Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Safe Gradient Flow for Bilevel Optimization

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper turns bilevel optimization into a single continuous-time dynamical system in which a safety filter enforces the lower-level optimality condition while a gradient flow minimizes the upper-level objective.

desk verdict A genuinely useful safety-filter/gradient-flow idea for bilevel problems, with a fixable gap in Theorem 2's statement and a typo in Theorem 1's proof. read the letter →

arxiv 2501.16520 v2 pith:LS62N5LD submitted 2025-01-27 math.OC cs.LGcs.SYeess.SY

Sina Sharifi , Nazanin Abolfazli , Erfan Yazdandoost Hamedani , Mahyar Fazlyab This is my paper · ORCID
classification math.OCcs.LGcs.SYeess.SY MSC 90C3090C2634D20
keywords bileveloptimizationsafetyfiltercontrolbarrierfunctiongradientflowLyapunovanalysishypergradientsingle-loopalgorithmimplicitdifferentiation
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 proposes solving bilevel optimization with a single continuous-time dynamical system built from two parts: a gradient flow that minimizes the upper-level objective and a safety filter that enforces the lower-level optimality condition as an invariant constraint. The filter is a quadratic program that minimally modifies the flow so that the KKT residual $\nabla_y g(x,y)$ contracts exponentially, or, in a relaxed variant, stays within a user-chosen tolerance $\varepsilon$. The central result is a Lyapunov proof that the time-averaged squared norm of the hypergradient converges to $O(\varepsilon^2)$ plus a term decaying like $1/t$, so the trajectory reaches a neighborhood of a stationary point of the relaxed problem. The paper's claim matters because it offers a principled single-loop alternative to nested AID/ITD schemes and avoids inverting matrices in the relaxed variant.

What carries the argument

The load-bearing object is the safety-filtered velocity field. A base gradient flow $(\dot{x},\dot{y}) = (-\nabla_x f, -\nabla_y f)$ is modified by the solution of a convex quadratic program whose constraint is the control-barrier condition. For the exact filter, the condition $\frac{d}{dt}\nabla_y g + \alpha\nabla_y g = 0$ forces exponential contraction to the KKT manifold and yields a closed-form dual variable $\lambda$. For the relaxed filter, the half-space constraint $\frac{d}{dt}(h-\varepsilon^2) + \alpha(h-\varepsilon^2) \le 0$ with $h = \|\nabla_y g\|^2$ keeps the trajectory inside the $\varepsilon$-suboptimality set while avoiding the matrix inversion in $\lambda$. The same projection structure, projecting velocities rather than decision variables onto a set defined by lower-level optimality, carries all the convergence proofs.

What would settle it

Run the proposed flows on a two-dimensional bilevel problem with an inner objective that is convex but not strongly convex, such as $g(x,y) = (y-x)^2$ on a bounded domain, and measure whether the time-averaged hypergradient norm still converges to the predicted $O(\varepsilon^2)$ neighborhood; if convergence fails or the $\varepsilon$-sublevel set is violated, the central claim is false. Alternatively, for a strongly convex example, numerically compute $\frac{1}{t}\int_0^t\|\nabla\ell(x(\tau))\|^2\,d\tau$ and compare to the bound $\frac{2M_1^2\varepsilon^2}{\mu_g^2} + \frac{2(f(x(0),y(0))-f^*_\varepsilon)}{ct}$; an exceedance would contradict Theorem 2.

