PaperlayerView exact revision on arXiv

Source-linked AI summary

Theory for Equivariant Quantum Neural Networks

Quynh T. Nguyen, Louis Schatzki, Paolo Braccia, Michael Ragone, Patrick J. Coles, Frederic Sauvage, Martin Larocca, M. Cerezo

arXiv:2210.08566v2quant-phcs.LGstat.ML

TL;DR

Quantum neural networks with weak inductive biases face trainability and generalization challenges. This paper develops a theoretical framework for equivariant quantum neural networks, including methods for constructing equivariant channels and group-equivariant QCNNs. It demonstrates an SU(2)-equivariant QCNN advantage on phase classification and discusses applications and potential benefits for broader QML challenges.

  • Problem

    Quantum neural networks with little-to-no inductive biases exhibit trainability and generalization challenges, while systematic symmetry-encoded model design remains needed.

  • Method

    The paper develops a general EQNN framework with generalized Fourier-space analysis, representation-dependent layers, and nullspace, twirling, and Choi-operator construction methods.

  • Results

    The methods construct unitary and non-unitary equivariant layers efficiently for exponentially large symmetry groups, and an SU(2)-equivariant QCNN outperforms a symmetry-agnostic QCNN on Heisenberg-model phase classification.

  • Takeaways & Limitations

    EQNNs provide representation-theoretic blueprints for symmetry-informed QML models applicable beyond supervised learning, including generative and reinforcement-learning settings.

Abstract

from arXiv · show

Quantum neural network architectures that have little-to-no inductive biases are known to face trainability and generalization issues. Inspired by a similar problem, recent breakthroughs in machine learning address this challenge by creating models encoding the symmetries of the learning task. This is materialized through the usage of equivariant neural networks whose action commutes with that of the symmetry. In this work, we import these ideas to the quantum realm by presenting a comprehensive theoretical framework to design equivariant quantum neural networks (EQNN) for essentially any relevant symmetry group. We develop multiple methods to construct equivariant layers for EQNNs and analyze their advantages and drawbacks. Our methods can find unitary or general equivariant quantum channels efficiently even when the symmetry group is exponentially large or continuous. As a special implementation, we show how standard quantum convolutional neural networks (QCNN) can be generalized to group-equivariant QCNNs where both the convolution and pooling layers are equivariant to the symmetry group. We then numerically demonstrate the effectiveness of a SU(2)-equivariant QCNN over symmetry-agnostic QCNN on a classification task of phases of matter in the bond-alternating Heisenberg model. Our framework can be readily applied to virtually all areas of quantum machine learning. Lastly, we discuss about how symmetry-informed models such as EQNNs provide hopes to alleviate central challenges such as barren plateaus, poor local minima, and sample complexity.

I. INTRODUCTION

The paper addresses trainability and generalization challenges in quantum machine learning by incorporating task symmetries into equivariant quantum neural networks. It develops a general framework, construction methods, and QCNN implementations, while discussing broader applications and potential benefits.

  • Quantum models with little-to-no inductive biases face poor trainability and generalization, limiting their scalability.
  • Geometric quantum machine learning uses task symmetries to construct quantum models with sharper inductive biases, but systematic design methods remain limited.
  • The paper provides a theoretical framework for building EQNNs by extending classical equivariant neural networks to quantum models.
  • EQNN layers admit a generalized Fourier-space interpretation that enables parameter counting and representation choices as architectural hyperparameters.
  • Three construction methods—nullspace, twirling, and Choi-operator approaches—can efficiently find unitary or non-unitary equivariant layers for exponentially large symmetry groups.
  • The framework generalizes QCNNs to group-equivariant architectures and introduces an SU(2)-equivariant QCNN that outperforms a symmetry-agnostic QCNN on Heisenberg-model phase classification.
  • The authors discuss applying EQNNs beyond supervised learning and potentially alleviating barren plateaus, poor local minima, and sample-complexity challenges.

B. Symmetry groups and representation theory

The section defines label symmetries, representations, equivariant maps, and invariant models using group representation theory. It then connects equivariance to quantum learning models and illustrates the framework with SU(2)-symmetric phase classification.

  • A QML problem has symmetry group G when a representation of G acts on input states without changing their labels.
  • A representation maps group elements to invertible linear operators while preserving group multiplication, and its commutant contains operators commuting with every represented group element.
  • Irreducible representations contain no non-trivial invariant subspaces and serve as fundamental building blocks for decomposing finite-dimensional representations.
  • Inner symmetries act locally on subsystems, whereas outer symmetries permute subsystems; SU(2) acting as g⊗n is an inner symmetry and qubit permutations form an outer symmetry.
  • An equivariant QNN combined with equivariant measurements produces a G-invariant model.
  • Equivariant maps commute with group actions across input and output representations, while a trivial output representation yields an invariant map.
  • For the bond-alternating Heisenberg model, SU(2) preserves ground-state labels, enabling an EQNN that significantly outperforms a conventional QCNN on phase classification.

