REVIEW 5 major objections 5 minor 27 references
Beyond Quadratic Costs: A Bregman Divergence Approach to H$_\infty$ Control
T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (5)
- [§IV-B.2, Eqs. (45)–(49)] 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.
- [§IV-B.1, Eqs. (39)–(40)] 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.
- [Appendix D, Theorem 3, Eq. (53)] 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.
- [Theorems 4 and 5; Appendix D] 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 VI and Assumption 1] 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.
minor comments (5)
- [Theorem 2 statement] 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.
- [Assumption 1] 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.
- [Eq. (16)] 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.
- [Proof of Theorem 2, around Eq. (27)] 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.
- [Figures 1–3] Figures 1–3 are referenced but not described in the text; include axes, legends, and disturbance models so the reported comparisons are reproducible.
Circularity Check
No circularity: the central claim is a conditional functional-equation condition, the controller is derived from the storage function rather than fitted, and the only self-citation is not load-bearing.
full rationale
I walked the derivation chain of Theorem 2. Lemma 2's Bregman three-point identity is a purely algebraic completion-of-squares: once p is defined by the Riccati-like equation (30), the objective (12) transforms into the divergence form (27), with the stationarity conditions (28)-(29) canceling the cross terms. The main theorem is conditional on the existence of a strictly convex p satisfying the dual functional equation (15); it does not smuggle the controller into the premise, because (15) is a functional equation in p alone and the controller u* is subsequently computed from p via (18). The necessity direction derives condition (45) from positivity of (39), and sufficiency uses (13) to make gamma^2 s - g(Ax + .) convex with stationary point w_hat, so choosing u = u* gives JT >= 0. No fitted parameter is renamed as a prediction: the design parameters M and G in Section V are free synthesis choices, and the derived q, r, or s are constructed to satisfy the conditions, which is legitimate design freedom rather than circular reasoning. The only self-citation, the companion LQR paper [21], is mentioned as related work and is not used in the proof; the Bregman identities are attributed to [22,24]. There are genuine correctness gaps, which I flag but do not count as circularity: the claim that strict concavity in w and strict convexity in x imply strict convexity of m(x) = max_w[-gamma^2 s(w) + g(Ax+w)] is asserted without proof in Section IV-B.2 and is not generally valid; Theorems 4 and 5 are labeled as 'proof idea' with proofs deferred to Appendix D and not supplied; and Appendix C explicitly says the derivation of condition (53) 'will be ommitted.' These are unproved mathematical assertions, not reductions of a claimed result to its own inputs. Because no central equation or controller formula is equivalent by construction to its inputs, the circularity score is 0.
Assumptions & free parameters
free parameters (3)
- M (positive definite matrix shaping m(x)=x^T M x)
- G (positive definite matrix shaping g(x)=x^T G x)
- Scalar shaping parameters in examples (m, g, t, s) =
m=0.11, t=0.1, s=1 (Fig.1); m=0.2, t=0.2, s=1 (Fig.2); g=15 (Fig.3)
assumptions (4)
- domain assumption q, r, s are strictly convex, even, nonnegative, with zero value and gradient at 0 (Assumption 1)
- domain assumption A is invertible and B is full rank (Assumption 2)
- ad hoc to paper There exists a strictly convex Lyapunov function p satisfying p*(ξ)+r*(B^T ξ) = (p−q)*(A^T ξ)+γ²s*(γ^{−2}ξ) (15)
- ad hoc to paper m = p−q is strictly convex
Cite this review
Pith. "Pith review of Beyond Quadratic Costs: A Bregman Divergence Approach to H$_\infty$ Control." pith.science (2026). https://pith.science/paper/KDMAYLAO
@misc{pith2026250500319,
author = {Pith},
title = {Pith review of: Beyond Quadratic Costs: A Bregman Divergence Approach to H$_\infty$ Control},
year = {2026},
howpublished = {\url{https://pith.science/paper/KDMAYLAO}},
note = {Machine review of arXiv:2505.00319}
}
abstract
In the past couple of decades, non-quadratic convex penalties have reshaped signal processing and machine learning; in robust control, however, general convex costs break the Riccati and storage function structure that make the design tractable. Practitioners thus default to approximations, heuristics or robust model predictive control that are solved online for short horizons. We close this gap by extending $H_\infty$ control of discrete-time linear systems to strictly convex penalties on state, input, and disturbance, recasting the objective with Bregman divergences that admit a completion-of-squares decomposition. The result is a closed-form, time-invariant, full-information stabilizing controller that minimizes a worst-case performance ratio over the infinite horizon. Necessary and sufficient existence/optimality conditions are given by a Riccati-like identity together with a concavity requirement; with quadratic costs, these collapse to the classical $H_\infty$ algebraic Riccati equation and the associated negative-semidefinite condition, recovering the linear central controller. Otherwise, the optimal controller is nonlinear and can enable safety envelopes, sparse actuation, and bang-bang policies with rigorous $H_\infty$ guarantees.
