{"id":"01a90462-0b5a-46fd-a90d-bc2758b62992","arxiv_id":"2608.12757","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"The DCR method approximates the Laplacian pseudoinverse via regularized MLE, but its claimed solution reconstruction is a supervised fit to the CVXPY reference, making the numerical validation circular.","lead":"The paper proposes a two-phase framework that learns a graph Laplacian pseudoinverse approximation with a Tyler-type regularized MLE, then reconstructs nonnegative least squares solutions by fitting dual variables to the convex-solver output. The reported speedups over CVXPY are not solver comparisons, because the fitting phase requires the CVXPY solution as its training label.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phase 2 of Algorithm 1 optimizes Eq. (58) against the CVXPY reference solution, so the reported accuracy and speedups are measures of supervised fitting; without x*_cvxpy, no rule selects (λ, μ, c) at inference.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: Phase 2 is supervised by x*_cvxpy through Eq. (58) and Algorithm 1 lines 14–18. This is not a speculative risk; the method section states that the variables are learned to induce a primal candidate matching x*_cvxpy, and the experimental section counts only Phase 2 runtime. Because the reconstructed x_DCR is a trained affine map evaluated at fitted dual variables, its near-machine-precision agreement with CVXPY (e.g., relative solution error 3.3e-9 at n=1000) is expected from overfitting rather than from any property of the approximation L̃†. The Phase 1 Tyler-style estimation may provide a useful spectral approximation, and the CCCP derivations appear internally coherent, so the concern is not about the algebra; it is about the meaning of the numerical evaluation. A test that removes the reference from the loss would settle the question. The recommendation is therefore to reject the paper's central claims in their current form, while noting that a revised version with an inference-time objective for (λ, μ, c) might support a conditional acceptance.","tokens_in":22939,"tokens_out":5719,"duration_ms":55303,"concrete_test":"Re-run the Section V-B protocol with Eq. (58) replaced by the primal objective of (3) evaluated at x(λ, φ(μ̃), c), i.e., choose (λ, μ̃, c) by minimizing (1/2)||A x(λ, φ(μ̃), c) − b||² + (1/2)x(λ, φ(μ̃), c)ᵀ L x(λ, φ(μ̃), c) with no access to x*_cvxpy, using the same Adam schedule and Phase 1 output. Record relative solution error vs CVXPY and time-to-accuracy. If the error no longer reaches 10⁻³ or the time-to-accuracy exceeds the CVXPY time, the reported performance is attributable to the supervised target in Eq. (58).","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that DCR solves (3) faster than CVXPY is undercut by the Phase 2 objective (58): min over (λ, μ̃, c) of ||x(λ, φ(μ̃), c) − x*_cvxpy||², with x(λ, μ, c) = L̃†(μ − Aᵀλ) + c1. Algorithm 1 line 15 evaluates this loss against the reference solution, and line 17 again fits c to x*_cvxpy; line 18 returns the projected fit. Therefore x_DCR is optimized to reproduce x*_cvxpy, not to minimize the primal objective (3) or to satisfy the KKT conditions, so the relative solution errors and objective gaps in Table III and Figs. 5–7 are training errors on the same instance. At inference time, for a new b, there is no mechanism to choose λ, μ, c without either x*_cvxpy or an additional optimization, so DCR is not a standalone solver. The runtime comparison counts only Phase 2 time (Section V-B), yet Phase 2 requires the CVXPY solution as its training target; the full pipeline therefore costs at least the CVXPY solve time plus Phase 2. Separately, the abstract's 'unique fixed point' claim exceeds Theorems 2–3, which establish existence but not uniqueness; this is secondary to the Phase 2 circularity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Difference-of-Convex Regularizer (DCR) framework for Laplacian-regularized nonnegative least squares (LR-NNLS), with the Laplacian given. The method has two phases. Phase 1 learns an approximation of the Laplacian pseudoinverse on the range space via a regularized Tyler-type MLE solved by a CCCP fixed-point iteration. Phase 2 learns auxiliary dual variables (λ, μ, c) by minimizing the squared distance between the reconstruction x(λ, μ, c) = L~†(μ − A^T λ) + c1 and a high-accuracy CVXPY reference solution x*_cvxpy, followed by a final nonnegativity projection. The paper states theoretical guarantees on fixed-point existence and reports numerical experiments showing