{"id":"d8842644-ba98-4de5-9fdc-9b4c8c5b89e2","arxiv_id":"2411.18127","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A collaborative neurodynamic optimizer (recurrent networks plus particle swarm) is proposed for nonnegative canonical polyadic decomposition, with a claimed almost-sure global convergence guarantee and experiments on collinear and real-world tensors.","lead":"This paper replaces the usual alternating-update solver for nonnegative tensor decomposition with a swarm of recurrent neural networks that share solutions through particle swarm optimization. The authors claim this collaborative neurodynamic scheme converges to a global optimum with probability one, and they test it on random, face-image, and hyperspectral tensors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3's almost-sure global convergence claim rests on Assumption II, which the proof asserts via support coverage without establishing the required positive-measure hitting condition; PSO contraction makes the condition doubtful.","rationale":"The reader's weakest-assumption analysis correctly identifies Assumption II as the point where the global convergence proof fails. My reading agrees: Theorem 4.3 reduces to an unproved probabilistic claim about the PSO-wavelet search process, and the proof's support-coverage argument does not establish the product condition (4.1). The numerical experiments may show good performance, but they do not compensate for the missing guarantee, especially since the algorithm's parameters (small beta1, beta2) create strong attraction to the current global best, making exploration of the full feasible region implausible. The secondary issue in Lemma 5.1 is also real: on the nonnegative orthant's boundary, the fixed-point condition of the projected dynamics permits nonzero gradient components, so the claimed equivalence between continuous and discrete equilibria is false. This reinforces the rejection but is not the primary load-bearing concern. Therefore the reader's REJECT verdict should stand unchanged.","tokens_in":21685,"tokens_out":3442,"duration_ms":32799,"concrete_test":"Run Algorithm 3.1 on a small nonnegative CPD instance (e.g., 5x5x5, R=3) with the paper's PSO parameters alpha=0.5, beta1=beta2=0.01, using a fixed target set B equal to a small ball around a known global minimizer. Over M=1000 independent runs, record the empirical fraction of runs in which the swarm ever enters B as kmax increases. If this fraction does not approach 1, Assumption II's condition (4.1) is violated in practice. Alternatively, analytically compute the support of one wavelet-mutation update from a fully collapsed swarm at a single point x and show it is a bounded, lower-dimensional set, contradicting the proof's claim that the search support covers Omega at every iteration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 4.3: CNO-CPD is globally convergent with probability one. The proof invokes Lemma 4.2, which requires Assumption II: for every positive-measure Borel subset B, the product of (1 - mu_k(B)) over k equals zero. The proof attempts to verify this by stating that wavelet mutation and re-initialization make the support cover the feasible region, so the search space S covers Omega. This conflates support coverage with the conditional-probability condition (4.1). Algorithm 3.1 reinitializes RNNs only at initialization; during iterations, positions are updated by PSO (2.6)-(2.7) with beta1 = beta2 = 0.01 and are wavelet-mutated only when diversity DI < delta. The mutation map (2.11)-(2.12) is deterministic once phi is drawn, has bounded image (with l=0, x' = (1+kappa)x), and for a collapsed swarm may move all particles to nearby or identical points. It does not imply positive probability of hitting every positive-measure Borel subset, so mu_k(B) need not satisfy (4.1). Without Assumption II, Lemma 4.2 does not apply, and Theorem 4.3 is unproved. This is load-bearing because the paper's headline guarantee of global convergence w.p.1 rests on this step. A secondary gap in Lemma 5.1 further weakens the discrete-time analysis: the equilibrium condition -x + [x - grad f]_+ = 0 is not equivalent to grad f = 0 when x is on the boundary of the nonnegative orthant, so the equivalence and stability statements in Section 6 inherit this error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a collaborative neurodynamic optimization method for nonnegative canonical polyadic decomposition (CNO-CPD): each of several recurrent neural networks solves a projected-gradient-type continuous-time system, and the networks exchange information through particle swarm optimization with wavelet mutation. A discrete-time projection neural network (DTPNN) with backtracking steps and Hessian preconditioning is also introduced. The main theoretical claims are that CNO-CPD is globally convergent with probability one (Theorem 4.3), that the discrete-time model is Lyapunov stable and