acceptodds
Withdrawn from ICLR 2027

Boundary Robustness Is Nonlocal: A Kirchhoff Law for Balanced Assignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local margins do not determine boundary robustness once a partition must give each region a fixed share of the data, as balanced Sinkhorn self-labelling, balanced token-to-expert matching and balanced transductive inference all do. We build two such partitions whose interface is identical in every local quantity, to machine precision in margin, density and conductance, yet which respond to the same shift by amounts differing by ; the ratio can be made arbitrarily large. The mechanism is exactly electrical: differentiating the mass constraints, the known implicit derivative of the semi-discrete dual, turns a shift into a current injection on the region-adjacency graph, so the potentials re-equilibrate through a Kirchhoff system and each interface moves by the voltage drop across its own edge over its local margin. Read one interface at a time this gives a per-unit-shift susceptibility , an effective resistance over the local margin, whose network term is tight, and a sharp impossibility: local data does give a valid certificate for a current across the interface, the isolated-edge quantity with the interface conductance, but no local functional can tell whether it is informative and no local point predictor is uniformly accurate. The same sensitivity and the same impossibility hold for entropic Sinkhorn assignments, whose conductances are the closed-form co-assignment mass. Exact re-solution confirms the law through three perturbation channels and a constraint-off knockout removes the effect. On a self-labelling head, natural-image partitions and annotated tumour sections the law's trapezoidal form orders the interfaces that move at Spearman per unit and beats the local current-conditioned competitor on every unit, while calibrated magnitudes degrade outside the linear regime, the shift-agnostic susceptibility ranks no better than a local score before the shift is seen, and warm-started Sinkhorn iteration wins on magnitude at higher cost.

Primary area:
learning theory
Keywords:
optimal transport, semi-discrete optimal transport, balanced assignment, Sinkhorn, effective resistance, graph Laplacian, distribution shift, sensitivity analysis, impossibility result, decision boundaries

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_YkpCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how assignment boundaries respond to changes in data, assignment costs, or target group proportions when each group must receive a prescribed share of the data. It argues that local boundary properties alone cannot determine this response because maintaining the mass constraints couples the entire partition. Building on established optimal-transport differentiation, the authors express boundary movement through a graph-Laplacian system and derive an effective-resistance sensitivity bound and limitations of local predictors. They extend the analysis to entropic Sinkhorn assignments and evaluate it using synthetic partitions, image features, and spatial transcriptomics data.

Critical strengths
  • The paper investigates a potentially useful question: how fixed mass constraints move an assignment boundary depends on the surrounding partition.
  • The supplementary controls and explicit reporting of negative results are useful. In particular, the authors distinguish first-order guarantees from finite-shift behavior, acknowledge the poor real-data performance of the shift-agnostic score, and compare the current-conditioned predictor with warm-started Sinkhorn iterations.
Critical weaknesses
  • My main concern with this paper is the writing. I am generally in favor of using LLMs to improve the quality of papers and research. But it is so hard to understand this paper as a human without using an LLM. For example, the abstract refers to “boundary robustness,” “the same shift,” and an interface “identical in every local quantity” before adequately explaining the measured response, the perturbation, or the information held fixed. It then states that “the mechanism is exactly electrical” and introduces current injections, potentials, and voltage drops before making the underlying learning problem sufficiently clear. (I thought I was assigned to the wrong paper here!!) The introduction’s scope paragraph references Equation (1), Assumption 1, and several subsequent results, making its technical qualifications difficult to understand without first reading the method. The reader should not need the whole paper as context to understand its motivation and scope.
  • Confusing terminology The paper contains a lot of clauses, poetic terminology, and defensive sentences, and it seems to be a generated output from an LLM chat. For instance, “The same thing without a hard interface,” “Is it really the effective resistance?”, “The circuit on real systems, against the honest competitor,” and “A head that is trained, and an encoder that drifts.” In particular, “honest competitor” seems interesting!
  • Novelty and contribution of the paper is not clear The paper explicitly acknowledges that the transport mass Jacobian, its Laplacian structure, and implicit differentiation are established. Theorem 2 interprets that derivative at an individual interface, Theorem 3 applies Cauchy–Schwarz in the Laplacian metric, and Proposition 5 uses Rayleigh monotonicity. These connections may provide a useful perspective, but the paper needs to establish more clearly what substantial advance follows from them. Repeatedly describing the “per-interface reading” as new does not adequately explain its technical significance relative to existing sensitivity analysis.
  • The empirical results do not support the main contribution of the paper. The shift-agnostic score is not significant on any of the three real systems and is outperformed by the local certificate on the two vision systems. Its strongest synthetic results use an energy-isotropic perturbation ensemble constructed from the same Laplacian, under which its RMS prediction is theoretically correct by construction. The useful real-data result is primarily ranking boundary movements after the shifted data are available. Even there, strong rank correlation does not imply accurate displacement magnitudes: the paper reports negative R^2 outside the small-shift regime.
Questions for authors
  • Can you explain the learning problem, perturbation, boundary-motion observable, and practical consequence in one paragraph without electrical terminology, and use that explanation to restructure the abstract and introduction?
  • What is the main technical advance beyond known implicit differentiation and effective-resistance results?
  • Do the findings hold when balanced targets are recomputed during training, with different random seeds and many more prototypes?
Reviewer_PZsFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how assignment boundaries move when each class must retain a fixed mass. The authors derive the response of an individual interface from the voltage drop across an edge of a graph. Two interfaces with the same margin and conductance can then respond quite differently to the same shift. The differentiation step builds on prior work in semi-discrete transport and Sinkhorn balancing.

