{"id":"b395a20a-3966-4e76-8837-ecab34ccff9b","arxiv_id":"2505.00319","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Bregman-divergence reformulation of discrete-time H∞ control yields closed-form nonlinear central controllers for strictly convex nonquadratic costs, reducing to the classical Riccati solution when costs are quadratic.","lead":"The paper derives a way to design robust controllers for linear systems when the penalties on state, control, and disturbance are convex but not quadratic, using Bregman divergences to keep the math closed-form. If it works, it would let engineers encode hard safety limits, saturating actuators, and sparse inputs into H∞ control without online optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main iff rests on unproved existence of strictly convex p (and strict convexity of m=p−q); without it the claimed extension to arbitrary strictly convex costs is not established.","rationale":"The reader's weakest assumption and my concern coincide: the main theorem is only an equivalence between solvability and a pair of conditions that include an unknown strictly convex p satisfying (15). The proof must show that such p exists whenever a controller exists, and that m=p−q is strictly convex enough to invert ∇m. Neither is established. This matters because the paper's stated contribution is to extend H∞ control to arbitrary strictly convex costs; if the only known route is to fix m or g quadratic, the promised generality is not delivered. The quadratic recovery in Corollary 1 gives independent support for the framework in the classical limit, but it cannot certify the nonquadratic case. The input-limited example's controller also appears to violate its own hard limit, but that is secondary. Since the missing existence/strict-convexity argument is a proof gap rather than a demonstrated counterexample, the paper remains conditionally acceptable pending a proof or a constructive existence procedure. I therefore do not change the reader's CONDITIONAL verdict.","tokens_in":19254,"tokens_out":15664,"duration_ms":156522,"concrete_test":"Independently re-derive equation (15) from equation (45) without using the assertion that m=p−q is strictly convex. The critical step is §IV-B.2, where 'by strict convexity of m, ∇m is invertible' is used to pass to m*; check whether (15) follows if m is only convex. If it does not, construct a one-dimensional instance satisfying Assumption 1 and concavity condition (13) for which the maximizer in (45) is non-unique over an interval (so m is not strictly convex); such an instance would invalidate the necessity proof of Theorem 2. If no such instance exists, prove that (13) plus (15) enforces strict convexity of m, which would resolve the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2 states an iff between solvability of Problem 2 and conditions (13)–(15), where p is a strictly convex Lyapunov function solving the dual functional equation (15). The proof of necessity in §IV-B.2 needs m := p−q to be strictly convex so that ∇m is invertible and the Legendre transform m* can be introduced. This strict convexity is not part of Assumption 1, and the difference of two strictly convex functions is not generally strictly convex. The only justification offered is the sentence 'the objective ... strictly concave in w and strictly convex in x. This implies m(·) will be strictly convex.' That implication is unproved and not generally valid: m(x) = max_w [g(Ax+w) − γ²s(w)] is a supremum of convex functions in x; strict convexity of the summands in x and strict concavity in w do not by themselves guarantee strict convexity of the supremum or uniqueness of the maximizer. Moreover, Lemma 2 and the derivation of (30) presuppose the existence of a stationary point (û, ŵ) and a storage p satisfying the Riccati-like equation; no existence theorem for such p under Assumption 1 is given. The synthesis in §V sidesteps the issue by fixing m(x) = xᵀMx or g(x) = xᵀGx, so the promised generality to arbitrary strictly convex q, r, s is not actually delivered. Theorems 4 and 5 are also stated with only 'proof idea' references and their proofs are not supplied, but the central gap is the missing existence/strict-convexity result for p.