that DCR achieves high solution accuracy and speedups over CVXPY and Chebyshev baselines across graph topologies and problem sizes.","tokens_in":23282,"tokens_out":3629,"duration_ms":36428,"significance":"The paper contains a few genuinely useful components: the KKT-based primal reconstruction formula in Theorem 1 is clean and correctly identifies the role of the null-space constant; the connection between CCCP and Frank-Wolfe in Remark 1 is a nice observation; the Phase 1 MLE formulation for approximating the spectral action of the Laplacian pseudoinverse is interesting and is validated on a small example against a CVXPY baseline. The authors also state that source code is available. However, the central claim that DCR is a fast solver for (3) that outperforms convex solvers is not supported by the evidence: Phase 2 is a supervised fitting procedure trained on the CVXPY solution of the very same instance, so the reported accuracy and speedups do not demonstrate that (3) is solved faster. The theoretical contribution is also weaker than advertised, as only existence, not uniqueness, of fixed points is proved. If the paper were reframed as a learning-to-predict method conditioned on reference solutions, the significance would be different, but as a solver the contribution is not established.","major_comments":[{"comment":"The Phase 2 objective is min_{λ, μ~, c} ||x(λ, φ(μ~), c) − x*_cvxpy||², where x*_cvxpy is the CVXPY solution of the same instance and x(λ, μ, c) = L~†(μ − A^T λ) + c1. Consequently, the reported relative solution errors and objective gaps in Table III and Figs. 5–7 are training errors, not measures of how accurately (3) is solved. The method also provides no rule for selecting (λ, μ, c) at inference time for a new instance without a reference solution, so DCR, as presented, is not a standalone solver for (3). This circularity is the central issue and it invalidates the paper's main claim of improved performance over convex solvers.","section":"Section IV-C, Eq. (58), Algorithm 1 lines 15–18"},{"comment":"The runtime comparison counts only Phase 2 time, with the text stating 'For DCR, only Phase 2 (online) time is counted; Phase 1 is excluded as it is amortized across instances with the same graph.' However, Phase 2 requires the CVXPY solution of the same instance as its training target (Algorithm 1 line 15), so the full pipeline cost is at least the CVXPY solve time plus the Phase 2 optimization time. The reported speedups of up to 13.2× over CVXPY therefore do not reflect end-to-end wall-clock performance and are not a fair comparison with a direct solver.","section":"Section V-A, runtime metric, and Fig. 6"},{"comment":"The abstract claims 'the existence of a unique fixed point' for the DCR algorithm, but Theorem 2 only establishes existence of an eigenvector fixed point α*Σ* for M_γ and hence a fixed point of the normalized map; it does not prove uniqueness. Theorem 3 similarly establishes at least one fixed point via Brouwer's theorem. No uniqueness result is proved anywhere in the manuscript, so the abstract overstates the theoretical guarantee. This is secondary to the Phase 2 circularity but still requires correction.","section":"Abstract and Theorems 2–3"},{"comment":"The number of random samples used in Phase 1 is set to K(n) = min(Kmax, max(Kmin, ⌈2√n⌉)) in Eq. (60), but Section IV-A2 justifies the method only by saying that 'with sufficiently large K, the set {r_k} spans R(L) with high probability.' For n = 2000, this gives K ≈ 89 samples to approximate a pseudoinverse on an (n−1)-dimensional range space, which is far below any standard sample-complexity guarantee for covariance or Tyler-type estimation. Because Phase 2 fits the reconstruction to the reference solution, the experiments do not isolate whether such a small K actually provides a faithful approximation of L†; this weakens the claimed decoupling of pseudoinverse learning from instance-specific inference.","section":"Section IV-A2 and Eq. (60)"}],"minor_comments":[{"comment":"The symbol L is used both for the graph Laplacian and for the training loss in Algorithm 1 line 15 and in the text 'Update (λ, μ~, c) by one gradient step on L'; this overloading is confusing and should be resolved by renaming the loss, for example as J or ℓ.","section":"Section IV-C, line 16"},{"comment":"The caption states that the vertical dashed line marks the CVXPY solve time, but the DCR curves begin at Phase 2 start; a note that Phase 1 time is excluded from the DCR curves would help the reader interpret the comparison.","section":"Fig. 