acceptodds
Under review as a conference paper at ICLR 2027

Symmetry Descent Neural Operators: From Local Equivariance to Global Consistency

Abstract

Equivariant neural networks commonly assume that a group action and its representations are known and consistent over the entire domain. Heterogeneous materials, local defects, and domains with holes make symmetry valid only locally; the relations between local frames vary across system instances, and loops can also carry nontrivial topological information. We propose Symmetry Descent Neural Operators (SDNO), which treat local symmetries and frame alignment as episode-conditioned structures determined from operator observations. Given multiple input–output observations from the same operator, SDNO learns the feature transformations that can be shared within each local region and the feature alignment between neighboring regions. These constraints define a cross-region weight-sharing space, while curvature and holonomy represent local defects and global topology, respectively. A differentiable projection parameterizes this space as trainable neural operator layers. Our theoretical results characterize when local equivariant layers compose into a global operator and establish stability and identifiability under structure-estimation error. The experiments cover local transport with the unit complex phase group and the noncommutative special unitary group , operator prediction and topology transfer under hidden local transport, polycrystalline local frames, and a public Structure-from-Motion rotation averaging benchmark. SDNO maintains accurate structure recovery and operator prediction under unseen changes in holonomy, curvature, topology, and local frames, and improves global orientation estimation on public noisy relative-rotation data.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Neural Operators, Geometric Deep Learning, Cellular Sheaves, Category Theory, Gauge Equivariance

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pKq3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes Symmetry Descent Neural Operators (SDNO), which address the following important gap of common equivariant neural networks: ordinary models assume that one single symmetry rule (group) and one single coordinate frame work everywhere in the domain. The proposed SDNO aim to achieve global consistency for local equivariance, which is not enough. The main idea behind SDNO consists of three steps: learn the hidden local symmetries from the observed operator; estimate the transport between neighboring patches and measure curvature and holonomy; constrain the weights so that they respect the obtained symmetries and transport rules. SDNO remains accurate under unseen changes in holonomy, curvature, topology, and local frames.

Critical strengths

The identified problem of local and global consistency of equivariance seems to be significant and new. The paper develops the framework (and the corresponding theory) on how to address it, and the ideas seem to be interesting for the community.

Critical weaknesses

I believe the main weakness is that the approach relies on lots of required assumptions (in particular, multiple observations per operator, Jacobians, spectral gap, etc.), which is mentioned in the Limitations section.

Comment on the content presentation: I would like to share with the authors my opinion that the abstract and introduction is quite dense and introduces many terms quickly (I was a bit shocked when reading it, so you can easily lose the broad audience).

I also would like to comment why I put “1: Difficult to follow despite careful reading.” on “Is the paper clearly presented?”. I believe the problem might be in dense terminology for the broad ICLR community (and for me as well). So, I have not noticed any inconsistency in notation or definitions, or smth like this, but it took me a lot of effort to understand (at least partially) the underlying terminology.

Questions for authors

I have some questions regarding the novelty of the approach, because the description of main contributions in the introduction seem to me vague.

Has anyone previously enforced loop holonomy constraints on network layers? If not, is that because it is genuinely new, or because practitioners found it unnecessary for some reason? Is the soft null-space projection itself new, or is the novelty in applying it to symmetry constraints? Does Theorem 5.1 require anything beyond standard covering-space / local-system / sheaf-cohomology theory (I mean, is this fact smth standard, restated for the neural-network domain)?

Another question is about the significance of the work: Could the authors please provide one real application or dataset (not a synthetic one) where ignoring loop holonomy causes a measurable failure that SDNO fixes?

4 comments

W1. The approach relies on many assumptions: multiple observations per operator, Jacobians, and a spectral gap.

Thank you for raising this concern. The Jacobian and sufficient-excitation assumptions apply specifically to local symmetry discovery, an optional step needed only when the structure group is unknown. The Aharonov-Bohm ring, qubit-array, polycrystalline, power-grid, and rotation-averaging experiments take the group from the physics (U(1)\mathrm{U}(1) phases, planar grain rotations, or SO(3)\mathrm{SO}(3)) and skip this step, so none uses a surrogate, Jacobian, or spectral gap. The step is described in revised Appendix C.