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers discrete-time LTI systems x_{k+1}=Ax_k+Bu_k+w_k and proposes an infinite-horizon H∞ formulation with strictly convex, even, nonnegative costs q,r,s on state, input, and disturbance. Using Bregman divergences, it replaces quadratic completion of squares by a three-point identity and claims a necessary and sufficient condition for existence of a causal time-invariant controller achieving level γ: a Riccati-like equation (15) for a strictly convex storage p, a concavity condition (13), and a closed-form central controller u*=-∇r*(B^T∇g(Ax+w)). It proves recovery of the classical quadratic H∞ results, offers three offline synthesis routes based on fixing m or g, and illustrates the method with input-limited, safety-envelope, and exponential-cost examples. The main theorems are Theorem 2 and Theorems 3–5.","tokens_in":19645,"tokens_out":13422,"duration_ms":143412,"significance":"If the claimed results hold, the paper would give a substantial and useful extension of H∞ control to nonquadratic penalties while preserving an explicit, offline-computable controller. The choice of Bregman divergence is natural, the quadratic-limit corollary is worked out in detail, and the controller formula (18) is explicit. The paper is also honest in pointing to some limitations (underactuated case, dependency on companion [21]). However, the current manuscript does not fully support the main iff theorem: several load-bearing steps are asserted without proof or rest on unstated assumptions, and the examples fall outside the stated assumptions. The significance is therefore prospective rather than established.","major_comments":[{"comment":"The proof of necessity requires m := p - q to be strictly convex so that ∇m is invertible and the Legendre transform m* can be introduced. After Eq. (45) the text asserts this from the fact that the objective is strictly concave in w and strictly convex in x; that implication is not generally valid, since a pointwise supremum of strictly convex functions need not be strictly convex and strict concavity in w does not guarantee a unique maximizer. Assumption 1 does not imply strict convexity of m, because the difference of two strictly convex functions need not be strictly convex. The derivation of Eqs. (47)–(50) therefore rests on an unstated hypothesis. Please add strict convexity (and differentiability) of m to the assumptions of Theorem 2, or supply a proof from the stated assumptions.","section":"§IV-B.2, Eqs. (45)–(49)"},{"comment":"The step claiming that D_{γ²s}(w_k, ŵ_k) - D_g(Ax_k + w_k, Ax_k + ŵ_k) ≥ 0 for all w_k holds if and only if γ²s(·) - g(Ax_k + (·)) is convex is not justified. The inequality only says that ŵ_k is a global minimizer of h(w)=γ²s(w)-g(Ax_k+w); a function can have a global minimum without being convex (e.g., h(w)=(w²-1)²). Since condition (13) of Theorem 2 is exactly this convexity/concavity condition, the necessity direction of the theorem is incomplete. Please prove that the global-minimizer property together with the Riccati-like identity forces convexity, or state the necessary condition in the weaker global-minimizer form.","section":"§IV-B.1, Eqs. (39)–(40)"},{"comment":"Condition (53) is part of the if-and-only-if statement of Theorem 3, and the design recipe in §V-A depends on it, but the proof text says the verification 'will be ommitted.' Omitting the proof of a load-bearing condition means Theorem 3 is not established as stated. Please supply the proof or explicitly mark this as a conjecture/partial result.","section":"Appendix D, Theorem 3, Eq. (53)"},{"comment":"Theorems 4 and 5 each state that the proof idea is in Appendix D, but Appendix D is titled 'Proof of Theorem 3' and contains no proof of either theorem. The second and third synthesis routes in §V-B and §V-C therefore rest on unproved results. Please provide complete proofs or clearly separate conjectured design conditions from proven ones.","section":"Theorems 4 and 5; Appendix D"},{"comment":"The examples in Section VI use extended-value indicator-like penalties, e.g., r(u)=∞ for |u|≥t in (68) and q(x)=∞ for |x|≥t in (70). These functions are not differentiable finite-valued strictly convex functions, so they do not satisfy Assumption 1, and the Bregman divergences, gradient inversions, and Legendre transforms used in the proofs are not directly defined for them. The application section therefore goes beyond the theory as stated; the paper should either restrict to smooth barrier approximations with a limiting argument or extend the theory to allow nondifferentiable extended-value convex functions.","section":"Section VI and Assumption 1"}],"minor_comments":[{"comment":"Theorem 2 refers to 'Problem (2)', but Eq. (2) is the finite-horizon quadratic background problem; the target is Problem 2 in Section