acceptodds
Withdrawn from ICLR 2027

Certified Selective Tracking of Moving KKT Systems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A small KKT residual can arrive after a deadline or correspond to the wrong active set or stationary branch. We introduce Certified Selective Tracking (CST), which delivers a result only when both deadline and branch-identity certificates hold. Its Geometry-aware Certified Trackability (GCT) instantiation encloses active-set events under an explicit uncertainty set. Strict separation of the event tubes certifies event membership, order, and endpoint identity. Within a finite candidate family, selection maximizes delivery availability under the same strict certificate contract; simultaneous envelope coverage bounds false-delivery risk. PTIF-KKT supplies inversion-free candidate updates. On 470 held-out PJM-driven arrivals in a constructed separable QP, public interval residual verification retains 134 deliveries at 5% curvature uncertainty with zero observed false identities. A true-model verifier retains 186, but 52 of these fail at an admissible curvature corner. Point-event selection makes 32 endpoint-label errors. A separate eight-dimensional heuristic measures agreement with offline multistart labels, not nonlinear branch certification or general solver superiority.

Primary area:
optimization
Keywords:
time-varying optimization, KKT tracking, branch identification, interval verification, selective certification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ppHsOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

I believe the goal of the paper is to establish some sufficient condition under which a convex optimiser can guarantee that it will converge faster than the rate at which the problem optimum moves due to change in the problem parameters (objective or constraints, or both, I'm not sure). The paper also talks about something called "branch identity", but this concept is never defined.

Critical strengths

I am not able to identify any.

Critical weaknesses

The paper is incomprehensible (to me). It does not define or introduce any of the terminology it uses, not even that used to state what it aims to accomplish.

Questions for authors

Not a question for the authors, really, but is there anyone on the PC who is interested in reviewing this paper?

Reviewer_eHE6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors aim to address:

  1. An optimisation problem that varies over time, requiring a solution that satisfies the KKT conditions to be found before a ddl
  2. A small KKT residual alone is insufficient, as it may indicate that an incorrect activity set or stationary branch has been tracked
  3. SO the authors verify separately whether the computation can be completed on time and whether the tracked branch is correct
  4. A result is only output when both proofs hold, otherwise, no decision is made.
Critical strengths

The paper addresses a meaningful issue in time-varying constrained optimization: a small KKT residual alone does not guarantee timely convergence or correct branch identification. Considering both requirements is potentially useful for real-time optimization applications.

Critical weaknesses

1. Firstly, the authors have not identified a sufficiently specific research gap.

The Introduction begins by stating: A small KKT residual may occur after a dll or correspond to the wrong active set or stationary branch.

Whilst this is a valid issue, the authors have not specifically demonstrated which existing methods are prone to this problem, nor why current approaches are unable to detect or address it.

In particular, both deadline and active-set tracking have a long history of research.

The authors must clarify which technical challenge, one not yet adequately addressed, they resolved.

2. The authors have organised many relatively straightforward conclusions into a complex theoretical framework. For example:

  • Theorem 8 proves that when the drift margin is sufficiently large, the residual can be reduced before the deadline.
  • Theorem 9 proves that when both the deadline and identity certifications are passed simultaneously, the result satisfies both requirements.
  • Proposition 10 proves that under strict certification rules, selecting candidate solutions that pass certification can maximise delivery availability. Theorem 8 includes a practical dynamical analysis and cannot simply be regarded as a trivial conclusion.

However, the core logic of Theorem 9 and Proposition 10 is relatively straightforward. It is difficult for me to understand why these need to be elevated to the status of significant theoretical contributions of the paper.

3. There is another issue: is this a theoretical paper or an algorithmic paper? This paper attempts to provide the following simultaneously:

  • A new form of residual dynamics
  • An active-set event certification method
  • A non-convex branch identification method
  • a theory of selective delivery
  • A framework for global root verification.... Each of these involves a relatively independent field of research.

However, the main experiment consists of only one 6-variable separable QP, and the various theoretical sections have not been validated to the same extent.

