Smooth globally PLI functions are nonlinear least-squares, and so are their gradient-dominated cousins

Eduardo D. Sontag

Smooth globally PLI functions are nonlinear least-squares, and so are their gradient-dominated cousins

T0 review · 0 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Under the semiglobal Polyak–Łojasiewicz inequality, smooth loss landscapes are still nonlinear least-squares functions.

desk verdict Sontag gives a careful, mostly clean extension of BCR's nonlinear least-squares normal form to semiglobal PŁ; the reparametrization equivalence is the real novelty, and the applications are honest but modest. read the letter →

arxiv 2608.08849 v2 pith:F5RDCOSF submitted 2026-08-09 eess.SY cs.SYmath.DGmath.DS

Eduardo D. Sontag This is my paper · ORCID

classification eess.SYcs.SYmath.DGmath.DS MSC 90C2637C1049N1093D30

keywords semiglobalPolyak–Łojasiewiczinequalitygradientdominationnonlinearleast-squaresnormalformdesingularizercomparisonfunctioncontinuous-timeLQRlogisticregressionRiemannianoptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper establishes that the global Polyak–Łojasiewicz inequality can be relaxed to its 'semiglobal' version without losing any of the structural conclusions of the recent normal-form theorem for smooth optimization landscapes. The semiglobal condition only asks that the gradient norm be bounded below by a positive-definite function of the excess loss, with that function growing at least like the square root near zero. Under that weaker hypothesis, a smooth function on a contractible complete Riemannian manifold is still a nonlinear least-squares function: there is a diffeomorphism $\psi=(\pi,\phi)$ onto $S \times \mathbb{R}^k$ with $f = f^* + \|\phi\|^2$. The paper shows that the two motivating cases that fail the global inequality — continuous-time LQR policy optimization and logistic regression — satisfy the semiglobal one, so their loss landscapes are globally diffeomorphic to paraboloids. A reader should care because this separates structural shape conclusions from rate and robustness conclusions: landscape geometry survives the weakening, while global quadratic growth does not.

What carries the argument

The load-bearing object is the desingularizer $\Psi(h) = \int_0^h \frac{ds}{\alpha(s)}$, built from a positive-definite comparison function $\alpha$ witnessing the gradient lower bound $\|\nabla f\| \ge \alpha(f-f^*)$. Condition (A), namely $\alpha(s)^2 \ge 2\mu s$ for small $s$, makes $\Psi$ finite and of order $\sqrt{h}$ at the origin; finite $\Psi$ forces every negative-gradient trajectory to reach the minimizer set in finite length, and the $\sqrt{h}$ behavior gives the local PŁ inequality from which the Morse–Bott structure of the minimizer set follows. The reparametrization route sharpens the witness to $c\sqrt{s}$ near zero and sets $\theta = (c^2/4)\Psi^2$, so that $g = \theta(f-f^*)$ is globally PŁ; applying the known theorem to $g$ and pulling back through a radial diffeomorphism transfers the normal form $f = f^* + \|\phi\|^2$ back to $f$. All structural conclusions flow through these two objects; the comparison-function class conditions at infinity, which govern robustness, never enter.

What would settle it

Look for a smooth function on $\mathbb{R}^n$ that attains its minimum, is unbounded above, has a unique nondegenerate minimizer (so local PŁ holds), and has gradient norm bounded away from zero on every level set, but for which no diffeomorphism $\phi$ exists with $f = f^* + \|\phi\|^2$; the paper's open case, where the sharp witness $\alpha_f$ is positive on every level yet its infimum over some band is zero, is the natural candidate. Producing such a function would show the continuous-witness hypothesis is essential; proving the normal form for all such functions would show that hypothesis can be weakened to a pointwise condition.

