{"id":"29d754f3-c4f7-4800-be06-928938171837","arxiv_id":"2603.03017","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MGU recurrent networks can be certified input-to-state and incrementally input-to-state stable under newly derived parametric conditions, and the paper shows training strategies that make those conditions achievable.","lead":"This paper derives mathematical conditions that guarantee two kinds of stability for Minimal Gated Unit (MGU) recurrent networks, and tests training tricks that push MGU networks into the stable region. If the conditions hold, the network's internal state stays bounded and similar inputs produce similar trajectories, which matters for control and embedded systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 1 lacks a uniform contraction bound: the λ_j used in (A.5)-(A.6) may depend on k, so the ISS recursion is not justified as written; the same gap propagates to the network-level ISS claim in Theorem 3.","rationale":"The paper's central claim is the sufficiency of (9) and (10) for ISS/δISS. The most load-bearing issue is the unjustified uniform-contraction step in the proof of Theorem 1, exactly as the reader's rationale flags. This is a genuine logical gap as written, but it is readily repaired by a compactness argument, so it does not falsify the theorem. The reader's formal weakest_assumption was about Assumptions 1-2 and input normalization, which is a different concern; the reader did, however, mention the uniform-contraction step in the rationale, hence 'partial' agreement. A secondary gap in Theorem 4's KL function is similarly repairable. Since both are addressable proof repairs rather than demonstrated falsehoods, the CONDITIONAL verdict remains appropriate; no change to the reader's verdict is warranted.","tokens_in":26253,"tokens_out":15040,"duration_ms":133620,"concrete_test":"Analytical repair check for Theorem 1: For a layer satisfying (9), compute M_min = min_{h∈[-1,1]^n, u∈[-1,1]^m} (w_f^T u + r_f^T h + b_f) by evaluating all vertices, set f_min = σ(M_min), and define λ = 1 − f_min(1 − \\barσ_f\\|R_{\\tilde h}\\|). Verify that λ<1 whenever (9) holds; then re-derive the recursion (A.5)-(A.6) with this explicit time-independent λ. If λ<1 for all such parameter sets, the uniform-contraction step can be made rigorous and the concern is a repairable proof gap; if any parameter set satisfying (9) yields λ=1, the ISS proof (and hence Theorem 3) fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central ISS certificate rests on the proof of Theorem 1 (Appendix A). After (A.4), the authors assert that for each j there exists λ_j∈(0,1) with 1−f_{j,k}+f_{j,k}\\barσ_f\\|R_{\\tilde h}\\| ≤ λ_j, then define λ=max_j λ_j and use this constant λ in the backward recursion (A.5)-(A.6) and in the KL functions (A.8)-(A.10). But f_{j,k}=σ(w_f^T \\tilde u_k + r_f^T h_k + b_f) varies with time and with the state/input; the pointwise λ_j need not be time-independent. Without a uniform λ<1 valid for all k and all admissible trajectories, the geometric-series bound (A.7) and the resulting exponential decay do not follow. The gap is repairable: because (h_k,u_k) lie in a compact box, f_{j,k} is bounded below by σ(−\\|[W_f\\ R_f\\ b_f]\\|)>0, so a uniform λ exists. But the paper never states this compactness argument, so the proof as written is incomplete. A related technical gap occurs in Theorem 4's KL function (A.26): the binomial polynomial cannot be strictly decreasing in k for all k, so it may not satisfy Definition 2's KL requirement; the standard Jordan-bound repair (\\|A_δ^k\\|≤ Cρ^k) is not provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the Minimal Gated Unit (MGU) recurrent network from a stability viewpoint. It derives sufficient parametric inequalities—condition (9) for ISS and condition (10) for incremental ISS—for an MGU layer, extends them by a cascade argument to multilayer MGU networks (Theorems 3 and 4), and proposes three training strategies (loss augmentation, parameters warm-start, projected gradient optimization) together with a stability-driven early stopping rule. The theoretical conditions are validated on the pH reactor, four-tank, and Silverbox benchmarks, where MGU is reported to match GRU accuracy with fewer parameters and to outperform a stable GRU on Silverbox. The appendix contains proofs of forward invariance and of the stability theorems.","tokens_in":26671,"tokens_out":17221,"duration_ms":149089,"significance":"If the proofs are repaired, the paper makes a useful contribution. The stability certificates are simple, layer-wise infinity-norm checks on the learned weights, so a practitioner can validate a trained MGU by inspecting a few row norms. The δISS condition is parametric, and the paper shows it implies ISS, allowing the stricter property to be used while inheriting the weaker one. The training methods, especially the warm-start and projected-gradient schemes, are reasonable, and the authors promise