{"id":"1099cab6-8058-491b-9a64-51f7d9e3ace4","arxiv_id":"2502.01180","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For positive linear systems, the minimax optimal control problem with unconstrained disturbances is solvable by linear programming exactly when the disturbance penalty dominates F^T p, in which case it reduces to the disturbance-free minimization problem.","lead":"This paper gives an explicit minimax optimal controller for positive systems with unbounded disturbances, and shows that the robust control problem has a finite solution exactly when the disturbance penalty is large enough. A smart generalist would read it because it reduces a worst-case robust control design to a simple linear program, making large-scale positive systems tractable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 omits a finite-maximum condition for LP (5); the LP can be unbounded under the stated assumptions, leaving the threshold γ ≥ F^T p undefined.","rationale":"In good faith, the paper's core idea is plausible: when γ ≥ F^T p, every disturbance w has non-positive net effect on the total cost, so the adversary chooses w=0 and the minimax problem collapses to the disturbance-free minimization problem. The sufficiency argument is sound given a stabilizing optimal policy, and the necessity follows from the Bellman recursion if p is well-defined. The most load-bearing gap is not the positivity-invariance condition (2), which is an explicit scope assumption, but the implicit assumption that LP (5) has a finite maximum. The theorem's 'if and only if' statement silently depends on p existing; the paper never states or proves boundedness of the LP. A simple admissible counterexample (A=I, B=0, E=0) shows the LP can be unbounded while the problem has no finite value, so the characterization as written is incomplete. This supports the reader's CONDITIONAL verdict: the mathematics is likely correct once the missing well-posedness condition is added, but the theorem statement and proof sketch are not yet fully rigorous. Agreement is partial because the reader identified condition (2) as the weakest assumption, whereas the more pointed concern is the undefined p in the absence of LP boundedness.","tokens_in":5151,"tokens_out":27130,"duration_ms":246221,"concrete_test":"Run LP (5) on the admissible data A=1, B=0, F=1, E=0, s=1, r=0. The feasible set is {p ≥ 0, ζ ≥ 0} and the objective 1^T p is unbounded. Confirm the original problem value is +∞ for any γ by noting x(t) = x0 + Σ w ≥ x0, so the cost Σ s^T x(t) diverges. Then decide whether the theorem's wording handles this case; if not, the missing boundedness condition is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central characterization rests on p being the solution of LP (5). The theorem omits a boundedness/well-posedness condition for this LP. Under the stated assumptions (2)-(3), the LP can be unbounded: take n=m=l=1, A=1, B=0, F=1, E=0, s=1, r=0. Then (2) is 1 ≥ 0, (3) is 1 > 0, and the LP constraints reduce to p ≤ 1+p, ζ ≥ 0, so every p ≥ 0 is feasible and 1^T p is unbounded. Meanwhile the minimax problem (state x^+ = x+w, no control) has infinite value for every x0 > 0 regardless of γ, since the cost Σ s^T x(t) diverges even with w=0. Thus the 'if and only if' statement, as written, does not characterize existence in this perfectly admissible case; the condition γ ≥ F^T p is undefined because no p is 'obtained'. The proof sketch, which says only that Lemma 5 is used to deduce (4), does not address whether p exists. The theorem should be amended to assume the LP has a finite optimum, or to state that an unbounded LP corresponds to no finite value. This is load-bearing because the entire minimax-to-minimization reduction is expressed in terms of the LP solution p.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the discrete-time, infinite-horizon minimax optimal control problem (1) for positive systems with nonnegative state, componentwise control constraint |u| ≤ Ex, and nonnegative disturbance w with no upper bound. The main result (Theorem 1) states that, under conditions (2) and (3), the problem has a finite value for every initial state x0 ∈ R^n_+ if and only if the disturbance penalty γ satisfies γ ≥ F^T p, where p is obtained from the linear program (5); in that case the value is p^T x0 and an optimal policy is the linear feedback u = -Kx with K given by (7). The proof outline argues via a dynamic-programming lemma from prior work that, under condition (4), the worst-case disturbance is attained at w = 0, so the minimax problem reduces to the corresponding disturbance-free minimization problem.","tokens_in":5414,"tokens_out":11299,"duration_ms":111704,"significance":"If correct, the result gives an explicit, computationally tractable LP characterization of the value function and of an optimal sparse linear controller for a class of robust positive-system problems, extending the