{"id":"e9f739cb-a15b-4216-b3ca-a99d7c1be3c1","arxiv_id":"2411.11303","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Block RSCN adds reservoir nodes in blocks and claims universal approximation, echo state, and online convergence.","lead":"This paper builds a recurrent neural network that grows by adding small groups of reservoir nodes at once, instead of one node at a time. It claims faster training and more compact models, but the proof of the main convergence guarantee has a gap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's universal approximation guarantee is not tied to Algorithm 1's block acceptance criterion: Eq. (12) is undefined for matrix X and can reject blocks whose projection satisfies Eq. (24).","rationale":"The reader correctly locates the risk in the Eq. (12)-to-Eq. (24) connection. Closer inspection shows two distinct sub-issues. First, X is an N x n_max matrix while e is a 1 x n_max vector, so the inner products in Eqs. (12)-(14) have no stated meaning. If one defines <e,X> = eX^T and <X,X> = tr(XX^T), then the Cauchy argument P >= C closes the implication in the direction 'accepted => theorem assumption', so the reader's literal concern is repairable by a short lemma. Second, and more seriously, the normalization in Eq. (12) is not equivalent to the projection norm: a block containing one informative node and several large-norm uninformative nodes has C < delta even when P >= delta, so Algorithm 1 may reject the very blocks the theorem needs; no existence argument is supplied to show some accepted block always exists at each step. The paper also claims Theorem 3's persistent excitation condition, whose proof incorrectly converts the window integral in Eq. (34) into pointwise bounds; and the validation sets in Sections V.C-D are constructed by adding noise to the test set, leaking test information into model selection. These are independent weaknesses, but the Theorem 2 gap is the most load-bearing because the universal approximation property is the paper's central theoretical claim. The manuscript could potentially be repaired by importing the missing block-SCN lemma from [28], defining X consistently, and rewriting the PE argument, but as submitted the central guarantees are not established. Hence the reader's REJECT stands.","tokens_in":19238,"tokens_out":24166,"duration_ms":255428,"concrete_test":"Use m=100, N=5. Let e be a unit vector, choose g with cos^2(e,g)=0.9, and choose h1,...,h4 mutually orthogonal to e and g with sum(||h_i||^2) = 100||g||^2. Form X=[g,h1,...,h4] (n_max x N). Compute C = ||X^T e||^2 / tr(X^T X) and P = e^T X (X^T X)^+ X^T e. For delta=0.5, this gives C < delta while P >= delta, showing Eq. (12) is not equivalent to Eq. (24). Then run Algorithm 1's search on a synthetic residual with this e and the same block size; if none of the G_max random blocks has C_q >= delta for all q, the theorem's assumption (Eq. 24) is not guaranteed by the algorithm's acceptance test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central guarantee, Theorem 2, is not connected to the selection rule Algorithm 1 actually uses. Theorem 2 assumes Eq. (24): for residual row e_{j,q} and new block state matrix X, the squared projection norm e_{j,q}X^T(XX^T)^{-1}X e_{j,q}^T is at least delta*_{j+1,q}. Algorithm 1 accepts blocks via Eq. (12), which bounds a normalized correlation C_q = <e_{j,q},X>^2/<X,X>. Since X is an N x n_max matrix and e_{j,q} is a 1 x n_max vector, the inner products in Eqs. (12)-(14) are never defined. Under the most charitable convention, <e,X> = eX^T and <X,X> = tr(XX^T), one can show C_q <= P_q, so the implication 'accepted => Eq. (24)' is salvageable with a short lemma, but the paper supplies neither the definition nor the lemma. Under a scalar Frobenius reading the implication can fail. Moreover, even under the charitable convention the normalization dilutes useful directions: appending N-1 large-norm nodes orthogonal to e leaves the projection P_q essentially unchanged but can drive C_q below delta, so Algorithm 1 can reject a block whose subspace projection satisfies the theorem's condition. The paper provides no density/existence argument showing that at every step some candidate block with C_q >= delta* for all q will be found. Thus the universal approximation property is not established for the blocks actually selected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes block recurrent stochastic configuration networks (BRSCNs), which add subreservoirs of N reservoir nodes at a time. The construction is guided by a supervisory inequality (Eq. (12)) for accepting candidate blocks, a spectral-radius scaling rule (Eq. (13)) is intended to enforce the echo state property, and the output weights