convergent (Theorems 6.2 and 6.3), and that the collaborative discrete version inherits the almost-sure global convergence guarantee. The paper also reports extensive numerical comparisons on synthetic collinear tensors, real image datasets, and hyperspectral data, with code made publicly available.","tokens_in":22088,"tokens_out":3851,"duration_ms":36256,"significance":"If the almost-sure global convergence claim were established, the paper would make a substantial contribution to nonnegative tensor decomposition by offering a distributed stochastic search method with a rigorous global guarantee in a nonconvex setting. The experimental work, particularly on tensors with highly collinear factors, suggests that the proposed scheme can be empirically competitive, and the availability of code is a practical strength. However, the mathematical core is not sound as stated: the proof of the headline global-convergence theorem rests on an unverified probabilistic assumption, and the equilibrium analysis of the discrete-time model contains a boundary-error that propagates into the stability and convergence claims. As a math.NA paper, these are load-bearing issues, so the theoretical contribution as written cannot be accepted.","major_comments":[{"comment":"The proof of Theorem 4.3 does not establish Assumption II. The argument states that wavelet mutation and re-initialization make the support of the search distribution cover the feasible region, and concludes that the product condition (4.1) holds. This conflates support coverage with the conditional-probability condition: Assumption II requires, for every positive-measure Borel subset B, that the infinite product of (1 - μ_k(B)) is zero, where μ_k(B) is the conditional probability of hitting B at iteration k. The paper gives no lower bound on these conditional probabilities. In Algorithm 3.1, re-initialization from the whole feasible region occurs only at the initial step; thereafter the states are moved by the PSO update (2.6)-(2.7) with β1 = β2 = 0.01, and wavelet mutation is applied only when the diversity falls below δ. The map (2.11)-(2.12) has bounded image within [l, u] and need not assign positive probability to every positive-measure Borel set, especially when the swarm is contracted near a current best. Thus Lemma 4.2 cannot be invoked, and the claim that CNO-CPD is globally convergent with probability one is unsupported.","section":"§4, Theorem 4.3 and Assumption II"},{"comment":"The equivalence claimed in Lemma 5.1 is false on the boundary of the nonnegative orthant. The equilibrium condition -x + [x - ∇F(x)]_+ = 0 does not imply ∇F(x) = 0: if x_i = 0 and ∇_i F(x) < 0, then [x_i - ∇_i F(x)]_+ = 0 = x_i, so the fixed-point equation holds while the gradient component is nonzero. Consequently the continuous-time neurodynamic (3.5)-(3.7) and the discrete-time update (5.2)-(5.4) need not have the same equilibrium points. This error is load-bearing because the subsequent stability analysis in §6, including Theorem 6.2 and the use of equilibrium points in the proof of Theorem 6.3, and the conclusion that CNO-DTPNN inherits the global convergence of Theorem 4.3, all rely on the equilibrium characterization.","section":"§5, Lemma 5.1"},{"comment":"The proof of Theorem 6.3 asserts that because the sequence {f(x_k)} is monotone non-increasing and bounded, and because (5.11) holds with ∇f(x_k)^T (x_{k+1} - x_k) ≥ 0, it follows that ∇f(x_k) = 0 in the limit. This does not follow: monotone convergence and Armijo-type inequalities with variable step sizes do not by themselves imply convergence to a stationary point unless additional conditions on the step sizes or the descent direction are supplied. The paper does not provide such conditions. Together with the boundary error in Lemma 5.1, this leaves the convergence claim for DTPNN and hence the global-convergence claim for CNO-DTPNN unproved.","section":"§6, Theorem 6.3"}],"minor_comments":[{"comment":"The text refers to the 'continues CPD' and later to 'continues neurodynamic'; these should read 'continuous'.","section":"§1, Abstract and §8"},{"comment":"The backtracking updates are written as λ_k ← -β λ_k, which makes the step size negative. The surrounding text says the step size is multiplied by a positive constant 0 < β < 1, so the minus signs appear to be typographical errors; as written, the algorithm does not implement the described backtracking procedure.","section":"Algorithm 5.1, lines 6, 9, 12"},{"comment":"The comparison in Figure 4 reports a single run for several baselines (notably ANLS 'for a successful run'), and no details are given on how many restarts, initializations, or random seeds were used. This makes it difficult to assess whether the reported speedups are statistically meaningful.","section":"§7, Example 1 and Figure 4"},{"comment":"The text states that the population size is q = 5 for the collinearity experiments, while the captions of Figure 