5 caption"},{"comment":"The Chebyshev baseline is described as being 'integrated into a gradient-based iterative solver for (3)', but no details are given on the step-size rule or number of inner iterations, making it hard to assess whether the comparison is favorable to DCR or the baseline.","section":"Section V-A, Chebyshev baseline"},{"comment":"Several equations and displayed expressions are typeset with missing or ambiguous spacing (e.g., Eq. (17) and the normalization in Theorem 2), which makes the mathematics harder to read; a careful proofreading pass would improve clarity.","section":"Throughout"},{"comment":"References [29] and [31] are arXiv preprints dated 2026; if these are not yet published, the authors should indicate whether they are under review or should cite accessible versions.","section":"References"}],"recommendation":"reject","confidential_remarks":"The central problem is that Phase 2 of Algorithm 1 is a supervised fit to the CVXPY reference solution of the same instance. This makes the accuracy results circular and the runtime comparison unfair, and it means DCR is not a solver for (3) at inference. The Phase 1 pseudoinverse-learning idea and the KKT reconstruction formula have some independent interest, but the paper's framing as a faster solver cannot be fixed by a local revision; it would require either a different inference mechanism for choosing dual variables or a substantially rewritten paper that drops the solver claims. The uniqueness claim in the abstract should also be corrected regardless of venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [colleague],\n\nThe headline: the paper has a genuinely interesting Phase 1, but the evaluation of Phase 2 is circular. DCR learns an approximation of L† on the range of L from random directions Lz via a regularized Tyler estimator. That part is new relative to the cited pseudoinverse-learning papers and seems to work: the spectral alignment experiments are convincing. The DC/CCCP reformulation is standard but clean, and the fixed-point analysis is fine as far as it goes.\n\nThe problem is that Phase 2 does not solve anything. The differentiable reconstruction minimizes Eq. (58), which is ||x(λ,φ(μ̃),c) − x*_cvxpy||², the distance to the CVXPY reference solution. Algorithm 1 line 15 evaluates exactly that loss, and line 17 fits c again to x*_cvxpy. So the reported relative solution errors and objective gaps are training errors, measured against the training target. The runtime comparison in Section V-B counts only Phase 2 time and explicitly excludes Phase 1, but Phase 2 requires the CVXPY solution as its labels; the full pipeline costs at least the CVXPY solve plus Phase 2. For a new instance with new A,b, there is no rule for choosing (λ,μ,c) without the reference solution or another optimization. So DCR as presented is not a solver for (3); it is a supervised fitting scheme that reproduces the reference.\n\nThere are smaller issues. The abstract claims a \"unique fixed point\" but Theorems 2 and 3 prove existence only. The large-scale experiments run in the undersampled regime K = 2√n << n, where the classical Tyler theory does not apply; the paper includes an adaptive shrinkage step and Theorem 3 covers the shrunk map, but the link to \"learning L†\" is looser there. Table III also shows DCR is slower than CVXPY at small n and only \"wins\" at larger n when the clock starts after the reference solution has been computed.\n\nWho is this for? Someone working on spectral approximation of graph operators could get value from Phase 1. The evaluation methodology should make anyone working on learned optimizers pause. I'd send it to review — the Phase 1 idea deserves referee scrutiny — but the paper needs a major rewrite: either reposition it as pseudoinverse approximation with a non-circular evaluation, or remove the solver claims entirely.","headline":"A novel Phase 1 for learning L† via Tyler estimation, but Phase 2 fits the CVXPY reference directly, so the speedup claims don't hold.","tokens_in":23799,"tokens_out":2247,"would_cite":false,"duration_ms":21384,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C25","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"DCR replaces the dense graph Laplacian pseudoinverse with a learnable spectral approximation and reconstructs nonnegative least-squares solutions faster than direct convex solvers.","keywords":["graph Laplacian pseudoinverse","difference-of-convex regularization","Tyler's estimator","CCCP","Laplacian-regularized nonnegative least squares","Toland