open-source code. The conditions are derived from the architecture equations rather than fitted, so the guarantees are not circular. The main limitations are technical: the proof of Theorem 1 lacks a uniform contraction argument, the KL function in Theorem 4 is not a valid KL function as written, one equality in the proof of Theorem 2 is false, and the Silverbox comparison is confounded by unequal model sizes. These issues are repairable without changing the main statements.","major_comments":[{"comment":"After (A.4), the existence of λ_j∈(0,1) is asserted pointwise for each k. Because f_{j,k} depends on k through (h_k,u_k), the λ_j may depend on k; defining λ=max_j λ_j and using it in the backward recursion (A.5) and the geometric series (A.7) is unjustified. The fix is standard: with (h_k,u_k) in compact boxes, f_{j,k} ≥ σ(−||[W_f R_f b_f]||) =: f_min > 0, so 1 − f_{j,k} + f_{j,k} c ≤ 1 − f_min(1−c) < 1 uniformly in k when c = σ̄_f||R_h̃|| < 1. This lower-bound/compactness argument should be stated explicitly. The gap propagates to the network-level ISS claim in Theorem 3, which uses the same λ.","section":"Appendix A, proof of Theorem 1 (Eqs. (A.4)–(A.7))"},{"comment":"The function defined in (A.26) is not a KL function as required by Definition 2: ψδ(r,k) = ω∑_{z=0}^{m−1} C(k,z)(max α)^{k−z} r need not be strictly decreasing in k for all k. For m=2 and max α > 1/2, the factor (1+k)(max α)^k increases initially. The standard repair is to use the Jordan bound only to conclude ||Aδ^k|| ≤ C ρ^k for some ρ ∈ (max_l αδ^(l), 1), so a valid KL function is ψδ(r,k) = C ρ^k r. Please replace (A.26) accordingly.","section":"Appendix A, proof of Theorem 4 (Eq. (A.26))"},{"comment":"The displayed equality max_{u∈U,h∈H} ||W_h u + R_h(f∘h) + b_h|| = ||W_h 1_{nu} + R_h f + b_h|| is false when W_h has mixed-sign entries (e.g., W_h = [1 −1], R_h = 0, b_h = 0). The row-wise maximum is the sum of absolute values, not the value at the all-ones vector. The final bound by ||[W_h R_h b_h]|| still follows from a row-sum argument, so the δISS condition is not invalidated, but the displayed equality should be replaced by a valid row-wise inequality.","section":"Appendix A, proof of Theorem 2 (Eq. (A.15))"},{"comment":"The Silverbox comparison varies architecture capacity simultaneously with architecture type: MGU_WS has 64 units per layer while GRU_LA has 8 units per layer (both depth 3). The conclusion that the stable MGU effectively captures the system dynamics while other stable RNNs fail is therefore not established, because the comparison confounds the MGU architecture with a much larger model. Please report a matched-capacity or matched-parameter comparison, or at least add the performance of a stable GRU with a comparable parameter count.","section":"Section 6.5, Table 1"}],"minor_comments":[{"comment":"The chain σ̄_f||R_h̃|| < ||R_h̃|| fails when R_h̃ = 0. Use a non-strict inequality: σ̄_f||R_h̃|| ≤ ||R_h̃|| ≤ 1−ε < 1.","section":"Appendix A, proof of Proposition 4"},{"comment":"The proof cites [13, Theorem 2] for cascades, but in (3) the input to layer l>1 is h^{(l−1)}_{k+1}, a one-step shifted version of the preceding layer's state. Please spell out the composition argument or re-index the systems so that the cited cascade theorem is directly applicable.","section":"Theorem 3 proof"},{"comment":"Line 3 says 'Project θ0 onto a stable δISS-compliant region', but Problem (14) is an unconstrained minimization of a violation penalty, not a projection onto the feasible set. Rename this step as a warm-start or define an actual distance-minimizing projection.","section":"Algorithm 1 / Section 5.3"},{"comment":"There is a typo: 'Multisinesequence' should be 'Multisine sequence'. Also, 'a-posteriori' in the abstract should be 'a posteriori'.","section":"Section 6.5"},{"comment":"Assumption 2 is said to be readily satisfied by normalizing the input vector, but the stability certificate is only valid when the same normalization is used at deployment. This caveat should be stated explicitly in the validation protocol or conclusion.","section":"Section 6.1 / Assumption 2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of Automatica and the main theoretical statements appear repairable. I recommend major revision rather than rejection: the uniform-contraction gap in Theorem 1, the KL-function issue in Theorem 4, the false equality in the proof of Theorem 2, and the unequal-capacity Silverbox comparison all need to be addressed. No concerns about citation or novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The genuinely new thing is the stability analysis of the MGU architecture itself. The MGU's tighter coupling between forget gate and candidate state means the parametric conditions are not just a GRU/LSTM transcription; the proofs have to handle f∘h inside the hyperbolic tangent. The delta-ISS condition in Theorem 2 and its proof are careful and, as far as I can tell, correct. That alone is a real contribution for people doing certified stable RNNs on embedded hardware. The training folklore is also useful: loss augmentation alone is weaker for MGU than for GRU, and the warm-start/PGM recipes address that.