{"id":"bbdd2b19-cf71-4334-9580-1ce77e599d8a","arxiv_id":"2411.14077","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a class of nonlinear capacity-limited networks satisfying a monotonicity assumption, a decentralized anti-windup PI controller globally stabilizes a unique equilibrium that minimizes a weighted l1 tracking error, and a rank-1 coordinated version minimizes the l-infinity error.","lead":"The paper proves that a simple, fully decentralized anti-windup PI controller stabilizes a unique equilibrium in nonlinear resource-sharing networks and that this equilibrium minimizes a weighted sum of tracking errors; a rank-1 coordinated variant instead minimizes the worst-case error.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core proofs of Theorems 1 and 2 contain sign/inequality errors as written; these must be corrected before the central claims are established.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditional acceptance is appropriate. However, the reader identified Assumption 1 and the unproved Proposition 1 as the weakest load-bearing premise, whereas my concern is more direct: as typeset in the manuscript, the main proofs of Theorems 1 and 2 contain sign/inequality errors. The bound in (14) uses '≥' where the text's own conclusion 'this expression is strictly negative' requires '≤', and Lemma 1 indeed implies the opposite sign for the right-hand side. Similarly, Eq. (28) in Theorem 2 has the wrong sign of the saturation increment; the saturation argument shows sign(dz) equals sign(v0-v_dagger), not sign(v_dagger-v0). These are internal inconsistencies, not disagreements with consensus. I believe they are typos rather than fatal flaws, because the surrounding arguments are otherwise coherent and the results extend the authors' linear M-matrix work in [18,19] and match the simulation. Nevertheless, the central proofs as written do not rigorously establish the theorems, so the paper should be accepted only after the sign errors are fixed and the derivations are verified. I also note the acknowledged gaps (unproved Proposition 1, sketched Lemma 3) but those are secondary to the proof issues in the core theorems.","tokens_in":16701,"tokens_out":35749,"duration_ms":315753,"concrete_test":"Re-derive the two flagged steps with corrected direction: (i) in Theorem 1, replace the '≥' in (14) with '≤' and verify the subsequent negativity argument remains valid; (ii) in Theorem 2, replace sign(v_i_dagger - v0_i) with sign(v0_i - v_i_dagger) in (28) and re-run the chain (28)-(32). If either corrected step does not restore the claimed conclusion, the corresponding theorem is unproved; if both are simple typos, the proofs go through. As an independent cross-check, simulate a two-agent linear M-matrix instance and compare the decentralized PI equilibrium's weighted l1 cost to the true minimum.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 1 (Section 5.2), term (13d) is bounded via (14) with '≥', but the following sentence claims the bounding expression is 'strictly negative' when the argument is nonzero. Lemma 1, however, gives sum_{i in J±} eta_i sign(sat_i~) b_i~ > sum_{j in J0} eta_j |b_j~|, so the right-hand side of (15)-(16) is strictly positive. For (13d) to be negative, the inequality in (14) must be '≤', not '≥'. In the proof of Theorem 2, Eq. (28) asserts sign(dz_i(u0)) = sign(v_i_dagger - v0_i), but saturation logic gives sign(dz_i(u0)) = sign(v0_i - v_i_dagger): if dz_i>0 then v0_i is the upper bound and v_i_dagger < v0_i; if dz_i<0 then v0_i is the lower bound and v_i_dagger > v0_i. The final expression in (32) uses sign(v0 - v_dagger), so (28) needs the opposite sign. As written, these steps do not yield the claimed negative definiteness and strict inequality; they are likely typos, but a corrected derivation is required for the central stability and optimality claims.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a class of multi-agent systems in which each agent has scalar first-order dynamics ẋ_i = −a_i x_i + b_i(sat(u)) + w_i, with a nonlinear, bounded, monotone-competition interconnection b satisfying Assumption 1. For a fully decentralized PI controller with anti-windup (2), Theorem 1 claims global asymptotic stability of a unique equilibrium, and Theorem 2 claims that this equilibrium minimizes the weighted l1 cost Σ η_i a_i |x_i| among all open-loop equilibria with different saturated input. For a rank-1 coordinating anti-windup variant (4), Theorem 3 claims that any equilibrium minimizes the l∞ tracking error, and Theorem 4 establishes global asymptotic stability in the disturbance-rejection regime b(l)+w > 0 and b(l̄)+w < 0. The paper also presents a district heating motivating example, a numerical simulation, and a proposition (Proposition 1) asserting that the hydraulic flow map satisfies