Projected Gradient Descent for Constrained Decision-Dependent Optimization

Changxin Liu; Karl H. Johansson; Michael M. Zavlanos; Thomas Parisini; Zifan Wang

Projected Gradient Descent for Constrained Decision-Dependent Optimization

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper proves that repeated projected gradient descent converges to a constrained equilibrium in decision-dependent optimization, where both the objective and the feasible sets shift in response to decisions.

desk verdict Abstract-only read: a plausible new algorithm and a Lipschitz-projection theorem that I can't verify, and whose hinge condition is exactly the part the abstract doesn't show. read the letter →

arxiv 2508.08856 v1 pith:Z5SCDLI2 submitted 2025-08-12 math.OC

Zifan Wang , Changxin Liu , Thomas Parisini , Michael M. Zavlanos , Karl H. Johansson This is my paper

classification math.OC MSC 90C3090C47

keywords decision-dependentoptimizationprojectedgradientdescentconstrainedequilibriumLipschitzcontinuityvaryingfeasiblesetsdualascentdynamicpricingmarket

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

The paper proposes repeated projected gradient descent (RPGD) for decision-dependent optimization problems, in which the data distribution reacts to decisions and reshapes both the objective function and the linear constraints. It establishes sufficient conditions under which RPGD converges to a constrained equilibrium point, even though the feasible set changes at every iteration. The method keeps all iterates feasible throughout the process and has a lower computational burden than the existing dual ascent approach. This matters for practical problems such as market equilibria and dynamic pricing, where constraints themselves respond to the agent's actions.

What carries the argument

The key mechanism is a Lipschitz continuity property of the projection map onto varying sets, with an explicit Lipschitz constant that depends on how the feasible sets change. This property quantifies how much the projected point can move when the constraint set itself shifts, and it is what makes the coupled decision-distribution dynamics tractable in the convergence proof.

What would settle it

Construct a decision-dependent problem with two feasible sets that switch discontinuously as a function of the decision, for which the projection map is not Lipschitz, and demonstrate that RPGD either cycles or diverges instead of approaching a constrained equilibrium.

Watch

Extended reading notes

Core claim

The central claim is that projecting onto a sequence of evolving feasible sets does not prevent convergence, provided the projections satisfy a Lipschitz continuity property with an explicitly given Lipschitz constant. RPGD performs a gradient step and then projects onto the current feasible set; the paper bounds the distortion caused by the moving projection sets and uses that bound to give sufficient conditions on step size and set variation for convergence to the constrained equilibrium. This contrasts with dual ascent, which may leave the feasible region during iteration and is computationally heavier.

Load-bearing premise

The convergence guarantee relies on the evolving feasible sets being regular enough that the projection map has a finite Lipschitz constant; if real-world constraints shift discontinuously or too sharply, this assumption may fail and the guarantee does not apply.

Editorial extensions

If this is right

If the sufficient conditions hold, RPGD can be used in decision-dependent settings while maintaining feasibility at every iterate, unlike dual ascent.
The explicit Lipschitz constant gives practitioners a way to check whether their problem's constraint variation is mild enough for the convergence guarantee to apply.
RPGD lowers computational cost by avoiding dual variable updates, making it attractive for large-scale market and pricing problems.
The proof framework suggests that similar guarantees could be derived for other first-order methods that rely on repeated projections onto changing constraint sets.

Reading between the lines

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

The Lipschitz constant of projections onto varying sets may connect to sensitivity analysis of variational inequalities with parameter-dependent feasible regions, potentially offering a unified view of equilibrium convergence.
A natural extension would be to stochastic gradients or inexact projections; the explicit Lipschitz bound could predict how noise or approximation error propagates under changing constraints.
One testable implication is that when the Lipschitz constant is large, convergence slows in practice; experiments could verify whether the bound is tight on the market and pricing problems considered.
The approach may extend to nonconvex objective functions if the equilibrium is understood as a game-theoretic fixed point rather than an optimum, although the paper's sufficient conditions likely target convex or well-behaved settings.

