pith. sign in

arxiv: 2605.02791 · v1 · submitted 2026-05-04 · 🧮 math.OC · cs.SY· eess.SY

Risk-Averse Ensemble Control for Control-Affine Systems

Alessandro Scagliotti , Thomas M. Surowiec This is my paper

Pith reviewed 2026-05-08 17:37 UTC · model grok-4.3

classification 🧮 math.OC cs.SYeess.SY
keywords risk-averse ensemble controlcontrol-affine systemsFréchet differentiabilityadjoint state of bounded variationoptimal controlquantum controlensemble control
0
0 comments X

The pith

Risk-averse ensemble control for control-affine systems yields continuously Fréchet differentiable control-to-state maps and first-order optimality conditions via bounded-variation adjoints.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Ensemble control requires a single deterministic input for dynamical systems driven by random parameters, as in quantum control or Neural ODE training. Standard formulations optimize expected performance and therefore tolerate poor behavior on low-probability realizations. The paper adopts control-affine dynamics to secure lower semi-continuity of the problem and then proves that the control-to-state operator is weakly-to-strongly continuous, continuously Fréchet differentiable, and that its derivative inherits the same continuity. These regularity properties produce primal and dual first-order necessary conditions expressed through an adjoint state of bounded variation. The same properties also guarantee the functional setting needed for convergence of infinite-dimensional optimization algorithms.

Core claim

We provide a rigorous characterization of the control-to-state mapping for risk-averse ensemble control of control-affine systems, establishing its weak-to-strong continuity, continuous Fréchet differentiability, and weak-to-strong continuity of the derivative operator. This regularity yields primal and dual first-order optimality conditions characterized by an adjoint state of bounded variation and fulfills the functional prerequisites required for the convergence of infinite dimensional optimization algorithms. The developments are illustrated by a numerical experiment in quantum control.

What carries the argument

The control-to-state mapping under control-affine structure, whose weak-to-strong continuity and continuous Fréchet differentiability carry the argument from existence to optimality conditions.

If this is right

  • Existence of minimizers follows directly from the lower semi-continuity induced by the affine structure.
  • Primal and dual first-order optimality conditions hold and are expressed through an adjoint state of bounded variation.
  • Infinite-dimensional optimization algorithms converge under the established differentiability and continuity properties.
  • The framework applies to quantum control instances where risk aversion penalizes outlier trajectories.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same differentiability could be used to derive gradient-based training rules for Neural ODEs that explicitly penalize high-risk trajectories.
  • The bounded-variation adjoint suggests that the optimality system remains well-posed under certain nonsmooth risk measures not treated in the paper.
  • Discretization schemes that preserve the weak-to-strong continuity property may be derived for reliable numerical solution of the optimality system.

Load-bearing premise

The system dynamics must be control-affine to guarantee the lower semi-continuity needed for existence of optimal solutions.

What would settle it

A concrete control-affine ensemble problem in which the numerically computed optimal control violates the stated first-order conditions involving the bounded-variation adjoint would falsify the regularity claims.

Figures

Figures reproduced from arXiv: 2605.02791 by Alessandro Scagliotti, Thomas M. Surowiec.

Figure 1
Figure 1. Figure 1: Comparison of control strategies. Left: final-state overlap |hψ tar | ψ θ u (T)i| as a function of the parameter θ ∈ Θ, for controls obtained by minimizing Iavg, Iworst, and IAVaR, together with the constructed control from [52]. Right: corresponding control inputs u ⋆ avg, u ⋆ worst, and u ⋆ AVaR over the time interval [0, T]. The AVaR-based control achieves the best uniform performance across the ensembl… view at source ↗
read the original abstract

A number of important modern applications in optimal control can be formulated as open loop control problems in which the underlying dynamical systems are subject to random inputs. These so-called ensemble control problems require the corresponding optimal control to be deterministic, as it must be computed before the realization of uncertainty and the passage of time. Practical applications of ensemble control include quantum control and the training of Neural ODEs. However, the standard approach to ensemble control treats the uncertainty in the objective function via the expectation, which provides optimal controls that only work well on average while ignoring critical outlier phenomena. This study provides a comprehensive mathematical treatment of risk-averse ensemble control. Within this setting, we adopt a control-affine structure that ensures the lower semi-continuity needed for proving the existence of optimal solutions. The central analytical contribution of this paper is a rigorous characterization of the control-to-state mapping in which we establish weak-to-strong continuity, continuous Fr\'echet differentiability, and weak-to-strong continuity of the derivative operator. Furthermore, this regularity yields primal and dual first-order optimality conditions characterized by an adjoint state of bounded variation, and it fulfills the functional prerequisites required for the convergence of infinite dimensional optimization algorithms. We conclude by validating these theoretical developments through a numerical experiment in quantum control.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops a risk-averse formulation of ensemble optimal control for control-affine systems. It proves existence of optimal controls via lower semicontinuity induced by the control-affine structure, establishes weak-to-strong continuity of the control-to-state map F, continuous Fréchet differentiability of F, and weak-to-strong continuity of the derivative DF, derives primal and dual first-order optimality conditions involving an adjoint state of bounded variation, and validates the theory with a quantum-control numerical example.