are updated online with a projection algorithm. Theorem 1 states a condition for the echo state property, Theorem 2 claims universal approximation under the inequality in Eq. (24), and Theorem 3 claims parameter convergence under persistent excitation. Experiments compare BRSCN with ESN, DESN, GESN, and RSCN on Mackey-Glass prediction, nonlinear system identification, debutanizer soft sensing, and short-term power-load forecasting.","tokens_in":19579,"tokens_out":12471,"duration_ms":119571,"significance":"If the three theorems were fully established, BRSCN would be a useful block-incremental extension of RSCN: block construction would reduce the number of iterations while preserving the universal approximation argument, and the online projection update would carry a convergence guarantee. The experimental comparison is broad, includes several block sizes, and reports compact reservoirs with competitive NRMSE on four tasks. The paper does not provide code or machine-checked proofs, and the main theory is an adaptation of the earlier RSCN proof [26] and the block-SCN framework [28], so the incremental contribution is moderate. Nevertheless, the extension is natural and the experiments suggest practical value; most of the technical gaps are local and appear repairable.","major_comments":[{"comment":"The connection between Algorithm 1's acceptance rule and Theorem 2's hypothesis is missing. In Eq. (12), X^{(j+1),i} is an N x n_max matrix while e_{j,q} is a 1 x n_max row vector, so the expressions <e_{j,q}, X>^2 / <X, X> are undefined unless a convention is supplied, and none is stated. If one adopts <e,X> = eX^T and <X,X> = tr(XX^T), then Eq. (12) is equivalent to C_q >= delta*_q with C_q <= eX^T(XX^T)^{-1}Xe^T = P_q, so the test can be sufficient for the projection condition only after proving a bridging lemma; conversely, the test can reject blocks with P_q >= delta*_q when the block contains large-norm components orthogonal to e_{j,q}. In addition, Theorem 2 states Eq. (24) with W_out evaluated by Eq. (16), while the proof uses the new-block-only least-squares solution Eq. (22); these are different objects. Without repaired definitions and a bridging lemma, the universal approximation conclusion is not established for the blocks the algorithm actually accepts.","section":"Sections III-A and III-C, Eqs. (12), (14), (24)"},{"comment":"No argument shows that, at every residual stage, a candidate block satisfying Eq. (12) exists within the Gmax random draws. Algorithm 1's fallback of increasing r and repeating is not proved to terminate with a valid block. The density assumption on span(Gamma) in Theorem 2 is not used to prove such an existence statement for block states; in the scalar SCN/RSCN theory this is supplied by a construction or probabilistic density argument, and the block analogue is absent. Therefore, even if the inequality implication in the previous comment were repaired, the theorem would not yet guarantee that Algorithm 1 produces an infinite sequence of accepted blocks whose residuals converge to zero.","section":"Section III-C, Theorem 2 and Algorithm 1"},{"comment":"Theorem 1 requires the strict scaling condition alpha < rho_max / sigma_max to ensure that the maximum singular value of each scaled subreservoir is below one, but Remark 2 and Eq. (13) specify only 0 < alpha < 1 in the scaling W <- (alpha / rho_max) W. The experimental setting alpha in [0.5, 1] (Section V) does not enforce the theorem's bound, so the echo state property guarantee as stated does not cover the models actually trained. The proof of Theorem 1 is correct under Eq. (17), but the condition needs to be incorporated into Remark 2 and Algorithm 1.","section":"Section III-B, Remark 2, Eq. (13) and Theorem 1"},{"comment":"The persistent-excitation proof contains two load-bearing errors. First, Eq. (34) is a window-integral inequality, but the text immediately asserts the pointwise bound eta1 >= g(n)^T g(n) >= eta2, which does not follow from an integral bound. Second, Eq. (37) treats the parameter error update as (I - P gamma g^T g) E(n-1), but with g a column vector and E a row vector the update is E(n) = E(n-1)(I - gamma P g g^T), not the scalar form used; the subsequent Lyapunov difference in Eqs. (39)-(40) is therefore dimensionally inconsistent. As written, the convergence claim of Theorem 3 is not proven.","section":"Section IV-B, Theorem 3"}],"minor_comments":[{"comment":"The symbol Gamma in 'span(Gamma)' is never defined, and the superscript asterisk on e^*_j is not explained; the reader must infer that it denotes the residual after the optimal output weights are used.","section":"Section III-C, Theorem 2"},{"comment":"The text after Fig. 3 states that the optimal BRSCN