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.

Desk Editor's Note

read the letter

Quick take: I'm reading only the abstract, so my verdict is "needs full text," not "flawed." The paper proposes RPGD, a repeated projected gradient descent for decision-dependent optimization, and claims a Lipschitz continuity property for projections onto varying feasible sets with an explicit constant. That is a real technical contribution if it holds. It also promises continuous feasibility and lower computational burden than dual ascent. Those are meaningful claims.

Where the soft spot sits: the convergence guarantee hangs on a Lipschitz constant for the map P_{C(x)}(z) as x moves. That map is not automatically Lipschitz. It can jump at points where the active set changes or the constraint Jacobian loses rank. You need something like a uniform Slater condition or bounded curvature to control it. The abstract says the constant is "explicitly given" but doesn't state the conditions. So the central theorem's hypotheses are the blind spot. If they're narrow, the method works on a limited class; if they're broad, that's good but unproven from what I can see. The numerical experiments on a market problem and dynamic pricing might be benign cases, so they wouldn't settle it either.

I can't assess the citation pattern or the derivation without the full text. Nothing in the abstract raises a red flag there. The paper reads like honest work—the claims are precise enough to check.

Bottom line: deserves a serious referee, not a desk reject. The referees need to press on the regularity conditions and ask for the explicit Lipschitz constant and the theorem statement before the numerical section. If those hold, this is a solid subfield contribution. If they don't, it's a narrow result overclaimed. Either way, I'd want to see the theorem statement before citing it.

Referee Report

3 major / 3 minor

Summary. The paper proposes a repeated projected gradient descent (RPGD) method for decision-dependent optimization, where both the objective and linear constraints respond to the decision. The core technical claim is a Lipschitz continuity property for projections onto decision-dependent feasible sets, with an explicitly given constant. From this, the authors derive sufficient conditions for convergence of RPGD to a constrained equilibrium point. The paper also claims that, unlike dual ascent, RPGD maintains continuous feasibility and has lower computational burden. Numerical experiments on a market problem and a dynamic pricing problem are reported in the abstract as validation. The full text was not available for this review; consequently the assessment is based solely on the abstract and the accompanying reviewer notes.

Significance. If the results are correct, the paper would make a useful contribution to decision-dependent optimization by providing a projection-based method with an explicit Lipschitz constant for varying feasible sets and a convergence guarantee to an equilibrium. The promise of continuous feasibility is practically appealing relative to dual-ascent methods, which typically produce infeasible iterates. The explicit Lipschitz constant is a notable strength if the derivation is valid, as it would make the sufficient conditions checkable. However, because the abstract contains no equations, no theorem statements, and no experimental details, the significance cannot currently be assessed beyond this conditional statement. The numerical claims in particular are unverified from the submitted material.

major comments (3)

[Abstract] The central convergence claim is stated without any of the hypotheses needed to substantiate it. The paper asserts a Lipschitz continuity property of projections onto varying sets with an explicit constant, but does not state the required regularity conditions on the family of feasible sets. For constraints of the form C(x) = {y : g(y,x) <= 0}, the projection operator P_{C(x)}(z) is not Lipschitz in x in general; failures occur when active sets change discontinuously or the constraint Jacobian loses rank. Without conditions such as a uniform Slater condition, bounded curvature, or active-set stability, the stated Lipschitz constant may be infinite. The abstract must either state these conditions or point to the precise theorem in the full text. As submitted, the central claim is unverifiable.
[Abstract] The notion of a 'constrained equilibrium point' is not defined. Convergence to an equilibrium is only meaningful if the equilibrium is defined independently of the algorithm, e.g., as a solution to a variational inequality or a fixed point of a decision-dependent best-response map. If the equilibrium is instead defined as the limit of RPGD, then the convergence claim becomes circular. Please provide a formal definition and clarify the relationship between the equilibrium and the original optimization problem.
[Abstract] The claim that RPGD 'reduces the computational burden' relative to dual ascent is not supported by any experimental detail in the submitted material. No mention is made of the number of iterations, per-iteration cost, projection cost, or benchmark implementations. The numerical experiments on a market problem and a dynamic pricing problem are listed but not described. At minimum, the full text should include a cost comparison with dual ascent under identical stopping criteria and accuracy levels, and the abstract should summarize the key comparison result if it is to be claimed.

