QHNet: Equivariant & Quaternion Networks

• QHNet denotes a set of architectures including SE(3)-equivariant graph networks for quantum Hamiltonian prediction and quaternion-valued defenses against adversarial attacks.
• It emphasizes structure preservation by reducing costly tensor operations and leveraging symmetry properties such as SO(3) and SO(2) equivariance.
• Variants like QHNetV2 and Quaternion-Hadamard Network demonstrate improved speed, memory efficiency, and robustness across quantum chemistry and low-level vision tasks.

QHNet is not a single canonical architecture in the recent arXiv literature. The name denotes, in one line of work, a SE(3)-equivariant graph network for predicting the Hamiltonian matrix in quantum chemistry and condensed matter physics (Yu et al., 2023); in a later extension, QHNetV2, an efficient network that achieves global SO(3) equivariance without the costly SO(3) Clebsch-Gordan tensor products (Yu et al., 11 Jun 2025); and, in a separate quaternion-imaging line, the Quaternion-Hadamard Network, a model-agnostic defense against first-order white-box adversarial attacks for rain, snow, and haze removal (Frants et al., 12 Feb 2025). A nearby but distinct acronym is CQH-Net, the Lean Classical-Quantum Hybrid Neural Network for image classification (Liu et al., 2024). The resulting literature is unified less by a single method than by recurrent commitments to structure-preserving representations, efficiency constraints, and domain-specific inductive bias.

1. Disambiguation and scope

In the arXiv record represented here, the label “QHNet” is reused across distinct subfields.

Name	Domain	Defining description
QHNet	Quantum Hamiltonian prediction	SE(3)-equivariant network with 92% reduction of tensor products (Yu et al., 2023)
QHNetV2	Quantum Hamiltonian prediction	SO(2) local-frame model eliminating SO(3) tensor products (Yu et al., 11 Jun 2025)
Quaternion-Hadamard Network (QHNet)	Adversarial defense for weather removal	Encoder-decoder with QHDCB, QDRB, and polynomial thresholding (Frants et al., 12 Feb 2025)
CQH-Net / LCQHNN	Image classification	Lean classical-quantum hybrid neural network with a 4-qubit VQC (Liu et al., 2024)

The overlap in nomenclature can obscure substantive differences. In the Hamiltonian-prediction papers, QHNet refers to symmetry-aware graph architectures operating on atom types, coordinates, irreps, and block-structured matrix outputs. In the weather-removal paper, QHNet refers to a quaternion-valued encoder-decoder used as an image pre-processing defense. This suggests that “QHNet” functions as a reused acronym rather than a standardized model family.

2. QHNet for predicting quantum Hamiltonians

The 2023 QHNet addresses prediction of the Hamiltonian matrix, which is central in computational quantum chemistry and condensed matter physics, while treating efficiency and equivariance as “important, but conflicting factors” (Yu et al., 2023). The model inputs atom types $Z$ and coordinates $\mathbf{p}$ and uses 5 node-wise interaction layers to generate SE(3)-equivariant irreducible representations for atoms. After these layers, both diagonal and off-diagonal atom pair representations are constructed using specialized modules, and an expansion module converts pairwise representations into intermediate blocks with a fixed shape per node pair, assembling the final Hamiltonian.

Its equivariance mechanism is based on Tensor Field Networks, with each layer transforming and combining irreducible representations using tensor products with Clebsch–Gordan coefficients:

$$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$

The architecture’s main efficiency innovation is explicit minimization of both the total number and the maximum sequence of tensor product operations. PhiSNet uses 121 tensor products with maximum sequence 76, whereas QHNet uses only 9 tensor products with maximum sequence 6, yielding a 92% reduction. The same paper reports that QHNet consumes 50% less memory due to its streamlined architecture and is 3–6.5× faster in training.

A second architectural issue is channel growth. Prior models such as PhiSNet require tracking orbitals per atom, so the number of channels grows rapidly as more atom types and orbital types are added. QHNet decouples the channel dimension from orbital/type combinatorics by using an expansion module that outputs intermediate matrices with a fixed, dataset-dependent shape for each node pair, while the actual atom-type–orbital correspondence is selected in post-processing. The channel dimension is therefore fixed as a hyper-parameter, unaffected by the number of atom or orbital types.

The paper reports experiments on MD17 datasets, including water, ethanol, malondialdehyde, and uracil, with double-precision floating point throughout because single-precision error is comparable to Hamiltonian prediction error magnitude. On these datasets, QHNet achieves comparable performance to state-of-the-art methods at significantly faster speed, with wavefunction cosine similarity greater than 0.999 for all cases. The code is publicly available as part of the AIRS library.

3. QHNetV2 and the shift from SO(3) tensor products to SO(2) local frames

QHNetV2 is presented as a successor architecture for Hamiltonian prediction that is motivated by the relationship between the off-diagonal blocks of the Hamiltonian matrix and the SO(2) local frame (Yu et al., 11 Jun 2025). Its central claim is architectural rather than merely incremental: QHNetV2 achieves global SO(3) equivariance without the costly SO(3) Clebsch-Gordan tensor products. It does so by introducing efficient and powerful SO(2)-equivariant operations and performing all off-diagonal feature updates and message passing within SO(2) local frames.

