acceptodds
Withdrawn from ICLR 2027

FENCE: Enforcing Mixed Functional Constraints in Neural Fields

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Enforcing constraints at sampled points does not guarantee that a neural field remains admissible throughout the domain. We introduce Functional Equality– iNequality Constraint Enforcement (FENCE), which assembles linear field functionals, basis envelopes, and cover-dependent variation margins into a finite convex feasible set enforced by a standard differentiable QP layer. For representations with valid enclosure or variation bounds, this jointly enforces equalities and continuous-domain inequalities without a nonlinear inner solve. On heat design, sufficient constraints pass all independent dense and finite-difference checks across held-out tasks and seeds. Constraint exchange preserves full-system objectives while cutting final QP size by over two orders of magnitude on a development cover sweep. On motion retargeting, derivative envelopes meet continuous rate limits on every held-out clip where sampling misses inter-sample violations, with < 0.1% mean objective overhead relative to analytic exchange. Matched QP implementations confirm that coverage follows from the constraint construction

Primary area:
optimization
Keywords:
Hard Constraint Enforcement, Mixed Equality–Inequality Constraints, Differentiable Convex Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B2ZbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper concerns constrained neural fields, where constraints are expected to hold throughout the domain described by the field, rather than only at sampled points. The authors propose QP projections in the field's coefficient space, using basis bounds and cover-dependent margins to provide continuous-domain feasibility under representation-specific assumptions. The approach is tested on several problems, including motion-trajectory correction, temperature matching, and topology design.

Critical strengths
  • The work addresses an interesting problem. The authors apply bounding principles to neural-field representations, providing a handle for enforcing constraints over an entire continuous domain.
  • Independent checks and matched QP controls help distinguish the effects of the constraint construction from those of the solver.
Critical weaknesses

In its current form, I find the paper inadequate for publication at ICLR, primarily because of the presentation issues below.

  • Overall, the exposition is unclear. Concepts and notation are presented with insufficient context, and the paper does not read as self-contained.

    • The abstract devotes much of its text to experimental results before making the problem and proposed approach clear.
    • The introduction is very short and does not sufficiently explain the problem, how it has been addressed before, or the remaining gap. Equations and notation in lines 033 and 040 appear before adequate explanation.
    • The related work section is dense and reads largely as a list of papers, with insufficient technical positioning of the proposed work.
    • Section 3 is difficult to follow. Some notation is used without being explicitly defined in the main text, leaving the reader to infer its meaning or consult the appendix. Several key concepts are also used with insufficient explanation. The accompanying figures do not sufficiently clarify the construction.
    • The comparisons in Section 4 are confusing, particularly regarding which experiments use pointwise constraints, which use continuous-domain bounds, and which methods receive identical constraints.
    • What I could follow relied heavily on my own familiarity with convex-hull bounds and cover-based arguments. The paper uses these concepts without enough explanation to make the exposition self-contained.
  • Overall, the contribution is unclear. The authors acknowledge that the bounding principles, QP projection, and exchange methods are established. What is new in how they are combined or used? The contribution list should distinguish the formulation, application-specific constructions, interface, and empirical analysis more clearly.

  • Convex-hull bounds and enclosures can require stricter constraints, reducing the feasible space. The paper acknowledges and measures this conservatism, but its implications for the results and comparisons need clearer explanation. In particular, the cost of continuous feasibility should be distinguished from the additional cost of a conservative approximation.

  • The benefit of learning is not demonstrated. Direct heat optimization achieves a lower objective, with no consistent inference speedup from the learned encoder. The motion experiments do not train an encoder. The results therefore support the constraint constructions more clearly than the learning component.