Questions for authors
  1. Poor readability and presentation. The paper introduces numerous definitions, certification conditions, and theoretical results without a sufficiently clear explanation of how they constitute a methodological framework. The relationships among PTIF-KKT, GCT, branch-lock certification, and the proposed delivery router are particularly difficult to follow.

  2. The experiments are highly inadequate. I have the following concerns regarding the current, woefully limited set of experiments:

Q1: the certification rate drops too rapidly. A 5% uncertainty level results in approximately 71.5% of instances being rejected, whilst at 20%, certification becomes entirely impossible.

If the robot must continuously make control decisions, this abstention would pose practical difficulties in real-world applications. The paper should discuss how to handle these uncertified instances and whether the certification mechanism is overly conservative.

Q2: The experimental problems are too simple.

The main experiment use a QP problem with only six variables. Although it uses real PJM load data to drive problem variations, the optimisation task itself is artificially constructed. And i think the additional 8-dimensional experiment is merely a heuristic evaluation and does not provide a branch certificate for general non-convex problems.

Q3: the computational advantages of PTIF-KK

  • Is the solution speed faster compared to traditional optimisation solvers?
  • How much additional computation time does the certification itself require?
Reviewer_2tvyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes tracking KKT residuals under time-varying objective and constraints (both equality and inequality) and certifying the results under a maximum time budget.

Critical strengths

Time varying constrained problems are important class of problems as they are very general, especially under black-box time varying constraint functions, where obtaining certificates is an interesting research topic.

Critical weaknesses

Unfortunately, the paper is badly written and hard to parse.

It start with the first sentence in the abstract ("A small KKT residual can arrive after a deadline or correspond to the wrong active set or stationary branch."). it is both unclear what it means for a KKT residual to arrive after a deadline, what deadline refers to, what is meant with active set or stationary branch. It continues in the next sentence, where it is unclear what branch-identity certificates are. The abstract continues with sentences that are unclear, like "Within a finite candidate family, selection maximizes delivery availability under the same strict certificate contract;".

The same theme continues in the introduction:

  • page 1, line 35: unclear what is meant for a constrained problem to be "tracked" before a prescribed deadline. Also, what is meant with deadline is not clear here (a max time-budget)
  • page 1, line 38: Fischer–Burmeister residual E(y, t) is mentioned without description/definition
  • page 1, line 43: "static nonlinear programming studies convergence on a frozen instance" unclear what a frozen instance is
  • page 1, line 46: PTIF-KKT is mentioned without explanation
  • page 1, line 46: unclear what generalized Jacobian selection refers to
  • page 1, line 51: "Ψκ is a norm-scaled prescribed-time residual correction" not explained or defined

and also the main part of the paper is similarly unclear:

  • page 2, line 102: "our target is residual tracking of KKT points, not global optimality" despite being central to the paper, this is not defined to vaguely stated in prose
  • page 2, line 104: unclear what "inequality complementarity" refers to
  • page 3, line 111: the relationship to FB is not given and never explained
  • page 3, line 118: FB-KKT residual is never formally defined, but used in the proposition
  • page 4, line 166: the differential inclusion appears out of nowhere and is never explained or derived and is central to the paper
  • the entire 4.2 (continuous dynamics) is unclear. It needs to be properly derived and explained
  • page 4, alg 1: the algorithm is not proper pseudo-code
  • page 4, line 215: "Geometry-aware Certified Trackability (GCT) implements the event-complete identity requirement for a piecewise-affine KKT path." is it unclear what event-complete is, what the identity requirement refer to and what is meant with KKT-path
  • page 4, line 215: "public path" is not defined

This theme continues into the experimental section, where the main problem is not properly explained ("We evaluate CST on 470 chronological held-out PJM load arrivals driving a constructed six-variable separable QP. "). The term PJM is never defined.

The vague, unclear writing makes it hard to follow and understand the paper, what it is about and to review it.

Questions for authors

see above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.