The key representation-theoretic move is local-frame reduction. For a local direction vector $\hat{r} \in \mathbb{R}^3$, a rotation $h \in SO(3)$ aligns $\hat{r}$ to a canonical direction, and minimal frame averaging gives

$$\langle \Phi \rangle_{\mathcal{F}(\hat{r})} = h \cdot \Phi(h^{-1} \hat{r}).$$

Within each SO(2) local frame, QHNetV2 defines SO(2) Linear, SO(2) Gate, SO(2) LayerNorm, SO(2) Tensor Product, and a continuous SO(2) tensor product for multi-body feature fusion. The computational consequence is explicit: SO(3) CG/TP-based operations have complexity $O(L_{\max}^6)$, whereas SO(2)-based operations have complexity $O(L_{\max}^3)$.

The overall architecture combines node-wise embeddings, pair embeddings, node-wise equivariant message passing, an SO(2) tensor product node update, off-diagonal feature updates in SO(2) local frames, and a final output layer that rotates off-diagonal SO(2) features back to the global frame to reconstruct the full Hamiltonian matrix. Off-diagonal Hamiltonian blocks are treated as especially natural objects for SO(2) framing, since each block corresponds to a unique SO(2) local frame defined by the atomic pair.

The empirical results are reported on the large QH9 and MD17 datasets. On QH9-stable-id, QHNetV2 attains $31.50$ H MAE, compared with $\mathbf{p}$0 for QHNet and $\mathbf{p}$1 for SPHNet; on QH9-stable-ood, it attains $\mathbf{p}$2 H MAE, compared with $\mathbf{p}$3 for QHNet and $\mathbf{p}$4 for SPHNet. The paper states that QHNetV2 reduces the error on off-diagonal Hamiltonian blocks by up to 47% compared to previous best models. It is also 4.34× faster than full TP-based QHNet and slightly faster than sparse TP-based SPHNet. On MD17, the gains are molecule-dependent: for ethanol, QHNetV2 reports $\mathbf{p}$5 H MAE versus $\mathbf{p}$6 for QHNet and $\mathbf{p}$7 for SPHNet; for water, the reported H MAE is $\mathbf{p}$8 for QHNetV2 and $\mathbf{p}$9 for QHNet. The paper therefore characterizes its performance as superior across a wide range of molecular structures and trajectories, while also emphasizing scalability and strong generalization capability.

4. Quaternion-Hadamard Network for adversarially robust weather removal

A separate 2025 paper uses QHNet to denote the Quaternion-Hadamard Network, a model-agnostic defense against first-order white-box adversarial attacks targeting deep-learning models for rain, snow, and haze removal (Frants et al., 12 Feb 2025). The motivation is low-level vision under attack: adversarial perturbations can compromise weather-artifact removal systems, while traditional defenses such as adversarial training and model distillation require extensive retraining and denoising or super-resolution pipelines can introduce visual artifacts and high computational demand.

This QHNet is a model-agnostic, training-free, efficient defense mechanism operating as an image pre-processing step. It combines Quaternion Neural Networks with the Walsh-Hadamard Transform, augmented by a polynomial thresholding layer that is non-differentiable during inference. Color images are encoded as quaternions,

$$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$0

where $$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$1, $$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$2, and $$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$3 are normalized color channels. The network uses an encoder-decoder architecture loosely inspired by UNet, with skip connections.

Its main blocks are the Quaternion Hadamard Denoising Convolutional Block (QHDCB) and the Quaternion Denoising Residual Block (QDRB). QHDCB suppresses adversarial noise via Walsh-Hadamard Transform, quaternion convolutions based on Hamilton products, polynomial thresholding in the frequency domain, and inverse Hadamard transform. QDRB combines a quaternion convolutional layer, a QHDCB layer, a parallel QConv identity branch, channel attention, spatial attention, and residual addition. A Quaternion Feature Aggregation and Refinement Block performs final feature refinement to recover texture and details and remove residual noise. QHNet outputs an estimate of the adversarial perturbation, which is then subtracted from the input image.

The polynomial thresholding layer is the paper’s main anti-attack mechanism:

$$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$4

Training uses a smooth surrogate to allow gradient-based learning, whereas inference uses hard thresholding, so gradients do not reliably propagate through the layer.