duality","differentiable programming","graph learning"],"falsifier":"Take a fixed graph and a new pair $(A,b)$ not used in Phase 2 training, run only Phase 2 from random dual initialization with the learned $\\tilde L^\\dagger$, and compare $x_{\\mathrm{DCR}}$ against the true optimum from an independent solver: the decoupling claim collapses if accuracy degrades materially across instances. A sharper test is to check the KKT residual of the learned dual variables—if $\\|L x_{\\mathrm{DCR}} + A^\\top \\lambda^\\star - \\mu^\\star\\|$ stays far from zero while the objective gap is small, the reconstruction is fitting the reference rather than solving the problem.","tokens_in":22726,"feed_emoji":"🧮","tokens_out":10138,"duration_ms":76387,"temperature":0.7,"pith_summary":"The paper tries to show that the dense, ill-conditioned graph Laplacian pseudoinverse can be replaced by a learnable approximation that is spectrally accurate on the Laplacian's range space, so that Laplacian-regularized nonnegative least squares can be solved without ever forming $L^\\dagger$. The proposed two-phase scheme first learns $\\tilde L^\\dagger$ through regularized maximum likelihood estimation written as a difference-of-convex program and solved by CCCP iterations, then uses a differentiable dual-guided map to reconstruct the primal solution. If the claim holds, the practical payoff is a graph-learning method whose online cost is a few matrix-vector products instead of an $O(n^3)$ pseudoinverse, with speedups over a standard convex solver that grow with problem size.","feed_headline":"Learning the Laplacian pseudoinverse speeds up graph-regularized LS","feed_subtitle":"A two-phase differentiable scheme learns the inverse operator and hits solver-grade accuracy in seconds.","key_machinery":"The central object is the two-phase DCR pipeline. Phase 1 learns $\\tilde L^\\dagger = Q \\sqrt{\\Sigma^{-1}_\\star} Q^\\top$, where $Q$ is an orthonormal basis of the Laplacian range space and $\\Sigma_\\star$ solves a scale-invariant regularized MLE of Tyler type; the objective is written as a difference of convex functions, $f(X) - h(X)$, and solved by CCCP, whose update is the fixed-point iteration (34), equivalently a Frank–Wolfe step. Phase 2 uses the dual representation $x(\\lambda,\\mu,c) = \\tilde L^\\dagger(\\mu - A^\\top \\lambda) + c\\mathbf{1}$ as a fully differentiable map, trains $(\\lambda,\\tilde\\mu,c)$ against a reference solution with loss (58), then returns the projected reconstruction (59). The load-bearing identities are the KKT solution formula (11) and the Tyler-type fixed-point mapping (43), whose nonlinear Perron–Frobenius structure gives existence and stability guarantees.","core_discovery":"For the LR-NNLS problem (3), the paper derives from the KKT conditions that any primal solution has the form $x^\\star = L^\\dagger(\\mu^\\star - A^\\top \\lambda^\\star) + c\\mathbf{1}$, with the constant $c$ determined by complementary slackness. It then constructs $\\tilde L^\\dagger$ as an approximation of $L^\\dagger$ on the range space of $L$, learned by a scale-invariant regularized MLE of Tyler type whose stationary point is reached by a CCCP fixed-point iteration, and shows that Phase 2's differentiable training of $(\\lambda, \\mu, c)$ against a high-accuracy reference solution yields a reconstructed $x_{\\mathrm{DCR}} = [\\tilde L^\\dagger(\\mu^\\star - A^\\top\\lambda^\\star) + c^\\star \\mathbf{1}]_+$ that matches the convex solver's solution to high accuracy. The paper also establishes fixed-point existence and stability guarantees for the iteration. Numerically, on grid2d and Erdős–Rényi graphs with $n$ up to 2000, DCR reaches a $10^{-3}$ relative objective gap in seconds and achieves speedups up to 13.2$\\times$ over CVXPY, while the Chebyshev polynomial baseline fails to reach the threshold.","pith_inferences":["Because Phase 2 supervises the dual variables with a high-accuracy reference solution, the pipeline is not yet a standalone solver for a fresh instance; a natural test is zero-shot transfer of the trained dual map to unseen $(A,b)$ on the same graph.","The same KKT structure extends beyond NNLS: any Laplacian-regularized problem whose solution is characterized through $L^\\dagger$ and constraint duals could reuse the learned operator, so DCR may port to ranking, nonnegative matrix factorization, and hyperspectral unmixing.","One could replace the reference-supervised loss with an unsupervised objective that minimizes the LR-NNLS objective directly in the dual parameters, removing the dependence on