IV. THEORY OF EQUIVARIANT LAYERS FOR EQNNS

This section explains EQNN layers through representation-dependent generalized Fourier actions. The representation determines which information layers can access and constrains their parameterization, while larger representations reduce equivariant-unitary freedom.

  • EQNN layers can process different information depending on their input and output representations, making intermediate representation choices important.
  • An equivariant unitary acts in an irrep basis by transforming into generalized Fourier space, operating on multiplicity-space components, and transforming back.
  • The generalized Fourier interpretation extends to quantum channels and connects equivariant quantum layers with group convolution in the Fourier basis.
  • Fixing the representation fixes the information accessible equivariantly, although changes of representation can provide access to otherwise inaccessible components.
  • The Fourier-space structure enables counting the free parameters of equivariant unitaries.
  • Larger group representations produce smaller commutants and therefore fewer parameters for fully characterizing equivariant unitaries.

2. Equivariant channels

The paper develops equivariant quantum channels through representation-aware layer design, parameter counting, and channel constructions. Intermediate representations determine how EQNN layers access information and support standard, embedding, and pooling operations.

  • Equivariant channels are characterized through their Choi operators, which lie in the commutant of Rin*⊗Rout.
  • The parameter utilization metric compares all CPTP maps with their equivariant subspace to quantify the benefit of symmetry constraints.
  • At most 14 free parameters remain for SU(2)-equivariant 2-to-2 qubit channels, giving a parameter-utilization reduction of µ ≥ 240/14 ≈ 17.
  • A layered EQNN is defined by a sequence of group representations connected by equivariant layers and followed by equivariant measurements for invariant outputs.
  • Intermediate representations act as hyperparameters because their irrep multiplicities determine the generalized Fourier spaces and information accessible to each layer.
  • Standard, embedding, and pooling layers preserve, increase, and decrease Hilbert-space dimension, respectively, while embedding layers can access higher-dimensional irreps.

2. Projection and lifting

Projection and lifting layers classify equivariant transitions by how representation kernels change, while repeated input copies provide equivariant nonlinear embeddings. The construction methods impose equivariance through nullspaces, group generators, twirling, or Choi representations.

  • Projection and lifting: Projection layers satisfy ker(Rl−1) < ker(Rl), whereas lifting layers satisfy ker(Rl−1) > ker(Rl).
  • Projection and lifting: Projection layers are useful with pooling under outer symmetries because reduced systems often lack faithful representations.
  • Projection and lifting: Lifting to a larger group may reduce expressibility and improve trainability, but the actual benefit of lifting layers is not known.
  • Nonlinearities: Applying ρ → ρ⊗k creates an order-k non-linear equivariant embedding when repeated access to the input state is available.
  • Construction methods: The nullspace method solves equivariance constraints on a generating set to obtain a basis of equivariant linear maps.
  • Construction methods: For compact Lie groups, equivariance can be imposed using a Lie-algebra generating set, provided exponentiation is surjective.
  • Construction methods: The exponential-map surjectivity assumption can be relaxed by adding further constraints.

2. Twirling method

The twirling method projects arbitrary quantum maps onto the space of equivariant maps, extending from unitary channels to general channels and Lie groups. It preserves CPTP structure but is most convenient for obtaining individual equivariant channels rather than a complete basis.

  • The framework extends twirling from equivariant unitary channels to general non-unitary channels with possibly different input and output representations.
  • For finite groups twirling averages over group elements, while Lie-group twirling replaces the sum with an integral over Haar measure.
  • Twirling is the projection onto the space of equivariant maps, leaving equivariant channels unchanged.
  • The anti-symmetric component satisfies TG[ϕA] = 0, so its norm quantifies how symmetric a map ϕ is.
  • Direct twirling is efficient for small groups but becomes cumbersome for large finite or Lie groups, motivating analytical, in-circuit, and approximate implementations.
  • Unlike the nullspace method, twirling guarantees that the twirl of a CPTP channel remains CPTP, but it finds maps one at a time rather than directly producing a complete basis.

C. Parametrizing the layers of an EQNN

The paper develops parameterization and optimization procedures for equivariant unitaries and general channels, then applies them to locality-constrained SU(2)-equivariant QCNN layers. These constructions expose expressivity–locality trade-offs while yielding concrete equivariant layer spaces and pooling maps.