reservoir size for the MG task is 60, while Table I reports a reservoir size of 50 for the same task; please reconcile this discrepancy.","section":"Section V.A, Fig. 3 and Table I"},{"comment":"Algorithm 1 is titled 'DeepRSC', although the paper introduces BRSCN; if this title is not intentional, it should be changed to match the proposed method.","section":"Algorithm 1"},{"comment":"Equation (34) is written as an integral over [n0, n0+nw] in a setting where the projection update is discrete-time; please clarify whether the integral denotes a summation and define n0 and nw.","section":"Section IV-B, Eq. (34)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans heavily on the authors' own RSCN paper [26] and on Dai et al.'s block SCN [28]; the genuinely new theoretical content is largely a notational extension. The gaps identified in the major comments are fixable in principle, but the current version should not be accepted without a corrected proof of Theorem 3 and a clear bridge between Eq. (12) and Eq. (24)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable block-increment extension of RSCN, but the central approximation theorem is disconnected from the algorithm's acceptance rule, and the validation protocol leaks test information. I would send it back for major revision, not desk-reject it.\n\nWhat's new: the combination of block increments from Dai et al.'s block SCN with the recurrent SCN from the authors' earlier RSCN paper. That specific combination is new as far as the cited literature goes. The paper is clearly organized, Theorem 1 on the echo state property for the block-diagonal reservoir is correct, and the empirical study is broad: Mackey-Glass, nonlinear system identification, and two industrial tasks. The reported gains over RSCN are modest but consistent.\n\nSoft spots, in order of severity. First, Theorem 2's universal approximation guarantee is not established for the blocks Algorithm 1 actually selects. The theorem assumes inequality (24), a lower bound on the squared norm of the least-squares projection of the residual onto the block state subspace. The algorithm accepts blocks via (12), a normalized correlation that is never defined for matrix-valued X. Under the most charitable Frobenius convention, accepted implies (24) only through a short lemma the paper does not supply, and even then the normalization dilutes useful directions, so the algorithm can reject a block whose projection would satisfy the theorem. No density/existence argument shows that a candidate satisfying the acceptance test exists at every step. So the main theoretical claim does not cover the construction used in practice. Second, Theorem 3's persistent excitation argument moves from a window integral in (34) to pointwise bounds on g(n)^T g(n), and the Lyapunov difference is not convincingly tied to the integral condition. That is secondary, but still a gap. Third, in both industrial cases the validation set is made by adding Gaussian noise to the test set, which leaks test information into model selection and makes the generalization numbers optimistic. Fourth, no code or data are released, so the empirical comparisons are not independently checkable.\n\nThis is not a takedown. The block-increment idea is sensible, and readers working on stochastic configuration networks and reservoir computing will find the construction worth knowing. But the load-bearing theory is not connected to the algorithm, and the validation leak is a real procedural flaw. If the authors add the missing lemma, adapt the acceptance criterion to match the projection condition, provide a density argument, and fix the validation protocol, this could become a citable contribution. As it stands, I wouldn't rely on the universal approximation claim or quote the reported test numbers.\n\nRecommendation: send to peer review. The paper deserves a serious referee, but I would expect major revision before acceptance. The gap between Eq. (12) and Eq. (24) is exactly the kind of thing a good referee should catch.","headline":"A reasonable block-increment extension of RSCN whose core approximation theorem does not cover the algorithm's acceptance rule, with a test-leaking validation design.","tokens_in":20077,"tokens_out":3953,"would_cite":false,"duration_ms":37148,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Block-by-block reservoir growth preserves convergence while training faster.","keywords":["recurrent stochastic configuration network","block increments","subreservoir","echo state property","universal approximation property","persistent excitation","projection algorithm","reservoir computing"],"falsifier":"Run a single block-acceptance step on the Mackey-Glass data: generate many random