and Remark 3.2] should be stated or precisely described in the paper rather than only cited.","section":"Definition 2.2 and references"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the unstated external convergence result [21, Proposition 3.1 and Remark 3.2] used in the proof of Theorem 3.3's sufficiency direction. Since [21] is from the same research group and one of the present authors is a co-author of that paper, it should be straightforward for the authors to supply a self-contained statement and proof, or to narrow the assumptions of Theorem 3.3 accordingly. Please request this in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe main result is a real step forward. Theorem 3.3 gives an iff geometric condition for stable recovery, and the condition is genuinely second-order: Ker Φ ∩ T_{∂R*(Im Φ*)}(x0) = {0}. That is new. The prior literature offered sufficient first-order conditions (sharp minima, null space property, nondegeneracy source condition with restricted injectivity), and the paper shows by example that stable recovery can happen without sharp minima and can fail even at a strong minimum. The examples are worked out carefully and I did not find an internal inconsistency. The corollaries for analysis group sparsity and total variation give checkable sufficient conditions that are independent of sharp minima, and Corollary 3.8 (piecewise linear-quadratic regularizers) is a nice extension of the ℓ1 result.\n\nThe proof of necessity in Theorem 3.3 is self-contained and sound. The sufficiency direction has a load-bearing external dependency. After proving uniqueness via [20], the paper asserts xk → x0 by citing [21, Proposition 3.1 and Remark 3.2]. That convergence is what lets the contradiction argument extract the unit vector w in the tangent-cone intersection. The paper never states the proposition or its assumptions. If that lemma requires coercivity of R, boundedness of the feasible set, or some additional structure, then Theorem 3.3 is not established for all continuous convex R as claimed. This is a dependency risk, not a demonstrated contradiction, but a referee must resolve it before the theorem can be taken as proven.\n\nThe numerics are the weaker part. The non-sharp population is selected by a hand-picked source-coefficient window (0.95 to 1.05), the verification problems are solved by Gurobi's nonconvex mode with a 5-second time limit, and no code or data are released. For a theory paper, the experiments are acceptable as illustrations, but they should not be read as a systematic validation.\n\nWho this is for: anyone working in variational analysis, regularization theory, or sparse recovery. The paper deserves a serious referee despite the caveats. My recommendation: send it to review, ask the referee to check [21, Prop. 3.1] carefully or require the authors to include a proof. If the lemma holds, the paper is a solid contribution; if not, the main theorem narrows to whatever class the lemma actually covers.","headline":"Genuinely new geometric characterization of stable recovery, but the sufficiency direction leans on an unstated external convergence lemma that a referee must verify.","tokens_in":28586,"tokens_out":4064,"would_cite":false,"duration_ms":38174,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49J52","49J53","49K40","52A41","90C25","90C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"For continuous convex regularizers, stable recovery is exactly a second-order tangent-cone condition: the kernel of the observation map meets the conjugate subdifferential's tangent cone only at zero.","keywords":["stable recovery","linear inverse problems","regularization methods","second-order analysis","tangent cone","group sparsity","isotropic total variation","convex regularizers"],"falsifier":"Compute, for a continuous convex regularizer $R$ and a unique minimizer $x_0$ of the noiseless problem, a sequence of noisy data $y_k$ with $\\|y_k - y_0\\| \\le \\delta_k \\to 0$ whose regularized minimizers $x_k$ satisfy $\\|x_k - x_0\\|/\\delta_k \\to \\infty$ while condition (3.8) holds; exhibiting such a sequence would refute the characterization, and the explicit matrices in Examples 3.2 and 3.4 provide small cases where this can be checked by hand. Similarly, a case where (3.8) fails but all regularized minimizers still converge at the linear rate would refute the converse direction.","tokens_in":27629,"feed_emoji":"🎯","tokens_out":10548,"duration_ms":84568,"temperature":0.7,"pith_summary":"This paper tries to settle exactly when a convex regularized linear inverse problem recovers the true signal at the same rate as the noise level, the property called stable recovery. The central claim is a characterization: for any continuous convex regularizer, an optimal solution $x_0$ is stably recoverable if and only if the kernel of the measurement operator meets a certain tangent cone of the conjugate regularizer's subdifferential image only at zero. This is a