7 report q = 30; the discrepancy should be resolved.","section":"§7, Example 2, Figure 7"},{"comment":"The description of Case Study I says one factor matrix is highly collinear while 'two other factor matrices (A(2), A(3))' have collinearity 0.4 ≤ μ ≤ 0.6; presumably the first of these should be A(1), since A(3) was already designated as the highly collinear factor. Please clarify.","section":"§7, Example 2, Case Study I"},{"comment":"The acknowledgment thanking 'the editor and two reviewers' is inappropriate for a submitted manuscript and should be removed regardless of the review outcome.","section":"§9, Acknowledgment"}],"recommendation":"reject","confidential_remarks":"The reader's report and my own reading agree on the decisive point: the almost-sure global convergence theorem is not proved, and the discrete-time convergence analysis contains a genuine boundary error. These are not presentation issues; they concern the central mathematical claims. The empirical part may have heuristic value, but a math.NA paper cannot rest on an unverified probabilistic condition that is close in content to the desired conclusion. The paper's repeated acknowledgement of editor and reviewers also suggests the submission process was not handled in the usual anonymous manner, which the editor may wish to check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the algorithmic core is interesting and the experiments, especially on high-collinearity tensors, show real gains. But the paper's headline claim — global convergence with probability one — does not survive contact with the proof. Theorem 4.3 relies on Assumption II, and the verification of that assumption in the proof is a support-coverage argument that does not imply the required probability condition. PSO with tiny acceleration constants and a wavelet mutation that only fires when diversity drops does not give positive probability of hitting every positive-measure Borel subset of the feasible set. So the almost-sure result is unproved. This is the load-bearing wall, and it is cracked.\n\nWhat is genuinely new: the CP-ALS-based continuous neurodynamic system with Hessian preconditioning (equations 3.8-3.10), the semi-implicit discrete projection network (5.5-5.10), and the PSO-swarm version for nonnegative CPD. The paper is honest that [15] already applied collective neurodynamic to nonnegative tensor decomposition; the contribution is the specific ALS-based formulation and the discrete network. The code is available, and the experiments cover random hard cases, hyperspectral data, and real face datasets. The Monte Carlo results with increasing swarm size are a nice touch. Credit where due: this is a reproducible numerical study.\n\nThe soft spots, in order: (1) The main theorem, as above. (2) Lemma 5.1 claims equilibrium of the projection dynamics implies zero gradient. That is false on the boundary of the nonnegative orthant: -x + [x - grad f]_+ = 0 does not force grad f = 0 when some components of x are zero. The same error leaks into Theorem 6.2 and Theorem 6.3. (3) Experimental comparisons: no error bars on the main relative-error figures, and the paper reports a single successful ANLS run for the difficult case, which is not a fair baseline comparison. Minor: a few typos and an appendix with inconsistent notation; these don't change the verdict.\n\nWho gets value from this paper: researchers working on hard nonnegative CPD instances, especially with collinear factors, could use the neurodynamic framework as a heuristic. The convergence theory needs to be either corrected or explicitly downgraded to 'empirical improvement' before the paper is publishable. A serious referee should be assigned, because the algorithmic contribution is worth developing even if the current theorems are not.","headline":"Worth reading for the algorithm and experiments, but the global-convergence theorem is unproved and the discrete-time stability proof has a boundary error.","tokens_in":22625,"tokens_out":2345,"would_cite":false,"duration_ms":20658,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","65F99","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a swarm of collaborating recurrent neural networks, coordinated by particle swarm optimization, solves nonnegative canonical polyadic decomposition with almost-sure global convergence.","keywords":["nonnegative tensor decomposition","canonical polyadic decomposition","collaborative neurodynamic optimization","recurrent neural network","particle swarm optimization","discrete-time projection neural network","Hessian preconditioning","global convergence"],"falsifier":"Take a one-dimensional feasible interval $[l,u]$ and iterate the wavelet-mutation map (2.11)-(2.12) from a collection of current states; if the reachable set after any number of steps omits a positive-measure interval of $[l,u]$, then the claimed support $\\mathcal{H}(\\mu_k) = \\Omega$ used in Theorem 4.3 is false for