Watch

Extended reading notes

Core claim

The central claim is that a safe gradient flow, defined as gradient descent on the outer objective protected by a projection-based safety filter, solves the bilevel problem without solving the inner problem exactly. Concretely, the paper proves that when the inner objective is strongly convex, the flow that projects the velocities onto the constraint manifold $\mathcal{M} = \{(x,y) : \nabla_y g(x,y)=0\}$ renders $\mathcal{M}$ exponentially stable and forward invariant, and its equilibrium satisfies the KKT conditions of the single-level reduction. For the relaxed, inversion-free flow, the paper proves forward invariance of the sublevel set $\{h(x,y) \le \varepsilon^2\}$ with $h = \|\nabla_y g\|^2$ and a non-asymptotic bound on the time-averaged hypergradient norm: $\frac{1}{t}\int_0^t \|\nabla\ell(x(\tau))\|^2\,d\tau \le \frac{2M_1^2}{\mu_g^2}\varepsilon^2 + \frac{2}{ct}(f(x(0),y(0)) - f^*_\varepsilon)$. Thus a user-chosen $\varepsilon$ directly controls the accuracy of the approximate stationarity, and the rate is $O(1/t)$ up to that tolerance.

Load-bearing premise

The inner problem must be strongly convex in $y$ (Assumption 2-i): without it the lower-level solution map may not be single-valued or differentiable, and the inequality $\mu_g\|y-y^*(x)\| \le \|\nabla_y g(x,y)\|$ used throughout the proofs fails.

Editorial extensions

If this is right

  • Bilevel solvers can be built as single-loop ODEs without nested lower-level solves, since the safety filter enforces inner optimality concurrently with the outer descent.
  • The relaxed filter scales to high-dimensional inner problems because it avoids the $m \times m$ matrix inversion required by the exact filter.
  • The time-averaged hypergradient norm is certified to reach $O(\varepsilon^2)$ with explicit constants, giving a quantitative trade-off between inner-optimality tolerance and achieved accuracy.
  • Starting from any feasible initial condition, the relaxed flow remains feasible for all time, so the method provides anytime safety for the approximate bilevel problem.
  • The prediction-correction variant analyzed in the appendix achieves an $O(1/t)$ bound on the hypergradient norm from an arbitrary initialization, extending the rate-based guarantee beyond the safety-filter formulation.

Reading between the lines

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

  • A testable extension is to replace strong convexity by a weaker regularity condition: the proofs use the inequality $\mu_g\|y - y^*(x)\| \le \|\nabla_y g(x,y)\|$, so any setting where that inequality holds with a computable constant could inherit the same guarantees, though the paper does not establish this.
  • The continuous-time formulation invites a natural discretization question: the experiments use RK-4, but it remains open whether an explicit Euler implementation preserves the safety constraint and the $O(1/t)$ rate exactly, since the filter is designed on the continuous velocity field.
  • Because the relaxed flow is forward invariant for all time, the method could plausibly be embedded in online or time-varying settings where the outer problem changes, a scenario not analyzed in the paper.
  • The prediction-correction dynamics in the appendix connect the safety-filter approach to classical tracking of a time-varying optimizer; one could compare the two on problems where the inner problem is only locally strongly convex, a regime the paper does not treat.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a control-theoretic approach to bilevel optimization. The main construction is a 'safe gradient flow' (SGF): a base gradient-flow dynamics for the upper-level objective is minimally modified by a safety filter that enforces the lower-level KKT manifold M = {(x,y) | ∇_y g(x,y)=0}. A relaxed, matrix-inversion-free variant (RXGF) enforces the inequality constraint h(x,y) = ||∇_y g(x,y)||^2 ≤ ε^2 instead. The authors prove Lyapunov-based convergence rates for both methods, present a prediction-correction variant in Appendix B, and validate the methods on a synthetic problem and a data hyper-cleaning task on MNIST.

Significance. If the results hold after revision, the paper makes a useful contribution by connecting control-barrier-function safety filters to bilevel optimization and providing single-loop continuous-time dynamics with explicit non-asymptotic rates. The relaxed RXGF method is particularly interesting because it avoids m×m matrix inversions and gives a user-controlled ε-approximate stationarity guarantee. The proofs are largely self-contained, with Lemma 1 using a standard hypergradient bound from the authors' prior work [42], and the numerical code is publicly available. However, the current proof of Theorem 1 contains an incorrect substitution in the application of Lemma 4, and Theorem 2 omits a necessary feasibility assumption, so the theoretical claims are not yet fully supported as stated.

