Relaxed Unitary Convolutions

• Relaxed unitary convolutions are operations that relax strict norm-preserving requirements to enable smoothness-guided transformations in both random matrix and neural network models.
• They use methods such as Taylor-series truncation and encoder–decoder architectures to achieve controlled deviations from strict unitarity, facilitating diffusion and PDE dynamics.
• Empirical results show these techniques outperform traditional unitary models in applications like graph-based PDE simulation and global weather forecasting by balancing stability with flexibility.

Relaxed unitary convolutions are a class of convolutional operations, primarily motivated by two independent strands of research: probabilistic calculations involving random matrix averages and neural models biased toward smoothness preservation in learning on irregular domains such as graphs and meshes. Their defining property is the controlled relaxation of strict unitarity constraints—either by manipulating the group-theoretic averaging domain or by constructing neural layers that interpolate between exact norm-preserving (unitary or orthogonal) and more flexible, smoothness-guided transformations. This concept addresses the limitations of strict unitarity, especially in contexts where natural dynamics are not exactly norm-preserving but exhibit some intrinsic degree of smoothing, as is common in physical diffusion or partial differential equation (PDE) dynamics.

1. Mathematical Foundations and Definitions

Relaxed unitary convolutions have distinct but conceptually related realizations in random matrix theory and in neural network architectures.

In the context of polynomial convolutions and random matrices, the classical "finite free convolutions"—introduced by Marcus, Spielman, and Srivastava—measure, for instance, the average characteristic polynomial after conjugating by a random unitary matrix. Campbell and Yin demonstrated that, by using subgroups of the unitary group (such as the orthogonal group or the signed permutation group), the averaging can be restricted to these subgroups without changing the resulting polynomial convolution law, provided they satisfy the so-called "quadrature property" (Campbell et al., 2019). This yields the notion of relaxed unitary convolutions—matrix averages over tractable subgroups that preserve the polynomial laws.

In neural dynamics modeling on graphs and meshes, unitary (or orthogonal) convolutions arise from enforcing invariance of smoothness measures such as the Rayleigh quotient. Let $G=(V,E)$ be an undirected graph with normalized adjacency $\widehat{A}$, and signal matrix $X\in\mathbb{R}^{n\times d}$. The smoothness of $X$ is quantified by the Rayleigh quotient

$$R_g(X) = \frac{\operatorname{Tr}\left[X^\top (I - \widehat{A}) X\right]}{\|X\|_F^2}.$$

Unitary graph convolutions are designed so that $R_g(J_{\text{uni}}(X;\widehat{A})) = R_g(X)$ for all $X$ (Berman et al., 5 Feb 2026).

Relaxed unitary convolutions enable controlled departures from this strict invariance. Two main neural relaxations are:

• Taylor-series truncation: The finite Taylor expansion of the matrix exponential in the Lie convolution allows small norm changes and minor increases in smoothness.
• Encoder–decoder structure: A sequence of unitary (or orthogonal) layers acts as a smoothness-preserving encoder, followed by a decoder (often an unconstrained MLP), which enables expressive and smoothing transformations as required by the task dynamics.

Group-theoretically, relaxed unitary convolutions in random matrix models involve replacing integration over the Haar measure on the unitary group $U(d)$ with an average over a subgroup $G\subset U(d)$ that satisfies the quadrature property up to a certain order (Campbell et al., 2019).

2. Theoretical Rationale and Constraints

Imposing unitarity is theoretically motivated by the desire to lock in invariances such as $\ell_2$ norm or graph-signal smoothness. For graphs and meshes, this equates to exact preservation of the Rayleigh quotient, which controls over-smoothing in graph neural networks (GNNs).

However, strict unitarity can be overconstraining in physical systems. For example, linear diffusion driven by the Laplacian operator intrinsically increases smoothness over time:

$$\partial_t u = -\Delta u \implies u(t+\Delta t) = e^{-\Delta t L} u(t)$$

and hence $R_g(u(t+\Delta t)) > R_g(u(t))$. Unitary transformations cannot realize such increases in smoothness.

A lower bound on the expressivity of strictly unitary models can be formalized: For any function $f:\mathbb{C}^n\to\mathbb{C}^n$ and any unitary approximation $u$, the mean squared error is bounded below by the angular variance of $\|f(z)\|$ over spheres. If $f$ varies in norm depending on the direction, as is typical for PDE flows, strict unitarity introduces a nontrivial approximation bottleneck (Berman et al., 5 Feb 2026).

In polynomial convolution theory, the quadrature property restricts the allowable group averages: only subgroups $G$ for which first $d$ moments match the Haar measure over $U(d)$ are valid. Not every subgroup qualifies; for example, generic tori in $U(d)$ fail at orders higher than $O(1)$.