minor comments (3)

[Abstract] The phrase 'explicitly given Lipschitz constant' is promising but the constant itself is not shown. Consider stating in the abstract the dependence of the constant on problem data (e.g., Lipschitz constant of the constraint map, Slater bound, condition number), so that readers can judge whether the constant is practically usable.
[Abstract] The claim of 'continuous feasibility throughout the optimization process' should be qualified. It may mean feasibility at every iteration after the projection step, or feasibility of the entire continuous-time trajectory. Clarify which notion is intended.
[Abstract] The comparison 'existing dual ascent method' would benefit from a specific reference or a brief statement of the dual ascent variant used. This would help readers situate the claimed advantage.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detectable from the abstract alone; no derivation or equations are available to inspect.

full rationale

This is an abstract-only review. The abstract states the paper proposes repeated projected gradient descent (RPGD), proves a Lipschitz continuity property for projections onto varying sets, and gives sufficient conditions for convergence to a constrained equilibrium point. No equations, definitions, or proof steps are included, so there is no text by which any claimed derivation can be shown to reduce to its own inputs. The concern that the Lipschitz constant may require strong regularity conditions is a correctness/robustness question, not a circularity question, and the equilibrium point is not described in a way that suggests it is defined in terms of the algorithm's convergence. Without the full text, no specific circular step can be exhibited, and per the review rules, absence of visible circularity should be recorded as a non-finding rather than a suspicion-based positive score.

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

The abstract provides no explicit parameters, axioms, or invented entities. The axioms listed here are inferred from the abstract's claims and from standard requirements of projected gradient descent methods; the full paper may state additional or different assumptions. The 'explicitly given Lipschitz constant' is a derived quantity, not a fitted parameter.

assumptions (3)

domain assumption The feasible sets evolve with a projection that is Lipschitz continuous with a finite constant uniformly over time
The abstract states this property is shown, but its validity depends on the specific geometry of the feasible sets, which is not described in the abstract. Without this property, the convergence proof would not hold.
domain assumption A constrained equilibrium point exists and the algorithm's iterates remain in the domain where gradients are defined
Convergence to an equilibrium presumes the equilibrium is a well-defined target, and projected gradient descent requires gradients of the objective to be available at iterates. Neither is specified in the abstract.
standard math Standard convexity or smoothness conditions typically required by projected gradient descent
PGD convergence usually requires the objective to be convex or smooth and the feasible sets to be convex. The abstract does not list these, so they are inferred as background assumptions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Projected Gradient Descent for Constrained Decision-Dependent Optimization." pith.science (2026). https://pith.science/paper/Z5SCDLI2

@misc{pith2026250808856,
  author       = {Pith},
  title        = {Pith review of: Projected Gradient Descent for Constrained Decision-Dependent Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5SCDLI2}},
  note         = {Machine review of arXiv:2508.08856}
}

read the original abstract

This paper considers the decision-dependent optimization problem, where the data distributions react in response to decisions affecting both the objective function and linear constraints. We propose a new method termed repeated projected gradient descent (RPGD), which iteratively projects points onto evolving feasible sets throughout the optimization process. To analyze the impact of varying projection sets, we show a Lipschitz continuity property of projections onto varying sets with an explicitly given Lipschitz constant. Leveraging this property, we provide sufficient conditions for the convergence of RPGD to the constrained equilibrium point. Compared to the existing dual ascent method, RPGD ensures continuous feasibility throughout the optimization process and reduces the computational burden. We validate our results through numerical experiments on a market problem and dynamic pricing problem.

Discussion (0). Sign in to comment.

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.