major comments (3)
  1. [Appendix A.2, proof of Theorem 1] The proof invokes Lemma 4 with d = -2λ(x,y)∇_y g(x,y). This substitution is dimensionally incorrect and does not yield the claimed bound ||F(x,y)|| ≤ ||ẋ|| + (L_g^yx/µ_g)||ẏ||. The correct substitution is d = -λ(x,y), because then A(x,y,d) = ∇_x f - ∇²_yx g^T d = ∇_x f + ∇²_yx g^T λ = -ẋ and B(x,y,d) = ∇_y f - ∇²_yy g d = ∇_y f + ∇²_yy g λ = -ẏ. Please correct this step and the domain of d in Lemma 4 (it should be R^m, not R^n).
  2. [Theorem 2 and Appendix A.5] Theorem 2 as stated does not assume (x(0),y(0)) ∈ L^-_{ε²}(h), but the proof uses Lemma 6, which requires (x(t),y(t)) ∈ L^-_{ε²}(h) for all t. By Proposition 1(ii), this holds only if the initial condition is feasible. Without this assumption the theorem is false: for f(x,y)=0.5(x-1)², g(x,y)=0.5(y-x)², ε=0.1, α=1, starting from (x,y)=(1,0), one has E(0)=0, yet the safety filter's projection forces ẋ<0 and hence E(t)>0. Please add the feasibility assumption to the theorem statement and to any related claims.
  3. [Abstract and Section 5] The abstract and conclusion state that the dynamics 'converge to a neighborhood of the optimal solution', but Theorem 2 (and Theorem 1) establish only a time-averaged Cesàro bound on ||∇ℓ(x(t))||². An L² time-average bound does not imply pointwise convergence of the trajectory or of the hypergradient. The results should be described as time-averaged convergence rates, with the abstract and conclusion adjusted to match the actual theorems.
minor comments (5)
  1. [Lemma 4] The lemma states 'For any vector d ∈ Rn', but from the definitions of A and B the vector d must lie in R^m; please correct the dimension.
  2. [Lemma 6] Lemma 6 states the assumption (x,y) ∈ L^-_ε(h), while the proof and its use in Theorem 2 require (x,y) ∈ L^-_{ε²}(h). Please make the sublevel-set notation consistent (likely L^-_{ε²}(h)).
  3. [Appendix B, Proposition 2] Proposition 2 is stated without proof ('omitted here for brevity'), but Theorem 3 directly depends on it. Please include the proof or a precise, self-contained derivation following [47], or explicitly state the exact result from [47] that implies (22).
  4. [Section 4, experiments] The experimental section does not report the RK-4 step size, the exact values of α and ε used in each figure, or any measure of variability (e.g., standard deviations over seeds). Please include these details for reproducibility.
  5. [Throughout] There are small typographical issues: 'Lyapanouv' in Remark 2 should be 'Lyapunov', Theorem 1 refers to 'the global optimal value of 3' which should be 'of (3)', and the caption of Figure 2 has a formatting issue with 'left/middle/right' labels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Lyapunov convergence analysis is self-contained, and the cited prior lemmas are standard supporting results rather than the paper's own conclusions.

full rationale

The paper's derivation chain does not reduce to its inputs by construction. The safe gradient flow is defined as a QP-based safety filter that enforces invariance of the KKT manifold or its relaxation; this is the mechanism, not a fitted prediction. Theorem 2's time-averaged bound follows from a Lyapunov function E(t) with c chosen to cancel the quadratic velocity terms, and the ε^2 term is the approximation error inherent in the relaxed feasible set L_{ε^2}(h), not a parameter fitted to the algorithm's outputs. Lemma 1 is cited from the authors' earlier work [42] and used throughout, but it is a standard Lipschitz bound on the hypergradient estimation error, it is not a uniqueness theorem, and it does not assume the theorem conclusions; the citation is therefore real supporting evidence rather than circularity. Likewise, Proposition 2 cites [47] for a standard exponential contraction argument. No 'prediction' is obtained from data or from a fitted parameter. Some claims are stronger than what is proven—Theorem 2 gives a time-averaged bound, not pointwise convergence, and its statement omits the feasibility initialization used in the proof—but those are correctness or presentation gaps, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the strong convexity of the lower-level problem, bounded-gradient assumptions on the upper-level objective, and a set of standard mathematical tools. No new entities are postulated.