The paper also introduces the Adversarial Weather Conditions Vision Dataset (AWCVD), built by attacking state-of-the-art weather removal models on haze, rain, and snow tasks. The construction uses RESIDE-6K for haze, Rain-13k for rain, and CSD for snow, with FGSM and I-FGSM attacks. The scale is 11,190 images for training and 2,100 images for evaluation. Using PSNR and SSIM metrics, the reported improvements are substantial. For haze removal with DehazeFormer on Reside6K under I-FGSM with $$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$5, attack drops PSNR from 29.28 to 10.96 and SSIM from 0.97 to 0.623, while QHNet boosts performance to 27.24 PSNR and 0.957 SSIM. For rain-streak removal with M3SNet on Rain100H, attack drops PSNR from approximately 29 to approximately 15, and QHNet restores it to above 26. For snow removal with DSANet on CSD, attack drops PSNR from 29 to 7, and QHNet restores it to 28.5. The paper further states that QHNet works with a single set of weights across different architectures, including CNNs and transformers.

5. Related nomenclature: quaternion recurrent networks and CQH-Net

The reuse of the label QHNet sits alongside a broader family of adjacent acronyms. In hypercomplex associative memory, quaternion-valued Hopfield Neural Networks (QHNNs) and Quaternion-valued Recurrent Projection Neural Networks (QRPNNs) are distinct earlier models rather than instances of QHNet (Valle et al., 2020). Quaternions are defined there as 4-dimensional hypercomplex numbers,

$$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$6

and the unit quaternions form the state space $$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$7. QHNNs update neuron states via

$$(u^{\ell_1} \otimes v^{\ell_2})_{m_3}^{\ell_3} = \sum_{m_1=-\ell_1}^{\ell_1} \sum_{m_2=-\ell_2}^{\ell_2} C^{(\ell_3, m_3)}_{(\ell_1, m_1), (\ell_2, m_2)} u_{m_1}^{\ell_1} v_{m_2}^{\ell_2}.$$8

while QRPNNs are obtained by combining the non-local projection learning with the quaternion-valued recurrent correlation neural network. The paper shows that QRPNNs overcome the cross-talk problem of QRCNNs and are appropriate to implement associative memories. This lineage is relevant because the Quaternion-Hadamard QHNet also relies on quaternion-valued processing, but the tasks, architectures, and guarantees are different.

A different nearby acronym is CQH-Net, introduced in the paper on the Lean Classical-Quantum Hybrid Neural Network model for image classification (Liu et al., 2024). LCQHNN, in the paper called CQH-Net, tightly integrates classical CNN layers with a quantum variational circuit. The architecture uses two convolutional layers, pooling, ReLU, dropout, and flattening to reduce data from 256 to 4 features, which are then passed to a 4-qubit VQC with Hadamard gates, U1 encoding, CNOT entanglers, parameterized RY rotations, and a final classical linear plus softmax layer. The paper reports 100\%, 99.02\%, and 85.55\% classification accuracy on MNIST, FashionMNIST, and CIFAR-10 in the abstract, while the detailed summary states that the paper does not report direct results on CIFAR-10. It also reports that CQH-Net reaches 100% accuracy by the 6th epoch on MNIST binary classification and 99.02% at 50 epochs on FashionMNIST, with faster convergence than comparable CNN baselines. This adjacency in naming is a recurrent source of confusion, but CQH-Net is a classical-quantum hybrid classifier rather than a Hamiltonian predictor or quaternion denoising defense.

6. Common misconceptions and recurring technical themes

A common misconception is that all QHNet variants are quaternion-valued. This is false for the Hamiltonian-prediction line. The 2023 QHNet and QHNetV2 are symmetry-aware graph networks whose defining machinery is SE(3), SO(3), or SO(2) equivariance, irreps, tensor products, local frames, and block-wise Hamiltonian reconstruction (Yu et al., 2023, Yu et al., 11 Jun 2025). By contrast, the Quaternion-Hadamard Network is explicitly quaternion-valued and frequency-domain denoising based (Frants et al., 12 Feb 2025).

A second misconception is that QHNetV2 is only a speed-tuned version of QHNet. The paper’s core claim is stronger: it eliminates SO(3) tensor products entirely in feature processing by moving computation into SO(2) local frames, while preserving global SO(3) equivariance through minimal frame averaging and frame-to-global mappings (Yu et al., 11 Jun 2025). The reported 4.34× speedup is therefore tied to a different equivariant factorization, not only to engineering optimization.

A third misconception is that acronym similarity implies methodological continuity. The data indicate only partial continuity. The Hamiltonian QHNet papers and the Quaternion-Hadamard QHNet both prioritize inductive bias and efficiency, but they solve different problems, use different mathematical objects, and evaluate on different metrics. A plausible implication is that the recurring appeal of the name “QHNet” comes from its compatibility with multiple expansions—quantum Hamiltonian, quaternion Hadamard, and neighboring labels such as CQH-Net—rather than from a single research lineage.

Across these usages, the recurring themes are explicit structure preservation, reduction of computational bottlenecks, and domain-specific design. In Hamiltonian prediction, the bottleneck is equivariant tensor algebra and orbital-type scaling; in adversarial weather removal, it is robust denoising under white-box first-order attacks without retraining; in CQH-Net, it is parameter and qubit efficiency on NISQ hardware. The shared pattern is architectural specialization rather than terminological uniformity.