authors' earlier constrained-disturbance result. The specific contribution is identifying the threshold condition (4) under which unconstrained disturbances do not increase the worst-case cost, so that the minimax value coincides with the L1-optimal control value. The double-tank example illustrates the design procedure. However, as detailed in the major comments, the theorem currently lacks a well-posedness condition for the LP (5), and the proof is only a sketch. These issues are load-bearing and must be resolved before the central claims are fully supported.","major_comments":[{"comment":"The theorem states that the problem has finite value iff γ ≥ F^T p, where p is obtained by solving (5), but it does not ensure that (5) has a finite optimum. Under the stated assumptions (2)–(3), the LP can be unbounded. For example, let n=m=l=1, A=1, B=0, F=1, E=0, s=1, r=0. Then (2) and (3) hold, and the constraints of (5) reduce to p ≤ 1+p and ζ ≥ 0, so every p ≥ 0 is feasible and maximize 1^T p is unbounded. At the same time, problem (1) has infinite value for every x0 > 0: even with w=0, the state remains x(t)=x0 and the cumulative cost Σ_{t=0}^∞ x0 diverges. Thus the phrase 'where p is obtained solving the linear program' is undefined in an admissible instance, and the 'if and only if' claim cannot be applied. The theorem should be amended to include a boundedness/well-posedness condition for the LP, or to state explicitly that an unbounded LP corresponds to no finite value, with a proof of that correspondence.","section":"Section 2, Theorem 1 and LP (5)"},{"comment":"The proof is explicitly only an outline: the text says 'In this extended abstract we only provide an outline of the proof.' The statements 'use induction over p_k^T x = J_k(x) ... to prove the equivalence' and 'use the equivalences in Lemma 5 to deduce the bound (4)' are not backed by a full induction or explicit derivations. In particular, the 'only if' direction requires showing that a finite value for (1) implies both that (5) has a finite optimal solution and that (4) holds; the outline does not address this. Since these omitted steps are the bridge between the minimax problem and the LP, a complete proof is necessary for a journal publication.","section":"Proof of Theorem 1"}],"minor_comments":[{"comment":"The reference to 'Hansson and Boydt (1998)' should be 'Hansson and Boyd'.","section":"References"},{"comment":"The second author's name appears as 'Emma T egling' with an extra space; it should be 'Emma Tegling'.","section":"Author byline"},{"comment":"Remark 3 is cryptic: 'The result in Theorem 1 is analogous for w < 0 and γ < F^T p respectively.' Since the problem is formulated only for w ≥ 0, the remark needs to define the modified problem for w < 0 and clarify what 'respectively' refers to.","section":"Remark 3"},{"comment":"The constraint '|u| ≤ Ex' should be written componentwise (e.g., |u_i| ≤ E_i x for i=1,...,m) to avoid ambiguity, since E is a matrix.","section":"Problem setup, Eq. (1)"},{"comment":"The phrase 'unconstrained disturbances' is used, but the problem has w ≥ 0 with no upper bound; the wording 'unbounded nonnegative disturbances' would be more precise.","section":"Abstract and Introduction"},{"comment":"The example says 'the disturbance and the control action are equally characterized, because a large-scale example cannot be tractably represented'; this sentence is unclear and should be rewritten.","section":"Example, Section 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript repeatedly describes itself as an extended abstract, and the proof is introduced as an outline. If the target venue accepts proof sketches, the proof-completeness concern could be softened, but the LP well-posedness issue in Theorem 1 is independent of the proof length and must be fixed. The paper may also need to clarify the selection of p if (5) has multiple optimal solutions, since the value p^T x0 and the threshold F^T p could then be ambiguous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main result is a genuine step forward. It removes the bounded-disturbance constraint from your earlier work and shows that, when the disturbance penalty γ is large enough, the minimax problem collapses to Rantzer's disturbance-free minimization. The explicit LP solution and the sparse feedback law are clean and useful, and the proof idea is sound: with assumptions (2)-(3), the max over w is nonnegative at w=0, and the Bellman equation reduces to the minimization case.\n\nThe soft spot is real. The statement of Theorem 1 says 'p is obtained solving the linear program (5)' but it never ensures that p exists. The LP can be unbounded under exactly the assumptions (2)-(3). Scalar example: A=1, B=0, F=1, E=0, s=1, r=0. Conditions (2) and (3) hold. The LP constraints become p ≤ 1+p and ζ≥0, so any p≥0 is feasible and 1^T p is unbounded. In the same example the minimax problem has infinite value for every x0>0 (the state is constant and the cost diverges even with w=0). So the 'if and only if' is undefined rather than false. The fix is easy: add the assumption that LP (5) is bounded, or state that an unbounded LP corresponds to no finite value. This omission doesn't kill the central insight — LP unboundedness exactly tracks the disturbance-free minimization problem being infinite — but it should be fixed in the statement.