free parameters (3)
  • alpha (α) = α > 0, user-tuned; experiments use α ∈ {0.001, 0.01, 0.1, 0.5, 1}
    Contraction rate of the safety filter. Chosen by hand; theorems hold for any positive α.
  • epsilon (ε) = ε > 0; experiments use ε ∈ {0.01, 0.05, 0.1, 0.2, 0.5}
    Tolerance for the relaxed lower-level optimality constraint. Chosen by hand; controls the size of the convergence neighborhood.
  • beta (β) in prediction-correction = β > 0; experiments use β ∈ {10^-2, 5×10^-2, 10^-1, 5×10^-1}
    Gain in the lower-level Newton correction dynamics of Appendix B. Not fitted to data.
assumptions (5)
  • domain assumption Assumption 2-i: g(x,·) is µ_g-strongly convex for all x
    Ensures y*(x) is unique and differentiable, and the KKT manifold (3) is an exact reformulation. Used throughout Theorems 1 and 2.
  • domain assumption Assumption 1: f is continuously differentiable with bounded gradients and Lipschitz derivatives
    Used to bound the hypergradient error (Lemma 1) and to establish Lipschitz continuity of the ODE right-hand side.
  • standard math Implicit function theorem for y*(x)
    Invoked in Section 1.3 to derive the hypergradient formula.
  • standard math Lyapunov stability theory and comparison principles
    Used to prove convergence rates in Theorems 1-3.
  • domain assumption Problems are unconstrained at both levels
    The paper explicitly states it only considers unconstrained bilevel problems, deferring constraints to future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safe Gradient Flow for Bilevel Optimization." pith.science (2026). https://pith.science/paper/LS62N5LD

@misc{pith2026250116520,
  author       = {Pith},
  title        = {Pith review of: Safe Gradient Flow for Bilevel Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LS62N5LD}},
  note         = {Machine review of arXiv:2501.16520}
}
read the original abstract

Bilevel optimization is a key framework in hierarchical decision-making, where one problem is embedded within the constraints of another. In this work, we propose a control-theoretic approach to solving bilevel optimization problems. Our method consists of two components: a gradient flow mechanism to minimize the upper-level objective and a safety filter to enforce the constraints imposed by the lower-level problem. Together, these components form a safe gradient flow that solves the bilevel problem in a single loop. To improve scalability with respect to the lower-level problem's dimensions, we introduce a relaxed formulation and design a compact variant of the safe gradient flow. This variant minimizes the upper-level objective while ensuring the lower-level decision variable remains within a user-defined suboptimality. Using Lyapunov analysis, we establish convergence guarantees for the dynamics, proving that they converge to a neighborhood of the optimal solution. Numerical experiments further validate the effectiveness of the proposed approaches. Our contributions provide both theoretical insights and practical tools for efficiently solving bilevel optimization problems.

Figures

Figures reproduced from arXiv: 2501.16520 by the authors.