Multiple observations are required because the connection is hidden and must be inferred from operator responses. Following your comment, we vary the number of context pairs per episode and compare errors (revised Appendix Table 8). The AB ring reaches an error of 0.036 with 10 context pairs, an order of magnitude below the best neural baseline. The controlled U(1)\mathrm{U}(1) task needs 2 pairs, and the qubit data use one trace per episode. A wrong symmetry assumption can be detected from the fitting residual, which rises from 0.04 to 0.29 as the symmetry is broken (revised Appendix Figure 9).

W2. The abstract and introduction are dense and introduce many terms quickly, and the stated contributions are vague.

We agree that the original presentation introduced too many concepts before establishing the problem. We have rewritten the abstract and introduction, added a conceptual illustration of transport and holonomy (revised Figure 1), and introduced the necessary background and compatibility conditions in revised Section 3. Revised Appendix A also provides a terminology guide.

Following your comment, the revised contribution list states three items. (a) Method: estimate a hidden gauge connection from input-output pairs of one operator and enforce the resulting constraints through a differentiable projection and a per-episode corrector. (b) Theory: the holonomy criterion for weight sharing, the containment between holonomy and full-group constraints, and stability under estimation error. (c) Experiments: controlled connection recovery, Aharonov-Bohm ring, measured qubit-array interference, polycrystalline local frames, and public SO(3) rotation averaging.

We have also moved the specialized sheaf-theoretic and categorical formulations to revised Appendix D. The main text now develops the compatibility problem directly, using elementary examples before introducing its mathematical interpretation.

Q1. Has anyone enforced loop-holonomy constraints on network layers before? If not, is this new, or was it found unnecessary?

Thank you for this question. The mathematical role of holonomy in global compatibility is classical, as we now acknowledge explicitly in revised Proposition 5.1. Conventional gauge-equivariant architectures generally impose equivariance under the full structure group GG. Since the holonomy group HH is a subgroup of GG, full-group equivariance can ensure the required loop compatibility without separately imposing holonomy constraints. However, it can also be more restrictive than necessary.

When a connection has holonomy group H⊆GH\subseteq G, the corresponding holonomy-intertwiner space can be larger than the space of fully GG-equivariant maps. Revised Corollary 5.2 illustrates this with an SU(2)\mathrm{SU}(2) representation: considering the holonomy constraints alone, the allowable dimensions are 9 under the full group, 18 under a diagonal torus, and 36 under trivial holonomy. Additional local-admissibility constraints must also be respected when constructing globally compatible layers.

SDNO uses this distinction when the connection is unknown and varies between episodes. It estimates the connection from operator observations and determines the resulting compatibility constraints. To our knowledge, previous operator-learning methods have not used holonomy estimated from episode-specific context pairs to determine the shareable layer space.

The revised experiments also demonstrate the value of recovering the connection. On the Aharonov–Bohm ring, a gauge-equivariant network constructed following [1, 2] has new-holonomy error 0.703 with identity links. Supplying SDNO's estimated connection reduces the error to 0.449, close to the true-connection reference of 0.433 (revised Table 4(a) and Appendix Table 10). Under a random gauge applied to the in-distribution episodes, the errors are 0.929, 0.414, and 0.402. These comparisons show that estimating the connection is beneficial even for an independently constructed gauge-equivariant architecture. Further comparisons appear in our response to Reviewer 6rrr (W1/Q1).

References

[1] Cohen et al. Gauge Equivariant Convolutional Networks and the Icosahedral CNN. ICML 2019.

[2] Favoni et al. Lattice Gauge Equivariant Convolutional Neural Networks. Physical Review Letters 128, 032003, 2022.

Q2. Is the soft null-space projection new, or is the novelty in applying it to symmetry constraints?

The soft projection itself is the ridge form of the hard null-space projection in EMLP [3]. The distinction in SDNO is that the constraint matrix CC combines local equivariance with edge compatibility determined by an estimated connection, and therefore changes between episodes. Its rank may change with holonomy, making a hard null-space projection discontinuous. The soft projection provides a differentiable relaxation. Revised Theorem 5.3 bounds the error by τ/(γ2+τ)\tau/(\gamma^2+\tau) and the residual by τ/2\sqrt{\tau}/2.

Q3. Does Theorem 5.1 of the original submission need anything beyond standard covering-space, local-system, or sheaf theory?

We agree that the result follows from established local-system theory. Following your comment, we have restated it as Proposition 5.1 (Holonomy criterion) and explicitly identified its connection to the classical holonomy principle. Building on the original setting and method, we bring out three further results. These are the containment relation with its equality conditions (revised Corollary 5.2), a stability bound under approximate edge compatibility (revised Theorem 5.4), and bounds for the soft projection (revised Theorem 5.3).