III-A.","section":"Theorem 2 statement"},{"comment":"Assumption 1 does not explicitly state differentiability or coercivity, although Bregman divergences, gradients, and Legendre transforms throughout the paper require them. Please make these hypotheses explicit.","section":"Assumption 1"},{"comment":"The 'parameterization of all controllers' in (16) is a verification inequality involving the unknown policy; unlike Eq. (6) in the quadratic case, it does not give an explicit construction of the family. The wording should be adjusted.","section":"Eq. (16)"},{"comment":"The identification of \\u005cu005cu005cwedge u_k with u_k^* (x_k, \\u005cu005cu005cwedge w_k) is made in the proof of Lemma 3 but used earlier in the main proof; state it before Eq. (27) to avoid confusion.","section":"Proof of Theorem 2, around Eq. (27)"},{"comment":"Figures 1–3 are referenced but not described in the text; include axes, legends, and disturbance models so the reported comparisons are reproducible.","section":"Figures 1–3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a promising extension but currently overclaims: the main iff theorem and two synthesis theorems are not fully proven, and the examples sit outside the stated assumptions. I recommend major revision rather than rejection because the core Bregman machinery appears sound and the quadratic recovery is carefully executed. The authors should also clarify the relation to the companion LQR paper [21], since the completion-of-squares setup in Theorem 2 borrows directly from that line of work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Josh—one thing you should know up front: this paper actually does something new. It extends discrete-time H-infinity control to strictly convex nonquadratic costs by replacing the quadratic completion-of-squares with Bregman divergences, producing a closed-form nonlinear central controller and a Riccati-like dual identity plus a concavity condition. The quadratic specialization recovers the classical H-infinity ARE, the negativity test, and the linear controller in a detailed and correct-looking way. That's a real contribution, and the Bregman machinery is used in a genuinely clever manner.\n\nThe soft spots are in proportion to how central they are. Theorem 2 is stated as an if-and-only-if for arbitrary strictly convex costs, but the proof silently assumes m = p − q is strictly convex in order to invert ∇m and introduce the Legendre transform. The only justification is a sentence claiming that because the objective is strictly concave in w and strictly convex in x, the supremum m will be strictly convex. That's not generally true, and no existence theorem for the storage function p under Assumption 1 is supplied. The synthesis sections dodge the issue by fixing m(x)=xᵀMx or g(x)=xᵀGx, so the advertised generality over arbitrary strictly convex q, r, s is not actually delivered. Theorems 4 and 5 are stated with only 'proof idea' references, and Appendix D proves Theorem 3 but explicitly defers condition (53). There's also a concrete bug in the input-limited example: the second branch of the controller is −a·sign(x), whose magnitude exceeds the stated limit t. That's a real error, not a nitpick.\n\nNone of this kills the core idea. The quadratic recovery in Appendix B is thorough, the framework is coherent, and the dependence on the companion LQR paper is acceptable because the H-infinity result is a genuine extension. But the main theorem's necessity direction is not established as written, and the missing proofs matter because they are the recipes users would actually follow.\n\nThis paper deserves a serious referee—it's novel enough and technically interesting enough to spend the time. I'd send it to review with a clear request for major revision: either prove the existence and strict convexity of p, or restate the theorem with those as explicit assumptions; supply the missing proofs for Theorems 4 and 5; and fix the input-limited example. After that, it could be a solid contribution.","headline":"Strong extension of H-infinity control to strictly convex costs via Bregman divergences, with a correct quadratic recovery, but the main iff theorem rests on an unproved strict-convexity/existence assumption and the synthesis theorems are incomplete.","tokens_in":20163,"tokens_out":4306,"would_cite":true,"duration_ms":40539,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B36","93C55","49N15","52A41"],"pacs":[],"model":"deepseek-v4-flash","headline":"For strictly convex costs on state, input, and disturbance, a