arxiv: 2604.10588 · v1 · submitted 2026-04-12 · 💻 cs.LG · cs.SY· eess.SY

Recognition: 2 theorem links

· Lean Theorem

Distributionally Robust PAC-Bayesian Control

Domagoj Herceg , Duarte Antunes

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:11 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SY

keywords PAC-Bayesdistributionally robust optimizationSystem Level Synthesiscontrol theorygeneralization boundssim-to-real gapsafety certificateslinear time-invariant systems

0 comments

The pith

PAC-Bayesian control derives safety certificates that account for sim-to-real distribution shifts by bounding loss via closed-loop operator norms.

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

The paper builds a framework that certifies the performance of learning-based finite-horizon controllers even when losses are unbounded and the deployment environment differs from the training data. It merges PAC-Bayesian generalization bounds with distributionally robust optimization that uses the type-1 Wasserstein distance to model shifts. The key step is to reparametrize the controller with System Level Synthesis so that a sub-Gaussian proxy for the loss and an explicit bound on shift-induced degradation both become functions of the operator norm of the closed-loop map. For linear time-invariant systems this reparametrization turns the certification task into a tractable optimization problem that returns high-probability safety guarantees usable in real-world settings.

Core claim

By employing the System Level Synthesis reparametrization, the authors obtain a sub-Gaussian loss proxy together with a performance-loss bound under type-1 Wasserstein distribution shifts; both quantities are controlled directly by the operator norm of the closed-loop map. For linear time-invariant plants the resulting program is computationally tractable and supplies PAC-Bayesian certificates that remain valid when the real environment differs from the training distribution.

What carries the argument

The System Level Synthesis (SLS) reparametrization, which rewrites the controller so that the closed-loop maps are explicit decision variables whose operator norms directly govern both the sub-Gaussian loss proxy and the shift bound.

If this is right

High-probability safety certificates become available for controllers deployed in environments that differ from training data.
The optimization problem remains tractable for finite-horizon linear time-invariant systems.
Unbounded losses are handled without requiring artificial clipping or boundedness assumptions.

Where Pith is reading between the lines

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

Minimizing the closed-loop operator norm during design would simultaneously tighten both the generalization gap and the shift sensitivity.
The same certificates could be used to decide when a simulation-trained policy is safe enough for physical deployment.
Similar norm-based bounds might be derivable for other controller parametrizations beyond SLS.

Load-bearing premise

The loss function admits a sub-Gaussian proxy once the controller is written in SLS form and every possible distribution shift stays inside a type-1 Wasserstein ball whose radius is known in advance.

What would settle it

A numerical test on an LTI system in which the measured performance degradation after a shift inside the assumed Wasserstein ball exceeds the bound predicted from the closed-loop operator norm.

Figures

Figures reproduced from arXiv: 2604.10588 by Domagoj Herceg, Duarte Antunes.

**Figure 2.** Figure 2: shows the effectiveness of our method in the presence of distribution shifts. We can observe that vanilla PAC-Bayes, which stands for the PAC-Bayes term without the Wasserstein part, cannot account for environmental shifts properly. It can clearly be seen from the figure that the reported bound is violated for all sizes of the dataset n. On the other hand, our robustified method provides a correct upper-bo… view at source ↗

read the original abstract

We present a distributionally robust PAC-Bayesian framework for certifying the performance of learning-based finite-horizon controllers. While existing PAC-Bayes control literature typically assumes bounded losses and matching training and deployment distributions, we explicitly address unbounded losses and environmental distribution shifts (the sim-to-real gap). We achieve this by drawing on two modern lines of research, namely the PAC-Bayes generalization theory and distributionally robust optimization via the type-1 Wasserstein distance. By leveraging the System Level Synthesis (SLS) reparametrization, we derive a sub-Gaussian loss proxy and a bound on the performance loss due to distribution shift. Both are tied directly to the operator norm of the closed-loop map. For linear time-invariant systems, this yields a computationally tractable optimization-based framework together with high-probability safety certificates for deployment in real-world environments that differ from those used in training.