Q4. Is there a real application or dataset where ignoring loop holonomy causes a measurable failure that SDNO fixes?

Yes, on measured data from a superconducting qubit array [4], ignoring the loop phase raises the error by 53% to 85% relative to SDNO, and the loop phase estimated by SDNO agrees with the magnitude of the applied flux, which is never provided to the model. Following your request and Reviewer 6rrr's call for more realistic validation, we add this dataset (revised Table 3 and Section 6.4). Trivial connection in the table below is the same model with the loop phase fixed to zero, ignoring loop holonomy.

RingTrivial connectionSDNOBest unstructured
4 sites0.5800.3130.769
8 sites0.7430.4850.759
12 sites0.7040.4370.679

The correlation is at least 0.989 on all three rings, and the median error is 1% to 3% of the phase range [0,π][0,\pi]. On real power-grid cases [5, 6] (revised Table 4(b) and Appendix I.7), removing the loop component of the phase-shifter angles changes active power by up to 74% (about 1.7 GW). In probing simulated on these cases, estimating the connection lowers the estimation error by more than a factor of 10 relative to a reciprocal model without the phase.

References

[3] Finzi et al. A Practical Method for Constructing Equivariant Multilayer Perceptrons for Arbitrary Matrix Groups. ICML 2021.

[4] Rosen et al. A Synthetic Magnetic Vector Potential in a 2D Superconducting Qubit Array. Nature Physics 20, 1881–1887, 2024.

[5] Josz et al. AC Power Flow Data in MATPOWER and QCQP Format: iTesla, RTE Snapshots, and PEGASE. arXiv:1603.01533, 2016.

[6] Zimmerman et al. MATPOWER: Steady-State Operations, Planning, and Analysis Tools for Power Systems Research and Education. IEEE Transactions on Power Systems 26(1), 12–19, 2011.

Please take a moment to respond to the rebuttal. Thank you!

I thank the authors for their detailed response. I apologize for the delayed reply, as I was ill over the past few days; however, I kept in mind that I needed to respond. I thank the authors for the corrections made and would like to raise my score to 3.

Reviewer_6rrrOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a symmetry descent neural operator (SDNO) that estimates local symmetries and inter-patch transport relations from operator observations. It employs a differentiable soft projection to constrain the learnable network parameters, encouraging local equivariance and cross-patch compatibility. The proposed method is evaluated on several downstream tasks, including controlled structure-recovery problems, polycrystalline material modeling, and 3D rotation averaging.

Critical strengths
  1. Interesting problem formulation & model design. The paper addresses an interesting challenge in neural operator learning: how to discover local symmetries from observations and ensure compatibility between local models at the global level. The proposed framework connects local symmetry discovery, inter-patch transport estimation, and constrained parameterization in a coherent manner. I find this perspective interesting because it provides a principled way to incorporate inferred geometric structure into model design, rather than relying solely on unconstrained learning.
  2. The paper evaluates the proposed method on several tasks, including controlled structure recovery, operator prediction under changing structural conditions, polycrystalline material modeling, and 3D rotation averaging. This range of experiments provides useful evidence for the potential applicability of the framework beyond a single benchmark. In particular, the polycrystalline material experiment illustrates how local coordinate systems and heterogeneous interfaces can be incorporated into operator learning, offering a meaningful connection to scientific and physical modeling.
Critical weaknesses
  1. Insufficient comparison with closely related methods. The experimental evaluation lacks comparisons with some recent methods closely related to the paper's central contributions, such as AtlasD (ICML 2025) for local symmetry discovery and Foundations of Equivariant Deep Learning (ICML 2026) for unifying equivariant and sheaf-based architectures. Although these methods may not be directly applicable to all the tasks considered, a clearer discussion of their differences, or comparisons where feasible, would help establish the novelty and benefits of SDNO's joint structure recovery and constrained parameterization.
  2. Limited evidence for general applicability. Most operator-learning experiments rely on controlled synthetic settings with prescribed structural assumptions. In particular, the polycrystalline experiment provides an interesting proof of concept, but does not establish the method's effectiveness on real material data. Further validation under more realistic conditions, along with a clearer discussion of the assumptions underlying the theoretical guarantees, would strengthen the paper's claims regarding generality.
Questions for authors
  1. Distinction from closely related methods. Could the authors clarify the key methodological differences between SDNO and existing approaches to local symmetry discovery and sheaf-based equivariant learning, such as AtlasD (ICML 2025) and Foundations of Equivariant Deep Learning (ICML 2026)? In particular, which capabilities of SDNO cannot be readily obtained by combining existing methods, and is there a feasible comparison or ablation that isolates the benefits of jointly recovering local symmetries and inter-patch transports?
  2. Scope of the theoretical and empirical claims. Could the polycrystalline experiment be interpreted primarily as a proof of concept under controlled synthetic conditions, rather than evidence of effectiveness on real material systems?
4 comments