subreservoir states, compute both the Eq. (12) quantity and the Eq. (24) projection norm of the residual onto the block state space, and check whether any state that passes Eq. (12) fails Eq. (24); if so, adding that block should produce a residual that does not contract by the claimed factor, contradicting Theorem 2 for the algorithm's actual selection rule.","tokens_in":18979,"feed_emoji":"🧠","tokens_out":3507,"duration_ms":36084,"temperature":0.7,"pith_summary":"This paper extends recurrent stochastic configuration networks (RSCNs) by adding reservoir nodes in whole blocks, or subreservoirs, rather than one node at a time. It claims that the resulting block RSCN retains the original network's universal approximation property and echo state property, and that its output weights can be updated online with guaranteed convergence under a persistent excitation condition. The central theoretical result is that under a supervisory inequality the training error converges to zero as the number of subreservoirs grows. A sympathetic reader would care because block increments promise the same accuracy guarantees with fewer construction iterations, a more compact reservoir, and practical online adaptability for industrial modeling.","feed_headline":"Block-by-block reservoir growth preserves convergence while training faster","feed_subtitle":"A recurrent stochastic configuration network that adds reservoir nodes in blocks retains universal approximation and echo-state stability.","key_machinery":"The central object is the block stochastic configuration algorithm: at each step a candidate subreservoir of N nodes is generated with random input weights, reservoir weights, and biases drawn from an adjustable uniform distribution, and it is accepted only if it passes the supervisory inequality in Eq. (12), which asks that the normalized correlation between the current residual and the subreservoir state exceed a threshold involving a contraction factor r and a decaying sequence µj+1. The accepted block that maximizes ξj+1 is added to a block-diagonal reservoir, whose weights are rescaled so that each subreservoir's maximum singular value is below one, guaranteeing the echo state property. The output weights are first set by global least squares and then updated online by the projection algorithm of Eq. (33). The proof mechanism for universal approximation is a residual contraction inequality: choosing output weights by least squares makes the new error satisfy ||e∗j+1||^2 ≤ (r + µj+1)||e∗j||^2, which forces the error to zero as j grows.","core_discovery":"The paper claims that a recurrent stochastic configuration network with block increments, called a BRSCN, simultaneously adds multiple reservoir nodes in the form of subreservoirs, each configured by a supervisory mechanism, and still guarantees the two properties that make such networks useful for temporal data. Theorem 1 states that scaling each subreservoir's weight matrix so that its maximum singular value is below one preserves the echo state property. Theorem 2 states that if the subreservoir state satisfies the projection inequality in Eq. (24), then the training error satisfies lim j→∞ ||T − F_{j+1}|| = 0, giving universal approximation. Theorem 3 states that the online projection update of the output weights converges when the state signal satisfies a persistent excitation condition. The experimental sections report that across Mackey-Glass time series, nonlinear system identification, a debutanizer column soft sensor, and short-term power load forecasting, BRSCNs attain lower NMSE than baseline ESNs, growing ESNs, decoupled ESNs, and RSCNs while using smaller or comparable reservoirs and fewer iterations.","pith_inferences":["A possible gap is that the algorithm accepts blocks using the scalar correlation inequality in Eq. (12), while Theorem 2 assumes the projection-norm inequality in Eq. (24); the paper does not prove these are equivalent, so the universal approximation guarantee might only hold for a subset of the blocks the algorithm actually selects.","If the implication between Eq. (12) and Eq. (24) holds, block increments could be extended beyond RSCNs to other stochastic-configuration architectures, such as deep or multi-scale reservoirs, without re-deriving the contraction argument.","The persistent excitation condition could be interpreted as a requirement on the spectral richness of the reservoir state, which would give practitioners a quantifiable diagnostic for when online weight updates are reliable.","A direct numerical test of the Eq. (12)-to-Eq. (24) implication on simple Mackey-Glass data, computing both quantities for randomly generated subreservoirs, would settle whether the theoretical guarantee