Assumption 1.","tokens_in":1774,"tokens_out":2704,"duration_ms":105536,"significance":"If the results hold, the paper meaningfully extends the linear M-matrix results of [18,19] to a broad nonlinear setting, requiring only qualitative monotonicity properties rather than an explicit model of the interconnection. The proofs are based on a coherent Lyapunov framework and a key inequality (Lemma 1), and the optimality claims are derived from the assumptions rather than fitted to pre-selected cost functions. The authors are honest about the main limitations: Theorem 3 is conditional on equilibrium existence, the global stability proof for the coordinating controller covers only the unsaturated regime, and Proposition 1 is stated without proof. These are real gaps, but the core ideas are attractive and likely correct after the proof issues below are fixed.","major_comments":[{"comment":"The inequality direction used to bound term (13d) is reversed. The text states that (13d) is bounded below by (15)–(16) via '≥' and then concludes that the expression is strictly negative; however, Lemma 1 gives sign( ˜sat (˜u))^T H ˜b( ˜sat (˜u)) > Σ_{j∈J0} η_j |˜b_j|, so the right-hand side of (15)–(16) is strictly positive when ˜sat ≠ 0. What is needed for negative definiteness of ˙V is an upper bound, i.e., (13d) ≤ −(15)+(16), with the opposite inequality. This is a load-bearing error in the proof of Theorem 1 and must be corrected before the global asymptotic stability claim is established.","section":"§5.2, Eq. (14)–(16)"},{"comment":"The identity sign(dz_i(u0)) = sign(v†_i − v0_i) is incorrect. Saturation logic gives sign(dz_i(u0)) = sign(v0_i − v†_i): if dz_i(u0) > 0 then v0_i is the upper saturation bound and v†_i < v0_i, while if dz_i(u0) < 0 then v0_i is the lower bound and v†_i > v0_i. The final expression in (32) uses sign(v0 − v†), so the sign in (28) must be reversed and the intermediate inequalities (27)–(29) revised accordingly.","section":"§5.3, Eq. (28)"},{"comment":"The existence of an equilibrium is essential for Theorem 1, but the proof of Lemma 3 is only a sketch. The map T is claimed to be forward-invariant on a sufficiently large box C, and then Brouwer's fixed-point theorem is invoked, but no argument is given for forward invariance or for the choice of α. A complete proof, or a precise reference, is required for this lemma.","section":"§5.2, Lemma 3"},{"comment":"Proposition 1 asserts that the hydraulic flow map q(v) satisfies Assumption 1, and the text states 'we can show the following' and 'Hence q satisfies Assumption 1', but the proof is omitted. This proposition is load-bearing for the district heating application and the simulation: without it, the numerical example is not connected to the theoretical results. It should either be proved in an appendix with the hydraulic model stated in detail, or explicitly formulated as an assumption on the simulation model.","section":"§4, Proposition 1"},{"comment":"The Lyapunov function (11) is not differentiable at points where a component of ˜ζ or ˜u is zero, and the proof uses the convention d/dt |x| = sign(x) ẋ as if it were an exact derivative. The brief remark about exchanging |·| by an arbitrarily close smooth approximation does not by itself justify the global asymptotic stability conclusion. A rigorous treatment via, e.g., Clarke generalized derivatives or an explicit smoothing argument with uniform negative-definiteness estimates is needed.","section":"§5.2, Eqs. (11)–(12)"}],"minor_comments":[{"comment":"The equivalence '2/kC − 1^T P 1 ≥ 0 ⇔ kC/2 1^T P 1 = kC/2 1^T kP ≥ 0' appears to contain a typo: the last expression should be '≤ 1', not '≥ 0'.","section":"§Appendix, Eq. (25)"},{"comment":"The condition 'kC 2 1⊤kP ≤ 1' is ambiguous; it should be written as (kC/2) 1^T kP ≤ 1.","section":"§3.2, Assumption 3"},{"comment":"The assertion that dz_i(u) > 0 implies b_i(sat(u)) + w_i > 0 via Assumption 1(i) is not immediate because sat_i(u) is then the upper bound, so Assumption 1(i) does not directly apply unless one first derives monotonicity of b_i in its own input from the combination of Assumption 1(i)–(ii). A short justification would make this step clear.","section":"§5.2, Lemma 4 proof"},{"comment":"The use of the same symbol 'l' for the lower and upper saturation bounds (in the definition of S and later in Assumption 3 and Lemma 4) is confusing; consider using l_low and l_high or l_i and l̄_i for clarity.","section":"§1.1, Notation"},{"comment":"The simulation comparison is qualitative. Reporting quantitative error metrics (e.g., final values of ||x||_1 and ||x||_infinity in each scenario) would make the illustrative comparison more informative.","section":"§4.1, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a promising contribution that extends linear resource-sharing results to a