Significance. If the claimed regularity properties hold, the work supplies the functional-analytic prerequisites for risk-averse ensemble control, enabling reliable handling of tail risks beyond expectation-based objectives. The explicit variation equation available under control-affine dynamics and the resulting adjoint of bounded variation are particularly useful for algorithm convergence in infinite-dimensional settings and for applications such as quantum control.

major comments (2)
  1. [§4] §4 (Regularity of the control-to-state map), proof of weak-to-strong continuity of DF(u): the integral estimates passing from weak convergence of controls to strong convergence of the derivative states rely on Lipschitz constants of f and g that are uniform with respect to the ensemble parameter. When the underlying measure space has unbounded support, these constants need not be uniform, so the claimed continuity of DF may fail. Please state the precise domination or compactness assumption on the parameter measure that restores uniformity and verify that it is satisfied in the subsequent optimality analysis.
  2. [§5] §5 (First-order optimality conditions): the dual formulation invokes an adjoint state of bounded variation whose existence rests on the risk measure being convex and lower semicontinuous. The manuscript does not specify which concrete risk measure (e.g., CVaR at level α) is adopted nor how its subdifferential interacts with the bounded-variation adjoint; this step is load-bearing for the dual optimality system.
minor comments (2)
  1. [§2] The notation for the ensemble state x(t,ω) and its dependence on the control u is introduced inconsistently between the problem statement and the variation equation; a uniform subscript or functional notation would improve readability.
  2. [§6] The numerical quantum-control example reports convergence but omits the discretization scheme for the ensemble measure and the precise risk parameter; adding these details would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. The points raised identify opportunities to strengthen the presentation of assumptions and the specialization of the risk measure. We address each major comment below.

read point-by-point responses

  1. Referee: [§4] §4 (Regularity of the control-to-state map), proof of weak-to-strong continuity of DF(u): the integral estimates passing from weak convergence of controls to strong convergence of the derivative states rely on Lipschitz constants of f and g that are uniform with respect to the ensemble parameter. When the underlying measure space has unbounded support, these constants need not be uniform, so the claimed continuity of DF may fail. Please state the precise domination or compactness assumption on the parameter measure that restores uniformity and verify that it is satisfied in the subsequent optimality analysis.

    Authors: We agree that uniformity of the Lipschitz constants of f and g requires an explicit assumption when the support of the ensemble measure μ may be unbounded. We will add a standing compactness assumption on the support of μ in Section 2 (Problem formulation), which guarantees the required uniformity for the integral estimates in the proof of weak-to-strong continuity of DF (Theorem 4.3). We will also verify that this assumption is satisfied by the quantum-control example in Section 6 and does not alter the subsequent optimality analysis in §5. revision: yes

  2. Referee: [§5] §5 (First-order optimality conditions): the dual formulation invokes an adjoint state of bounded variation whose existence rests on the risk measure being convex and lower semicontinuous. The manuscript does not specify which concrete risk measure (e.g., CVaR at level α) is adopted nor how its subdifferential interacts with the bounded-variation adjoint; this step is load-bearing for the dual optimality system.

    Authors: Our development is intentionally stated for a general convex lower-semicontinuous risk measure ρ to preserve broad applicability. The BV adjoint arises from the chain rule applied to the composition of ρ with the control-to-state map. To address the request for concreteness, we will add a remark in §5 that specializes the result to CVaR_α, where the subdifferential reduces to an expectation over the α-tail set; we will show that this specialization preserves the bounded-variation property of the adjoint without changing the general dual system. revision: partial

Circularity Check

0 steps flagged

No circularity: derivation uses standard functional-analytic arguments on control-affine dynamics

full rationale

The paper establishes weak-to-strong continuity, Fréchet differentiability, and optimality conditions for the control-to-state map of an ensemble control-affine system. These properties follow from the explicit linear dependence on the control in the dynamics, combined with standard estimates in Bochner spaces and adjoint equations of bounded variation. No step reduces a claimed result to a fitted parameter, self-definition, or self-citation chain; the control-affine structure is an explicit modeling assumption that supplies the needed lower semi-continuity, not a hidden tautology. The derivation is therefore self-contained against external functional-analysis benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper relies on the control-affine structure as a domain assumption to obtain lower semi-continuity and existence; standard results from functional analysis and optimal control are invoked without new free parameters or invented entities.