that $B$, and Assumption II is not satisfied by the implemented search.","tokens_in":21449,"feed_emoji":"🧠","tokens_out":7071,"duration_ms":61281,"temperature":0.7,"pith_summary":"Nonnegative canonical polyadic decomposition (CPD) is a nonconvex tensor factorization problem whose objective can have many local minima, and classical algorithms such as alternating least squares and HALS can stall on tensors with highly collinear factor matrices. This paper claims that a 'collaborative neurodynamic' model—a swarm of recurrent neural networks, each running a projected gradient-like differential equation, with particle swarm optimization (PSO) exchanging their best candidate solutions—solves this problem with a strong guarantee: the swarm's best iterate converges to a global minimum with probability one. A discrete-time version, based on an Euler discretization with adaptive step size, is shown to be Lyapunov stable and convergent to a partial optimum. If these convergence claims hold, the approach offers a way to push nonnegative tensor decompositions out of local minima without relying on tensor-specific structure, and the paper reports that it outperforms HALS, multiplicative updates, and other baselines on collinear and real-world data tensors, including hyperspectral unmixing tasks.","feed_headline":"Swarm of neural nets targets global tensor decomposition","feed_subtitle":"Collaborating recurrent networks share best solutions to escape local minima in nonnegative CPD","key_machinery":"The load-bearing object is the collaborative neurodynamic system: $q$ independent one-layer recurrent neural networks with dynamics $\\epsilon_1 \\dot A = -A + [A - \\nabla_A F\\, P_A^{-1}]_+$ (and analogous equations for $B,C$), where $[\\,\\cdot\\,]_+$ is projection onto the nonnegative orthant and $P_A = (C^TC)*(B^TB)$ is the Hessian preconditioner for factor $A$. Around this sits the PSO loop (2.6)-(2.9), a diversity measure $\\mathrm{DI}^{(k)}$ that triggers wavelet mutation (2.11)-(2.12) when particles cluster, and reinitialization that refreshes the swarm from the feasible region. The argument's logical joints are Lemma 4.1, identifying RNN equilibria with KKT points, and Lemma 4.2, the Solis–Wets random-search theorem that converts monotone non-increasing best values plus full-support sampling into almost-sure global convergence. The discrete counterpart is the Euler-based projection neural network (5.2)-(5.4) with Armijo backtracking, whose stability hinges on the step-size interval in Theorem 6.2.","core_discovery":"The paper's central claim is Theorem 4.3: the CNO-CPD algorithm is globally convergent with probability one, meaning the sequence of global-best positions $\\{p^{(k)}_{\\mathrm{best}}\\}$ generated by the swarm satisfies $\\lim_{k\\to\\infty} P(x^{(k)} \\in \\Omega_g) = 1$ for every initial point, where $\\Omega_g$ is the set of global minimizers of $\\|X - [[A,B,C]]\\|_F^2$ over $A,B,C \\ge 0$. The proof routes through two assumptions: the swarm's objective values never increase (Assumption I), and the stochastic search eventually visits any positive-measure region of the feasible set (Assumption II); then a random-search lemma gives almost-sure convergence. For the discrete model, Theorem 6.3 states that the backtracking DTPNN produces a monotonically non-increasing objective sequence that reaches an equilibrium point, and combining it with Lemma 4.2 gives the same probability-one guarantee for CNO-DTPNN. The paper also claims that Hessian preconditioning with $P_A = (C^TC) * (B^TB)$ accelerates both continuous and discrete dynamics.","pith_inferences":["Extension — Theorem 4.3's assumption that wavelet mutation makes the search support cover all of $\\Omega$ is not derived from the PSO and mutation formulas; a direct reachability check of (2.11)-(2.12) would determine whether the almost-sure guarantee genuinely holds for the implemented algorithm.","Extension — The same collaborative scheme could be transplanted to Tucker, tensor-train, and block-term decompositions, and to Kullback-Leibler or alpha-beta divergence objectives; the paper lists these as future work, but the convergence argument is largely objective-agnostic beyond smoothness and bounded-below conditions.","Extension — The comparison with HALS-PSO in Example 2 suggests the benefit is not merely PSO reinitialization; isolating the contribution of the ODE integrator versus the swarm communication in a controlled ablation would test that reading."],"forward_implications":["For tensors with high collinearity, where HALS and other baselines converge very slowly, CNO-CPD reaches relative error $10^{-4}$ in about 20 seconds while the best baseline stays near $10^{-1}$ after 100 seconds (Example 2).","A population of $q$ RNNs improves the attainable relative error monotonically in the Monte Carlo experiments (Figure 8), so adding networks trades computation for escaping local minima.","The log-barrier