nonlinear monotone setting. The sign errors in the proofs of Theorems 1 and 2 appear to be typographical and locally fixable, but the omitted proofs of Lemma 3 and Proposition 1 are more substantial and need to be supplied before publication. The non-smooth Lyapunov issue is also a genuine rigor gap. I see no circularity or fitting concerns: the optimality results are derived from the assumptions, and the controllers are not tuned to the objective functions. I recommend major revision and re-review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine nonlinear generalization of the authors' earlier M-matrix anti-windup PI results, and the main theorems are plausible—but the proofs as written contain sign/inequality errors in key steps. The paper needs correction before the central stability and optimality claims can be taken as established.\n\nWhat is actually new: the interconnection b is allowed to be fully nonlinear, with only a monotone-competition assumption (Assumption 1) instead of a linear M-matrix structure. On that basis, Theorems 1 and 2 show that a fully decentralized anti-windup PI controller yields global asymptotic stability and a weighted-l1 optimal equilibrium, and Theorem 3 gives l∞ optimality for a rank-1 coordinated variant. The district heating example is a nice real-world motivation, and the fact that no model of b is needed is a real strength.\n\nThe soft spots are real. The stress-test note is correct. In the proof of Theorem 1, term (13d) is bounded via (14) with '≥' and a plus sign in (15), but the correct bound has the opposite sign; Lemma 1 then gives strict negativity, but the printed chain does not. In the proof of Theorem 2, Eq. (28) asserts sign(dz_i(u0)) = sign(v†_i - v0_i), which is backwards; it should be sign(v0_i - v†_i). The later expression in (32) uses the correct sign, so the proof is likely repairable, but as printed the derivation is invalid. These are not minor typos in symbols only—they break the logical connection in the two main theorems. In addition, Proposition 1 (hydraulic monotonicity) is unproved and load-bearing for the simulation, Lemma 3 is only sketched, and the Lyapunov function uses a nonsmooth convention that is acknowledged but not fully resolved. Theorem 3's caveat about equilibrium existence is stated honestly.\n\nAll that said, the paper deserves a serious referee. The framework is interesting, the assumptions are clean, and the results, if corrected, would be a solid contribution. I would recommend a revise-and-resubmit with the request to fix the sign errors, provide a proof or reference for Proposition 1, and tighten the technical lemmas. I would not cite it in its current form.","headline":"Strong nonlinear extension of earlier anti-windup results, but the core proofs currently contain sign errors that need repair before the stability and optimality claims are established.","tokens_in":17448,"tokens_out":5163,"would_cite":false,"duration_ms":42464,"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 fully decentralized anti-windup PI controller drives capacity-limited networks to a unique globally stable equilibrium that minimizes a weighted sum of tracking errors, while a rank-1 coordinated variant minimizes the worst-case error.","keywords":["anti-windup control","PI control","multi-agent systems","capacity-limited networks","global asymptotic stability","equilibrium optimality","district heating","nonlinear interconnection"],"falsifier":"Run the decentralized controller on a two-agent plant whose coupling satisfies Assumption 1, then enumerate all open-loop equilibria and check whether any one has a strictly smaller $\\sum_i \\eta_i a_i |x_i|$ than the closed-loop equilibrium; a single such case would disprove Theorem 2. Alternatively, in a district-heating test rig hold one valve fixed while opening all others and measure that consumer's flow: if the flow does not strictly drop, Assumption 1(i) fails and the theory's application to that network is invalid.","tokens_in":16511,"feed_emoji":"⚖️","tokens_out":10688,"duration_ms":95383,"temperature":0.7,"pith_summary":"This paper establishes that two simple anti-windup PI controllers can optimally share a limited resource among many agents, even though the coupling between agents is unknown and nonlinear. In the fully decentralized version, each agent uses only local measurements and fixed gains; the closed loop has a unique, globally asymptotically stable equilibrium, and this equilibrium beats every other open-loop equilibrium with a different saturated input in a weighted sum of absolute tracking errors. Adding a single rank-1 coordination signal to the anti-windup makes any closed-loop equilibrium minimize the largest tracking error across agents. These theorems rest only on structural assumptions about the coupling: competition