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 distributionally robust PAC-Bayesian framework for finite-horizon LTI control that certifies performance under unbounded losses and Wasserstein-bounded distribution shifts (sim-to-real gap). Using the System Level Synthesis (SLS) reparametrization, it derives a sub-Gaussian loss proxy and a performance-loss bound, both expressed in terms of the closed-loop operator norm; the resulting optimization yields high-probability safety certificates that are computationally tractable for linear systems.

Significance. If the central derivations hold, the work supplies a concrete route to high-probability certificates for learning-based controllers that remain valid under both unbounded costs and modest distribution shift, by combining PAC-Bayes generalization with type-1 Wasserstein DRO and exploiting the SLS parametrization to obtain explicit dependence on the closed-loop map. This is a non-trivial technical contribution to safe sim-to-real control.

major comments (2)

[§4.2] §4.2 (sub-Gaussian loss proxy derivation): the claim that a sub-Gaussian proxy exists for the (unbounded) loss under the SLS closed-loop map is load-bearing for both the PAC-Bayes term and the Wasserstein performance-loss bound. For quadratic stage costs the loss random variable is sub-exponential rather than sub-Gaussian; the manuscript must explicitly construct a dominating sub-Gaussian proxy that remains uniform in the controller parameters (i.e., in the operator norm) and verify that the resulting constants do not explode with horizon length or noise variance.
[§5] §5 (Wasserstein DRO bound): the performance-loss bound is stated to hold inside a type-1 Wasserstein ball whose radius is chosen a priori. The paper should clarify whether this radius can be bounded from data or must be treated as a free hyper-parameter; if the latter, the high-probability safety certificate is conditional on an unverifiable modeling assumption and the practical utility of the certificate is reduced.

minor comments (2)

Notation for the closed-loop map Φ and its operator norm should be introduced once and used consistently; several early equations reuse Φ for both the full map and its blocks.
The finite-horizon assumption is used throughout; a brief remark on whether the same proxy construction extends to infinite-horizon or discounted settings would help readers assess generality.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough and constructive review. We address each major comment below and indicate the corresponding revisions we will make to the manuscript.

read point-by-point responses

Referee: [§4.2] §4.2 (sub-Gaussian loss proxy derivation): the claim that a sub-Gaussian proxy exists for the (unbounded) loss under the SLS closed-loop map is load-bearing for both the PAC-Bayes term and the Wasserstein performance-loss bound. For quadratic stage costs the loss random variable is sub-exponential rather than sub-Gaussian; the manuscript must explicitly construct a dominating sub-Gaussian proxy that remains uniform in the controller parameters (i.e., in the operator norm) and verify that the resulting constants do not explode with horizon length or noise variance.

Authors: We appreciate this observation on the tail behavior. The derivation in §4.2 bounds the loss moments via the closed-loop operator norm induced by the SLS parametrization. To strengthen the argument, we will add an explicit construction of a dominating sub-Gaussian proxy (via a suitable exponential-moment bound that majorizes the sub-exponential tail) that is uniform over all controllers whose closed-loop operator norm is bounded by a fixed constant. We will also include a short lemma showing that the resulting variance proxy scales at most linearly with horizon length and noise variance under the standard LTI stabilizability assumptions used in the paper; the constants therefore remain controlled and do not explode. These additions will be placed in the revised §4.2 and the associated appendix. revision: yes
Referee: [§5] §5 (Wasserstein DRO bound): the performance-loss bound is stated to hold inside a type-1 Wasserstein ball whose radius is chosen a priori. The paper should clarify whether this radius can be bounded from data or must be treated as a free hyper-parameter; if the latter, the high-probability safety certificate is conditional on an unverifiable modeling assumption and the practical utility of the certificate is reduced.

