- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper points out that edge-aligned convolutions such as eSCN and SLEM lower the tensor product cost from to but cannot tell apart irreps that share the same degree and differ in inversion parity . These models are therefore only equivariant. The authors propose EACO, which restricts the kernel in the edge frame to the stabilizer of the edge axis. Each irrep gets a reflection sign , and fixed masks let channels with the same sign couple through and channels with opposite signs couple through . Theorem 1 shows that local equivariance is necessary and sufficient for global equivariance in this construction, and Proposition B.3 shows that the masked kernel spans the parity allowed Clebsch-Gordan couplings. The method is built on SLEM and tested on Hamiltonian prediction for Si, GaN, HfO$_2$, graphene and MoS$_2$. Compared with SLEM it lowers the MAE by up to 22.1% with 17.2 to 29.6% fewer parameters, reduces the equivariance error under reflection and inversion by three to four orders of magnitude, and respects the selection rule for electron-phonon coupling in graphene.
Critical strengths
- The method gives exact equivariance at the cost of an convolution. EACO keeps the cost and strict locality of SLEM, needs fewer parameters, and reaches relative equivariance errors around . This weighs in its favor because it shows that parity awareness can be added to the most efficient family of Hamiltonian models at no extra cost, and that it improves accuracy over the baseline
- The graphene experiment connects the symmetry to physics. SLEM leaks coupling that the horizontal mirror forbids (0.129 meV for ZA and 0.324 meV for ZO), while EACO keeps it at numerical zero. This shows that the symmetry matters for real applications and not only for an equivariance metric.
Critical weaknesses
-
The novelty is limited and the paper does not position itself against related work. The argument behind Theorem 1 is the standard stabilizer restriction from steerable kernel theory. The paper cites Weiler et al. (2018) for the case but does not discuss the general constructions that already cover and kernels, such as Weiler & Cesa (2019) and Cesa, Lang & Weiler (2022). Theorem 1 is also close to an version of the paper's own Theorem A.3. In addition, recent work has studied the closely related problem of keeping parity-sensitive interactions in fast tensor products. Xie et al. (ICML 2026) recover the interactions missed by the Gaunt tensor product using vector spherical harmonics, and Liang et al. (arXiv 2607.01408) recover antisymmetric paths and parity-odd components using spin-weighted spherical harmonics. These works target Gaunt type products rather than edge-aligned convolutions, but they address the same question of keeping the correct parity behavior at low cost. For these reasons I see the main contribution as extending parity awareness to edge-aligned convolutions and showing that it helps Hamiltonian prediction, rather than the theory itself. I would ask the authors to discuss these works and adjust their claims.
-
It is unclear where the gains over parity-aware baselines come from. DeepH-E3 and MACE-H are already equivariant and also satisfy the mirror selection rule exactly in Table 3. Even so, the error of EACO on the allowed coupling is 2.6 to 3.9 times lower (0.100% compared with 0.260% and 0.280% for ZA, and 0.113% compared with 0.363% and 0.437% for ZO), and its overall MAE in Table 2 is lower on four of the five materials, with MACE-H better on MoS$_2$. Since all three models are parity aware, the parity constraint cannot explain these gains. The paper says they reflect "a more accurate learned Hamiltonian response" but gives no analysis. They could come from the SLEM backbone or from differences in training, such as 4 versus 3 layers, 1,000 versus up to 1,500 epochs, and different losses. The very large on-site errors of DeepH-E3 and MACE-H on GaN and HfO$_2$ (30 to 87 meV, compared with 0.3 to 2 meV for SLEM and EACO) also make me wonder whether these baselines were set up well. Without this analysis it is hard to tell how much of the improvement comes from EACO itself.
-
The efficiency numbers are given at a degree that is rarely used in practice. The memory and time reductions in Section 4.3 (Fig. 5a) are measured at . The paper's own experiments go up to degree 8 for the orbitals in HfO2 and use degrees 4 to 6 for the other materials, and is enough for bases up to orbitals, which covers most Hamiltonian datasets. Results at of 4, 6 and 8 would be more representative. The speedup over Clebsch-Gordan products also comes from SLEM, since EACO has the same cost.
-
Reproducibility could be improved. Appendix D.1 gives the optimizer, learning rate, number of epochs, batch size and channel counts. It does not report seeds, the number of runs, training hardware or training time. The baselines are only described as using their "native" settings. A full configuration table for all methods and released code would make the results easier to verify.
Questions for authors
-
How does EACO relate, both conceptually and empirically, to Xie et al., Liang et al., and the general and steerable kernel constructions? What does the edge-aligned approach offer beyond them?
-
Why does EACO outperform DeepH-E3 and MACE-H in Tables 2 and 3, given that both are also equivariant? For the MoS$_2$ in Table 2, why the EACO is not following the same trend with other materials to give the best result? Is there any specific reason for MoS$_2$? Were the baselines (DeepH-E3 and MACE-H) tuned for these datasets, and why are their on-site errors on GaN and HfO$_2$ so large?