between agents and a positive weighted monotonicity. A district-heating case study shows how the model covers real capacity constraints, and simulations indicate both controllers track their corresponding optimal equilibria.","feed_headline":"PI control provably finds fair equilibria in capacity-limited networks","feed_subtitle":"Two anti-windup schemes minimize weighted or worst-case tracking errors with no model of the coupling.","key_machinery":"The load-bearing object is the nonlinear coupling map $b:\\mathcal{S}\\to\\mathcal{B}$ in the agent dynamics, where $\\mathcal{S}$ is the saturation range. Assumption 1 gives it a competition property (if $v\\ge \\hat v$ and $v_i = \\hat v_i$, then $b_i(v)-b_i(\\hat v)<0$) and a positive weighted monotonicity ($\\eta^\\top(b(v)-b(\\hat v))>0$ for $v\\ge \\hat v$, $v\\ne \\hat v$). From these, Lemma 1 and Lemma 2 extract the sign-dominance and inverse-monotonicity features used in every proof. Stability of the decentralized loop is carried by the weighted $\\ell^1$ Lyapunov function $V=\\sum_i \\eta_i \\frac{d_i}{p_i}|\\tilde\\zeta_i|+\\eta_i|\\tilde u_i|$; optimality arguments compare any other open-loop equilibrium to the saturated control $u_0$ and use the competition structure to force a strict inequality. The coordinating controller's anti-windup term $k_C \\mathbf{1}\\mathbf{1}^\\top \\mathrm{dz}(u)$ is what forces $x$ to be parallel to $\\mathbf{1}$, which is the source of the min-max property.","core_discovery":"The paper's central claim is that global stability and equilibrium optimality survive a full nonlinear generalization of capacity-limited resource sharing. For a plant of $n$ agents with dynamics $\\dot x_i = -a_i x_i + b_i(\\mathrm{sat}(u)) + w_i$ and a bounded nonlinear coupling $b$ satisfying Assumption 1, the decentralized anti-windup PI controller (2) has a unique globally asymptotically stable equilibrium (Theorem 1), and this equilibrium strictly minimizes $\\sum_i \\eta_i a_i |x_i|$ among all open-loop equilibria whose saturated input differs (Theorem 2). The rank-1 coordinating controller (4) has the property that every closed-loop equilibrium satisfies $x = -k_C \\mathbf{1}\\mathbf{1}^\\top \\mathrm{dz}(u)$, so all agents share the same error, and any such equilibrium is the unique minimizer of the $\\infty$-norm of the tracking error among open-loop equilibria with different saturated input (Theorem 3). Global asymptotic stability for the coordinating controller is proved only when the disturbance can be fully rejected (Theorem 4); in the saturated regime existence and stability are not guaranteed.","pith_inferences":["One consequence the authors leave implicit is that when Assumption 1 holds for a whole family of weight vectors $\\eta$, the fully decentralized controller simultaneously optimizes all the corresponding weighted absolute-error costs, giving a multi-objective optimality statement beyond any single chosen $\\eta$.","The proof machinery suggests a model-free design recipe: verify only the competition and weighted monotonicity of the steady-state input-output map, then tune local PI and anti-windup gains independently; this could be tested in other saturated resource networks such as communication or building cooling systems.","Because the coordination signal is rank-1, it can be implemented by a single broadcast channel; a natural extension would replace the fixed vector $\\mathbf{1}$ by any positive weight vector to obtain weighted min-max equilibria, something the paper does not explore.","In the saturated regime, the coordinated controller's existence and stability remain open; a concrete next step is to search for a Luapunov function or a counterexample in the two-agent saturated case, which would settle whether the min-max equilibrium is actually reachable."],"forward_implications":["A decentralized anti-windup PI controller can be installed per agent with local gains satisfying $k_i^P a_i > k_i^I$ and $k_i^P k_i^A < 1$; it will globally stabilize the network and reach an equilibrium that is optimal in the weighted absolute-error sense, with no communication and no model of the coupling.","The decentralized equilibrium's optimality is strict: any other open-loop equilibrium with different saturated inputs has a strictly larger value of $\\sum_i \\eta_i a_i |x_i|$, so the controller is doing implicit real-time optimization.","The rank-1 coordinating controller forces all agents to share the same tracking error at equilibrium, and any such equilibrium is strictly better than every other open-loop equilibrium in the infinity-norm sense, giving an explicit fairness and worst-case guarantee.","Global asymptotic stability for the coordinating controller is proved only when the disturbance is small enough to reject; for larger disturbances, the paper guarantees