Critical strengths
  • The effective-resistance bound in Theorem 3 is clean. For current injected directly across the edge, the voltage drop is exactly the effective resistance. The synthetic experiments verify this case convincingly.

  • Useful distinction between ranking and magnitude. On CIFAR-10, the proposed method ranks interfaces well even when its displacement estimates have negative (R^2). Both results are reported.

Critical weaknesses
  • The definition of “local” is narrower than the main claim suggests. In Figure 1, changing the cap height HH changes the angle at which the third cell meets B12B_{12}. That angle is visible in any neighborhood of the endpoint. The two examples are indistinguishable under Definition 4 because it excludes the third cell. A local observer of the partition could distinguish them.

  • The claim of a common velocity for affine cost differences is incorrect under cost drift. Consider two cells of equal mass on the uniform unit disk, with c1(x,y,t)−c2(x,y,t)=x+tyc_1(x,y,t)-c_2(x,y,t)=x+ty. The boundary rotates about the origin. At t=0t=0, its normal velocity at (0,y)(0,y) is V12(0,y)=−yV_{12}(0,y)=-y. The barycentre gives zero, while the maximum speed is sup⁡∣y∣≤1∣V12(0,y)∣=1\sup_{|y|\leq 1}|V_{12}(0,y)|=1. Table 2 evaluates drift at one barycentre per interface. This can miss the motion the theorem bounds.

  • Theorem 8 gives an O(t2)O(t^2) remainder for the potentials. The extension to boundary position needs stronger spatial regularity. On [−1,1][-1,1], take the cost difference, density, and target mass to be

    q(x)=\frac{g'(x)}{4}, \qquad m_1(t)=\frac{1}{2}+\frac{t}{4}.$$ The conductance is constant and $w_1(t)-w_2(t)=t$ exactly. For $t>0$, the boundary position satisfies $x(t)+x(t)^{3/2}=t$. Its linear prediction is $t$, so the absolute error is $|x(t)-t|=\Theta(t^{3/2})$ as $t\to 0^+$. This occurs even though $g'(x)\geq 1$ throughout the domain.
  • The geometric interpretation of Theorem 7 is unclear. Consider squared-distance costs ck(x)=(x−μk)2c_k(x)=(x-\mu_k)^2 with prototypes (μ1,μ2,μ3)=(−1,0,1)(\mu_1,\mu_2,\mu_3)=(-1,0,1) and equal potentials. The outer pair has positive entropic conductance, h13ε>0h_{13}^{\varepsilon}>0. Its scores tie only at x=0x=0, where the middle class wins. Which pairs in the real-data rankings actually share a decision boundary? The reported 2.112.11-fold variation also leaves the arbitrary-factor claim for entropic assignments unproved.

Questions for authors

Can the response ratio in Theorem 6 diverge while all cell masses stay uniformly positive? The planar example allows small cells. The alternative bridge construction grows (K), so the smallest mass is at most (1/K).

How does the predictive trapezoid compare with one Newton step on the shifted problem? Equation 12 already computes the shifted Jacobian at the clean potentials. This would be a direct comparison at essentially the same cost.

Reviewer_K4adOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates boundary robustness in mass-constrained assignment systems, focusing on whether local boundary information can determine sensitivity to distribution shift. In balanced clustering, Sinkhorn self-labelling, and other constrained allocation settings, the assignments must satisfy prescribed group masses. A perturbation can therefore change the global balancing potentials and displace a boundary even when its local margin, density, and conductance are unchanged. The authors argue that robustness in these systems is a nonlocal property governed by the connectivity of the entire assignment network.