causal time-invariant H∞ controller exists exactly when a dual Riccati-like identity and a concavity condition hold, with a closed-form nonlinear central controller.","keywords":["H-infinity control","Bregman divergence","nonquadratic costs","Riccati-like equation","convex analysis","discrete-time linear systems","closed-form controller","robust control"],"falsifier":"Take the scalar system $x_{k+1}=ax_k+bu_k+w_k$ with quartic costs $q(x)=x^4$, $r(u)=u^4$, $s(w)=w^4$ and search numerically for a strictly convex $p$ satisfying $p^*(\\xi)+r^*(b\\xi)=(p-q)^*(a\\xi)+\\gamma^2s^*(\\gamma^{-2}\\xi)$ across a range of $\\gamma$; if any admissible $a,b,\\gamma$ admits no strictly convex $p$ with $p-q$ strictly convex, then the 'if' direction needs an additional existence condition beyond Assumption 1.","tokens_in":19015,"feed_emoji":"🎛️","tokens_out":8478,"duration_ms":79300,"temperature":0.7,"pith_summary":"Most H∞ control synthesis is built on quadratic penalties, where Riccati equations and completion-of-squares yield closed-form solutions; nonquadratic costs usually force online optimization or heuristics. This paper claims that strictly convex penalties on state, input, and disturbance can still support a closed-form, time-invariant, full-information controller meeting the worst-case bound γ, so long as one storage function satisfies a dual Riccati-like equation and a concavity condition holds. The reward is a general template for nonlinear H∞ controllers that can encode input saturation, hard state envelopes, or sparse actuation without online optimization, and the classical quadratic H∞ solution is recovered as a special case.","feed_headline":"Nonquadratic H∞ control gains a closed-form central controller","feed_subtitle":"Bregman divergences replace completion-of-squares; quadratic Riccati H∞ falls out as a special case.","key_machinery":"The machinery is the Bregman divergence $D_\\phi(x,y)=\\phi(x)-\\phi(y)-\\nabla\\phi(y)^\\top(x-y)$, together with its three-point identity $D_{\\phi_1}(x,y)+D_{\\phi_2}(x,z)=D_{\\phi_1+\\phi_2}(x,x^*)+D_{\\phi_1}(x^*,y)+D_{\\phi_2}(x^*,z)$, where $\\nabla(\\phi_1+\\phi_2)(x^*)=\\nabla\\phi_1(y)+\\nabla\\phi_2(z)$. This identity is the nonquadratic analogue of completion of squares, and its duality property $D_\\phi(x,y)=D_{\\phi^*}(\\nabla\\phi(y),\\nabla\\phi(x))$ turns the one-step storage balance into the Fenchel-dual Riccati-like equation (15). The same toolkit converts the worst-case disturbance test into the concavity condition (13) and identifies the closed-form control law as the unique minimizer of a convex subproblem.","core_discovery":"On the paper's own terms, the central claim is an equivalence: for the discrete-time linear system $x_{k+1}=Ax_k+Bu_k+w_k$ and strictly convex even penalties $q(x)$, $r(u)$, $s(w)$, a causal time-invariant full-information controller exists at level $\\gamma>0$ if and only if there is a strictly convex Lyapunov function $p$ satisfying $p^*(\\xi)+r^*(B^\\top\\xi)=(p-q)^*(A^\\top\\xi)+\\gamma^2 s^*(\\gamma^{-2}\\xi)$ and the function $-\\gamma^2 s(\\cdot)+g(Ax+(\\cdot))$ is concave in its second argument, where $g^*:=p^*+r^*(B^\\top(\\cdot))$. When these hold, the central controller is $u_k^*=-\\nabla r^*(B^\\top\\nabla g(Ax_k+w_k))$, it is stabilizing, and all level-$\\gamma$ controllers are characterized by a Bregman divergence inequality. Under quadratic costs, the dual identity and concavity condition reduce exactly to the standard discrete-time H∞ algebraic Riccati equation, the negative-semidefinite test, and the classical linear central controller.","pith_inferences":["Inference: the proof's use of strict convexity of $m=p-q$ to invert $\\nabla m$ is an extra regularity hypothesis beyond Assumption 1; a natural next step is an existence theorem for $p$ given arbitrary admissible $q,r,s$, rather than only designs that fix $m$ or $g$ as quadratic.","Inference: the Fenchel-dual form of the Riccati-like identity suggests that costs with simple duals (such as $\\ell^p$ penalties with $1<p\\le 2$) may yield closed-form nonlinear laws without ever computing $p$ explicitly.","Inference: the concavity condition could be certified empirically for a candidate design by sampling Hessians of $-\\gamma^2s+g(Ax+\\cdot)$ over the relevant state-disturbance region, providing a practical verification route when symbolic analysis is hard.","Inference: because the framework is algebraic in the Bregman divergences rather than tied to full-information structure, analogous nonquadratic H∞ filters or output-feedback