variant removes the projection but yields worse relative errors than the classical projected ODE, so the projected dynamics rather than the barrier is the recommended formulation.","The discrete CNO-DTPNN, with semi-implicit updates and backtracking step sizes, gives the practical algorithm for large tensors where integrating the continuous ODE is expensive.","If Theorem 4.3 is correct, the algorithm carries a formal almost-sure global optimality property that classical CPD algorithms lack."],"supporting_citations":[{"why":"Supplies Lemma 4.1, the one-to-one correspondence between RNN equilibria and KKT points, linking single-network dynamics to local optimality.","marker":"[52]"},{"why":"Supplies Lemma 4.2, the random-search convergence theorem that converts Assumptions I and II into almost-sure global convergence.","marker":"[40]"},{"why":"Establishes that collaborative neurodynamic models with sufficiently large diversity converge almost surely, the template for Theorem 4.3.","marker":"[51]"},{"why":"Provides the collective neurodynamic approach for nonnegative matrix factorization that the paper adapts, including the claim that wavelet mutation makes Assumptions I and II hold.","marker":"[14]"},{"why":"Gives the discrete-time projection neural network for NMF and the stability lemma used to derive the step-size condition in Theorem 6.2.","marker":"[4]"},{"why":"Provides the discrete-time neurodynamic framework for sparsity-constrained NMF, used in Lemma 5.1 for equilibrium equivalence and in the DTPNN stability proof.","marker":"[29]"},{"why":"Supplies the tensor decomposition identities, including the Khatri-Rao-to-Hadamard product simplification used in the gradient formulas (3.2)-(3.4).","marker":"[24]"},{"why":"Provides the reference algorithms and background for nonnegative matrix and tensor factorization, including HALS and multiplicative updating rules used as baselines.","marker":"[9]"}],"fun_headline_variants":["Neural swarm hits global tensor decomposition","Collaborative neurodynamics guarantee global CPD","Swarm of neural nets finds global tensor fit","Neurodynamic swarm converges to global CPD","Neural nets + PSO achieve global nonnegative CPD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee collapses if Assumption II fails: the proof asserts, without deriving it from the formulas, that wavelet mutation and reinitialization make the swarm visit every positive-volume part of the feasible region infinitely often.","fun_headline_variants_meta":{"raw":{"variants":["Neural swarm hits global tensor decomposition","Collaborative neurodynamics guarantee global CPD","Swarm of neural nets finds global tensor fit","Neurodynamic swarm converges to global CPD","Neural nets + PSO achieve global nonnegative CPD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1870,"prompt_tokens":879,"completion_tokens":991,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":921}},"tokens_in":495,"tokens_out":991,"duration_ms":7969,"temperature":1.0,"reasoning_tokens":921,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:29:58.098924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a one-dimensional feasible interval $[l,u]$ and iterate the wavelet-mutation map (2.11)-(2.12) from a collection of current states; if the reachable set after any number of steps omits a positive-measure interval of $[l,u]$, then the claimed support $\\mathcal{H}(\\mu_k) = \\Omega$ used in Theorem 4.3 is false for that $B$, and Assumption II is not satisfied by the implemented search.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 4.1, the one-to-one correspondence between RNN equilibria and KKT points, linking single-network dynamics to local optimality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 4.2, the random-search convergence theorem that converts Assumptions I and II into almost-sure global convergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that collaborative neurodynamic models with sufficiently large diversity converge almost surely, the template for Theorem 4.3."},{"cited_title":"F an and J","cited_arxiv_id":null,"evidence_quote":"Provides the collective neurodynamic approach for nonnegative matrix factorization that the paper adapts, including the claim that wavelet mutation makes Assumptions I and II hold."},{"cited_title":"Che and J","cited_arxiv_id":null,"evidence_quote":"Gives the discrete-time projection neural network for NMF and the stability lemma used to derive the step-size condition in Theorem 6.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time neurodynamic framework for sparsity-constrained NMF, used in Lemma 5.1 for equilibrium equivalence and in the DTPNN stability proof."},{"cited_title":"Cichocki, R","cited_arxiv_id":null,"evidence_quote":"Provides the reference algorithms and background for nonnegative matrix and tensor factorization, including HALS and multiplicative updating rules used as baselines."}],"review_version":1}