Figure 1
Figure 1. An overview of the proposed method. The gradient flow attempts to minimize the upper-level objective, while the safety filter ensures that the constraints induced by the lower-level problem are satisfied. the constraints imposed by the lower-level problem. We will then leverage the notion of set invari￾ance to design a safety filter, posed as a convex quadratic program (QP), that minimally modifies the dynamics of t… view at source ↗
Figure 2
Figure 2. Effect of ε on the convergence of the (left:) surrogate map and (middle:) lower-level problem with α = 0.01. (right:) Effect of α on lower-level behavior. 0 20 40 60 80 100 time 2 4 6 8 10 Validation Loss InversionFree: p= 0.25 AIDBio: p= 0.25 InversionFree: p= 0.4 AIDBio: p= 0.4 0 20 40 60 80 100 time 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Test Accuracy InversionFree: p= 0.25 AIDBio: p= 0.25 InversionFree: p= 0.4 AIDBio: p= 0… view at source ↗
Figure 3
Figure 3. The comparison of the validation loss and the test accuracy between our Inversion-free method and AIDBio with α, β ∈ {0.001, , 0.01, 0.1} for p ∈ {25%, 40%}. 4 Experiments In this section, we evaluate our method’s performance and analyze the impact of each hyperpa￾rameter on a synthetic problem and a data hyper-cleaning task using the MNIST dataset. All experiments employ the Fourth-order Runge-Kutta (RK-4) discreti… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison between our second order method and STABLE with α = 1 and β ∈ {10−2 , 5× 10−2 , 10−1 , 5×10−1}. Then, by integrating both sides we obtain E(t) − E(0) ≤ Z t 0 1 2 ∥F(x(τ ), y(τ )) − ∇ℓ(x(τ ))∥ 2 dτ ≤ M2 1 2 Z t 0 ∥y(τ ) − y ⋆ (x(τ ))∥ 2 dτ ≤ M2 1 2µ2 g ∥∇yg(x…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sequential QCQP for Bilevel Optimization with Line Search

    math.OC 2025-05 conditional novelty 5.0 of 10

    A bilevel optimization algorithm uses a tilted QCQP and a control-barrier line search to guarantee anytime feasibility and an O(1/k) ergodic convergence rate.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [47]

    Prediction- correction interior-point method for time-varying convex optimization

    Mahyar Fazlyab, Santiago Paternain, Victor M Preciado, and Alejandro Ribeiro. “Prediction- correction interior-point method for time-varying convex optimization”. In: IEEE Transac- tions on Automatic Control 63.7 (2017), pp. 1973–1986. 19

    work page 2017

  2. [42]

    An Inexact Conditional Gradient Method for Constrained Bilevel Optimization

    Nazanin Abolfazli, Ruichen Jiang, Aryan Mokhtari, and Erfan Yazdandoost Hamedani. “An inexact conditional gradient method for constrained bilevel optimization”. In: arXiv preprint arXiv:2306.02429 (2023)

    work page Pith review arXiv 2023

  3. [1]

    Investments in merchant energy storage: Trading-off between energy and reserve markets

    Hrvoje Pandˇ zi´ c, Yury Dvorkin, and Miguel Carri´ on. “Investments in merchant energy storage: Trading-off between energy and reserve markets”. In:Applied energy 230 (2018), pp. 277–286

    work page 2018

  4. [2]

    The theory of the market economy

    Heinrich Von Stackelberg and Stackelberg Heinrich Von. The theory of the market economy . Oxford University Press, 1952

    work page 1952

  5. [3]

    An iterative algorithm for two level hierarchical time minimization transportation problem

    Anuj Sharma, Vanita Verma, Prabhjot Kaur, and Kalpana Dahiya. “An iterative algorithm for two level hierarchical time minimization transportation problem”. In: European Journal of Operational Research 246.3 (2015), pp. 700–707. 16

    work page 2015

  6. [4]

    Model-agnostic meta-learning for fast adap- tation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. “Model-agnostic meta-learning for fast adap- tation of deep networks”. In: International conference on machine learning . PMLR. 2017, pp. 1126–1135

    work page 2017

  7. [5]

    Meta-learning with implicit gradients

    Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. “Meta-learning with implicit gradients”. In: Advances in neural information processing systems 32 (2019)

    work page 2019

  8. [6]

    One-shot learning of object categories

    Li Fei-Fei, Rob Fergus, and Pietro Perona. “One-shot learning of object categories”. In: IEEE transactions on pattern analysis and machine intelligence 28.4 (2006), pp. 594–611

    work page 2006

Show all 47 references
  1. [7]