W1 and Q1. The evaluation lacks comparison with AtlasD (ICML 2025) and Foundations of Equivariant Deep Learning (ICML 2026). Which capabilities of SDNO cannot be obtained by combining them, and is there an ablation that isolates the benefit of recovering local symmetries and transport together?

Thank you for suggesting these comparisons. AtlasD and Foundations of Equivariant Deep Learning (FEDL) address complementary aspects of equivariance, but neither estimates the inter-patch connection from operator observations. We have clarified these distinctions in revised Appendix B and the Related Work section, and added experiments that isolate the contribution of estimating transport (revised Table 2 and Appendix Table 10).

Relation to AtlasD and FEDL. AtlasD discovers local Lie-group symmetries on individual charts but does not estimate inter-chart transport or loop holonomy. FEDL characterizes equivariant layers on a specified vector bundle over a face poset. Combining these capabilities still requires a method for estimating how local frames are connected. In the physical systems considered here, this connection varies between episodes and is not determined by mesh geometry. SDNO estimates the connection from input–output context pairs and uses its holonomy to construct globally compatible layers through a differentiable projection. The holonomy criterion is classical, as acknowledged in revised Proposition 5.1; our contribution is its use with an estimated, episode-dependent connection.

Evidence for the role of transport. Revised Table 2 compares SDNO with a trivial-connection operator solver, representing the setting in which local symmetry is available but cross-patch transport is not. The trivial-connection solver has 23–49 times SDNO's prediction error, whereas SDNO is within 0.005 of the true-connection reference. The zero-flux control in revised Appendix Table 10 provides a complementary comparison: when the physical connection is trivial, the trivial-connection and true-connection solvers both achieve 0.021 error, compared with 0.024 for SDNO. Thus, the advantage of estimating transport is most apparent when the physical connection is nontrivial.

The estimated connection also benefits a different architecture. On the Aharonov–Bohm ring, a gauge-equivariant network constructed following Cohen et al. (2019) and Favoni et al. (2022) has new-holonomy error 0.703 with identity links, decreasing to 0.449 when supplied with SDNO's estimated connection, compared with 0.433 with the true connection (revised Table 4(a)). These rows compare transport settings. Running the AtlasD code on this ring recovers the local group U(1) when the search is restricted to the field channels, and it returns no transport between charts (revised Appendix I). We provide further test-time adaptation comparisons in our response to Reviewer KGvh (W2).

Local symmetry discovery and global compatibility. Revised Figure 6c illustrates how cross-patch compatibility reduces a four-dimensional local symmetry candidate space to a globally compatible one-dimensional u(1)\mathfrak{u}(1) subspace. In the SU(2)\mathrm{SU}(2) experiments, the compatible algebra dimensions are 4/2/1 depending on the connection's holonomy (revised Figure 13). Revised Appendix Table 6 further isolates the importance of group-valued transport: with two context pairs, direct transport estimation gives error 0.028, compared with 0.557 for unrestricted real matrices and 0.457 for complex-linear matrices.

These experiments isolate transport estimation and global compatibility. The prediction experiments in Section 6 use groups known from the physics and skip the optional local discovery procedure in revised Appendix C. The discovery and compatibility stages are examined separately in Appendix C and Figure 6.

W2. Most experiments use controlled synthetic settings. Validation under more realistic conditions and a clearer statement of the assumptions behind the theory would strengthen the claims.

Thank you for encouraging broader validation and clearer assumptions. The original submission included rotation averaging on public 1DSfM Structure-from-Motion data [1], with lower mean and median angular error than the evaluated baselines in five of six scenes (revised Appendix Table 18). This tests connection synchronization on real observations rather than operator prediction. We now add measured superconducting-qubit dynamics and published power-grid networks.