The authors express the derivative of the mass constraints as a weighted graph Laplacian and interpret the resulting implicit sensitivity equation as a Kirchhoff circuit law. The change in each boundary is then given by a difference of balancing potentials divided by its local margin, while effective resistance characterizes the network contribution. They further formulate an impossibility result for uniformly informative local predictors and extend the analysis to entropically regularized Sinkhorn assignments. Experiments use synthetic partitions, a CIFAR-10 self-labelling head, image-region partitions, and spatial transcriptomics data. The circuit predictions closely match infinitesimal changes and rank displaced boundaries well after a shift is observed. However, finite-shift displacement magnitudes become poorly calibrated, and the shift-agnostic resistance score does not consistently outperform local baselines on real data.

Critical strengths

The paper presents a coherent mathematical interpretation of how exact mass-balance constraints couple assignment boundaries. The electrical-network analogy makes the role of the global constraint explicit, and the controlled examples distinguish otherwise identical local interfaces by changing remote connectivity. The authors also acknowledge prior work on semi-discrete optimal transport derivatives rather than claiming the underlying Laplacian identity as new. The experiments report both successes and negative findings, including failure of the shift-agnostic score on real datasets. These aspects make the technical discussion informative, although they do not resolve the central concerns about the importance and novelty of the research objective.

Critical weaknesses

1.Limited problem significance and unclear practical implications. The central observation—that global mass constraints induce nonlocal coupling between assignment boundaries—is largely intuitive. While the paper provides a mathematically precise characterization through graph Laplacians and effective resistance, it remains unclear why predicting boundary displacement is an important objective in practical learning systems. The experiments primarily validate the proposed sensitivity equations, without demonstrating meaningful improvements in downstream robustness, assignment quality, or learning performance. Moreover, the shift-agnostic predictor fails to outperform local baselines on real datasets, further limiting the practical significance of the framework. 2.Limited theoretical novelty and a flaw in the stated impossibility bound. The mass-Jacobian Laplacian and the implicit differentiation of balancing potentials are established results, as the paper itself acknowledges. The effective-resistance bound follows from standard Laplacian identities and Cauchy–Schwarz; consequently, the new contribution mainly lies in its per-boundary interpretation. More importantly, Theorem 6 appears to overstate the multiplicative error of any shared point predictor. If two positive responses differ by a factor of ρ, a prediction at their geometric mean has worst-case multiplicative error √ρ, not ρ as the theorem asserts. The arbitrary-ratio graph construction also deserves a clearer connection to realizable balanced partitions. These issues affect the paper’s main claimed theoretical contribution. 3.The empirical validation does not establish practical utility. Much of the synthetic evaluation demonstrates that a first-order derivative agrees with sufficiently small perturbations of the same constrained optimization problem; this is an important correctness check but not, by itself, evidence of a useful learning method. On real data, the reported rank correlations are strong only after the shift is measured, while magnitude predictions degrade substantially for larger perturbations. Warm-started Sinkhorn updates are also more accurate on magnitude, and realistic large-scale settings with many prototypes or production mixture-of-experts routing are not evaluated. The paper does not connect boundary rankings to classification error, routing failures, robustness interventions, or measurable downstream gains.

Questions for authors

The paper characterizes which balanced-assignment boundaries will move under a measured distribution shift, but it does not establish why boundary displacement should be optimized or predicted in a deployed learning system. In several of the motivating applications, balance is imposed only during training, and the inference-time mechanism does not necessarily satisfy the assumptions of the analysis. Consequently, the relationship between the proposed diagnostic and a practically consequential failure remains unclear. The negative results for shift-agnostic prediction further narrow the practical use case.

Could the authors identify a concrete learning or deployment decision that benefits from predicting interface displacement, and demonstrate a resulting improvement in accuracy, robustness, assignment quality, or another relevant downstream metric? Why should this circuit-based analysis be preferred to simply warm-starting and updating the balanced assignment once the shifted data are available? In addition, could the authors revisit Theorem 6: under the stated multiplicative-error definition, a response ratio of ρ appears to imply only a √ρ minimax error for a common point prediction. Please clarify the claimed factor and distinguish what is genuinely new from the established implicit differentiation and graph-Laplacian results.

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

Related papers

Discussion (0)

Sign in to comment.