{"id":"d8dc623b-7b7e-4558-9afb-fd08b00b5cb4","arxiv_id":"2501.19059","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A bias-modulated neural controller with fixed connectivity can locally replicate arbitrary linear controllers, and its multi-task approximation error is bounded above and below.","lead":"This paper studies a single nonlinear controller, built like a small neural network, that can handle several different control tasks by changing only a low-dimensional bias vector while keeping its connection weights fixed. The authors prove this controller can match any single linear controller and give bounds on how well it can approximate a whole family of controllers at once.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The H2 cost in problem (5) is finite only when every -I + D_i W is Hurwitz, but the paper neither restricts (5) to that stable set nor guards gradient descent, leaving the core optimization problem and Theorem 3.1 formally ill-posed.","rationale":"We agree with the reader's weakest-assumption assessment. The H2-based formulation (5) is the central object: Theorem 2.1 establishes the parametrization, Theorem 3.1 gives gradients for (5), and Theorem 4.1 bounds the optimum of (5). All of these require the approximating systems to be stable. The paper explicitly acknowledges the lack of guarantees in Section III, but does not incorporate stability as a constraint, leaving the optimization problem formally undefined outside the stable region. The upper bound construction is stable, so the main existence claim survives; nevertheless, the formulation and the proposed training algorithm need modification. We also note the proof of Theorem 4.3 omits the cases r <= r_j and r >= r_l, as the reader observed; this is a secondary incompleteness in a lower bound for a scalar L1 problem and does not change the verdict. The numerical experiments are illustrative and Figs. 3-5 suggest the framework works in practice, but no code is released. Overall, these issues are correctable and do not invalidate the core idea, so the CONDITIONAL verdict stands.","tokens_in":13156,"tokens_out":27135,"duration_ms":251281,"concrete_test":"Run the gradient algorithm of Section III on the four-controller example of Fig. 2 with 100 random initializations, and at every iteration compute the eigenvalues of -I + D_i W for each task i. Record the first iteration at which any of these matrices fails to be Hurwitz. If such an event occurs, the H2 cost and the gradient from Theorem 3.1 are undefined at that point, confirming that the unconstrained problem (5) is not well-posed and that the algorithm needs an explicit stability safeguard.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Problem (5) minimizes the sum of H2 norms ||Sigma^D_i - Sigma^L_i||_2^2. The H2 norm of the error system is finite if and only if A^err_i = blkdiag(A_i, -I + D_i W) is Hurwitz; since the desired A_i are stable, this reduces to requiring each approximating linearization -I + D_i W to be stable. The paper states in Section III that the gradient procedure 'offers no stability or performance guarantees,' and no constraint or projection enforcing stability appears in (5). The gradient formulas in Theorem 3.1 use the Gramians P_i, Q_i of the error system, which exist only when the error system is stable; a gradient step that crosses the boundary destabilizes the approximation, and the objective and gradient are no longer defined. Thus the minimization problem is not well-posed over its nominal domain. This is load-bearing because the paper's multi-task guarantee is expressed through the optimum of (5). The existence side is partly rescued by Theorem 4.1, whose proof constructs stable approximations via balanced truncation, but the unconstrained formulation and the algorithm remain formally incomplete. The paper should add stability constraints (or define an extended-valued cost) and provide a safeguard for gradient descent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a nonlinear neural controller of the form ẋ = -x + Φ(Wx + d) + Bu, y = Cx, with a fixed connectivity matrix W and task-dependent bias vector d, and studies its ability to approximate a set of M desired linear controllers via the linearized dynamics around bias-dependent equilibria. The authors prove that any diagonal matrix D in (0,1)^N can be realized by a suitable choice of d (Theorem 2.1), derive analytical gradient expressions for the H2 approximation cost (Theorem 3.1), and provide upper and lower bounds on the multi-task approximation error (Theorems 4.1–4.3). Numerical experiments with aircraft, pendulum, and bicycle examples illustrate the approach.","tokens_in":13368,"tokens_out":13314,"duration_ms":124614,"significance":"The paper addresses a timely and interesting question: how a single static neural architecture with a low-dimensional task-dependent bias can approximate multiple linear controllers. Theorem 2.1 gives a clean and explicit parametrization of the linearized controller dynamics by the bias vector, and Theorem 3.1 provides usable gradient formulas. The upper and lower bounds, although partly technical, attempt to quantify how the approximation error depends on the number of tasks, the controller dimension, and the similarity of the desired systems, which is a useful step for this class of architectures. The numerical studies support the viability of the approach, though they do not by themselves establish the theoretical claims.","major_comments":[{"comment":"The H2 norm in problem (5) is finite only if the error system is stable; for stable desired systems A_i, this reduces to requiring that every linearized controller matrix -I + D_i W is Hurwitz. The paper does not restrict the optimization to this stable set, nor does it define the H2 norm as extended-valued (equal to +∞) for unstable error systems. Consequently, the gradient expressions in Theorem 3.1, which use the