Measured superconducting-qubit dynamics. Revised Section 6.4 and Table 3 evaluate Aharonov–Bohm interference using measured data from Rosen et al. [2]. On rings of 4, 8, and 12 sites, we train with ∣Φ∣≤π/2|\Phi|\leq\pi/2 and test on unseen flux magnitudes ∣Φ∣>π/2|\Phi|>\pi/2, without providing flux labels. Using the first 40 of 121 time steps as context, SDNO achieves prediction errors of 0.313/0.485/0.437, compared with 0.769/0.759/0.679 for the best unstructured baselines, a reduction of 36%–59%. Fixing the loop phase to zero increases error to 0.580/0.743/0.704, isolating the contribution of estimating holonomy.

We also evaluate loop-phase recovery from the full 121-step traces (revised Appendix Table 15). Without flux labels, the estimated phase magnitude has correlations of 0.999/0.989/0.997 with the applied flux magnitude and median errors of 0.035/0.087/0.052 radians. Using only the first 40 steps, the corresponding errors are 0.017/0.271/0.070 radians. The prediction experiments use this shorter context. Only ∣Φ∣|\Phi| is identifiable because the measured single-site populations are invariant under reversal of the flux sign.

Power-grid phase shifters. Revised Table 4(b) and Appendix I.7 use published MATPOWER network instances, including French RTE and Polish grids [3, 4]. Transformer phase shifts provide a physically meaningful connection whose loop component cannot be removed by changing bus-voltage angle references. On the four RTE grids, removing this component changes active power through phase-shifting branches by 46%–74%, with the largest change approximately 1.7 GW. Using 32 simulated input–output probe pairs on these published networks, estimating the connection reduces injected-current prediction error from 1.16%–2.71% for a reciprocal model to 0.08%–0.11%. The Polish case has smaller phase shifts, and connection estimation does not improve the reciprocal model's already low error. These experiments use real network configurations but simulated probing, rather than measured operational responses.

Theoretical assumptions. Following your request for a clearer statement, we separate two groups of assumptions. The theory in revised Section 5 makes three assumptions, each describing the physical setting. The cell complex is finite and connected because a specimen is a bounded domain divided into finitely many patches that form one piece. The fibers are finite-dimensional and unitary because each patch carries a finite feature vector and a frame change preserves its norm. Transport is representation-compatible because it moves features between frames by an element of the structure group, which is also the physical definition of a connection. The descent branch uses flat transport, and curvature is controlled separately by revised Theorem 5.4. The three conditions for local symmetry discovery (sufficient excitation, surrogate Jacobian accuracy, spectral gap) apply only when the structure group is unknown and are stated in revised Appendix C.

References

[1] Wilson and Snavely. Robust Global Translations with 1DSfM. ECCV 2014.

[2] Rosen et al. A Synthetic Magnetic Vector Potential in a 2D Superconducting Qubit Array. Nature Physics 20, 1881–1887, 2024.

[3] Josz et al. AC Power Flow Data in MATPOWER and QCQP Format: iTesla, RTE Snapshots, and PEGASE. arXiv:1603.01533, 2016.

[4] Zimmerman et al. MATPOWER: Steady-State Operations, Planning, and Analysis Tools for Power Systems Research and Education. IEEE Transactions on Power Systems 26(1), 12–19, 2011.

Q2. Is the polycrystalline experiment mainly a proof of concept under synthetic conditions?

Yes, the polycrystalline task uses simulated data, as clarified in revised Section 6.3. It provides a physically interpretable setting where hidden material frames and boundary defects can be compared directly with ground truth. The measured qubit-array and public Structure-from-Motion experiments provide complementary evidence from real observations, while the power-grid experiments use published physical network configurations with simulated probing.

In the polycrystalline task, each of the 70 grains has a local orientation determining its anisotropic stiffness, while weakened grain boundaries introduce localized defects. Transport between neighboring grain frames depends on these material orientations rather than the flat mesh geometry. The simulation allows us to evaluate recovery of both the frame structure and the defective boundaries.

Revised Appendix Figures 16–17 show that, with eight context pairs and 1% noise, SDNO recovers grain orientations with median error 0.008 radians, and detects weakened boundaries with AUC 1.00. With four context pairs and 5% noise, the corresponding values are 0.122 radians (approximately 7.0∘7.0^\circ) and AUC 0.87 (revised Appendix Figure 17). These experiments quantify recovery under reduced observation budgets and increased noise.

We also added test-time adaptation baselines to this task (revised Appendix Table 17). With four context pairs, SDNO achieves errors of 0.188 in distribution and 0.184 under a new grain-boundary density, compared with 0.207/0.215 for the best evaluated neural baseline. This corresponds to improvements of approximately 9% and 14%. Unconstrained per-episode fitting is substantially less accurate because the available observations leave its parameterization underdetermined.