    A two-timescale frame- work for bilevel optimization: Complexity analysis and application to actor-critic

    Mingyi Hong, Hoi-To Wai, Zhaoran Wang, and Zhuoran Yang. “A two-timescale frame- work for bilevel optimization: Complexity analysis and application to actor-critic”. In: arXiv preprint arXiv:2007.05170 (2020)

    2020 arXiv

  2. [8]

    Gradient-based optimization of hyperparameters

    Yoshua Bengio. “Gradient-based optimization of hyperparameters”. In: Neural computation 12.8 (2000), pp. 1889–1900

    2000

  3. [9]

    Bilevel coreset selection in continual learning: A new for- mulation and algorithm

    Jie Hao, Kaiyi Ji, and Mingrui Liu. “Bilevel coreset selection in continual learning: A new for- mulation and algorithm”. In: Advances in Neural Information Processing Systems 36 (2024)

    2024

  4. [10]

    Revisiting and advancing fast adversarial training through the lens of bi-level optimization

    Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, and Sijia Liu. “Revisiting and advancing fast adversarial training through the lens of bi-level optimization”. In: International Conference on Machine Learning . PMLR. 2022, pp. 26693–26712

    2022

  5. [11]

    New branch-and-bound rules for linear bilevel programming

    Pierre Hansen, Brigitte Jaumard, and Gilles Savard. “New branch-and-bound rules for linear bilevel programming”. In: SIAM Journal on scientific and Statistical Computing 13.5 (1992), pp. 1194–1217

    1992

  6. [12]

    Hyperparameter optimization with approximate gradient

    Fabian Pedregosa. “Hyperparameter optimization with approximate gradient”. In: Proceed- ings of the 33nd International Conference on Machine Learning (ICML) . 2016

    2016

  7. [13]

    Generic Methods for Optimization-Based Modeling

    Justin Domke. “Generic Methods for Optimization-Based Modeling”. In: Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics. 2012, pp. 318–326

    2012

  8. [14]

    Approximation methods for bilevel programming

    Saeed Ghadimi and Mengdi Wang. “Approximation methods for bilevel programming”. In: arXiv preprint arXiv:1802.02246 (2018)

    2018 arXiv

  9. [15]

    Truncated back- propagation for bilevel optimization

    Amirreza Shaban, Ching-An Cheng, Nathan Hatch, and Byron Boots. “Truncated back- propagation for bilevel optimization”. In: The 22nd International Conference on Artificial Intelligence and Statistics . PMLR. 2019, pp. 1723–1732

    2019

  10. [16]

    Bilevel Programming for Hyperparameter Optimization and Meta-Learning

    L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil. “Bilevel Programming for Hyperparameter Optimization and Meta-Learning”. In: ICML. 2018

    2018

  11. [17]

    On the iteration complexity of hypergradient computation

    Riccardo Grazzi, Luca Franceschi, Massimiliano Pontil, and Saverio Salzo. “On the iteration complexity of hypergradient computation”. In: International Conference on Machine Learn- ing. PMLR. 2020, pp. 3748–3758

    2020

  12. [18]

    A constrained optimization approach to bilevel optimization with multiple inner minima

    Daouda Sow, Kaiyi Ji, Ziwei Guan, and Yingbin Liang. “A constrained optimization approach to bilevel optimization with multiple inner minima”. In: arXiv preprint arXiv:2203.01123 (2022)

    2022 arXiv

  13. [19]

    On Bilevel Optimization without Lower-level Strong Convexity

    Lesi Chen, Jing Xu, and Jingzhao Zhang. “On Bilevel Optimization without Lower-level Strong Convexity”. In: arXiv preprint arXiv:2301.00712 (2023)

    2023 arXiv

  14. [20]

    A Method for Bilevel Optimization with Convex Lower-Level Problem

    Han Shen, Santiago Paternain, Gaowen Liu, Ramana Kompella, and Tianyi Chen. “A Method for Bilevel Optimization with Convex Lower-Level Problem”. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE. 2024, pp. 9426–9430. 17