Gramians of the error system, are undefined for iterates that yield an unstable linearization, and gradient descent has no safeguard to remain in the well-defined region. This is load-bearing because the upper bound in Theorem 4.1 is expressed through the optimum of problem (5). Please add an explicit stability constraint or an extended-value definition of the cost, and describe how the proposed algorithm handles the stability boundary.","section":"Section II, Eq. (5)"},{"comment":"The proof of Theorem 4.3 explicitly omits the cases r ≤ r_j and r ≥ r_l, stating that they are 'not included' and lead to analogous conclusions. These cases are part of the theorem's stated domain, and the theorem claims a lower bound over the full problem. The omitted cases must be treated explicitly, or an argument must be provided showing that they are dominated by the included cases; otherwise the proof is incomplete and the theorem is not fully established.","section":"Section IV-B, Theorem 4.3 proof"},{"comment":"The derivation of the upper bound is too terse at a critical point. The expression for J_B, which includes the term 2N ||Σ_aux||_∞, is introduced without explaining how it follows from the balanced-truncation error bound in reference [30]. The proof simply states that the bound is obtained by 'leveraging' the balanced truncation result. Since Theorem 4.1 is one of the main theoretical contributions, the proof should be expanded to show the origin of J_B and to specify which balanced-truncation result is used for the H2 norm.","section":"Section IV-A, Theorem 4.1 proof"}],"minor_comments":[{"comment":"The parameter count for storing M distinct linear controllers is given as 'M n^4 m p', which appears to be a typo; the number of free parameters in a state-space realization (A, B, C) is n^2 + nm + pn (or n^2 + nm + pn + pm if a direct feedthrough term is included), not n^4 m p.","section":"Remark 1"},{"comment":"The sentence 'Notice that φ_d is a injective function' should read 'an injective function'.","section":"Theorem 2.1 proof"},{"comment":"In the table after equation (26), the entry x_j = -∞ in Case 3 is not an admissible finite value; please clarify that this denotes a limiting value as x_j tends to -∞, and state the sense in which the infimum is attained.","section":"Section IV-B, proof of Theorem 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for eess.SY and addresses a topic of current interest. The main theoretical ideas are sound, but the formal well-posedness of the optimization problem and the incomplete proof of Theorem 4.3 should be addressed before acceptance. I also suggest that the authors double-check the definition and derivation of the upper bound's constant J_B, as it is currently difficult to verify."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real contribution with a fixable gap. The core idea—one fixed connectivity matrix W with task-specific bias vectors d—is new relative to the cited single-model approximation work [22], and Theorem 2.1's universal realizability of any diagonal D is clean and correct. The gradient formulas in Theorem 3.1 are standard but correctly derived, and the balanced-truncation upper bound in Theorem 4.1 is a reasonable constructive argument. The lower bounds in Theorems 4.2 and 4.3 are more specialized, but they are honest attempts at fundamental limits.\n\nThe soft spots are real but not fatal. The main one is exactly what the stress-test note flagged: problem (5) minimizes a sum of H2 norms without restricting the approximating systems -I + D_i W to be stable. The H2 norm of the error system is finite only if those matrices are Hurwitz, so the domain of the optimization problem is not the set of all feasible variables, and the gradient formulas rely on Gramians that exist only for stable error systems. The paper even admits in Section III that the gradient procedure 'offers no stability or performance guarantees.' This is a formal ill-posedness in the problem statement, not just a numerical nuisance. It can be fixed by introducing stability constraints or an extended-valued cost, and the existence of stable approximants is partially rescued by the constructive proof of Theorem 4.1. But as written, the training problem is incomplete. Second, the proof of Theorem 4.3 explicitly omits two cases with a hand-wave; that is a minor completeness issue, but it needs to be filled in.\n\nThe numerical experiments are illustrative: no code or data are provided, and the plots are suggestive rather than conclusive. For a theory paper that is acceptable, though one reproducible example would strengthen it.\n\nOverall, the central claim—that a fixed connectivity matrix with varying bias can approximate multiple linear controllers with explicit bounds—holds up in the theory sections. The well-posedness issue is genuine, but it is repairable and does not undermine the main conceptual contribution. This paper deserves a serious referee; I would send it out with expectations of major revisions, primarily to address the stability constraint in (5) and complete the proof of Theorem 4.3.","headline":"A genuinely new multi-task controller formulation with a formal gap: the H2 cost in problem (5) is only defined over stable linearizations, but stability is not enforced, leaving the optimization problem ill-posed as stated.","tokens_in":13929,"tokens_out":2622,"would_cite":true,"duration_ms":26804,"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":"A recurrent neural controller with fixed connectivity can replicate any single linear controller and approximate whole families of them by tuning only its bias vector.","keywords":["multi-task control","neural controller","bias adaptation","linearized dynamics","smooth ReLU","H2 norm approximation","balanced truncation","lower bounds"],"falsifier":"Run a dense numerical search over $(W,D_1,D_2,B,C)$ with $N=1$ for two stable scalar systems with residues $r_1