Watch

Extended reading notes

Core claim

The central claim is that, under the hypothesis $\|\nabla f(x)\| \ge \alpha(f(x)-f^*)$ with $\alpha$ positive definite and satisfying $\alpha(s)^2 \ge 2\mu s$ near $0$, every structural result of the global PŁ normal-form theory holds verbatim; equivalently the same is true under semiglobal PŁ ($sgl$-PŁI). In particular, when $M$ is contractible there is a diffeomorphism $\psi=(\pi,\phi)\colon M \to S \times \mathbb{R}^k$ with $f = f^* + \|\phi\|^2$ and $\phi$ a submersion, so the function is a nonlinear least-squares objective in new coordinates. The proof works by showing that the desingularizer $\Psi(h)=\int_0^h ds/\alpha(s)$ is finite, which bounds gradient-flow trajectories and yields the Morse–Bott structure of the minimizer set; a second, shorter proof reparametrizes the loss through $\theta(f-f^*)$ to manufacture a globally PŁ companion and imports the existing theorem verbatim. The paper also proves sharpness: dropping either the square-root behavior at the origin or the uniform positivity on every level set destroys the conclusions in explicit examples, and it records exactly what survives — the normal form and fiber-bundle structure — and what does not — global quadratic growth and quantitative control of the diffeomorphism away from the minimizer set.

Load-bearing premise

The whole argument rests on one premise: a single continuous lower-bound function $\alpha(s)$, positive at every positive excess loss and at least square-root in $s$ near zero, must bound the gradient norm at every point — if the best possible bound has infimum zero on some band of levels, or exists only pointwise and not continuously, the conclusions can fail.

Editorial extensions

If this is right

For the continuous-time LQR loss on the stabilizing-gain set $D$, the theorem yields a global diffeomorphism $\phi\colon D \to \mathbb{R}^{mn}$ with $L = L^* + \|\phi\|^2$; in particular $D$ is diffeomorphic to Euclidean space.
For logistic regression with non-separable data, the cross-entropy loss is globally of the form $L^* + \|\phi\|^2$ with $\phi$ a global change of parameters, so sublevel sets are diffeomorphic images of round balls.
On a contractible complete manifold, the minimizer set of any semiglobal PŁ function is connected and properly embedded, and the endpoint map is a smooth fiber bundle trivial over contractible neighborhoods of the minimizer set.
There is a complete metric, depending on the function, that makes it geodesically convex and globally 1-PŁ; conversely, quantifying over complete metrics, the semiglobal PŁ class and the global PŁ class coincide.
The family of minimizer sets realizable by semiglobal PŁ functions is exactly the family realizable by globally PŁ functions, namely the contractible submanifolds; no new geometry is added.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A consequence the paper leaves implicit is that the normal form is metric-independent while the robustness hierarchy is controlled by the comparison function at infinity; one can therefore test whether other policy-gradient losses with saturating but not global gradient dominance still have benign global landscape geometry.
The reparametrization equivalence — a function is semiglobal PŁ if and only if some monotone reparametrization of its excess loss is globally PŁ — suggests that algorithmic conclusions invariant under monotone loss reparametrization, such as rank-based or line-search analyses, automatically extend from the global PŁ class to the semiglobal one.
A natural testbed the paper does not pursue is the overparametrized LQR formulation, whose positive-dimensional critical sets contain strict saddles: restricting to the uniformly imbalanced invariant sets that do satisfy gradient dominance, the fiber-bundle picture may describe the low-rank approximation landscape.
The narrow open gap in the sharpness analysis — $\alpha_f$ positive at every level but with infimum zero on some band — is the right place to decide whether the hypothesis can be weakened from a continuous witness to a pointwise condition; either outcome sharpens the boundary of the theorem.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Eduardo Sontag's note does exactly what the title says: BCR's structural conclusions survive when global PŁ is weakened to semiglobal PŁ, i.e. (GD_alpha) with alpha positive definite plus condition (A). The direct proof replaces the three PŁ-dependent lemmas — trajectory bound, Morse–Bott property, fiber coercivity — with clean analogues under the weaker hypothesis. The second proof is the more interesting one: a reparametrization result (Prop. 4.12) showing that any sgl-PŁI function is a smooth monotone reparametrization of a globally PŁ function, with theta a diffeomorphism of [0,∞). The converse in Remark 4.13 pins down the class exactly. That is new and genuinely useful.

The sharpness examples are well chosen. Example 6.4 (x^2 y^2) shows that the square-root behavior at the origin is indispensable for the manifold structure of the minimizer set; Example 6.5 (tanh(x^2)) shows the level-wise uniformity is also needed. The applications to continuous-time LQR and logistic regression verify the hypotheses and read off the normal form; the author is careful about what the diffeomorphism does not give — no metric consequences, no algorithmic content, no robustness. That scoping is a strength, not a weakness.