    2024

  15. [21]

    Bome! bilevel optimization made easy: A simple first-order approach

    Bo Liu, Mao Ye, Stephen Wright, Peter Stone, and Qiang Liu. “Bome! bilevel optimization made easy: A simple first-order approach”. In: Advances in neural information processing systems 35 (2022), pp. 17248–17262

    2022

  16. [22]

    A Generalized Alternating Method for Bilevel Learning under the Polyak-{\L} ojasiewicz Condition

    Quan Xiao, Songtao Lu, and Tianyi Chen. “A Generalized Alternating Method for Bilevel Learning under the Polyak-{\L} ojasiewicz Condition”. In: arXiv preprint arXiv:2306.02422 (2023)

    2023 arXiv

  17. [23]

    Alternating projected sgd for equality- constrained bilevel optimization

    Quan Xiao, Han Shen, Wotao Yin, and Tianyi Chen. “Alternating projected sgd for equality- constrained bilevel optimization”. In: International Conference on Artificial Intelligence and Statistics. PMLR. 2023, pp. 987–1023

    2023

  18. [24]

    Linearly constrained bilevel optimization: A smoothed implicit gradient ap- proach

    Prashant Khanduri, Ioannis Tsaknakis, Yihua Zhang, Jia Liu, Sijia Liu, Jiawei Zhang, and Mingyi Hong. “Linearly constrained bilevel optimization: A smoothed implicit gradient ap- proach”. In: International Conference on Machine Learning . PMLR. 2023, pp. 16291–16325

    2023

  19. [25]

    Efficient gradient approximation method for constrained bilevel optimization

    Siyuan Xu and Minghui Zhu. “Efficient gradient approximation method for constrained bilevel optimization”. In: Proceedings of the AAAI Conference on Artificial Intelligence . Vol. 37. 10. 2023, pp. 12509–12517

    2023

  20. [26]

    First-Order Methods for Linearly Constrained Bilevel Optimization

    Guy Kornowski, Swati Padmanabhan, Kai Wang, Zhe Zhang, and Suvrit Sra. “First-Order Methods for Linearly Constrained Bilevel Optimization”. In:arXiv preprint arXiv:2406.12771 (2024)

    2024 arXiv

  21. [27]

    Control barrier function based quadratic programs for safety critical systems

    Aaron D Ames, Xiangru Xu, Jessy W Grizzle, and Paulo Tabuada. “Control barrier function based quadratic programs for safety critical systems”. In: IEEE Transactions on Automatic Control 62.8 (2016), pp. 3861–3876

    2016

  22. [28]

    Control barrier functions: Theory and applications

    Aaron D Ames, Samuel Coogan, Magnus Egerstedt, Gennaro Notomista, Koushil Sreenath, and Paulo Tabuada. “Control barrier functions: Theory and applications”. In: 2019 18th European control conference (ECC). IEEE. 2019, pp. 3420–3431

    2019

  23. [29]

    Zero-order Control Barrier Functions for Sampled-Data Systems with State and Input Dependent Safety Constraints

    Xiao Tan, Ersin Das, Aaron D Ames, and Joel W Burdick. “Zero-order Control Barrier Functions for Sampled-Data Systems with State and Input Dependent Safety Constraints”. In: arXiv preprint arXiv:2411.17079 (2024)

    2024 arXiv

  24. [30]

    On constraints in first-order optimization: A view from non-smooth dynamical systems

    Michael Muehlebach and Michael I Jordan. “On constraints in first-order optimization: A view from non-smooth dynamical systems”. In: Journal of Machine Learning Research 23.256 (2022), pp. 1–47

    2022

  25. [31]

    Control barrier function-based design of gradient flows for constrained nonlinear programming

    Ahmed Allibhoy and Jorge Cort´ es. “Control barrier function-based design of gradient flows for constrained nonlinear programming”. In: IEEE Transactions on Automatic Control (2023)

    2023

  26. [32]