axioms (1)
  • domain assumption The underlying dynamical system is control-affine
    Invoked to guarantee lower semi-continuity of the objective and existence of optimal controls.

pith-pipeline@v0.9.0 · 5525 in / 1297 out tokens · 46943 ms · 2026-05-08T17:37:00.862678+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages · 1 internal anchor

  1. [1]

    Agrachev and C

    A. Agrachev and C. Letrouit. Generic controllability of equivariant systems and applications to particle systems and neural networks. Annales de l’Institut Henri Poincaré C, Analyse non linéaire, 43(3):639–668, 2025

    work page 2025

  2. [2]

    Agrachev and A

    A. Agrachev and A. Sarychev. Control in the spaces of ensembles of points. SIAM J. Control Optim., 58:1579–1596, 2020

    work page 2020

  3. [3]

    Agrachev and A

    A. Agrachev and A. Sarychev. Control on the manifolds of mappings with a view to the Deep Learning. J. Dyn. Control Syst. , 28:989–1008, 2022

    work page 2022

  4. [4]

    Ambrosetti and G

    A. Ambrosetti and G. Prodi. A primer of nonlinear analysis . Number 34. Cambridge University Press, 1995. 35

    work page 1995

  5. [5]

    M. S. Aronna, G. de Lima Monteiro, and O. Sierra Fonseca. A verage optimal control of uncertain control-affine systems. Set-Valued and Variational Analysis , 33(4), 2025

    work page 2025

  6. [6]

    M. S. Aronna, M. Palladino, and O. Sierra Fonseca. Dynamic programming principle and Hamilton–Jacobi–Bellman equation for optimal control problems with uncertainty. ESAIM: Control, Optimisation and Calculus of Variations , 32:5, 2026

    work page 2026

  7. [7]

    Artzner, F

    P. Artzner, F. Delbaen, J. Eber, and D. Heath. Coherent measures of risk. Mathematical Finance, 9(3):203–228, July 1999

    work page 1999

  8. [8]

    Augier, U

    N. Augier, U. Boscain, and M. Sigalotti. Adiabatic ensemble control of a continuum of quantum systems. SIAM Journal on Control and Optimization , 56(6):4045–4068, Jan. 2018

    work page 2018

  9. [9]

    Bartsch, A

    J. Bartsch, A. Borzì, F. Fanelli, and S. Roy. A theoretical investigation of brockett’s ensemble optimal control problems. Calculus of Variations and Partial Differential Equations , 58(5), 2019

    work page 2019

  10. [10]

    Bartsch, A

    J. Bartsch, A. Borzì, F. Fanelli, and S. Roy. A numerical investigation of brockett’s ensemble optimal control problems. Numerische Mathematik , 149(1):1–42, 2021

    work page 2021

  11. [11]

    Beauchard, J.-M

    K. Beauchard, J.-M. Coron, and P. Rouchon. Controllability issues for continuous-spectrum systems and ensemble controllability of Bloch equations. Comm. Math. Phys. , 296(2):525– 557, 2010

    work page 2010

  12. [12]

    Beauchard and E

    K. Beauchard and E. Pozzoli. Small-time approximate controllability of bilinear Schrödinger equations and diffeomorphisms. Annales de l’Institut Henri Poincaré C, Analyse non linéaire , 2025

    work page 2025

  13. [13]

    Belhadj, J

    M. Belhadj, J. Salomon, and G. Turinici. Ensemble controllability and discrimination of per- turbed bilinear control systems on connected, simple, compact Lie groups. Eur. J. Control , 22:23–29, 2015

    work page 2015

  14. [14]

    Bettiol and N

    P. Bettiol and N. Khalil. Necessary optimality conditions for average cost minimization problems. Discete Contin. Dyn. Syst. - B , 24(5):2093–2124, 2019

    work page 2093

  15. [15]

    Bettiol and N

    P. Bettiol and N. Khalil. A verage cost minimization problems subject to state constraints. SIAM J. Control Optim. , 62(3):1884–1907, 2024

    work page 1907

  16. [16]

    Bonalli and B

    R. Bonalli and B. Bonnet. First-order Pontryagin Maximum Principle for risk-averse stochas- tic optimal control problems. SIAM Journal on Control and Optimization , 61(3):1881–1909, 2023

    work page 1909

  17. [17]