Questions for authors

Please address the issues above. In addition, some more granular questions and comments:

  • Line 016: What are “enclosure or variation bounds,” and what assumptions do they require?
  • Line 031: What does “what part of the field is covered” mean here?
  • Line 033: What do these coefficient equalities represent? Please introduce the notation and clarify the restriction to affine coefficient constraints.
  • Line 040: What are the mixed field constraints? Please explain the notation.
  • Line 043: What does “supply linear rows” mean?
  • Line 044: Please explain how “a cover radius times a field-variation bound” supplies sufficient constraints.
  • Line 102: Please explain “a QP layer with identical rows has the same feasible set; our experiments include this control.” What does this comparison establish?
  • Line 106: Please briefly explain exchange methods and distinguish finite-row exchange from continuous constraint separation.
  • Line 107: Please explain the Bernstein convex-hull enclosure and why its constraints are sufficient rather than necessary.
  • Table 1: Please explain what is presented, particularly the distinction between field interfaces and algebraic QP rows, and what the feasibility column establishes.
Reviewer_JAJqOpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FENCE represents neural fields through basis coefficients, assembles linear functional equalities and inequalities, and enforces them using a standard differentiable QP projection. For suitable representations, basis envelopes and cover-dependent variation margins provide sufficient conditions for continuous-domain feasibility. Experiments cover heat design, motion correction, and topology optimization.

Critical strengths
  1. Continuous coverage is treated carefully. The paper distinguishes sampled feasibility from continuous-domain guarantees and explicitly states the assumptions behind its sufficient bounds.
  2. Numerical validation is thorough. Independent dense and finite-difference checks, analytic spline-extrema verification, and matched QP implementations strengthen confidence in the constructions.
  3. The application-specific constructions are useful. The sine-to-Bernstein transformation, temporal derivative envelopes, and coverage–conservatism tradeoffs provide reusable examples. Constraint exchange substantially reduces the working QP size while checking the complete finite constraint system.
Critical weaknesses
  1. Linear-functional assembly follows CNF, mixed algebraic constraints are handled by established differentiable QP layers, and Bernstein/B-spline envelopes, variation margins, and constraint exchange are standard techniques.The manuscript fails to demonstrate a sufficiently distinct new theoretical finding, algorithmic capability, or class of constraints that cannot be handled by prior approaches. This constitutes the primary reason for my inclination to reject the paper under the novelty-first criterion.

  2. The learned predictor has no demonstrated advantage over direct optimization. Predict-and-project shows no consistent inference advantage: reported median times are 7.69 ms versus 7.57 ms for direct optimization, excluding common assembly. Thus, the experiments establish the value of the constraint construction more clearly than the value of learning.

Questions for authors
  1. What specific technical obstacle remains unresolved by combining CNF-style assembly, standard differentiable QP, and established sufficient bounds, and how does FENCE overcome it?

  2. In which demonstrated regime does predict-and-project outperform direct task optimization after accounting for objective quality, inference cost, and training cost?

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

This paper investigates how to make neural fields satisfy both equality and inequality constriants, and guarantee that a neural field remains admissible throughout the domain rather than only on sampled points. The authors propose FENCE, a framework represents neural fields as linear combinations of basis functions, converts functional constraints into linear constraints on the coefficients, and enforces them through a differentiable quadratic programming. To satisfy constraints on the whole continuous domain, FENCE leverages basis envolops and cover-dependent variation bounds. The method was evaluated on three tasks: motion retargeting, heat design and topology optimization. It can effectively enforce the constraints throughout the entire domain.

Critical strengths
  1. Theoretical Guarantees. The authors provide mathematical justification for continuous-domain feasibility using basis envelopes and cover-dependent variation bounds.

  2. The experiments show good support for continuous-domain constraint preserving.

Critical weaknesses

I think most weakness has been discussed in the paper.

  1. Generalizability. Although the authors discuss the possibility of using learned basis functions, it remains unclear how these guarantees can be extended to general learned neural fields.

  2. Conservatism of Sufficient Constraints. The continuous-domain guarantees may exclude feasible solutions which could limit the quality of the optimized neural fields. Although the authors investigate the trade-off between conservatism and density, it remains unclear if the trade-off would be manageable for more complex neural fields.