\n\nThe proof is explicitly an outline, as admitted; the induction and the use of Lemma 5 are not shown. For an extended abstract that is acceptable, but a full version needs to supply them. The numerical example is too simple to test the threshold condition; it just asserts the bound.\n\nBottom line: this is a plausible, useful result for positive-systems robust control. It deserves a serious referee; I would send it out. The authors need to patch the LP well-posedness clause and, for a journal version, fill in the proof.","headline":"Clean, plausible reduction of minimax to disturbance-free minimization for positive systems, but the theorem as stated misses an LP-boundedness condition that a simple counterexample exposes.","tokens_in":5916,"tokens_out":3786,"would_cite":true,"duration_ms":38060,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","93B52","90C05","49L20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A minimax optimal control problem for positive systems is solved explicitly by a linear program, with optimal linear feedback.","keywords":["minimax optimal control","positive systems","Bellman equation","linear programming","disturbance penalty","linear feedback","sparsity","robust control"],"falsifier":"Take the double-tank example in the paper with $\\gamma = 1.0$, below the computed threshold $F^{\\top}p \\approx 1.32$, and run value iteration on a long finite horizon for problem (8): if the worst-case value remains bounded as the horizon grows, the if-and-only-if statement of Theorem 1 is false. Alternatively, keep $\\gamma$ above the threshold and search numerically over nonlinear periodic policies for a value below $p^{\\top}x_0$, which would disprove optimality of the linear law.","tokens_in":4952,"feed_emoji":"🎛️","tokens_out":8945,"duration_ms":81982,"temperature":0.7,"pith_summary":"This paper asks when a worst-case optimal control problem for positive linear systems—systems whose state stays nonnegative—has a finite solution when disturbances are unconstrained and adversarial. It establishes an if-and-only-if condition: the problem has a finite minimax value exactly when the disturbance penalty vector $\\gamma$ satisfies $\\gamma \\geq F^{\\top}p$, where $p$ comes from an explicit linear program. When this condition holds, the optimal value is $p^{\\top}x_0$ and the optimal controller is a linear feedback law $u = -Kx$, so no nonlinear policy can do better. The result matters because it turns a robust game-theoretic control problem into a tractable linear program with a controller whose sparsity structure is prescribed by the constraint matrix $E$.","feed_headline":"A linear program decides when minimax control is finite","feed_subtitle":"For positive linear systems, the optimal worst-case policy is linear feedback, and feasibility hinges on a disturbance-penalty bound.","key_machinery":"The argument is carried by three pieces. First, the elementwise inequality $A \\geq |B|E$ keeps trajectories in the nonnegative orthant under all admissible controls and nonnegative disturbances, so the state space $\\mathbb{R}^n_+$ is invariant. Second, a linear value function $J^*(x)=p^{\\top}x$ is substituted into the Bellman equation, reducing the minimax problem to the linear program (5) whose solution $p$ and auxiliary variable $\\zeta$ produce the disturbance-penalty threshold $F^{\\top}p$. Third, the optimal control is chosen pointwise from the sign of $r_i^{\\top} + p^{\\top}B_i$, which yields the linear feedback matrix $K$ with rows $\\operatorname{sign}(r_i^{\\top} + p^{\\top}B_i)E_i$ and therefore inherits the sparsity pattern of $E$. A supporting lemma from earlier work by the same authors supplies the equivalence between finite value, convergence of value iteration, and existence of a nonnegative Bellman solution, which is what turns these pieces into an if-and-only-if proof.","core_discovery":"The central claim is Theorem 1: for discrete-time positive systems $x(t+1)=Ax(t)+Bu(t)+Fw(t)$ with $w\\geq 0$, running cost $s^{\\top}x + r^{\\top}u - \\gamma^{\\top}w$, and control constraint $|u|\\leq Ex$, the assumptions $A \\geq |B|E$ and $s > E^{\\top}|r|$ imply that the minimax problem has a finite value for every $x_0\\in\\mathbb{R}^n_+$ if and only if $\\gamma \\geq F^{\\top}p$, where $p$ solves the linear program maximizing $\\mathbf{1}^{\\top}p$ subject to $p \\leq s + A^{\\top}p - E^{\\top}\\zeta$ and $|r + B^{\\top}p| \\leq \\zeta$. Under that condition the optimal value equals $p^{\\top}x_0$ and the