second-order condition, and it explains why earlier first-order sufficient conditions (null space property, nondegeneracy source condition with restricted injectivity, minimum gain property) are not necessary: they imply sharp minima, and sharp minima imply the new condition but not conversely. The paper then derives new sufficient conditions for analysis group sparsity and isotropic total variation problems and reports numerical evidence that most strong non-sharp solutions of group sparsity problems are stably recoverable. This matters because it determines when linear convergence is guaranteed even in non-polyhedral problems where uniqueness alone is not enough.","feed_headline":"Stable recovery boils down to one tangent-cone condition","feed_subtitle":"Linear noise-rate recovery holds exactly when a tangent cone meets the kernel only at zero.","key_machinery":"The load-bearing object is the tangent/contingent cone $T_{\\partial R^*(\\operatorname{Im}\\Phi^*)}(x_0)$, the set of directions $w$ for which there are times $t_k \\downarrow 0$ and vectors $w_k \\to w$ with $x_0 + t_k w_k$ inside the image of the conjugate subdifferential restricted to $\\operatorname{Im}\\Phi^*$. The paper uses this set as a second-order derivative object: it records the directions in which the subdifferential of the conjugate regularizer, evaluated along the adjoint range, bends at $x_0$, and its intersection with $\\operatorname{Ker}\\Phi$ decides stability. Fenchel duality and subdifferential calculus set up the correspondence between primal feasible directions and dual certificates, while the critical cone and the set $W(x_0)$ are used to specialize the condition to analysis group sparsity problems, eventually yielding nonconvex quadratic conditions that can be checked numerically.","core_discovery":"The paper's central claim is Theorem 3.3: with $R$ a continuous convex function and $x_0$ an optimal solution of $\\min R(x)$ subject to $\\Phi x = y_0$, stable recovery occurs if and only if $\\operatorname{Ker}\\Phi \\cap T_{\\partial R^*(\\operatorname{Im}\\Phi^*)}(x_0) = \\{0\\}$, where $T$ denotes the contingent/tangent cone, $\\partial R^*$ the subdifferential of the Fenchel conjugate of $R$, and $\\operatorname{Im}\\Phi^*$ the range of the adjoint operator. The proof builds explicit noisy problems to show that any nonzero common direction would let noise push regularized minimizers away at a nonlinear rate, and conversely uses the condition to rule out any sequence of minimizers that fails to approach $x_0$ linearly. The paper argues that this condition is genuinely second-order: it records directions in which the conjugate subdifferential image bends at $x_0$, not just first-order descent directions. It shows by example that stable recovery can hold without sharp minima, that strong minima do not always imply stable recovery, and that for piecewise linear-quadratic regularizers stable recovery is equivalent to solution uniqueness. For smooth convex regularizers the condition becomes $\\operatorname{Ker}\\Phi \\cap \\operatorname{Ker}\\nabla^2 R(x_0) = \\{0\\}$, which is exactly the strong minimum condition.","pith_inferences":["The paper does not pursue this, but the tangent-cone condition could be turned into a numerical certificate: an optimization over $\\operatorname{Ker}\\Phi$ and the tangent cone whose zero optimal value verifies stable recovery without solving the regularized problems.","Beyond the paper, the characterization suggests that measurement thresholds for group-sparse recovery are governed by this second-order cone rather than by sharp-minimum conditions, so random matrix theory for cones could yield direct bounds on the number of measurements needed for stable recovery.","For isotropic total variation the equality in Theorem 4.3 fails because the discrete gradient is not surjective, so the paper's sufficient condition leaves a gap; closing it likely requires a direct computation of the tangent cone for the TV regularizer, which the paper leaves open.","The same tangent-cone strategy may transfer to nuclear norm regularization, where the conjugate subdifferential image has known structure, potentially characterizing stable recovery in low-rank inverse problems."],"forward_implications":["If the theorem is right, stable recovery for any continuous convex regularizer is a second-order property, and all earlier first-order sufficient conditions are merely routes to the tangent-cone condition.","For piecewise linear-quadratic regularizers (elastic net, Huber norm, discrete Blake-Zisserman), stable recovery and solution uniqueness coincide, extending the known $\\ell^1$ equivalence to a broad class.","For smooth convex regularizers, stable recovery is equivalent to strong minimality, so the condition is checkable as $\\operatorname{Ker}\\Phi \\cap \\operatorname{Ker}\\nabla^2 R(x_0) = \\{0\\}$.","For analysis group sparsity problems, conditions (4.45) and (4.50) certify stable recovery at