    Bonalli, B

    R. Bonalli, B. Bonnet-Weill, and L. Pfeiffer. A characterization of law-invariant and coherent risk measures through optimal transport. arXiv preprint arXiv:2512.19157 , 2025. 36

    work page arXiv 2025

  18. [18]

    J. F. Bonnans and A. Shapiro. Perturbation Analysis of Optimization Problems . Springer New York, 2000

    work page 2000

  19. [19]

    Bressan and B

    A. Bressan and B. Piccoli. Introduction to the mathematical theory of control , volume 1. American Institute of Mathematical Sciences, Springfield, 2007

    work page 2007

  20. [20]

    Carter and B

    M. Carter and B. van Brunt. The Lebesgue-Stieltjes Integral . Springer New York, 2000

    work page 2000

  21. [21]

    R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud. Neural ordinary differen- tial equations. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, pages 6571–6583, 2018

    work page 2018

  22. [22]

    F. C. Chittaro and J. P. Gauthier. Asymptotic ensemble stabilizability of the Bloch equation. Sys. Control Lett. , 113:36–44, 2018

    work page 2018

  23. [23]

    G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals and Systems , 2(4):303–314, 1989

    work page 1989

  24. [24]

    Dal Maso

    G. Dal Maso. An Introduction to Γ-convergence. Birkhäuser, Boston MA, 1993

    work page 1993

  25. [25]

    Danhane, J

    B. Danhane, J. Lohéac, and M. Jungers. Conditions for uniform ensemble output control- lability, and obstruction to uniform ensemble controllability. Math. Control Rel. Fields , 14(3):1128–1175, 2024

    work page 2024

  26. [26]

    Dirr and M

    G. Dirr and M. Schönlein. Uniform and lq-ensemble reachability of parameter-dependent linear systems. J. Differ. Eq. , 283:216–262, 2021

    work page 2021

  27. [27]

    Föllmer and A

    H. Föllmer and A. Schied. Convex measures of risk and trading constraints. Finance and Stochastics, 6(4):429–447, Oct. 2002

    work page 2002

  28. [28]

    Fonseca and G

    I. Fonseca and G. Leoni. Modern Methods in the Calculus of Variations: Lp spaces. Springer, New York NY, 2007

    work page 2007

  29. [29]

    Goldberg, W

    H. Goldberg, W. Kampowsky, and F. Tröltzsch. On Nemytskij operators in Lp-spaces of abstract functions. Mathematische Nachrichten , 155(1):127–140, 1992

    work page 1992

  30. [30]

    Haber and L

    E. Haber and L. Ruthotto. Stable architectures for deep neural networks. Inverse Problems, 34(1):014004, Dec. 2017

    work page 2017

  31. [31]

    J. Hale. Ordinary Differential Equations . Krieger Publishing Company, 1980

    work page 1980

  32. [32]

    Hille and R

    E. Hille and R. S. Phillips. Functional analysis and semi-groups , volume 31. American Mathematical Soc., 1996

    work page 1996

  33. [33]

    Hofmann and A

    S. Hofmann and A. Borzì. The Pontryagin Maximum Principle for Training Convolutional Neural Networks. SIAM Journal on Mathematics of Data Science , 7(4):1616–1642, 2025. 37

    work page 2025

  34. [34]

    K. Hornik. Approximation capabilities of multilayer feedforward networks. Neural Networks, 4(2):251–257, 1991

    work page 1991

  35. [35]

    Imaz and Z

    C. Imaz and Z. Vorel. Generalized ordinary differential equations in Banach space and applications to functional equations. Boletín de la Sociedad Matemática Mexicana , 11:47– 59, 1966

    work page 1966

  36. [36]

    D. P. Kouri and A. Shapiro. Optimization of PDEs with Uncertain Inputs , page 41–81. Springer New York, 2018

    work page 2018

  37. [37]

    D. P. Kouri and T. M. Surowiec. Risk-averse PDE-constrained optimization using the con- ditional value-at-risk. SIAM Journal on Optimization , 26(1):365–396, 2016

    work page 2016

  38. [38]

    D. P. Kouri and T. M. Surowiec. Existence and optimality conditions for risk-averse PDE- constrained optimization. SIAM/ASA Journal on Uncertainty Quantification , 6(2):787–815, 2018

    work page 2018

  39. [39]

    D. P. Kouri and T. M. Surowiec. Epi-regularization of risk measures. Mathematics of Operations Research, 45(2):774–795, May 2020

    work page 2020

  40. [40]

    D. P. Kouri and T. M. Surowiec. A primal–dual algorithm for risk minimization. Mathe- matical Programming, 193(1):337–363, Feb. 2021

    work page 2021

  41. [41]

    existence and optimality conditions for risk- averse PDE-constrained optimization

