Quasi-Symmetric Nets (QS-nets)

• Quasi-Symmetric Nets (QS-nets) are frameworks that enforce exact symmetry in key network components while allowing controlled deviations to model perturbations.
• In quantum many-body physics, QS-nets use a two-stage architecture—combining a symmetry-breaking subnetwork with a strictly invariant block—to achieve near-optimal energy accuracy.
• In discrete geometry, QS-nets construct equimodular elliptic Kokotsakis polyhedra by imposing linear symmetry relations, enabling explicit and flexible 3D designs.

Quasi-Symmetric Nets (QS-nets) denote two distinct but conceptually parallel frameworks: one in condensed matter physics and machine learning, where they provide parameter-efficient, scalable, approximately symmetric neural architectures for quantum many-body problems (Kufel et al., 2024); and another in discrete differential geometry, where they enable the explicit algebraic construction and flexion of equimodular elliptic Kokotsakis polyhedra in real three-dimensional space (Nurmatov et al., 24 Nov 2025). Both usages are characterized by enforcing or leveraging quasi-symmetry—respecting exact invariance in certain network components or geometric substructures, while permitting controlled deviations elsewhere to accommodate physical perturbations or enhance tractability.

1. Formal Definitions and Context

QS-nets in Quantum Many-Body Physics

A Quasi-Symmetric Net is a neural quantum state (NQS) architecture designed for variational studies of systems with large or emergent symmetry groups $G$ acting on spin or bit-string configurations $s \in \{\pm 1\}^N$. A strictly $G$-invariant NQS satisfies $\psi(g s) = \psi(s)$ for all $g \in G$. The QS-net decomposes the wavefunction as:

$$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$

where $\chi_\mu$ is a generally non-equivariant, symmetry-breaking subnetwork, $\sigma$ is a $G$-invariant nonlinearity (often related to group-theoretic observables, e.g., Wilson-loop operators), and $\Omega_\phi$ is a strictly $s \in \{\pm 1\}^N$0-invariant network block that restores full symmetry (Kufel et al., 2024). The primary objective is to accurately capture ground states with (possibly) only approximate symmetries induced by quantum perturbations.

QS-nets in Discrete Geometry

A Kokotsakis polyhedron with a quadrangular base is called a quasi-symmetric net if the planar ("flat") angles at its sixteen corners satisfy specific linear symmetry relations: opposite corners are grouped so that two are equal and the other two are complements to $s \in \{\pm 1\}^N$1. Explicitly,

$s \in \{\pm 1\}^N$2

where $s \in \{\pm 1\}^N$3 enumerate the flat angles around the central quadrilateral (Nurmatov et al., 24 Nov 2025). These symmetry constraints drastically reduce the design space, enabling constructive classification and explicit solution of flexion in $s \in \{\pm 1\}^N$4.

2. Mathematical Structure and Architectural Ingredients

Quantum QS-nets: Variational Principles and Architecture

The learning objective is to minimize the variational loss

$s \in \{\pm 1\}^N$5

using Monte Carlo sampling of $s \in \{\pm 1\}^N$6, subject to hard symmetry-enforcement ($s \in \{\pm 1\}^N$7) via exact parameter-tying in $s \in \{\pm 1\}^N$8, and with "softness" (approximate symmetry) introduced only by the initial and trainable choice of $s \in \{\pm 1\}^N$9. In some contexts, a "soft" regularization term for approximate symmetry,

$G$0

may be added, though in the main construction the architecture enforces symmetry exactly in its output block (Kufel et al., 2024).

Key architectural details:

• $G$1: local convolutional layers with small receptive fields, initialized to identity mapping; breaking symmetry to learn effective local transformations.
• $G$2: nonlinearity that enforces local gauge invariants (e.g., $G$3 Wilson-loops).
• $G$4: deep, translation-equivariant (and possibly point-group-equivariant) convolutional layers on transformed features, with shared parameters to guarantee strict $G$5-invariance, followed by pooling and exponentiation.