optimal policy is linear, with row $i$ of $K$ equal to $\\operatorname{sign}(r_i^{\\top} + p^{\\top}B_i)E_i$. A direct corollary is that, whenever a finite solution exists, the worst-case disturbance problem collapses to the disturbance-free minimization problem, and $\\gamma \\geq F^{\\top}p$ is exactly the threshold on the disturbance penalty that makes the game finite.","pith_inferences":["Implicit in the proof is a reading of $\\gamma \\geq F^{\\top}p$ as an induced-gain condition; one could test numerically how far it is from necessary when the invariance condition $A \\geq |B|E$ is violated but states spend only a little time outside the orthant.","The linear-value-function approach suggests a continuous-time analogue for Metzler matrices, with the linear program replaced by a suitable linear program or linear matrix inequality; the paper flags the continuous setting as ongoing rather than proving it.","A finite-horizon version of the problem would likely have finite value for any $\\gamma$, with the threshold pinning down when the infinite-horizon limit stays finite; this could be checked by value iteration on the double-tank example for $\\gamma$ below the bound."],"forward_implications":["Any control policy, including nonlinear and nonsparse ones, achieves worst-case cost no lower than the linear policy $u=-Kx$.","Below the threshold $\\gamma < F^{\\top}p$, no controller can make the worst-case cost finite; at or above it, the closed loop is positively asymptotically stable with performance level $\\gamma$.","The sparsity structure chosen through $E$ is preserved exactly in the optimal gain $K$, so large-scale control constraints can be designed up front and solved by linear programming.","When the threshold is met, minimax control of positive systems is no harder than disturbance-free linear-cost control, so existing linear-programming solvers apply."],"supporting_citations":[{"why":"It supplies the prior problem setting and Lemma 5, the equivalence between finite value, convergent value iteration, and a nonnegative Bellman solution on which the proof of Theorem 1 builds.","marker":"Gurpegui et al. (2023)"},{"why":"It gives the disturbance-free positive-system minimization problem whose explicit linear-programming solution the minimax problem is shown to coincide with when feasible.","marker":"Rantzer (2022)"},{"why":"It provides the double-tank process dynamics used in the illustrative example.","marker":"Hansson and Boydt (1998)"}],"fun_headline_variants":["Minimax control is finite iff this LP is feasible","A linear program decides whether minimax control is finite","Finite minimax control for positive systems: an LP check","Feasible LP means worst-case control becomes disturbance-free","Linear feedback is optimal when an LP condition holds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the elementwise inequality $A \\geq |B|E$, which guarantees the state never leaves the nonnegative orthant under any admissible control and nonnegative disturbance; if that inequality fails, the linear value-function proof and the LP characterization no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["Minimax control is finite iff this LP is feasible","A linear program decides whether minimax control is finite","Finite minimax control for positive systems: an LP check","Feasible LP means worst-case control becomes disturbance-free","Linear feedback is optimal when an LP condition holds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2524,"prompt_tokens":874,"completion_tokens":1650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1571}},"tokens_in":490,"tokens_out":1650,"duration_ms":13196,"temperature":1.0,"reasoning_tokens":1571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:17:46.456445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the double-tank example in the paper with $\\gamma = 1.0$, below the computed threshold $F^{\\top}p \\approx 1.32$, and run value iteration on a long finite horizon for problem (8): if the worst-case value remains bounded as the horizon grows, the if-and-only-if statement of Theorem 1 is false. Alternatively, keep $\\gamma$ above the threshold and search numerically over nonlinear periodic policies for a value below $p^{\\top}x_0$, which would disprove optimality of the linear law.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the prior problem setting and Lemma 5, the equivalence between finite value, convergent value iteration, and a nonnegative Bellman solution on which the proof of Theorem 1 builds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It gives the disturbance-free positive-system minimization problem whose explicit linear-programming solution the minimax problem is shown to coincide with when feasible."},{"cited_title":"and Boydt, S","cited_arxiv_id":null,"evidence_quote":"It provides the double-tank process dynamics used in the illustrative example."}],"review_version":1}