applies to the actual construction procedure."],"forward_implications":["BRSCNs inherit the asymptotic training-error guarantee of RSCNs, so block-based construction does not sacrifice the theoretical basis for randomized reservoir design.","The block-diagonal reservoir with singular-value scaling preserves the echo state property, meaning the network's temporal behavior remains stable for long input sequences.","The online projection update with persistent excitation gives a convergence guarantee for output weights, supporting adaptive use in changing industrial processes.","Larger subreservoir sizes, such as Nsub = 10, reduce the number of construction iterations and yield the smallest reported NMSE across the tested benchmarks.","The reported results suggest that block increments produce more compact reservoirs, with testing NMSE that is often lower than point-by-point RSCN growth at similar or smaller reservoir sizes."],"supporting_citations":[{"why":"Introduces stochastic configuration networks with a supervisory mechanism for random parameter assignment, which is the basis for the node-growth strategy used here.","marker":"[25]"},{"why":"Introduces recurrent stochastic configuration networks, the point-incremental predecessor whose approximation and echo-state properties this paper extends to blocks.","marker":"[26]"},{"why":"Provides block-incremental stochastic configuration networks for process data modeling, the direct template for the block acceptance and universal approximation argument.","marker":"[28]"},{"why":"Supplies the projection algorithm and its convergence analysis, which the paper adapts for online update of the output weights.","marker":"[27]"},{"why":"Defines the echo state property and the maximum-singular-value condition for ESN reservoirs, which Theorem 1 relies on.","marker":"[9]"},{"why":"Provides the growing echo state network with multiple subreservoirs, one of the empirical baselines the BRSCN is compared against.","marker":"[19]"},{"why":"Provides the decoupled echo state network, another baseline with subreservoir structure used in the experiments.","marker":"[29]"}],"fun_headline_variants":["Block reservoir growth preserves echo state and convergence","BRSCN: add reservoir blocks, keep universal approximation","Block-wise RSCN training cuts iterations, keeps accuracy","Growing recurrent nets by blocks: faster and still stable","Subreservoirs added in blocks maintain echo state and learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the scalar correlation inequality used to accept a subreservoir is strong enough to guarantee the projection-norm condition that drives the convergence theorem, and the paper does not prove that implication; the notation for block states is also ambiguous between vectors and matrices.","fun_headline_variants_meta":{"raw":{"variants":["Block reservoir growth preserves echo state and convergence","BRSCN: add reservoir blocks, keep universal approximation","Block-wise RSCN training cuts iterations, keeps accuracy","Growing recurrent nets by blocks: faster and still stable","Subreservoirs added in blocks maintain echo state and learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1359,"prompt_tokens":940,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":341}},"tokens_in":556,"tokens_out":419,"duration_ms":5183,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:41:40.626778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a single block-acceptance step on the Mackey-Glass data: generate many random subreservoir states, compute both the Eq. (12) quantity and the Eq. (24) projection norm of the residual onto the block state space, and check whether any state that passes Eq. (12) fails Eq. (24); if so, adding that block should produce a residual that does not contract by the claimed factor, contradicting Theorem 2 for the algorithm's actual selection rule.","supporting_citations":[{"cited_title":"Stochastic configuration networks with block increments for data modeling in process industries,","cited_arxiv_id":null,"evidence_quote":"Provides block-incremental stochastic configuration networks for process data modeling, the direct template for the block acceptance and universal approximation argument."},{"cited_title":"Growing echo-state network with multiple subreservoirs,","cited_arxiv_id":null,"evidence_quote":"Provides the growing echo state network with multiple subreservoirs, one of the empirical baselines the BRSCN is compared against."},{"cited_title":"Decoupled echo state networks with lateral inhibition,","cited_arxiv_id":null,"evidence_quote":"Provides the decoupled echo state network, another baseline with subreservoir structure used in the experiments."}],"review_version":1}