Geometric QS-nets: Algebraic and Elliptic Constraints

The equimodular elliptic type is characterized algebraically via local moduli at each vertex:

• $G$6 (with $G$7, etc.), with all $G$8 equal;
• amplitude-matching conditions $G$9 etc.;
• ellipticity, i.e., absence of resonant angle sums modulo $\psi(g s) = \psi(s)$0;
• periodicity—elliptic function phase-shifts $\psi(g s) = \psi(s)$1 obeying a lattice period condition.

Underlying existence is thus reduced to a small system of trigonometric and polynomial equations, admitting solution procedures suitable for both symbolic and high-precision numerical approaches (Nurmatov et al., 24 Nov 2025).

3. Interpretational Frameworks and Physical Motivation

Quasi-Adiabatic Continuation Perspective (Physics)

At the exactly symmetric ("toric-code") point $\psi(g s) = \psi(s)$2, the system has an exact local gauge symmetry $\psi(g s) = \psi(s)$3. When the gauge is weakly broken (perturbed Hamiltonian), a quasi-adiabatic local unitary $\psi(g s) = \psi(s)$4 exists such that $\psi(g s) = \psi(s)$5 and the ground state admits exact symmetry in "dressed" operators. QS-nets implement $\psi(g s) = \psi(s)$6 as a finite-depth, learned approximation to $\psi(g s) = \psi(s)$7 that untwists perturbed loops so the subsequent block $\psi(g s) = \psi(s)$8 can effectively impose strict group invariance (Kufel et al., 2024). This two-stage approach—in which $\psi(g s) = \psi(s)$9 is trained at the symmetric point and $g \in G$0 is then optimized under perturbation—recovers nearly optimal ground-state energies.

Geometric Flexion Law (Geometry)

QS-nets in geometry yield one-degree-of-freedom (1-DOF) flexible Kokotsakis polyhedra whose motion is parameterized in closed form in terms of a real parameter $g \in G$1. The dihedral angles $g \in G$2 along the four central hinges are expressed via rational and radical trigonometric functions of $g \in G$3, with flexion blocked only at algebraic degeneracies (isolated zeros of a discriminant $g \in G$4) (Nurmatov et al., 24 Nov 2025). This explicit parameterization enables robust and visualizable realization of real 3D flexion paths.

4. Empirical Realizations and Benchmarks

Quantum Variational Performance

On the mixed-field toric code (Hamiltonian $g \in G$5), QS-nets achieve:

• Exact energy convergence to diagonalization accuracy ($g \in G$6 for $g \in G$7, $g \in G$8) in the sign-free case, outperforming RBMs which stagnate at $g \in G$9;
• For sizes up to $$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$0 ($$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$1), energies within $$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$2 of state-of-the-art DMRG and QMC;
• Stability and correct phase identification (via string order parameters and Rényi entropy) in the sign-problem regime ($$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$3), where QMC fails and DMRG becomes memory-limited. After $$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$4 extrapolation, the QS-net locations of confinement transitions match iPEPS/PCUT predictions to within $$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$5 (Kufel et al., 2024).

Explicit Polyhedral Construction and Numerical Pipeline

QS-nets yield the first explicit, constructible class of equimodular elliptic Kokotsakis polyhedra that flex in real space (not just complexified configurations). High-precision solvers (Newton–Raphson and homotopy methods at $$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$6 or better) are deployed to search, filter, and certify candidate solutions based on encoded trigonometric-polynomial system constraints. Reconstructed examples are non-self-intersecting and verified to belong exclusively to the equimodular elliptic class (Nurmatov et al., 24 Nov 2025).