The main soft spot is self-inflicted. Remark 4.9 shows that after quantifying over complete metrics the enlarged class collapses to global PŁ, and Corollary 4.10 says the realizable minimizer sets do not change. So the "generalization" buys a cleaner hypothesis, not a bigger universe. That is a fair trade, and the author says so plainly, but it does blunt the headline. The minor typo in Lemma 3.5 is cosmetic. The open gap in Remark 3.3 — no continuous positive definite witness on a band even though h(M)=[0,∞) — is real but narrow, and the author explicitly states no counterexample is known. I would not block acceptance on it. The reference list is appropriate; self-citations to the author's prior work on the comparison hierarchy are clearly flagged and used only where needed.

Who is this paper for? Anyone who wants to know which parts of PŁ-type assumptions drive the shape of an optimization landscape rather than the rates. It is a math paper wearing an eess.SY badge; that is fine. I would send it to a serious referee and accept after minor revisions.

Referee Report

0 major / 3 minor

Summary. The paper shows that the global Polyak–Łojasiewicz hypothesis in the recent structural theory of Boumal, Criscitiello and Rebjock (BCR) can be replaced by a strictly weaker semiglobal PŁ condition without losing any of the structural conclusions. The main theorem, under (GD_α) with α positive definite and satisfying condition (A), equivalently under sgl-PŁI, recovers verbatim the BCR normal form f = f* + ‖φ‖² on contractible manifolds, together with the fiber-bundle structure, the Morse–Bott property of the minimizer set, the characterization of possible minimizer sets, and the hidden-convexity theorem. Two independent proofs are given: a direct proof that replaces exactly the three PŁ-dependent steps of BCR, and a reparametrization proof that manufactures a genuinely global PŁ function θ(f−f*) and transports the BCR conclusions back. The paper also provides sharpness examples showing that neither the square-root behavior at the origin nor the level-wise uniformity can be dropped, and it applies the theory to continuous-time LQR policy optimization and to logistic regression.

Significance. If the result holds, and I found no load-bearing reason to doubt it, this is a substantial and clean extension of a recent structural theory. The paper is unusually careful in identifying exactly which ingredients of BCR use the global PŁ inequality and in proving replacements for precisely those ingredients. The reparametrization characterization (Proposition 4.12 and Remark 4.13) is a valuable contribution in its own right, and the sharpness examples in Section 6 are elementary but effective. The applications to LQR and logistic regression are worked in detail, and the paper is honest about what the normal form does and does not imply. It also explicitly flags the one remaining open gap in Remark 3.3, which lies outside the theorem rather than being a counterexample to it. Overall, the paper meets the standard for publication; the remaining issues are local and editorial.

minor comments (3)

[§1.4, Proposition 1.4] In the proof of (ii)⇒(i), the definition of γ(s) is written as γ(s):=∫_2^1 c(sv)dv; the following equality γ(s)=(1/s)∫_s^{2s}c(u)du shows the intended integral is ∫_1^2 c(sv)dv. Please correct the limits of integration.
[§3.2, Remark 3.6] Remark 3.6 uses the end-point map π and refers to Proposition 3.11 before π is formally defined and Proposition 3.11 is proved in §3.3. Since the remark is explicitly forward-looking, adding a pointer or moving it after §3.3 would improve readability.
[Typesetting] The extracted text contains many missing spaces between inline mathematics and prose, for example 'function𝑓:M→Rsatisfyingtheglobal' in the abstract and similar artifacts throughout. The final version should be typeset so that inline formulas are properly separated from surrounding text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sgl-PLI structural theorem is proved directly against BCR and via an independent reparametrization; self-citations are confined to flagged applications.