a convex solver.","A stress test is whether the learned dual variables satisfy the KKT conditions exactly or only approximately in a way the final nonnegativity projection hides."],"forward_implications":["For a fixed graph, Phase 1's learned $\\tilde L^\\dagger$ can be reused across many instances $(A,b)$, amortizing the expensive pseudoinverse approximation.","The dual-guided reconstruction turns primal recovery into a differentiable program whose dimension is $O(m+n)$, not $O(n^2)$.","Accuracy holds on both grid2d and Erdős–Rényi topologies, indicating the learned operator adapts to spectral structure rather than relying on a fixed polynomial approximation.","Time-to-accuracy grows slowly with $n$, and speedups over CVXPY increase with problem size, so the method becomes more attractive exactly where direct inversion is prohibitive.","The Chebyshev baseline's large, irreducible error suggests fixed polynomial approximations of $L^\\dagger$ carry a bias that a learned spectral approximation can remove."],"supporting_citations":[{"why":"Supplies the high-accuracy reference solutions $x^\\star_{\\mathrm{cvxpy}}$ that supervise Phase 2 training and define the accuracy metrics.","marker":"[41]"},{"why":"Tyler's estimator provides the fixed-point likelihood model that Phase 1's regularized MLE builds on.","marker":"[42]"},{"why":"Regularized Tyler's scatter estimator supplies the existence and algorithmic foundation for the shrinkage-stabilized iteration.","marker":"[43]"},{"why":"CCCP is the iterative scheme that linearizes the concave part and yields the fixed-point update for the DC objective.","marker":"[26]"},{"why":"Toland duality frames the DC formulation and gives the optimality condition used to derive the CCCP stationarity.","marker":"[25]"},{"why":"Nonlinear Perron–Frobenius theory is used to prove the existence of the fixed point of the Tyler-type mapping.","marker":"[50]"},{"why":"Shrinkage stabilization of the iteration is borrowed from robust covariance shrinkage estimation.","marker":"[53]"},{"why":"General solution of linear systems via the pseudoinverse justifies the $+c\\mathbf{1}$ term in the KKT reconstruction.","marker":"[8]"}],"fun_headline_variants":["Learn the Laplacian pseudoinverse to speed up graph-regularized LS","Difference-of-convex training learns the pseudoinverse fast","13x faster graph-regularized LS by learning the inverse","Dual-guided learning cracks dense Laplacian pseudoinverse","Pseudoinverse learning beats Chebyshev for graph LS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's online accuracy depends on having a high-accuracy reference solution from a convex solver to train the dual variables; without such a reference, the paper provides no objective for choosing $(\\lambda,\\mu,c)$ at inference, so DCR cannot yet act as a standalone solver for a new instance.","fun_headline_variants_meta":{"raw":{"variants":["Learn the Laplacian pseudoinverse to speed up graph-regularized LS","Difference-of-convex training learns the pseudoinverse fast","13x faster graph-regularized LS by learning the inverse","Dual-guided learning cracks dense Laplacian pseudoinverse","Pseudoinverse learning beats Chebyshev for graph LS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1502,"prompt_tokens":1022,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":638,"tokens_out":480,"duration_ms":4570,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:54:46.877117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed graph and a new pair $(A,b)$ not used in Phase 2 training, run only Phase 2 from random dual initialization with the learned $\\tilde L^\\dagger$, and compare $x_{\\mathrm{DCR}}$ against the true optimum from an independent solver: the decoupling claim collapses if accuracy degrades materially across instances. A sharper test is to check the KKT residual of the learned dual variables—if $\\|L x_{\\mathrm{DCR}} + A^\\top \\lambda^\\star - \\mu^\\star\\|$ stays far from zero while the objective gap is small, the reconstruction is fitting the reference rather than solving the problem.","supporting_citations":[{"cited_title":"A distribution-free M-estimator of multivariate scatter,","cited_arxiv_id":null,"evidence_quote":"Tyler's estimator provides the fixed-point likelihood model that Phase 1's regularized MLE builds on."},{"cited_title":"Wireless network optimization by Perron-Frobenius theory,","cited_arxiv_id":null,"evidence_quote":"Nonlinear Perron–Frobenius theory is used to prove the existence of the fixed point of the Tyler-type mapping."}],"review_version":1}