Representative angle assignments for a geometric QS-net:

Flat Angle	Value (degrees)	Related Angles by Symmetry
$$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$7	$$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$8	$$\psi_\theta(s) = \Omega_\phi \circ \sigma \circ \chi_\mu(s)$$9, $\chi_\mu$0, $\chi_\mu$1
$\chi_\mu$2	$\chi_\mu$3	$\chi_\mu$4, $\chi_\mu$5, $\chi_\mu$6
$\chi_\mu$7	$\chi_\mu$8	$\chi_\mu$9, $\sigma$0, $\sigma$1
$\sigma$2	$\sigma$3	$\sigma$4, $\sigma$5, $\sigma$6

These symmetry relations enforce automatic satisfaction of the equimodular algebraic conditions and elliptic period constraint.

5. Algorithmic Construction and Design Guidelines

Quantum QS-nets

• Identify the relevant symmetry group $\sigma$7 (e.g., gauge, loop, or Hamiltonian invariance).
• Construct $\sigma$8 as a strictly $\sigma$9-invariant equivariant network (using group convolutions, equivariant MLPs, or gauge-invariant nonlinearities).
• Insert a small symmetry-breaking subnetwork $G$0 prior to $G$1, initialized near-identity to preserve invariance at training onset.
• Optionally enforce "soft" symmetry by an explicit loss term.
• The representational bottleneck is typically $G$2—increasing its depth or channel count improves performance.
• Markov chain Monte Carlo (MCMC) chains with both single-spin and local gauge-flip proposals assure ergodic and accurate sampling (Kufel et al., 2024).

Geometric QS-nets

• Prescribe half the 16 flat angles (e.g., $G$3); the remaining are fixed via the linear QS-net relations.
• Pose the trigonometric-polynomial existence system (eight algebraic equations) and solve for the internal variables $G$4 under admissible initializations ($G$5, proper sign patterns).
• Filter and certify solutions numerically, ensuring all geometric and algebraic constraints are satisfied to high precision.
• Reconstruct the 3D coordinates via hinge relations and visualize for ranges of the flexion parameter $G$6.
• Closed-form flexion law permits animation and explicit geometric offsets, suited to CAD and deployable mechanism design (Nurmatov et al., 24 Nov 2025).

6. Applications and Broader Impact

Quantum Many-Body and Machine Learning

QS-nets are competitive with or superior to tensor networks and quantum Monte Carlo for both stoquastic and sign-problem Hamiltonians, scaling to hundreds of spins. Their architecture enables interpretable, parameter-efficient detection and study of quantum spin liquids and other symmetry-rich many-body systems, even when dominant symmetries are emergent or only approximately realized (Kufel et al., 2024). A plausible implication is that similar hybrid architectures can be transferred to other models exhibiting large-scale or hidden symmetries.

Geometry, Mechanism Design, Architecture, and Metamaterials

QS-nets provide the first constructive, closed-form flexible mechanisms of the equimodular elliptic Kokotsakis type. Design workflows become efficient: selecting a reduced set of free angles prescribes the entire geometry, and the algebraic system is suitable for integration into CAD solvers, with real-time flexion computation. Applications encompass deployable shells, architectural panels, origami-inspired metamaterials, flexible molds for fabrication, and actuation-constrained robotic linkages. Flexion remains isometric except at discrete "locking" configurations determined by algebraic degeneracies (Nurmatov et al., 24 Nov 2025).

7. Comparison, Limitations, and Generalization

In both domains, QS-nets serve as bridges between strictly symmetric constructions (often too rigid or unrepresentative for perturbed or emergent regimes) and fully unconstrained models (which lack interpretability or may be inefficient). In neural architectures, their main limitation is the expressivity bottleneck rooted in the invariant block; deeper or more expressive $G$7-equivariant backbones or transfer learning strategies (pretraining at symmetric points) can systematically improve accuracy. In geometric contexts, explicit flexion fails only at isolated algebraic degeneracies, and non-self-intersecting QS-nets can be robustly designed using the provided pipeline.

The dual-development of QS-nets in these distinct, technical regimes underscores the utility of quasi-symmetry as an organizing principle: balancing strict invariance and local adaptation enables powerful, tractable solutions for problems in both variational physics and geometric design (Kufel et al., 2024, Nurmatov et al., 24 Nov 2025).