    Anytime solution of constrained nonlinear programs via control barrier functions

    Ahmed Allibhoy and Jorge Cort´ es. “Anytime solution of constrained nonlinear programs via control barrier functions”. In: 2021 60th IEEE Conference on Decision and Control (CDC) . IEEE. 2021, pp. 6527–6532

    2021

  27. [33]

    Orthogonal Directions Constrained Gradient Method: from non-linear equality constraints to Stiefel man- ifold

    Sholom Schechtman, Daniil Tiapkin, Michael Muehlebach, and Eric Moulines. “Orthogonal Directions Constrained Gradient Method: from non-linear equality constraints to Stiefel man- ifold”. In: The Thirty Sixth Annual Conference on Learning Theory . PMLR. 2023, pp. 1228– 1258

    2023

  28. [34]

    A geometric method in nonlinear programming

    Kunio Tanabe. “A geometric method in nonlinear programming”. In: Journal of Optimization Theory and Applications 30 (1980), pp. 181–210

    1980

  29. [35]

    A single-timescale method for stochastic bilevel optimization

    Tianyi Chen, Yuejiao Sun, Quan Xiao, and Wotao Yin. “A single-timescale method for stochastic bilevel optimization”. In: International Conference on Artificial Intelligence and Statistics. PMLR. 2022, pp. 2466–2488. 18

    2022

  30. [36]

    Two-timescale stochastic gradient descent in continuous time with applications to joint online parameter estimation and optimal sensor placement

    Louis Sharrock and Nikolas Kantas. “Two-timescale stochastic gradient descent in continuous time with applications to joint online parameter estimation and optimal sensor placement”. In: Bernoulli 29.2 (2023), pp. 1137–1165

    2023

  31. [37]

    On a continuous time stochastic approximation problem

    G Yin and Ishita Gupta. “On a continuous time stochastic approximation problem”. In: Acta Applicandae Mathematica 33 (1993), pp. 3–20

    1993

  32. [38]

    Connection between continuous and discrete Riccati equations with applications to Kalman filtering

    M Salgado, R Middleton, and Graham C Goodwin. “Connection between continuous and discrete Riccati equations with applications to Kalman filtering”. In: IEE Proceedings D (Control Theory and Applications) . Vol. 135. 1. IET. 1988, pp. 28–34

    1988

  33. [39]

    Least squares identification for continuous-time systems

    HF Chen. “Least squares identification for continuous-time systems”. In: Analysis and Op- timization of Systems: Proceedings of the Fourth International Conference on Analysis and Optimization of Systems Versailles, December 16–19, 1980 . Springer. 2005, pp. 264–277

    1980

  34. [40]

    Stochastic gradient descent in continuous time

    Justin Sirignano and Konstantinos Spiliopoulos. “Stochastic gradient descent in continuous time”. In: SIAM Journal on Financial Mathematics 8.1 (2017), pp. 933–961

    2017

  35. [41]

    Principles of Mathematical Analysis

    Walter Rudin. Principles of Mathematical Analysis . 1976

    1976

  36. [43]

    Bilevel optimization: Convergence analysis and enhanced design

    Kaiyi Ji, Junjie Yang, and Yingbin Liang. “Bilevel optimization: Convergence analysis and enhanced design”. In: International Conference on Machine Learning . 2021, pp. 4882–4892

    2021

  37. [44]

    Directional derivatives of the solution of a parametric nonlinear program

    Daniel Ralph and Stephan Dempe. “Directional derivatives of the solution of a parametric nonlinear program”. In: Mathematical programming 70.1 (1995), pp. 159–172

    1995

  38. [45]

    Lectures on convex optimization

    Yurii Nesterov. Lectures on convex optimization . Vol. 137. Springer, 2018

    2018

  39. [46]

    Nonlinear dynamical systems and control: a Lyapunov-based approach

    Wassim M Haddad and VijaySekhar Chellaboina. Nonlinear dynamical systems and control: a Lyapunov-based approach. Princeton university press, 2008

    2008

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.