controllers are plausible immediate extensions."],"forward_implications":["With quadratic $q,r,s$, the main theorem reproduces the standard H∞ ARE, the negativity condition, and the linear central controller, so the extension contains the classical theory as a limiting case.","The central controller $u_k^*=-\\nabla r^*(B^\\top\\nabla g(Ax_k+w_k))$ makes the closed loop stable: with no disturbance, $p$ decreases along trajectories, and with bounded disturbance the state remains bounded with the advertised $\\gamma$-level gain.","All controllers that achieve level $\\gamma$ are captured by a single Bregman divergence inequality, so the central controller sits inside a complete parameterization rather than being an isolated construction.","Synthesis can be done offline: fixing the shape of $m=p-q$ or of $g$ reduces feasibility to matrix inequalities and, under strong-convexity/smoothness assumptions, to a convex feasibility program; no online optimization is needed.","The recipe produces explicit nonlinear laws for input-limited control, hard state safety envelopes, and exponentially penalized actuation, each with a guaranteed H∞ performance ratio."],"supporting_citations":[{"why":"Supplies the classical quadratic H∞ framework the paper extends: Theorem 1, the Riccati recursion, the negativity condition, and the central linear controller.","marker":"[10]"},{"why":"Defines the Bregman divergence that replaces squared-error completion-of-squares throughout the proofs.","marker":"[22]"},{"why":"Provides the duality and three-point identities for Bregman divergences on which the main theorem's derivation rests.","marker":"[24]"},{"why":"Companion generalization of LQR to nonquadratic Bregman costs, setting the pattern the present paper applies to H∞ control.","marker":"[21]"}],"fun_headline_variants":["Bregman H∞: explicit controller for convex costs","Closed-form H∞ controller for convex costs via Bregman","H∞ control with convex costs: closed-form via Bregman","Nonquadratic H∞ gets closed-form Bregman control","Beyond quadratic: Bregman-based closed-form H∞ control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the existence of a strictly convex storage function $p$ that satisfies the dual Riccati-like equation (15), with $p-q$ also strictly convex so gradients can be inverted; the paper proves no such $p$ exists for every admissible choice of costs, and its synthesis sections avoid the issue by fixing $m$ or $g$ to be quadratic.","fun_headline_variants_meta":{"raw":{"variants":["Bregman H∞: explicit controller for convex costs","Closed-form H∞ controller for convex costs via Bregman","H∞ control with convex costs: closed-form via Bregman","Nonquadratic H∞ gets closed-form Bregman control","Beyond quadratic: Bregman-based closed-form H∞ control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":5065,"prompt_tokens":1014,"completion_tokens":4051,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":3965}},"tokens_in":630,"tokens_out":4051,"duration_ms":29808,"temperature":1.0,"reasoning_tokens":3965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:46:11.442911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the scalar system $x_{k+1}=ax_k+bu_k+w_k$ with quartic costs $q(x)=x^4$, $r(u)=u^4$, $s(w)=w^4$ and search numerically for a strictly convex $p$ satisfying $p^*(\\xi)+r^*(b\\xi)=(p-q)^*(a\\xi)+\\gamma^2s^*(\\gamma^{-2}\\xi)$ across a range of $\\gamma$; if any admissible $a,b,\\gamma$ admits no strictly convex $p$ with $p-q$ strictly convex, then the 'if' direction needs an additional existence condition beyond Assumption 1.","supporting_citations":[{"cited_title":"Hassibi, A","cited_arxiv_id":null,"evidence_quote":"Supplies the classical quadratic H∞ framework the paper extends: Theorem 1, the Riccati recursion, the negativity condition, and the central linear controller."},{"cited_title":"The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming,","cited_arxiv_id":null,"evidence_quote":"Defines the Bregman divergence that replaces squared-error completion-of-squares throughout the proofs."},{"cited_title":"Clustering with bregman divergences,","cited_arxiv_id":null,"evidence_quote":"Provides the duality and three-point identities for Bregman divergences on which the main theorem's derivation rests."},{"cited_title":"Beyond quadratic costs in lqr: Bregman divergence control,","cited_arxiv_id":null,"evidence_quote":"Companion generalization of LQR to nonquadratic Bregman costs, setting the pattern the present paper applies to H∞ control."}],"review_version":1}