Figures
Reference graph
Works this paper leans on
-
[21]
Beyond quadratic costs in lqr: Bregman divergence control,
B. Hassibi, J. Hajar, and R. Ghane, “Beyond quadratic costs in lqr: Bregman divergence control,” 2025. Submitted to arxiv
work page 2025
-
[1]
G. Zames, “Feedback and optimal sensitivity: Model reference trans- formations, multiplicative seminorms, and approximate inverses,” IEEE TAC, vol. 26, pp. 301–320, Apr. 1981
work page 1981
-
[2]
Feedback, minimax sensitivity, and optimal robustness,
G. Zames and B. Francis, “Feedback, minimax sensitivity, and optimal robustness,” IEEE TAC, vol. 28, no. 5, pp. 585–601, 2003
work page 2003
-
[3]
Directional interpolation approach to h ∞-optimization and robust stabilization,
H. Kimura, “Directional interpolation approach to h ∞-optimization and robust stabilization,” IEEE TAC, vol. 32, no. 12, pp. 1085–1093, 1987
work page 1987
-
[4]
State-space solutions to standard h2 and h∞ control problems,
J. Doyle, K. Glover, P. Khargonekar, and B. Francis, “State-space solutions to standard h2 and h∞ control problems,” in 1988 ACC , (Atlanta, GA, USA), pp. 1691–1696, IEEE, June 1988
work page 1988
-
[5]
K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control . Prentice Hall, 1996
work page 1996
-
[6]
L2-gain analysis of nonlinear systems and nonlin- ear state feedback hl control,
A. van der Schaft, “L2-gain analysis of nonlinear systems and nonlin- ear state feedback hl control,” IEEE TAC, vol. 37, no. 6, pp. 770–784, 1992
work page 1992
-
[7]
H-infinity static output-feedback control for rotorcraft,
J. Gadewadikar, F. L. Lewis, K. Subbarao, K. Peng, and B. M. Chen, “H-infinity static output-feedback control for rotorcraft,” Journal of Intelligent and Robotic Systems , vol. 54, pp. 629–646, 2009
work page 2009
Show all 27 references
-
[8]
Nonlinear h-infinity feedback control for asynchronous motors of electric trains,
G. Rigatos, P. Siano, P. Wira, and F. Profumo, “Nonlinear h-infinity feedback control for asynchronous motors of electric trains,” Intelli- gent Industrial Systems , vol. 1, pp. 85–98, 2015
2015
-
[9]
Robust h-infinity control for connected vehicles in lattice hydrodynamic model at highway tunnel,
Y . Zhang, M. Zhao, D. Sun, X. Liu, S. Huang, and D. Chen, “Robust h-infinity control for connected vehicles in lattice hydrodynamic model at highway tunnel,” Physica A: Statistical Mechanics and its Applications, vol. 603, p. 127710, 2022
2022
-
[10]
Hassibi, A
B. Hassibi, A. H. Sayed, and T. Kailath, Indefinite-Quadratic Estima- tion and Control . SIAM, 1999
1999
-
[11]
Wasserstein distribution- ally robust regret-optimal control over infinite-horizon,
T. Kargin, J. Hajar, V . Malik, and B. Hassibi, “Wasserstein distribution- ally robust regret-optimal control over infinite-horizon,” in Proceed- ings of the 6th Annual Learning for Dynamics and Control Conference (A. Abate, M. Cannon, K. Margellos, and A. Papachristodoulou, e...