-
What are the memory and time reductions at of 4, 6 and 8?
-
How many runs are behind each reported number, and will the full training configurations for EACO and the baselines be released?
5 comments
Response to Reviewer o3pk (Part 2)
W2 & Q2: Sources of the gains over DeepH-E3 and MACE-H, MoS$_2$, and the baseline setup
A2: Tables 2 and 3 already separate the two sources. SLEM and EACO both beat DeepH-E3 and MACE-H on Si and graphene and both lose to MACE-H on MoS$_2$, and SLEM's allowed-EPC errors (0.22%) are already below the baselines' (0.26–0.44%), so the gains over the O(3) baselines come from the edge-aligned backbone. EACO's contribution is to make this backbone exactly O(3)-equivariant at the same cost, which no SO(2) model provides.
1. EACO itself is measured against SLEM. The two share the backbone, data, loss, optimizer, and 1,000 epochs, and differ only in the edge kernel. EACO lowers the reflection and inversion errors by 3–4 orders of magnitude, keeps the mirror-forbidden coupling at numerical zero (SLEM leaks 0.129/0.324 meV), and uses 17–30% fewer parameters while matching or improving SLEM's accuracy, with clear gains on Si and HfO$_2$.
2. The training differences favor the baselines. They use their own losses and schedules (Table 9) for 1,500 epochs, ours. SLEM and EACO are strictly local, so their 4 layers see one cutoff, while the 3 baseline layers see three.
3. On GaN and HfO$_2$, the gains over the baselines come from the angular degrees these materials require.
- The baselines cannot afford these degrees. At the required degrees and parities ( for GaN, for HfO$_2$), DeepH-E3 runs out of memory on an 80-GB A100 for both materials, and MACE-H needs over 1 TB of host memory to build for HfO$_2$ and around 10 days for a single run on GaN.
- Their large on-site errors come from their official angular degree, not from tuning. Both baselines use hidden degree with natural parity. Ga on-site blocks need degrees up to 6 with both parities and Hf blocks up to 8, as the MACE-H paper [5] itself states. Setting only these label components to zero gives on-site errors of 37.0 meV for Ga and 102.3 meV for Hf, 95–98% of the baselines' Ga and Hf errors (38.1–38.8 and 104.1 meV). Because the GaN and HfO$_2$ rows of Table 2 mixed this degree limit with model quality, we have moved these baseline results to App. D.4 and clarified this in Sec. 4.1.
- This further highlights EACO's advantage: it makes O(3)-equivariant training at these degrees practical. EACO trains at and at cost. It attains 0.126 and 0.326 meV on GaN and HfO$_2$, 17% below SLEM on HfO$_2$, and keeps the equivariance error below meV, while SLEM's is 0.10–0.24 meV.
4. On MoS$_2$, EACO stays comparable to MACE-H with far fewer parameters. MACE-H has EACO's parameters (2.79M vs. 0.38M). However, EACO's overall MAE is only 0.032 meV higher (0.441 vs. 0.409 meV), its on-site MAE is lower (0.327 vs. 0.966 meV), and it outperforms DeepH-E3 (0.576 meV) with fewer parameters.
We have added this decomposition to Sec. 4.1, and the cost and on-site analyses to App. D.4.
[5] C. Qian et al. Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing. npj Comput. Mater. 12, 169, 2026.
Response to Reviewer o3pk (Part 3)
W3 & Q3: Efficiency at
A3: The speedup indeed comes from the edge-aligned SO(2) structure. However, SO(2) convolutions are not O(3)-equivariant, so an O(3)-equivariant model had to pay the cost of CG tensor products or complete fast products. EACO brings exact O(3) equivariance to the cost of SO(2) convolution. The table compares the four operators on the same workload as Fig. 5(a): 4,096 edges, 16 channels per degree and parity, forward and backward on one idle A100. EACO, CG, and SpinGTP (official code) use the same CG paths with independent weights, except that SpinGTP lacks 4 paths at where its spin-weighted Gaunt factor vanishes, so its numbers are lower bounds. SLEM is the SO(2) reference.
| EACO | SLEM | CG | SpinGTP | |
|---|---|---|---|---|
| 4 | 18.0 / 165 | 13.9 / 141 | 94.4 / 676 | 145.9 / 1,796 |
| 6 | 31.0 / 326 | 25.9 / 279 | 269.3 / 2,353 | 783.3 / 7,197 |
| 8 | 48.9 / 653 | 42.6 / 653 | 584.2 / 6,072 | 3,135.9 / 21,878 |
(Time in ms / peak memory in MiB; medians of three runs.)