\n\nThe soft spots, in proportion. The proof of Theorem 1 has a genuine gap. After (A.4), the authors assert a time-independent λ_j < 1 from the pointwise convex combination bound. But f_{j,k} depends on k and on the trajectory, so the max over j could in principle depend on k. The fix is easy, and the stress-test note is right: because h_k and u_k live in compact boxes, f is bounded below by σ(−||[W_f R_f b_f]||) > 0, giving a uniform λ. But the paper never states that compactness argument, so the ISS recursion as written doesn't hold. The same gap propagates to Theorem 3. Repairable, but should be fixed before publication.\n\nSecond, the claimed KL function in Theorem 4, (A.26), is a polynomial times an exponential. It decays to zero, but the paper doesn't show it is strictly decreasing in k for all k as Definition 2 requires. That's a definitional detail; the standard Jordan bound repair works, but again, as written it's incomplete.\n\nEmpirically, the headline that 'other stable RNNs fail to converge' on Silverbox overstates the evidence. The comparison is MGU WS with 64 hidden units per layer versus a single GRU LA configuration with 8 units per layer. That's not a fair baseline, and no other stable RNN is tested. The paper's own Figure 4 shows a clear accuracy drop for PGM methods, which is honest, but the abstract's phrasing goes beyond what the data show.\n\nNet: the central mathematical machinery is sound and the contribution is new. The gaps are real but local and repairable. This deserves a serious referee; I'd send it out and ask for revision, not desk reject. I'd cite it if I worked on stable gated RNNs.","headline":"Genuinely new MGU stability conditions with a sound delta-ISS core; the ISS proof has a repairable uniformity gap and the empirical comparison overreaches.","tokens_in":27114,"tokens_out":2303,"would_cite":true,"duration_ms":22342,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93D25","93C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-gate recurrent neural network, the Minimal Gated Unit, can be certified input-to-state stable by checking two simple row-norm inequalities on its weights, and the paper proves this for both single-layer and multi-layer networks.","keywords":["Minimal Gated Unit","Input-to-state stability","Incremental input-to-state stability","Recurrent neural networks","System identification","Model predictive control","Projected gradient training","Sufficient parametric conditions"],"falsifier":"Find a set of MGU weights that satisfies condition (10) for every layer, initialize two trajectories from distinct states within [-1,1] with identical bounded inputs (both inside [-1,1]), and observe the state difference growing without bound; this would directly contradict Theorem 4. A complementary test would be to feed an input slightly outside [-1,1] to a certified-stable MGU and check whether the hidden state escapes the unit box, which would show the certificate is not robust to normalization violations.","tokens_in":26197,"feed_emoji":"🧮","tokens_out":3444,"duration_ms":34912,"temperature":0.7,"pith_summary":"The paper claims that the Minimal Gated Unit (MGU) recurrent network—a lightweight two-gate alternative to GRU and LSTM networks—can be made input-to-state stable (ISS) and incrementally input-to-state stable (δISS) under simple sufficient conditions on its weight matrices. The conditions are easy to check: for ISS, the product of an upper bound on the forget gate and the infinity norm of one recurrent weight matrix must stay below 1; for δISS, a slightly more elaborate combination of the same norm bounds must stay below 1. If these hold layer by layer, the whole MGU network is provably stable, which matters because formal stability is a prerequisite for using identified models inside model predictive control loops and state observers. The paper also proposes three training modifications—loss augmentation, parameter warm-start, and projected gradient optimization—to keep trained parameters inside the stable region, and demonstrates on three benchmarks that the stable MGU matches or beats stable GRU models with about two-thirds the parameters and faster inference.","feed_headline":"Two inequalities certify MGU recurrent networks are stable","feed_subtitle":"Stable MGU nets match GRU accuracy with two-thirds the parameters and faster inference.","key_machinery":"The central objects are two scalar bounds extracted from the network parameters: σ̄_f = σ(||[W_f R_f b_f]||), the maximal value the forget gate can take over the unit input and state box, and φ̄_˜h = tanh(||[W_˜h R_˜h b_˜h]||), the maximal candidate hidden state. These bounds convert the nonlinear recursion into a linear contraction inequality: the state norm (or the incremental