equilibrium optimality but leaves existence and stability of the saturated coordinating equilibrium open.","In district heating, the decentralized and coordinating laws reproduce the offline optimal allocations for average and worst-case temperature deviations respectively, suggesting that simple valve-level PI control can replace centralized optimization in practice."],"supporting_citations":[{"why":"Supplies the linear M-matrix case whose decentralized anti-windup PI controller this paper extends to a nonlinear setting.","marker":"[18]"},{"why":"Supplies the rank-1 coordinating anti-windup scheme and the fair-equilibrium optimality result that the paper generalizes.","marker":"[19]"},{"why":"Provides the definition and properties of M-matrices used to show that the linear case satisfies Assumption 1.","marker":"[22]"},{"why":"Gives the district-heating hydraulic and load-control model used to motivate the nonlinear coupling and the convex feasible flow set for benchmarks.","marker":"[10]"},{"why":"One of the static hydraulic network models cited to justify that the flow map $q(v)$ satisfies Assumption 1 in Proposition 1.","marker":"[24]"},{"why":"Another hydraulic-network steady-state model cited to support the same Proposition 1.","marker":"[25]"},{"why":"Provides the graph-based hydraulic modeling approach used in the district-heating simulation.","marker":"[26]"}],"fun_headline_variants":["PI control provably finds fair equilibria in capacity-limited networks","Decentralized anti-windup PI: global stability, weighted error minimum","No coupling model needed: PI control minimizes tracking errors","Rank-1 PI coordination stably minimizes worst-case tracking error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The single most load-bearing premise is Assumption 1 on the otherwise unknown coupling map: increasing another agent's input must lower your resource share, and some positive weighted sum of all shares must rise when everyone increases input; if a physical network lacks this competitive monotone structure, the stability and optimality conclusions do not follow.","fun_headline_variants_meta":{"raw":{"variants":["PI control provably finds fair equilibria in capacity-limited networks","Decentralized anti-windup PI: global stability, weighted error minimum","No coupling model needed: PI control minimizes tracking errors","Rank-1 PI coordination stably minimizes worst-case tracking error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1389,"prompt_tokens":952,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":568,"tokens_out":437,"duration_ms":5300,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:33:33.907568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the decentralized controller on a two-agent plant whose coupling satisfies Assumption 1, then enumerate all open-loop equilibria and check whether any one has a strictly smaller $\\sum_i \\eta_i a_i |x_i|$ than the closed-loop equilibrium; a single such case would disprove Theorem 2. Alternatively, in a district-heating test rig hold one valve fixed while opening all others and measure that consumer's flow: if the flow does not strictly drop, Assumption 1(i) fails and the theory's application to that network is invalid.","supporting_citations":[{"cited_title":"Decentralized pi-control and anti-windup in resource sharing net- works,","cited_arxiv_id":null,"evidence_quote":"Supplies the linear M-matrix case whose decentralized anti-windup PI controller this paper extends to a nonlinear setting."},{"cited_title":"Anti- windup coordination strategy around a fair equilibrium in resource sharing networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the rank-1 coordinating anti-windup scheme and the fair-equilibrium optimality result that the paper generalizes."},{"cited_title":"Combating district heating bottlenecks using load control,","cited_arxiv_id":null,"evidence_quote":"Gives the district-heating hydraulic and load-control model used to motivate the nonlinear coupling and the convex feasible flow set for benchmarks."},{"cited_title":"Output Regu- lation of Large-Scale Hydraulic Networks,","cited_arxiv_id":null,"evidence_quote":"One of the static hydraulic network models cited to justify that the flow map $q(v)$ satisfies Assumption 1 in Proposition 1."},{"cited_title":"On the Existence and Uniqueness of Steady State Solutions of a Class of Dynamic Hydraulic Networks via Actuator Placement,","cited_arxiv_id":null,"evidence_quote":"Another hydraulic-network steady-state model cited to support the same Proposition 1."},{"cited_title":"Hydraulic parameter estimation for district heating based on laboratory experiments,","cited_arxiv_id":null,"evidence_quote":"Provides the graph-based hydraulic modeling approach used in the district-heating simulation."}],"review_version":1}