1. EACO is on par with SLEM (slightly slower and larger at small , as grouping channels by parity adds index operations, although EACO needs only half of SLEM's multiplications).
2. CG and SpinGTP cost far more for the same kernel space. CG takes more time and more memory than EACO at , and SpinGTP takes more time and more memory.
3. SpinGTP is slower than CG at these degrees. Its bound needs fast spherical transforms, which bring no speedup at these degrees (Liang et al., Sec. 5). For a fully connected product, its code instead contracts precomputed coupling coefficients into a dense kernel for every edge, and its spin-1 features are complex, with about twice the real dimension.
These degrees are those of our experiments: for Si, MoS$_2$, and graphene, and and for GaN and HfO$_2$, where DeepH-E3 runs out of memory and MACE-H needs over 1 TB of host memory or around 10 days per run (See our response in A2).
Following your suggestion, we have replaced Fig. 5(a) with –, added these results to Sec. 4.3, and added the setup with the table to App. D.3.
W4 & Q4: Runs, configurations, and code
A4: Each number in Tables 2 and 3 comes from a single run with a fixed seed shared by the four methods. We have added a second run of SLEM and EACO with seed 42 (App. D.5). EACO's exact O(3) equivariance and 17–30% fewer parameters hold in every run by construction, and its accuracy gains on Si and HfO$_2$ hold in both runs, by 12–22%. On GaN, graphene, and MoS$_2$, the two methods are on par. Following your suggestion, Table 9 in App. D.1 now lists the full training configurations of all methods, including the seeds and hardware, and Table 12 in App. D.3 reports the training time of every run. We will release our code together with these configurations.
In summary, we hope the above responses can resolve your concerns and help you re-evaluate our work. We would appreciate it if you consider raising your score based on our reply. Thanks in advance for your consideration.
Response to Reviewer o3pk (Part 1)
Thank you for your detailed and constructive review. We have revised our paper based on your comments, with new text in blue. Below, we answer your questions one by one.
W1 & Q1: Relation to steerable kernels, Xie et al., and Liang et al.
A1: Theorem 1 is the O(3) extension of the standard reference-direction construction (Theorem A.3), as stated in App. A (Lines 1091–1092). However, this construction only gives the condition for equivariance, not an operator that satisfies it. Our theoretical contribution is the complete solution of this condition inside the SO(2) convolution (Prop. 1 and Prop. B.3). It brings complete O(3)-equivariant convolution to the O(L3) cost of SO(2) convolution (L=ℓmax), and the parity-aware edge-aligned convolution you identify as our main contribution is exactly its result.
1. Beyond steerable kernel theory. Weiler & Cesa [1] and Cesa et al. [2] characterize the O(2) and O(3) kernel spaces. In message passing, the O(3) space is realized by CG tensor products at O(L6) cost, or O(L5) with sparse coefficients; our theory gives its edge-aligned form at O(L3). An SO(2) convolution mixes each ±m pair by aI2+bJ2, where I2 carries the CG paths with an even sum of input, filter, and output degrees ℓi+ℓf+ℓo and J2 the odd ones, and O(3) keeps a path iff (−1)ℓi+ℓf+ℓo=τinτout with τ=p(−1)ℓ. Hence, equal-τ channels couple only through I2 and opposite-τ channels only through J2 (Prop. 1), and these kernels span exactly the parity-allowed CG couplings (Prop. B.3). Attaching p to SLEM channels without this solution either keeps forbidden couplings or loses allowed ones.
2. Different from Xie et al. [3] and Liang et al. [4] in two points. First, they complete Gaunt-type tensor products, which miss allowed parity-odd paths; EACO corrects edge-aligned convolutions, which admit forbidden paths. Second, their complete products cost O(L4log2L), while EACO costs O(L3).
Both works note that these bounds bring no speedup at the degrees used in equivariant networks (Xie et al., Sec. 7; Liang et al., Sec. 5).
3. Empirically, EACO keeps SLEM's efficiency at L=4,6,8, while CG and SpinGTP cost several times more for the same kernel space (See our response in A3). The released VSTP code provides only the pointwise product, not the complete product that recovers all CG paths, so we do not time it.
Following your suggestion, we have added this discussion to Sec. 2, the path-parity analysis and the table to App. B.5, and the contribution statement above to Sec. 1.
[1] M. Weiler and G. Cesa. General E(2)-Equivariant Steerable CNNs. NeurIPS, 2019.
[2] G. Cesa, L. Lang, and M. Weiler. A Program to Build E(N)-Equivariant Steerable CNNs. ICLR, 2022.
[3] Y. Xie et al. Asymptotically Fast Clebsch–Gordan Tensor Products with Vector Spherical Harmonics. ICML, 2026.
[4] C. Liang et al. Spin-Weighted Spherical Harmonics Enable Complete and Scalable E(3)-Equivariant Networks. arXiv:2607.01408, 2026.