Please take a moment to respond to the rebuttal. Thank you!

Reviewer_KGvhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper suggests Symmetry Descent Neural Operators (SDNO) that automatically discover local equivariance and transport between regions from operator observations, and enforce global consistency through shared parameters. The main contributions are a method to discover local symmetries and construct globally consistent neural operators, and theoretical guarantees for global compatibility and stability. The authors validate their method on neural operator prediction under different topologies and local frames showing better performance than existing baselines.

Critical strengths
  1. The paper addresses the limitation of existing equivariant neural operators that assume known global symmetries.
  2. The paper provides theoretical guarantees for global compatibility and stability.
Critical weaknesses
  1. Although I have a background in neural operators, equivariant models, and basic group theory, it is difficult to understand the main problem and theoretical contributions of the paper. In particular, a background section is absent, although the paper assumes substantial preliminary knowledge. A conceptual figure or illustrative example of the target issue (such as the compatibility condition) would also be helpful for understanding why the proposed method is necessary. The paper assumes readers have specialized knowledge across several different areas, making it difficult to assess the significance of the theoretical contributions. Reorganizing the paper may be necessary.

  2. The comparison in Table 2 seems unfair. The proposed model optimizes episode-specific parameters at test time, whereas all the other baselines are used without test-time parameter optimization (even when context is provided). The significant performance improvement may therefore come from explicit system identification rather than the proposed symmetry descent methodology. I recommend including baselines with comparable test-time optimization or an ablation that isolates the contribution of episode-specific parameter fitting. Without such comparisons, it is difficult to determine whether the performance improvement is attributable to the proposed methodology.

Questions for authors

None

6 comments

W1. The main problem and the theoretical contributions are hard to understand. There is no background section and no conceptual figure or example of the compatibility condition, and the paper assumes knowledge from several areas.

Thank you for identifying the difficulty of following the problem formulation and theoretical contributions. We agree that the original presentation introduced too many specialized concepts before establishing the central problem. We have rewritten the abstract and introduction, added a conceptual figure (revised Figure 1), and introduced a background section with concrete examples (revised Section 3).

The compatibility problem. The central issue is that locally equivariant layers need not define a globally consistent operator. When features are represented in different local frames, neighboring layers must satisfy

TijoutWi=WjTijin.T_{ij}^{\mathrm{out}}W_i=W_jT_{ij}^{\mathrm{in}}.

This requires applying a layer before transporting its output to agree with transporting the input before applying the neighboring layer. Revised Section 3 illustrates the condition with a scalar example: under identity transport, Wi=1W_i=1 and Wj=2W_j=2 are individually admissible but incompatible across their shared boundary. Existing gauge-equivariant architectures generally assume that the transport is specified by geometry or provided as input [1, 2]. In SDNO, the connection varies across physical systems and must be estimated from operator observations.

Conceptual illustration. Revised Figure 1 illustrates the distinction between local flatness and global holonomy using a truncated cone. Transport around the hole rotates vectors by an angle θ\theta, even though each patch is flat. Changing local frames redistributes the rotation along the loop but cannot eliminate its total effect. Isotropic kernels are preserved by this transport, whereas anisotropic kernels can be shared consistently only if they are invariant under the rotation. This illustrates the holonomy criterion of revised Proposition 5.1. Revised Section 5 provides a complementary algebraic example: for one-dimensional complex fibers, unequal input and output loop phases leave only the zero map globally compatible.

Theoretical contributions. We have reorganized Section 5 to distinguish the classical mathematical foundations from their use in SDNO. Revised Proposition 5.1 states the standard holonomy criterion for global compatibility and acknowledges its origin in local-system theory. Revised Corollary 5.2 characterizes how holonomy restricts the space of globally compatible layers, with an SU(2)\mathrm{SU}(2) example illustrating the effect on the available parameter dimensions. Revised Theorem 5.3 bounds the approximation error and constraint residual of the soft projection, while Theorem 5.4 bounds deviation from global consistency in terms of the edge-compatibility residual.

To make these results more accessible, revised Appendix A provides a terminology guide, Appendix C contains the optional local symmetry discovery procedure, and Appendix D develops the sheaf-theoretic and categorical formulations. Revised Appendix B also clarifies the relationship to existing local symmetry discovery [3] and sheaf-based equivariant frameworks [4]. These changes separate the elementary compatibility problem from its more specialized mathematical interpretations.