Questions for authors

None

1 comment

Thank you for your response. Currently I choose to maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarifying FENCE’s Contribution, Guarantees, and Scope

We thank the reviewers for recognizing the importance of continuous-domain feasibility and the care of our validation. The main concerns are the novelty beyond existing components, the role of learning, and the scope and conservatism of our guarantees.

Contribution beyond CNF and generic QP layers. We agree that FENCE does not introduce a new QP solver, Bernstein theorem, spline theorem, or exchange algorithm. Its contribution is the missing field-to-algebra interface. CNF compiles linear field functionals into equalities,

Aeqβ=b,A_{\mathrm{eq}}\boldsymbol{\beta}=\mathbf b,

but cannot represent one-sided bounds, budgets, or rate limits. Generic differentiable QPs accept mixed algebraic rows,

\qquad G\boldsymbol{\beta}\leq\mathbf h,$$ but do not determine how a continuum condition such as $$f_{\boldsymbol{\beta}}(\mathbf x)\leq u, \qquad \forall \mathbf x\in\Omega,$$ should be converted into a finite valid system. FENCE combines mixed functional assembly with explicit representation-dependent coverage conditions. This distinction is empirically necessary: in heat design, a QP satisfying all sampled rows passes only $1.4\%$ of independent dense/finite-difference checks, while the sufficient construction passes $100\%$. In continuous motion, both $30$-Hz and $120$-Hz sampling miss inter-sample violations on every held-out clip, whereas the spline-envelope construction passes all $7/7$ clips. Matched OptNet and direct-QP controls confirm that the gain comes from the constraint construction rather than the solver. We do not claim a new algebraic constraint class: an expert could manually derive application-specific QP rows. The contribution is a reusable mixed-functional contract, explicit coverage assumptions, and concrete spatial and temporal constructions demonstrating where sampled feasibility fails. **Learning versus direct optimization.** We agree that the current experiments do not show predict-and-project outperforming direct optimization. On heat design, median times are $7.69$ ms versus $7.57$ ms, excluding common assembly, and direct optimization obtains a lower objective. Training cost is additional. We have therefore removed any suggestion of an inference-speed or objective advantage. The encoder demonstrates that the contract can be embedded in a differentiable learning pipeline; it is not evidence that learning is preferable for these small, fully observed convex tasks. The supported claim is narrower: FENCE provides a differentiable field-level constraint layer whose feasibility is determined by the assembled contract. Whether learned priors help under incomplete observations or larger repeated problem families remains future work. **Scope of the guarantees.** The guarantees do not apply to arbitrary MLPs. They require a representation with a valid basis enclosure or variation bound. For a nonnegative partition-of-unity basis, coefficient bounds imply field bounds. Given a $\delta$-cover and a certified variation bound $L$, enforcing $$f(\mathbf x_i)+\delta L\leq u$$ is sufficient for $$f(\mathbf x)\leq u, \qquad \forall \mathbf x\in\Omega.$$ Our experiments use fixed bases. Learned features are covered only when the final field remains linear in $\boldsymbol{\beta}$ and the required bounds are established; we do not claim certification for general neural fields. **Conservatism.** These sufficient constraints are not necessary and can exclude feasible fields. We now state this explicitly and quantify the trade-off. In heat design, refining the cover improves the sufficient objective while increasing the row count; constraint exchange reduces the densest working QP to at most $89$ rows without changing the full finite-system optimum. In continuous motion, subdivision tightens the envelope, while the one- and two-subdivision variants remain below $0.1\%$ mean objective overhead relative to analytic exchange. These results show that conservatism is manageable in the evaluated representations, not universally negligible. Finally, we revised the exposition to separate three concepts clearly: functional assembly creates coefficient rows, a standard QP enforces the finite system, and a representation-specific enclosure determines continuous coverage. FENCE should therefore be read neither as a new convex solver nor as evidence that learning beats direct optimization, but as a field-level mixed-constraint interface that closes failures left invisible by sampled feasibility.

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

Related papers

Discussion (0)

Sign in to comment.