state norm) is bounded by a factor α < 1 times its previous value plus a gain times the input norm. Condition (9) makes that contraction factor for ISS, and condition (10) makes the analogous factor for δISS, both expressed directly as row-norm inequalities on the weight matrices.","core_discovery":"The author establishes sufficient parametric conditions under which an MGU layer is ISS (Theorem 1) and δISS (Theorem 2), and shows these conditions extend to the cascade of L layers that forms the full MGU network (Theorems 3 and 4). The ISS condition requires that σ̄_f ||R_˜h|| < 1, where σ̄_f is a worst-case upper bound on the forget gate's activation and ||R_˜h|| is the infinity norm of the candidate-state recurrent weight matrix. The δISS condition adds a contraction bound that involves these quantities plus the recurrent weight of the forget gate and an upper bound on the candidate-state nonlinearity. The paper proves that the δISS condition implies the ISS condition, so enforcing the","pith_inferences":["A similar row-norm certificate could likely be derived for other gated architectures, such as LSTM variants with peephole connections, by bounding the activation ranges of their gates and replacing the single forget-gate factor with an analogous worst-case product.","The sufficient conditions are likely conservative; the gap between the stable region they define and the true stability region of MGU dynamics is unknown, and a tighter analysis (for instance via integral quadratic constraints) might yield higher-accuracy stable models.","The stability guarantee is only as strong as the normalization assumed at deployment: if a deployed input exceeds the unit bound or the initial hidden state is outside [-1,1], the forward-invariance argument breaks. A runtime monitor that checks ||u_k|| ≤ 1 and clips or warns could enforce the assumption online.","The projected-gradient method, which restricts only the candidate recurrent weight to an L∞ ball, sacrifices significant accuracy; a projection onto the actual, nonconvex δISS set—computed approximately—might preserve more of the modeling capacity."],"forward_implications":["A user with a trained MGU can certify its stability by checking two explicit row-norm inequalities on each layer, without running a single trajectory simulation.","Because the δISS property also implies ISS, a network that passes condition (10) is automatically suitable for control applications that require incremental stability, such as observer design and robust MPC.","The paper's training methodology—loss augmentation, warm-start, and projected gradient—provides a practical path to obtain networks that provably lie in the stable region, while preserving competitive accuracy on benchmark identification problems.","The parameter efficiency of the MGU (roughly two-thirds the parameters of a GRU with the same hidden size) combines with formal stability, making this architecture attractive for embedded control with limited compute and memory.","The stability check is layer-local: as long as every layer satisfies the respective inequality, the cascade is stable, so the condition scales to deep networks without cross-layer coupling."],"fun_headline_variants":["MGU stability certified by two inequalities","Two inequalities guarantee MGU network stability","Stable MGU nets: simple conditions, rigorous proof","MGU nets proven stable with parametric checks","Certified ISS for lightweight MGU networks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The guarantees rely on inputs being normalized to the unit box and initial hidden states staying in [-1,1] at deployment; if real inputs or initial states leave those bounds, the forward-invariance and all derived norm bounds no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["MGU stability certified by two inequalities","Two inequalities guarantee MGU network stability","Stable MGU nets: simple conditions, rigorous proof","MGU nets proven stable with parametric checks","Certified ISS for lightweight MGU networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2088,"prompt_tokens":732,"completion_tokens":1356,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1289}},"tokens_in":476,"tokens_out":1356,"duration_ms":10438,"temperature":1.0,"reasoning_tokens":1289,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:12:04.863608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a set of MGU weights that satisfies condition (10) for every layer, initialize two trajectories from distinct states within [-1,1] with identical bounded inputs (both inside [-1,1]), and observe the state difference growing without bound; this would directly contradict Theorem 4. A complementary test would be to feed an input slightly outside [-1,1] to a certified-stable MGU and check whether the hidden state escapes the unit box, which would show the certificate is not robust to normalization violations.","supporting_citations":[],"review_version":1}