Relation to existing frameworks. Revised Appendix B clarifies how SDNO differs from gauge-equivariant networks [1, 2], local symmetry discovery [3], and sheaf-based equivariant frameworks [4]. These approaches either assume that transport is available or do not estimate it from operator observations. SDNO instead estimates the connection and uses its holonomy to determine globally compatible layers. We provide a detailed comparison, including the experimental results in revised Appendix Table 10 and the zero-flux control, in our response to Reviewer 6rrr (W1/Q1).

References

[1] Cohen et al. Gauge Equivariant Convolutional Networks and the Icosahedral CNN. ICML 2019.

[2] Favoni et al. Lattice Gauge Equivariant Convolutional Neural Networks. Physical Review Letters 128, 032003, 2022.

[3] Bhat et al. AtlasD: Automatic Local Symmetry Discovery. ICML 2025.

[4] Maruyama. Foundations of Equivariant Deep Learning: Unifying Graph and Sheaf Neural Networks. ICML 2026.

W2. The comparison in Table 2 of the original submission is unfair. SDNO fits episode-specific parameters at test time and the baselines do not, so the gain may come from system identification alone.

Thank you for identifying this confound. We agree that the original comparison did not fully distinguish the benefit of structured connection estimation from that of test-time parameter fitting. SDNO performs a form of system identification, but constrains the estimated parameters to represent group-valued transport between local frames. To test whether this structure matters, revised Table 2 introduces test-time fine-tuning and unconstrained per-episode fitting baselines on an Aharonov–Bohm ring prediction task (revised Section 6.3). All adaptive methods receive the same 16 input–output context pairs per episode. FNO-2d and DeepONet fine-tune their network parameters; free per-edge couplings are fitted using the same block-coordinate solver as SDNO; and SDNO estimates the connection while keeping its learned corrector fixed.

MethodIn distributionNew holonomyNew curvature
FNO-2d (query)0.3730.7200.403
FNO-2d + test-time fine-tuning0.3630.4240.389
DeepONet + test-time fine-tuning0.7670.8400.772
Free per-edge couplings without transport structure (2,800 params/ep.)0.1210.1140.110
SDNO (880 params/ep.)0.0230.0210.022

Structured versus unconstrained adaptation. Fine-tuning improves FNO-2d's new-holonomy error from 0.720 to 0.424, but its error remains approximately 20 times that of SDNO. Independently fitting complex edge couplings is more effective, reaching 0.114, but still has over five times SDNO's error despite using more than three times as many episode-specific parameters. Because the free per-edge baseline uses the same context pairs and solver, this comparison isolates the benefit of constraining the fitted couplings to group-valued transport rather than allowing unrestricted complex coefficients. The estimated-connection model also approaches the true-connection reference, with errors differing by at most 0.005 across the evaluation conditions.

Contributions of holonomy and curvature. The ablations in revised Table 2 further distinguish the components of the estimated connection. Removing unit-modulus refinement increases prediction error approximately threefold. For the structure model without the neural corrector, removing loop holonomy increases error by 16–21 times, while removing face curvature increases error by 7–10 times. The holonomy ablation removes the noncontractible-loop component while retaining local gauge and face curvature, isolating the contribution of global transport information. In the zero-flux control (revised Appendix Table 10), the true connection is trivial; the trivial-connection solver and true-connection reference both achieve 0.021 error, compared with 0.024 for SDNO. These results support the importance of estimating the connection when it is not specified by the geometry.

Transfer to other architectures. We also evaluate a gauge-equivariant network constructed following [1, 2] on the same episodes (revised Table 4(a) and Appendix Table 10). With identity links, its error is 0.703 under new holonomy and 0.929 when a random gauge is applied to the in-distribution episodes. Supplying SDNO's estimated connection reduces these errors to 0.449 and 0.414, respectively, within 0.02 of the corresponding true-connection references. Thus, the estimated connection is useful beyond SDNO's particular operator solver. The remaining performance difference reflects the distinct evaluated architectures: the gauge-equivariant network approximates the resolvent through fixed-depth convolution, whereas SDNO incorporates the estimated connection into the operator formulation.

Finally, we extend the adaptation comparisons to the simulated polycrystalline task (revised Appendix Table 17). With four context pairs, fine-tuning does not improve the evaluated neural baselines. Unconstrained per-episode fitting gives errors of 3.67/3.96, compared with SDNO's 0.188/0.184. SDNO also improves on the best neural baseline by approximately 9% in distribution and 14% under a new grain-boundary density. These results provide complementary evidence that structured per-episode estimation is beneficial beyond the Aharonov–Bohm setting.