Authors: We agree that the interpretation of the certificate depends on the choice of radius. In the current manuscript the radius is introduced as a modeling parameter that encodes the anticipated sim-to-real gap. In the revision we will explicitly state in §5 that the high-probability performance-loss bound holds conditionally on the true deployment distribution lying inside the chosen Wasserstein ball. We will also add a short discussion on practical selection of the radius, including (i) conservative a-priori bounds derived from system-identification error and (ii) data-driven estimates obtained from limited real-world rollouts via empirical Wasserstein distances. This clarifies the conditional nature of the certificate while preserving its utility as a safety certificate under a quantifiable modeling assumption. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper's central claims rest on leveraging the existing System Level Synthesis (SLS) reparametrization to derive a sub-Gaussian loss proxy and Wasserstein DRO performance-loss bound, both expressed in terms of the closed-loop operator norm. These steps draw on standard PAC-Bayes generalization theory and type-1 Wasserstein DRO results from the literature rather than on any fitted parameters, self-defined quantities, or self-citation chains internal to the present work. No equation or step reduces by construction to its own inputs; the sub-Gaussian proxy is presented as a derived object under the SLS parametrization for LTI systems, not as a tautological renaming or fit. The framework therefore remains self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The framework rests on standard properties of Wasserstein distances, sub-Gaussian concentration, and the SLS parametrization; no new entities are postulated.

free parameters (2)

Wasserstein ball radius
Models the size of the allowed distribution shift; its value must be chosen or bounded externally.
PAC-Bayes confidence parameter delta
Standard parameter controlling the probability of the certificate.

axioms (2)

domain assumption The loss function admits a sub-Gaussian tail bound under the SLS closed-loop parametrization
Invoked to obtain the loss proxy used in the PAC-Bayes bound.
domain assumption Distribution shifts can be captured by a type-1 Wasserstein ball of finite radius
Central modeling choice for the robust optimization component.

pith-pipeline@v0.9.0 · 5448 in / 1483 out tokens · 57242 ms · 2026-05-10T16:11:45.691182+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By leveraging the System Level Synthesis (SLS) reparametrization, we derive a sub-Gaussian loss proxy and a bound on the performance loss due to distribution shift. Both are tied directly to the operator norm of the closed-loop map.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_strictMono_of_one_lt unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the centered loss ℓ(θ,w)−R(θ) is σ(θ)-sub-Gaussian with σ(θ):=||M(θ)Σ_w^{1/2}||_op

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages

[1]

Catoni,PAC-Bayesian Supervised Classification: The Thermody- namics of Statistical Learning

O. Catoni,PAC-Bayesian Supervised Classification: The Thermody- namics of Statistical Learning. Institute of Mathematical Statistics, 2007

2007
[2]

Computing nonvacuous generaliza- tion bounds for deep (stochastic) neural networks with many more parameters than training data,

G. K. Dziugaite and D. M. Roy, “Computing nonvacuous generaliza- tion bounds for deep (stochastic) neural networks with many more parameters than training data,” inProceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, UAI 2017, Sydney, Australia, August 11-15, 2017. AUAI Press, 2017

2017
[3]

Pac-bayes control: Synthesizing controllers that provably generalize to novel environments,

A. Majumdar and M. Goldstein, “Pac-bayes control: Synthesizing controllers that provably generalize to novel environments,” inPro- ceedings of The 2nd Conference on Robot Learning, ser. Proceedings of Machine Learning Research, A. Billard, A. Dragan, J. Peters, and J. Morimoto, Eds., vol. 87. PMLR, 29–31 Oct 2018, pp. 293–305

2018
[4]

A pac-bayesian framework for optimal control with stability guarantees,

M. G. Boroujeni, C. L. Galimberti, A. Krause, and G. Ferrari- Trecate, “A pac-bayesian framework for optimal control with stability guarantees,” in2024 IEEE 63rd Conference on Decision and Control (CDC), 2024, pp. 8237–8244

2024
[5]

Pac-bayesian optimal control with stability and generalization guarantees,

——, “Pac-bayesian optimal control with stability and generalization guarantees,”arXiv preprint arXiv:2512.02858, 2025

work page arXiv 2025
[6]

Shapiro, D

A. Shapiro, D. Dentcheva, and A. Ruszczynski,Lectures on stochastic programming: modeling and theory. SIAM, 2021

2021
[7]

Distri- butionally robust control of constrained stochastic systems,

B. P. Van Parys, D. Kuhn, P. J. Goulart, and M. Morari, “Distri- butionally robust control of constrained stochastic systems,”IEEE Transactions on Automatic Control, vol. 61, no. 2, pp. 430–442, 2015

2015
[8]

Risk-averse model predictive control,

P. Sopasakis, D. Herceg, A. Bemporad, and P. Patrinos, “Risk-averse model predictive control,”Automatica, vol. 100, pp. 281–288, 2019