unique non-sharp minimizers and are independent of sharp minima.","Numerical experiments suggest that in random Gaussian group-sparsity and total-variation problems, most strong non-sharp optimal solutions pass the new stability test."],"supporting_citations":[{"why":"Supplies the radial-cone theorem that derives uniqueness of the noiseless solution from condition (3.8) in the converse proof.","marker":"[20]"},{"why":"Provides the cited convergence of regularized minimizers that carries the 'if' direction of Theorem 3.3, and the sharp-minimum framework the paper compares against.","marker":"[21]"},{"why":"Gives the $\\ell^1$ necessary and sufficient condition for linear convergence that the paper extends to piecewise linear-quadratic regularizers.","marker":"[25]"},{"why":"Formulates the minimum gain property, an earlier sufficient condition that the paper shows is implied by sharp minima.","marker":"[12]"},{"why":"Uses nondegeneracy source condition with restricted injectivity, an earlier sufficient condition the paper shows is not necessary.","marker":"[7]"},{"why":"Provides earlier null-space-type sufficient conditions for stable recovery with analysis decomposable priors.","marker":"[22]"},{"why":"Supplies the restricted injectivity condition (4.17) for positively homogeneous functionals that the new group-sparsity conditions avoid.","marker":"[24]"},{"why":"Provides the tangent-cone definition and the piecewise linear-quadratic calculus used in the specialization arguments.","marker":"[35]"}],"fun_headline_variants":["Stable recovery comes down to one tangent-cone condition","One tangent-cone condition decides stable recovery in inverse problems","Stable recovery iff a tangent cone meets kernel only at zero","Tangent-cone test: the only condition for stable recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is the cited result that once $x_0$ is the unique solution of the noiseless problem, every sequence of regularized minimizers converges to $x_0$ as the noise vanishes; the proof of the 'if' direction imports this theorem without re-deriving it.","fun_headline_variants_meta":{"raw":{"variants":["Stable recovery comes down to one tangent-cone condition","One tangent-cone condition decides stable recovery in inverse problems","Stable recovery iff a tangent cone meets kernel only at zero","Tangent-cone test: the only condition for stable recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1399,"prompt_tokens":951,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":567,"tokens_out":448,"duration_ms":4641,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:06:43.909752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for a continuous convex regularizer $R$ and a unique minimizer $x_0$ of the noiseless problem, a sequence of noisy data $y_k$ with $\\|y_k - y_0\\| \\le \\delta_k \\to 0$ whose regularized minimizers $x_k$ satisfy $\\|x_k - x_0\\|/\\delta_k \\to \\infty$ while condition (3.8) holds; exhibiting such a sequence would refute the characterization, and the explicit matrices in Examples 3.2 and 3.4 provide small cases where this can be checked by hand. Similarly, a case where (3.8) fails but all regularized minimizers still converge at the linear rate would refute the converse direction.","supporting_citations":[{"cited_title":"Solution uniqueness of convex optimization problems via the radial cone","cited_arxiv_id":"2401.10346","evidence_quote":"Supplies the radial-cone theorem that derives uniqueness of the noiseless solution from condition (3.8) in the converse proof."},{"cited_title":"Fadili, T","cited_arxiv_id":null,"evidence_quote":"Provides the cited convergence of regularized minimizers that carries the 'if' direction of Theorem 3.3, and the sharp-minimum framework the paper compares against."},{"cited_title":"Grasmair, O","cited_arxiv_id":null,"evidence_quote":"Gives the $\\ell^1$ necessary and sufficient condition for linear convergence that the paper extends to piecewise linear-quadratic regularizers."},{"cited_title":"Chandrasekaran, B","cited_arxiv_id":null,"evidence_quote":"Formulates the minimum gain property, an earlier sufficient condition that the paper shows is implied by sharp minima."},{"cited_title":"Cand `es and B","cited_arxiv_id":null,"evidence_quote":"Uses nondegeneracy source condition with restricted injectivity, an earlier sufficient condition the paper shows is not necessary."},{"cited_title":"Fadili, G","cited_arxiv_id":null,"evidence_quote":"Provides earlier null-space-type sufficient conditions for stable recovery with analysis decomposable priors."},{"cited_title":"Grasmair","cited_arxiv_id":null,"evidence_quote":"Supplies the restricted injectivity condition (4.17) for positively homogeneous functionals that the new group-sparsity conditions avoid."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the tangent-cone definition and the piecewise linear-quadratic calculus used in the specialization arguments."}],"review_version":1}