full rationale

The central derivation is self-contained against the external benchmark [1] (BCR). The direct route (Section 3) replaces exactly three PŁ-dependent ingredients — the trajectory-length bound (Lemma 3.5), the Morse–Bott property (Lemma 3.8), and fiber coercivity (Proposition 3.11) — using the positive-definite witness α and the desingularizer Ψ, with all estimates proved from (GD_α)+(A) rather than assumed. The reparametrization route (Proposition 4.12) constructs a new function g=θ(h) from the witness, proves ||∇g||^2 ≥ c^2 g by the explicit identity θ'(s)α(s)=(c^2/2)Ψ(s), and then applies BCR to g; the transfer map Λ is built from θ^{-1}, so the conclusion is not an input renamed. Proposition 1.4 is a transparent equivalence between (GD_α)+(A) and sgl-PŁI, used only to relabel the hypothesis. The boundary examples (especially Example 6.1(iv) and Example 6.5) show the normal form does not imply any PŁI, so the theorem is not equivalent to its hypothesis by construction. Section 7 cites the author's own [4], [5], [13] for the LQR and logistic applications; these are parameter-free published theorems with stated assumptions not containing the target result, they are explicitly flagged as external inputs, and they are not needed for the main structural claim. The open gap admitted in Remark 3.3 concerns a narrowed boundary case and is a scoped limitation, not a circular step. No circular reduction is exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The central claim rests on importing BCR's structural theorems and the local PL-to-Morse-Bott theorem of Rebjock-Boumal, plus standard Riemannian facts. No free parameters are fitted to data. The only new mathematical objects are constructions (the witness alpha, the desingularizer Psi, the reparametrization theta, the radial map Lambda), which are defined rather than postulated entities.

assumptions (7)

domain assumption BCR Theorem 3.3: if f is smooth and coercive on M with a unique critical point x* and positive definite Hessian at x*, then f = f(x*) + ||phi||^2 for a diffeomorphism phi: M -> R^n.
Imported from [1]. Used in Theorem 4.1 and Proposition 3.11 to put fibers into the nonlinear least-squares normal form; not re-proved in this note.
domain assumption Local PL inequality on a neighborhood of the critical set implies the critical set is a smooth Morse-Bott submanifold with grad^2 f positive definite in normal directions (Rebjock-Boumal [18], quoted as BCR Lemma 2.2).
Used in Lemma 3.8 to convert the local PL inequality obtained from condition (A) into the Morse-Bott property required for the fiber bundle theorems.
standard math Falconer's theorem: the limit mapping of a smooth flow with pseudo-hyperbolic attractor is smooth (Falconer [7]).
Cited in Proposition 3.10 to prove smoothness of the end-point map pi; inherited from BCR Proposition 4.3.
standard math Hopf-Rinow theorem and standard facts about complete Riemannian manifolds: closed bounded sets are compact, and finite-length curves converge.
Used in Lemma 3.5 to turn finite trajectory length into convergence in M.
domain assumption For the continuous-time LQR problem, the CJS-PL estimate ||grad L|| >= xi_1(L - L*) holds with xi_1 of class K on the stabilizing set D (Cui-Jiang-Sontag [4]).
Used in Proposition 7.3 to supply the level-wise uniform positivity half of sgl-PLI for the LQR application; stated as a theorem from [4], not re-proved.
domain assumption For logistic regression under the no-weak-separation condition (9), the cross-entropy loss is coercive, has a unique nondegenerate minimizer, and satisfies the K-PL condition (Cui-Jiang-Sontag [5]).
Used in Section 7.3 to verify sgl-PLI for logistic regression; results imported from [5].
domain assumption The algebraic Riccati equation has at most one stabilizing solution, so the LQR loss has a unique critical point.
Used in Proposition 7.3 to conclude the critical set is a single point; standard control theory result, not proved in this note.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Smooth globally PLI functions are nonlinear least-squares, and so are their gradient-dominated cousins." pith.science (2026). https://pith.science/paper/F5RDCOSF

@misc{pith2026260808849,
  author       = {Pith},
  title        = {Pith review of: Smooth globally PLI functions are nonlinear least-squares, and so are their gradient-dominated cousins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F5RDCOSF}},
  note         = {Machine review of arXiv:2608.08849}
}

abstract

Boumal, Criscitiello and Rebjock (BCR) proved that if $M$ is a contractible, connected and complete Riemannian manifold, then every smooth function $f\colon M\to R$ satisfying the global Polyak--\L{}ojasiewicz inequality (P\L{}I) is necessarily of the form $f = f^* + \|\phi\|^2$ with $\phi$ a submersion. Informally, minimizing such a function amounts to solving a nonlinear least-squares problem in new coordinates. The global P\L{}I hypothesis fails, however, in many problems of interest, among them continuous-time LQR policy optimization in optimal control and a standard formulation of logistic regression. A hierarchy of weakened P\L{} inequalities has been introduced in order to cover such problems, and more generally to study the effect of noise and adversarial perturbations on gradient flows. This note shows that, with minor modifications, the same reduction to a nonlinear least-squares problem holds under a substantially weaker hypothesis, ``semiglobal'' P\L{}I, which is satisfied in both of the examples just mentioned. That condition asks that $f$ satisfy an estimate $\|\nabla f(x)\| \ge \alpha\bigl(f(x)-f^*\bigr)$ for all $x$, with $\alpha$ merely positive definite and bounded below by a positive multiple of $\sqrt{s}$ for small $s>0$.