2019
[9]

The optimizer’s curse: Skepticism and postdecision surprise in decision analysis,

J. E. Smith and R. L. Winkler, “The optimizer’s curse: Skepticism and postdecision surprise in decision analysis,”Management Science, vol. 52, no. 3, pp. 311–322, 2006

2006
[10]

Frameworks and Results in Dis- tributionally Robust Optimization,

H. Rahimian and S. Mehrotra, “Frameworks and Results in Dis- tributionally Robust Optimization,”Open Journal of Mathematical Optimization, vol. 3, pp. 1–85, 2022

2022
[11]

Data-driven distributionally robust optimization using the wasserstein metric: Performance guar- antees and tractable reformulations,

P. Mohajerin Esfahani and D. Kuhn, “Data-driven distributionally robust optimization using the wasserstein metric: Performance guar- antees and tractable reformulations,”Mathematical Programming, vol. 171, no. 1, pp. 115–166, 2018

2018
[12]

Villaniet al.,Optimal transport: old and new

C. Villaniet al.,Optimal transport: old and new. Springer, 2009, vol. 338

2009
[13]

A system-level approach to controller synthesis,

Y .-S. Wang, N. Matni, and J. C. Doyle, “A system-level approach to controller synthesis,”IEEE Transactions on Automatic Control, vol. 64, no. 10, pp. 4079–4093, 2019

2019
[14]

Pac- bayes-chernoff bounds for unbounded losses,

I. Casado, L. A. Ortega, A. P ´erez, and A. R. Masegosa, “Pac- bayes-chernoff bounds for unbounded losses,”Advances in Neural Information Processing Systems, vol. 37, pp. 24 350–24 374, 2024

2024
[15]

Boucheron, G

S. Boucheron, G. Lugosi, and P. Massart,Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford: Oxford University Press, 2013

2013
[16]

Wasserstein distributionally robust optimization: Theory and appli- cations in machine learning,

D. Kuhn, P. M. Esfahani, V . A. Nguyen, and S. Shafieezadeh-Abadeh, “Wasserstein distributionally robust optimization: Theory and appli- cations in machine learning,” inOperations research & management science in the age of analytics. Informs, 2019, pp. 130–166

2019
[17]

User-friendly introduction to

P. Alquier, “User-friendly introduction to pac-bayes bounds,”arXiv preprint arXiv:2110.11216, 2021

work page arXiv 2021
[18]

G. H. Golub and C. F. Van Loan,Matrix computations. JHU press, 2013

2013
[19]

Vershynin,High-dimensional probability: An introduction with applications in data science

R. Vershynin,High-dimensional probability: An introduction with applications in data science. Cambridge university press, 2018, vol. 47

2018
[20]

Bayesian learning via stochastic gradient langevin dynamics,

M. Welling and Y . W. Teh, “Bayesian learning via stochastic gradient langevin dynamics,” inProceedings of the 28th international confer- ence on machine learning (ICML-11), 2011, pp. 681–688

2011
[21]

Stein variational gradient descent: A general purpose bayesian inference algorithm,

Q. Liu and D. Wang, “Stein variational gradient descent: A general purpose bayesian inference algorithm,”Advances in neural informa- tion processing systems, vol. 29, 2016

2016
[22]

Bezanson, A

J. Bezanson, A. Edelman, S. Karpinski, and V . B. Shah, “Julia: A fresh approach to numerical computing,”SIAM review, vol. 59, no. 1, pp. 65–98, 2017. [Online]. Available: https://doi.org/10.1137/141000671

work page doi:10.1137/141000671 2017
[23]

Don’t unroll adjoint: Dif- ferentiating SSA-Form programs

M. Innes, “Don’t unroll adjoint: Differentiating ssa-form programs,”CoRR, vol. abs/1810.07951, 2018. [Online]. Available: http://arxiv.org/abs/1810.07951

work page arXiv 2018
[24]

Jump: A modeling language for mathematical optimization,

I. Dunning, J. Huchette, and M. Lubin, “Jump: A modeling language for mathematical optimization,”SIAM Review, vol. 59, no. 2, pp. 295– 320, 2017

2017