acceptodds
Withdrawn from ICLR 2027

The Curvature of a Move Set: Dirichlet–Wilson Bounds and Certified Trivialization Costs for Non-Commutative Transport on Graphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local search improves a candidate solution by applying moves from a fixed set, but its theory usually describes the search space through the objective. Moves need not commute: the two orders of a pair of moves may end at different candidates. We ask whether the resulting loop defect can be bounded from both sides, computed cheaply and certified against true optima. We model the moves as non-commutative transport on a base graph that records which move sequences are assumed equivalent. Loop holonomy measures the mismatch, and we study the minimum cost of making every loop close. A fractional packing of cycle defects and a representation-spectrum bound certify this cost from below; a spanning-tree gauge and a cycle-balanced section bound it from above. The bracket is exact on a single cycle; with the discrete metric, the cost equals a weighted frustration index on finite groups and vanishes on the state graph of any move set. Thus curvature depends on the relations the base graph assumes. We also connect the cost to three other readings: the point-level non-commutativity defect equals plaquette holonomy, the connection Laplacian separates into cut and holonomy branches with a corrected Cheeger inequality, and an ordered cohomological certificate bounds the distance a local algorithm must travel without exceeding graph distance. On 450 rotation instances a dual certifies the optima to within , and interval arithmetic proves exactness for 490 instances. The cheap lower certificates recover on average 96 to 100 percent of the certified optimum. On the commutation lattice of two rotations, the certified cost per plaquette reaches eight times the single-plaquette value on a box. These results give a computable, certifiable curvature relative to the relations a method assumes.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph holonomy, connection Laplacian, group synchronization, local search, cycle packing, certified optimization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_pb64OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies the minimum inconsistency achievable when assigning group elements to the vertices of a graph with fixed edge transformations. It derives lower and upper bounds on a weighted squared-distance objective, including an exact solution for a single cycle, and develops numerical certificates to assess how closely these bounds approximate the global minimum. Synthetic experiments illustrate the bounds and related spectral properties.

Critical strengths
  1. Computable bounds for the stated optimization problem.

    For a given graph with fixed edge transformations and weights, the paper derives lower and upper bounds on the minimum weighted squared geodesic consistency error. It also obtains an exact expression for the single-cycle case, making the dependence on the cycle transformation and edge weights explicit. These are concrete theoretical results for the objective studied in the paper.

  2. A principled way to evaluate the quality of the bounds.

    The paper develops a procedure to place the global minimum within a narrow, mathematically justified interval. This makes it possible to measure how close the proposed inexpensive lower bounds are to the true optimum.

Critical weaknesses
  1. Motivation unclear / not grounded.

    The paper motivates its work through local search and the order in which search moves are applied, but its main result is a bound on a consistency error defined on a graph. It doesn't clearly show how this bound explains search behavior or leads to a better algorithm. The bound also depends on a base graph, and it is unclear how an actual search algorithm would determine this graph.

    As a synchronization optimization problem, the work has value. But the paper doesn't make clear why the new bounds and certificates matter compared with prior work.

    Also the purpose of the separate sheaf-distance construction is unclear, the paper does not explain how it relates to the motivating search problem.

    This makes me rate 2 instead of 3.

  2. Opaque writing and missing supplementary archive.

    • The paper is difficult to follow. E.g., the abstract lists fractional packing, spectral bounds, tree gauges, cycle-balanced sections, and cohomological certificates before clearly explaining the quantity being bounded. Listing these terms does not explain what each result contributes. This is just one example out of many.

    • The reproducibility statement and Appendix H describe a supplementary archive:

      (L507) The supplementary code release provides all JSON records, scripts and one-command verification.

      (SM, L1324) The supplementary archive contains the source (paperB/), all inputs, all JSON results with per-record inputs and certificates, the verification scripts for every certificate layer, the figure script that regenerates every figure of this paper from the JSON, and one-command verification scripts with pre- and post-run provenance manifests.

      I could not find this archive in the supplied materials. No anonymous download link either.

    These further lower my rating from 2 to 1.

Questions for authors
  1. Given a real search algorithm, how should the base graph be determined, rather than picked by the researcher?

  2. Beyond certifying the consistency objective, how do these results help us understand or improve a concrete search algorithm? More broadly speaking, why should ICLR readers care about these results?

  3. (L368) So far every statement has been a theorem; we now test each one, ... and a development batch precedes a single confirmation batch (Table 2, Appendix H). Fourteen consistency checks (V0), among them the fixed-space pitfall of Proposition 4, run first and all pass.

    What are “development batch” and “confirmation batch”? What is done in each?

  4. (SM, L1293) ... the master seed 20260904 ...

    What is the point of providing this seed without the code and environment details needed to use it? or, where can reviewers access the supplementary archive?

Reviewer_fZhoOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work is a theoretical paper analyzing the curvature of a move set. It modeled the moves as non-commutative transport on a base graph, derive the upper and lower bounds on the loop defect. The paper further extend to analyze three other aspects. These optima are further justified by simulations. This is a manuscript hard to read and understand, with little presentations on motivations and downstream applications. It is also hard to relate it to any topics of ICLR.

Critical strengths

This is a theoretical paper with analysis on the move set, with coverage of a series of analysis on the non-commutative transport on a base graph, the point-level non-commutativity defect, ordered cohomological certificate bounds, etc. These theories seem to be deep.

Critical weaknesses

I have serious concerns on this work. It seems to be a hallucinated paper with some isolated theories, but without any specific topic related to the ICLR and without any downstream useful/potential applications to the representation learning.

  1. What is the motivation and objective of this research? The theories in the paper are diverse and isolated. How to understand the relations and organizations of these analysis results?

  2. It is unclear how these analysis relate to the statement "local search improves a candidate solution by applying moves from a fixed set". The curvature a move set is not well defined and clearly related to the geometry curvatures.

  3. It is hard to understand the value and objectives of these simulation results. It should present experiments directly related to a downstream applications, instead of validating the optimal bounds.

  4. The "related work" section presents a number of related research topics. However, they are not compared with this research, and how this research positioned among these related works.

I seriously concern that this is a AI generated work, and the reviewing of this work is a practical hallucination to me. I strongly suggest rejection on this work.

Questions for authors

Please clearly respond whether this is a AI-generated work. This work is not a meaningful research contribution and did not bring any insights to the community of ICLR. If it does have merits in some sense, it should be submitted to other venues fitting this research topic.

Reviewer_fNPjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is centered on the study of a discrete bundle of groups over a graph with connection in that group and in particular the study of how far it is from trivializability through the introduction of an energy function which informs on the loop holonomy of the connection. A link is proposed to cellular sheafs, in particular with connection Laplacians through a representation of the compact group, however no clear link is made to sheaf neural networks.

Critical strengths

The different approches on loop holonomy for bundles of compact groups over a finite graphs.

Critical weaknesses

The appendix seems to be missing.

The paper does not explain how its framework and results can be applied to machine learning. For example, as stated in the paper, connection Laplacians are used for defining generalizations of Graph Neural Networks, known as Sheaf Neural Networks. However, the proposed results do not explain how they can be applied to the study of such architectures (e.g., learning, expressivity).

Furthermore, for the proposed bundle, the nodes take values in compact groups, whereas it seems to me that in traditional signal processing, they take values in vector spaces.

Questions for authors

Could the authors please explain why loop holonomy and in particular their bounds on E∗E^* and Laplacian decomposition can be used for the study of sheaf neural networks.

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

Related papers

Discussion (0)

Sign in to comment.