Thank you for the additional experiments.

The new test-time adaptation baselines substantially address my original concern about unfair comparisons. However, I think one important question remains. The comparison with unconstrained per-edge fitting demonstrates the benefit of structured connection estimation, but does not isolate the contribution of enforcing global compatibility constraints, which appears to be a central methodological contribution of SDNO. Although the holonomy and curvature ablations are encouraging, their largest improvements are reported for the structure-only model without the neural corrector. Could the authors clarify how much global compatibility enforcement contributes to the performance of the complete SDNO model, beyond group-valued connection estimation alone?

Also, I still find Figure 1 insufficient for intuitively understanding the central compatibility problem. The figure mainly illustrates geometric concepts such as holonomy and curvature, which themselves require specialized background knowledge.In particular, Figure 1(d) shows how holonomy affects isotropic and anisotropic kernels, but does not clearly illustrate why locally equivariant layers can be globally incompatible or how SDNO resolves this issue. A simple example explicitly illustrating the compatibility condition between neighboring neural operator layers would be more helpful.

Comments on the paper

1 comment

Thank you for the careful reviews and constructive suggestions. We have revised the manuscript to clarify the problem formulation and theoretical contributions, strengthen the experimental comparisons, and evaluate the role of holonomy on measured physical data. This revision preserves all original assumptions, methodology, algorithms, and empirical conclusions; our edits focus on improving presentation clarity and better articulating the underlying intuition and background.

Positioning and readability. SDNO addresses operator learning when local features are expressed in different frames and their alignment is governed by a connection that varies across system instances. Unlike architectures that assume this connection is known, SDNO estimates group-valued transport from input–output observations of the same operator and uses it to determine which local layers can be shared consistently across patches. In response to Reviewers KGvh and pKq3, we have rewritten the abstract and introduction, added a conceptual illustration of transport and holonomy (revised Figure 1), and introduced the necessary background and compatibility conditions in revised Section 3. Revised Appendix A provides a terminology guide. Following Reviewer 6rrr's comments, revised Appendix B clarifies the relationship to local symmetry discovery and sheaf-based equivariant frameworks. Following Reviewer KGvh's suggestions, revised Figure 19 (Appendix J, p. 46) gives a numerical example of Equation (3) on a two-channel fiber, shows how SDNO's estimated transport and projection make the paths agree and the weights consistent around the loop.

New experiments. To address Reviewer KGvh's concern about test-time adaptation, we introduce an Aharonov–Bohm ring prediction task with comparisons against neural operators fine-tuned at test time and independently fitted per-edge couplings (revised Table 2). SDNO estimates the connection separately for each episode while keeping its learned corrector fixed. It achieves relative errors of 0.021–0.023, compared with 0.110–0.121 for unconstrained per-edge fitting, supporting the benefit of structured connection estimation beyond test-time optimization alone. Additional ablations isolate the contributions of loop holonomy, face curvature, and group-valued transport. We also add test-time adaptation baselines to the polycrystalline experiment (revised Appendix Table 17). In response to Reviewers 6rrr and pKq3, we evaluate measured Aharonov–Bohm interference in a superconducting qubit array (revised Table 3), where ignoring the loop phase substantially increases prediction error, and examine phase-shifter effects on published power-grid networks using simulated probing (revised Table 4(b)). Further analyses examine sensitivity to the number of context observations (revised Appendix Table 8) and damping strength (revised Appendix Table 9).

Theory. In response to Reviewers KGvh and pKq3, we have reorganized the theoretical development around two questions: which local layers are globally compatible under a given connection, and how accurately these constraints can be enforced. Revised Proposition 5.1 states the classical holonomy criterion for global compatibility and explicitly acknowledges its mathematical foundations. Revised Corollary 5.2 characterizes how the holonomy group restricts the space of compatible layers relative to full-group equivariance. Revised Theorem 5.3 bounds the approximation error and constraint residual of the differentiable soft projection, while Theorem 5.4 bounds deviation from global consistency in terms of the edge-compatibility residual. We also clarify that the Jacobian and spectral-gap assumptions for local symmetry discovery apply only when the structure group is unknown (revised Appendix C).

All section, table, and figure references in our responses refer to the revised manuscript unless otherwise indicated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.