Figures

Figures reproduced from arXiv: 2608.08849 by the authors.

**Figure 1.** The scalar continuous-time LQR loss (1). (a) Relative loss along the gradient flow from 𝑘 (0) = 10 and 𝑘 (0) = 15: an almost linear decrease at rate ≈ 1 4 , then a soft switch to exponential decay. (b) The sharpest comparison function 𝛼𝑓 (𝑟) = inf{|∇L| : L − L = 𝑟}, which behaves like √ 2𝑟 at the origin but saturates at 1 2 ; the 𝑠𝑎𝑡-PŁI bound √︁ 𝑟/(1 + 4𝑟) (that is, 𝑎 = 𝑏 = 1 4 ) is valid and asymptotically tight, … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

[1]

Smooth, globally Polyak-{\L}ojasiewicz functions are nonlinear least-squares

N.Boumal,C.Criscitiello,andQ.Rebjock.Smooth,globallyPolyak–Łojasiewiczfunctionsarenonlinear least-squares. arXiv:2604.07972, 2026

work page Pith review arXiv 2026
[2]

J. Bu, A. Mesbahi, and M. Mesbahi. Policy gradient-based algorithms for continuous-time linear quadratic control. arXiv:2006.09178, 2020

arXiv 2006
[3]

J. Bu, A. Mesbahi, and M. Mesbahi. On topological properties of the set of stabilizing feedback gains. IEEETrans.Automat.Control,66(2):730–744,2021.AlsoarXiv:1904.08451,2019;and,fortheMIMO case, arXiv:1904.02737, 2019

work page Pith review arXiv 2021
[4]

Cui, Z.-P

L. Cui, Z.-P. Jiang, and E. D. Sontag. Small-disturbance input-to-state stability of perturbed gradient flows: applications to LQR problem.Systems & Control Letters, 188:105804, 2024

work page 2024
[5]

Small-covariancenoise-to-statestabilityofstochasticsystemsand its applications to stochastic gradient dynamics

L.Cui,Z.-P.Jiang,andE.D.Sontag. Small-covariancenoise-to-statestabilityofstochasticsystemsand its applications to stochastic gradient dynamics. In2026 American Control Conference (ACC), 2026. Also arXiv:2509.24277, 2025.doi:10.48550/arXiv.2509.24277. 33

work page doi:10.48550/arxiv.2509.24277 2026
[6]

ConvergenceanalysisofoverparametrizedLQR formulations.Automatica, 182: 112504, 2025

A.CastelloB.deOliveira,M.Siami,andE.D.Sontag. ConvergenceanalysisofoverparametrizedLQR formulations.Automatica, 182: 112504, 2025

work page 2025
[7]

Falconer

K. Falconer. Differentiation of the limit mapping in a dynamical system.J. London Math. Soc., s2- 27(2):356–372, 1983

work page 1983
[8]

Optimizingstaticlinearfeedback: gradientmethod.SIAMJ.ControlOptim., 59(5):3887–3911, 2021

I.FatkhullinandB.Polyak. Optimizingstaticlinearfeedback: gradientmethod.SIAMJ.ControlOptim., 59(5):3887–3911, 2021

work page 2021

Show all 19 references

[9]

Fazel, R

M. Fazel, R. Ge, S. Kakade, and M. Mesbahi. Global convergence of policy gradient methods for the linear quadratic regulator. InProc. ICML, pages 1467–1476, 2018

2018
[10]

Convergenceandsamplecomplexity ofgradientmethodsforthemodel-freelinear-quadraticregulatorproblem.IEEETrans.Automat.Control, 67(5):2435–2450, 2022

H.Mohammadi,A.Zare,M.Soltanolkotabi,andM.R.Jovanović. Convergenceandsamplecomplexity ofgradientmethodsforthemodel-freelinear-quadraticregulatorproblem.IEEETrans.Automat.Control, 67(5):2435–2450, 2022