3. Methodological Realizations

In Random Matrix and Polynomial Convolutions

Consider two polynomials $p(x) = \sum_{k=0}^d (-1)^k a_k x^{d-k}$ and $q(x) = \sum_{k=0}^d (-1)^k b_k x^{d-k}$ as characteristic polynomials of normal matrices $A, B \in M_d(\mathbb{C})$. Three principal convolutions are defined:

• Symmetric additive ($\boxplus_d$): $$(p\boxplus_d q)(x) = \mathbb{E}_{U\in U(d)} \det[xI - (A + UBU^*)]$$
• Symmetric multiplicative ($\boxtimes_d$): $$(p\boxtimes_d q)(x) = \mathbb{E}_{U\in U(d)} \det[xI - (A U B U^*)]$$
• Asymmetric additive ($\boxdot_d$): $$(p\boxdot_d q)(x) = \mathbb{E}_{U,V\in U(d)} \det[(A+UBV)(A+UBV)^*]$$

Replacing $U(d)$ with any subgroup $G$ satisfying the quadrature property yields identical results:

$$\frac{1}{|G|}\sum_{g\in G} \det(xI - (A + g B g^*)) = (p\boxplus_d q)(x),$$

and similarly for the other convolutions (Campbell et al., 2019). Notable choices include $G=O(d)$, the real orthogonal group, and $G=H_d$, the group of signed permutation matrices.

In Neural Models for Dynamics on Graphs and Meshes

Neural relaxed unitary convolutions take two complementary forms (Berman et al., 5 Feb 2026):

• Taylor truncation: Truncating the Taylor series of the matrix exponential in Lie convolutions, with

$$f^{(T)}_{\text{relaxed}}(X;\widehat{A},W) = \sum_{k=0}^T (\widehat{A}XW)^k / k!$$

where $T$ is selected by sensitivity analysis or knowledge of the underlying PDE. For finite $T$, norm and smoothness are not exactly preserved, but deviations are provably small and controllable.

• Encoder–decoder: Padding $X$ to higher dimension, applying $k$ strict unitary (or orthogonal) blocks, and decoding back to the target dimensionality with an unconstrained decoder layer. This hybrid scheme retains much of the smoothness bias while affording the flexibility needed for modeling dissipative dynamics.

For mesh domains, the graph Laplacian is replaced by the cotangent Laplacian, preserving the Rayleigh quotient for signals on the mesh (Berman et al., 5 Feb 2026).

4. Empirical Performance and Comparative Results

Empirical studies demonstrate the superiority of relaxed unitary convolutional models in problems where controlled smoothing is physically mandated, but excessive oversmoothing (typical in standard GNNs) degrades fidelity.

Graph and Mesh PDEs

In heat diffusion prediction over random grid-graphs:

• Relaxed Taylor-truncated convolution with $T=3$ achieves MSE of $0.11 \times 10^{-2}$ and mean Rayleigh-error (MRE) of $2.07 \times 10^{-2}$, outperforming both standard GCN and strict Lie-unitary models in smoothness fidelity and accuracy.

On complex triangular meshes for time-evolved PDEs (heat, wave, Cahn–Hilliard), the "R-UNIMESH" architecture—encoder of unitary mesh convolutions plus decoder—exhibits state-of-the-art performance versus mesh-aware transformers, equivariant graph nets, and gauge-convolutions. It achieves the best or competitive normalized RMSE (NRMSE), symmetric mean absolute percentage error (SMAPE), and Rayleigh-quotient matching across long rollouts.

Real-World Applications

On ERA5-based global weather forecasting, relaxed unitary mesh networks rival transformer-based and advanced graph-based baselines, achieving anomaly correlation coefficient $>0.6$ for 48-hour lead times, with much smaller parameter budgets.

5. Algorithmic and Practical Considerations

Hyperparameter selection for relaxed unitary convolutional networks involves:

• Taylor truncation order $T \sim 5$–$10$ by Rayleigh-sensitivity or domain knowledge.
• Encoder depth $k$ and padded dimension $d'$, balancing bias toward smoothness and capacity.
• Rayleigh-matching penalty $\lambda$, tuning how strictly output smoothness is matched to ground truth.

The complexity overhead versus standard GNNs is negligible, with norm-preserving activations such as GroupSort available in mainstream libraries.

Use cases favoring relaxed unitary convolutions include physical systems with diffusion or dispersion, engineering surrogate models for mesh-based PDEs, and geoscientific simulations (weather, climate, ocean, ice-sheet modeling).

6. Connections, Limitations, and Outlook

Relaxed unitary convolutions, whether via group-theoretic averaging or neural relaxation, embody a principled compromise between stability and flexibility. The quadrature property in random matrix models determines eligibility for subgroup replacement; not all subgroups maintain the full "finite-free window" of allowable convolution orders. In neural settings, theoretical lower bounds confirm that strict norm invariance constrains approximation power for targets with angular norm variance—a typical feature of dissipative dynamics.

A plausible implication is that future methods may further exploit intermediate symmetries and controlled relaxations to address a wider array of complex dynamics and learning scenarios, especially as problem sizes and mesh resolution increase. However, the fidelity of relaxed approximations always remains contingent on careful management of the deviation from strict unitarity—quantified by the error terms in Taylor truncation or encoder–decoder design (Berman et al., 5 Feb 2026, Campbell et al., 2019).