    D. P. Kouri and T. M. Surowiec. Corrigendum: “existence and optimality conditions for risk- averse PDE-constrained optimization” . SIAM/ASA Journal on Uncertainty Quantification , 10(3):1321–1322, Sept. 2022

    work page 2022

  42. [42]

    Li and N

    J.-S. Li and N. Khaneja. Control of inhomogeneous quantum ensembles. Phys. Rev. A , 73(3), 2006

    work page 2006

  43. [43]

    Q. Li, L. Chen, C. Tai, and W. E. Maximum principle based algorithms for deep learning. Journal of Machine Learning Research , 18(165):1–29, 2018

    work page 2018

  44. [44]

    Liang, U

    R. Liang, U. Boscain, and M. Sigalotti. Ensemble control of n-level quantum systems with a scalar control. Automatica, 185:112733, 2026

    work page 2026

  45. [45]

    Lohéac and E

    J. Lohéac and E. Zuazua. From averaged to simultaneous controllability of parameter de- pendent finite-dimensional systems. Annales de la Faculté des Sciences de Toulouse: Math- ématiques, 25(4):785–828, 2016

    work page 2016

  46. [46]

    Melnikov and J

    O. Melnikov and J. Milz. Convergence rates for ensemble-based solutions to optimal control of uncertain dynamical systems. arXiv preprint arXiv:2407.18182 , 2024

    work page internal anchor Pith review arXiv 2024

  47. [47]

    Milz and T

    J. Milz and T. M. Surowiec. Asymptotic consistency for nonconvex risk-averse stochastic optimization with infinite-dimensional decision spaces. Mathematics of Operations Research, 49(3):1403––1418, 2024. 38

    work page 2024

  48. [48]

    D. O’Regan. Existence theory for nonlinear ordinary differential equations , volume 398. Springer Science & Business Media, 1997

    work page 1997

  49. [49]

    Palladino, A

    M. Palladino, A. Pesare, and T. Scarinci. Convergence results for control problems with unknown dynamic and applications to reinforcement learning. Mathematical Control and Related Fields, 17(0):99–119, 2026

    work page 2026

  50. [50]

    G. C. Pflug and W. Römisch. Modeling, Measuring and Managing Risk . WORLD SCIEN- TIFIC, Aug. 2007

    work page 2007

  51. [51]

    Pozzoli and A

    E. Pozzoli and A. Scagliotti. Approximation of diffeomorphisms for quantum state transfers. IEEE Control Systems Letters , 9:571–576, 2025

    work page 2025

  52. [52]

    Robin, N

    R. Robin, N. Augier, U. Boscain, and M. Sigalotti. Ensemble qubit controllability with a single control via adiabatic and rotating wave approximations. J. Diff. Equ. , 318:414–442, 2022

    work page 2022

  53. [53]

    R. T. Rockafellar and S. Uryasev. The fundamental risk quadrangle in risk management, optimization and statistical estimation. Surveys in Operations Research and Management Science, 18(1–2):33–53, Oct. 2013

    work page 2013

  54. [54]

    Ruiz-Balet and E

    D. Ruiz-Balet and E. Zuazua. Neural ODE control for classification, approximation and transport. SIAM Review , 65(3):735–773, 2023

    work page 2023

  55. [55]

    Ruths and J.-S

    J. Ruths and J.-S. Li. Optimal control of inhomogenous ensembles. IEEE Trans. Aut. Control, 57(8):2021–2032, 2012

    work page 2021

  56. [56]

    Scagliotti

    A. Scagliotti. A gradient flow equation for optimal control problems with end-point cost. J. Dyn. Control Syst. , 29(2):521–568, 2023

    work page 2023

  57. [57]

    Scagliotti

    A. Scagliotti. Optimal control of ensembles of dynamical systems. ESAIM: Control Optim Calc. Var., 29, 2023

    work page 2023

  58. [58]

    Scagliotti

    A. Scagliotti. Minimax problems for ensembles of control-affine systems. SIAM J. Control Optim., 63(1):502–523, 2025

    work page 2025

  59. [59]

    Schönlein

    M. Schönlein. Polynomial methods to construct inputs for uniformly ensemble reachable linear systems. Math. Control Signals Syst. , 36:251–296, 2024

    work page 2024

  60. [60]

    Shapiro, D

    A. Shapiro, D. Dentcheva, and A. Ruszczynski. Lectures on Stochastic Programming: Mod- eling and Theory, Third Edition . Society for Industrial and Applied Mathematics, July 2021. 39

    work page 2021