2024
-
[12]
Infinite-horizon distributionally robust regret-optimal control,
T. Kargin, J. Hajar, V . Malik, and B. Hassibi, “Infinite-horizon distributionally robust regret-optimal control,” in Proceedings of the 41st International Conference on Machine Learning (R. Salakhut- dinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berke...
2024
-
[13]
The distributionally robust infinite-horizon lqr,
J. Hajar, T. Kargin, V . Malik, and B. Hassibi, “The distributionally robust infinite-horizon lqr,” in2024 IEEE 63rd Conference on Decision and Control (CDC) , pp. 6539–6546, 2024
2024
-
[14]
Wasserstein distributionally robust regret-optimal control under partial observability,
J. Hajar, T. Kargin, and B. Hassibi, “Wasserstein distributionally robust regret-optimal control under partial observability,” in 2023 59th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 1–6, 2023
2023
-
[15]
Min-max feedback model predictive control for constrained linear systems,
P. O. M. Scokaert and D. Q. Mayne, “Min-max feedback model predictive control for constrained linear systems,” IEEE TAC, vol. 43, no. 8, pp. 1136–1142, 1998
1998
-
[16]
J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Control: Theory, Computation, and Design . Madison, WI: Nob Hill Publishing, 2 ed., 2017
2017
-
[17]
Kouvaritakis and M
B. Kouvaritakis and M. Cannon, Model Predictive Control: Classical, Robust and Stochastic . Advanced Textbooks in Control and Signal Processing, Cham: Springer, 2015
2015
-
[18]
Analytical solution of input con- strained reference tracking problems by dynamic programming,
J. B. Mare and J. A. De Don’a, “Analytical solution of input con- strained reference tracking problems by dynamic programming,” in Proceedings of the 44th IEEE Conference on Decision and Control , p. 1701–1706, IEEE, 2005
2005
-
[19]
A generalization of bellman’s equation with application to path planning, obstacle avoidance and invariant set estimation,
M. Jones and M. M. Peet, “A generalization of bellman’s equation with application to path planning, obstacle avoidance and invariant set estimation,” Automatica, vol. 127, p. 109510, 2021
2021
-
[20]
Explicit solution to bellman equation for positive systems with linear cost,
A. Rantzer, “Explicit solution to bellman equation for positive systems with linear cost,” in 2022 IEEE 61st CDC, p. 6154–6155, IEEE, 2022
2022
-
[22]
The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming,
L. M. Bregman, “The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming,” USSR Computational Mathematics and Mathematical Physics, vol. 7, no. 3, pp. 200–217, 1967
1967
-
[23]
Bas ¸ar and P
T. Bas ¸ar and P. Bernhard, H-infinity optimal control and related minimax design problems: a dynamic game approach . Springer Science & Business Media, 2008
2008
-
[24]
Clustering with bregman divergences,
A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh, “Clustering with bregman divergences,” Journal of Machine Learning Research, vol. 6, pp. 1705–1749, Oct 2005. APPENDIX A. Extended Proof of Theorem 2 Our aim is to extend the argument in Section IV-B—which currently assumesB...
2005
-
[25]
Convexity via Hessian Positivity: A function f is convex if ∇2f(x)≻ 0 ∀x
-
[26]
By imposing Assumption 1 as a constraint, convexity implies positivity of the functions
-
[27]
This crucial relation shows that locally the Hessian of f is the inverse of the Hessian of f∗
Fenchel Duality and the Inverse Function Theorem: For a convex function f and its Fenchel conjugate f∗, one obtains (via the inverse function theorem) ∇2f ∇f∗(ξ) ∇2f∗(ξ) =I. This crucial relation shows that locally the Hessian of f is the inverse of the Hessian of f∗. Combinin...
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.