2022
[11]

E. D. Sontag. Input to state stability: basic concepts and results. InNonlinear and Optimal Control Theory, pages 163–220. Springer, 2007

2007
[12]

E. D. Sontag. Remarks on input to state stability of perturbed gradient flows, motivated by model-free feedback control learning.Systems & Control Letters, 161:105138, 2022

2022
[13]

SomeremarksongradientdominanceandLQRpolicyoptimization

E.D.Sontag. SomeremarksongradientdominanceandLQRpolicyoptimization. arXiv:2507.10452,

arXiv
[14]

Onthe(almost)globalexponentialconvergence of overparameterized policy optimization for the LQR problem, Proceedings of American Control Conference (ACC), 2026

M.Wafi,A.CastelloB.deOliveira,andE.D.Sontag. Onthe(almost)globalexponentialconvergence of overparameterized policy optimization for the LQR problem, Proceedings of American Control Conference (ACC), 2026

2026
[15]

Ongradientsoffunctionsdefinableino-minimalstructures.Ann.Inst.Fourier,48(3):769– 783, 1998

K.Kurdyka. Ongradientsoffunctionsdefinableino-minimalstructures.Ann.Inst.Fourier,48(3):769– 783, 1998

1998
[16]

Łojasiewicz

S. Łojasiewicz. Sur les trajectoires du gradient d’une fonction analytique.Seminari di Geometria, 1983:115–117, 1982

1983
[17]

B. Polyak. Gradient methods for the minimisation of functionals.USSR Comput. Math. Math. Phys., 3(4):864–878, 1963

1963
[18]

Rebjock and N

Q. Rebjock and N. Boumal. Fast convergence to non-isolated minima: four equivalent conditions for 𝐶2 functions.Math. Program., 213: 151-199, 2025. 34

2025
[2025]

Keynote, Learning for Dynamics & Control

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Smooth, globally Polyak-{\L}ojasiewicz functions are nonlinear least-squares

N.Boumal,C.Criscitiello,andQ.Rebjock.Smooth,globallyPolyak–Łojasiewiczfunctionsarenonlinear least-squares. arXiv:2604.07972, 2026

work page Pith review arXiv 2026

[2] [2]

J. Bu, A. Mesbahi, and M. Mesbahi. Policy gradient-based algorithms for continuous-time linear quadratic control. arXiv:2006.09178, 2020

arXiv 2006

[3] [3]

J. Bu, A. Mesbahi, and M. Mesbahi. On topological properties of the set of stabilizing feedback gains. IEEETrans.Automat.Control,66(2):730–744,2021.AlsoarXiv:1904.08451,2019;and,fortheMIMO case, arXiv:1904.02737, 2019

work page Pith review arXiv 2021

[4] [4]

Cui, Z.-P

L. Cui, Z.-P. Jiang, and E. D. Sontag. Small-disturbance input-to-state stability of perturbed gradient flows: applications to LQR problem.Systems & Control Letters, 188:105804, 2024

work page 2024

[5] [5]

Small-covariancenoise-to-statestabilityofstochasticsystemsand its applications to stochastic gradient dynamics

L.Cui,Z.-P.Jiang,andE.D.Sontag. Small-covariancenoise-to-statestabilityofstochasticsystemsand its applications to stochastic gradient dynamics. In2026 American Control Conference (ACC), 2026. Also arXiv:2509.24277, 2025.doi:10.48550/arXiv.2509.24277. 33

work page doi:10.48550/arxiv.2509.24277 2026

[6] [6]

ConvergenceanalysisofoverparametrizedLQR formulations.Automatica, 182: 112504, 2025

A.CastelloB.deOliveira,M.Siami,andE.D.Sontag. ConvergenceanalysisofoverparametrizedLQR formulations.Automatica, 182: 112504, 2025

work page 2025

[7] [7]

Falconer

K. Falconer. Differentiation of the limit mapping in a dynamical system.J. London Math. Soc., s2- 27(2):356–372, 1983

work page 1983

[8] [8]

Optimizingstaticlinearfeedback: gradientmethod.SIAMJ.ControlOptim., 59(5):3887–3911, 2021

I.FatkhullinandB.Polyak. Optimizingstaticlinearfeedback: gradientmethod.SIAMJ.ControlOptim., 59(5):3887–3911, 2021

work page 2021