  • EQNN layers are parameterized and optimized for both unitary channels and general quantum maps.
  • Equivariant unitary layers can use exponentials of equivariant Hermitian generators, with generators found through nullspace or twirling methods.
  • General-channel parameterization combines equivariant maps subject to trace-preserving and complete-positivity constraints, with updates projected back into the feasible region when necessary.
  • SU(2)-equivariant QCNN: The framework alternates local 2-to-2 equivariant channels with 2-to-1 equivariant pooling channels to construct SU(2)-equivariant QCNNs.
  • 2-to-2 layers via Choi operator: For 2-to-2 SU(2)-equivariant completely positive maps, the Choi representation has a 14-dimensional parameter space.
  • 2-to-1 layers via Nullspace: Five superoperators form a basis for 2-to-1 qubit SU(2)-equivariant maps, including trace, partial traces, SWAP-measurement, and a cross-product map.

B. Various examples and physical considerations

The paper applies its equivariant-layer construction to discrete symmetries, including Z2×Z2 and Zn, using twirling, local pooling, parameter sharing, and randomized channels.

  • Z2×Z2 symmetry: For Z2×Z2 symmetry, twirling identifies the full subalgebra of equivariant unitary generators in u(2^n).The generators are characterized through Pauli strings with parity constraints on Y and Z operators across even and odd qubits.
  • Z2×Z2 symmetry: Z2×Z2-equivariant pooling is constructed from local 2-to-1 maps whose input and output representations are specified on neighboring qubits.These local maps can be combined into a global equivariant channel.
  • Zn symmetry: For Zn symmetry, twirling local generators produces translationally repeated operators, connecting equivariance to correlated parameters across local gates.This construction is straightforward when the twirled terms mutually commute.
  • Zn symmetry: Noncommuting global generators can be difficult to implement, motivating randomized constructions that combine locally implementable Zn/2-equivariant unitaries into Zn-equivariant channels.The randomized channel applies odd- or even-site unitaries with equal probability.
  • Projection and pooling: Randomized projection layers can reduce n qubits to n/2 while preserving Zn equivariance, and analogous constructions extend to outer symmetries such as Sn.Hoeffding’s bound gives O(log |G|) samples to reach a specified error for this sampling method.

VII. NUMERICAL EXPERIMENTS

The numerical section evaluates an SU(2)-equivariant QCNN on phase classification in the bond-alternating XXX Heisenberg model, using symmetry-compatible layers and measurements. The model is compared with a symmetry-agnostic QCNN through supervised training and test accuracy.

  • Physical model: The bond-alternating XXX Heisenberg model alternates exchange couplings J1 and J2, with a phase transition at α = J2/J1 = 1.For α < 1 the system is trivial, while α ≥ 1 corresponds to the topologically protected phase.
  • Physical model: The Hamiltonian has SU(2) symmetry represented globally by g⊗n, and this symmetry also preserves the quantum phase labels.The ground space remains invariant under the symmetry because [H, g⊗n] = 0.
  • SU(2)-EQCNN: The SU(2)-EQCNN uses two-qubit gates U(θ) = e−iθSWAP, with parameter sharing within translation-symmetric convolutional sublayers.Each standard convolution layer therefore has two parameters.
  • SU(2)-EQCNN: The architecture ends with two qubits so that the nontrivial SWAP observable can distinguish the two phase labels.A single-qubit SU(2)-invariant measurement would be trivial, whereas SWAP has eigenvalues ±1.
  • Comparison model: The baseline QCNN uses hardware-efficient convolutional circuits and alternate partial traces for pooling, without symmetry-based inductive biases.Both models use the same SWAP-based measurement and threshold assignment for phase prediction.

D. Training results

The SU(2)-equivariant QCNN classified phases effectively, outperforming the symmetry-agnostic HEA-QCNN for odd system sizes, while equivariance was unnecessary for even sizes. The numerical study remains preliminary and excludes general 2-to-1 equivariant pooling because projection instability prevented training.

  • Implementation limitation: The study could not train general 2-to-1 equivariant pooling because projection onto the feasible CPTP region appeared to destabilize optimization.The reported numerics therefore use simpler 2-to-1 partial-trace pooling operations shared by the EQCNN and HEA-QCNN.
  • Performance comparison: The SU(2)-EQCNN separated the two phases with excellent accuracy, whereas HEA-QCNN predictions became cloudy for N = 13.Both architectures used 750 training epochs, with comparable training-parameter counts achieved by stacking EQCNN layers before pooling.
  • Performance comparison: For N = 12, the EQCNN did not significantly outperform the HEA-QCNN because the unique ground state is invariant under the Hamiltonian symmetry.Odd system sizes have degenerate ground states, allowing symmetry actions to rotate states within the ground space and making equivariance more useful.
  • Statistical analysis: Enforcing equivariance when unnecessary can reduce expressibility and produce training instabilities, as shown by large error bars for even system sizes.The statistical study compared EQCNN and HEA-QCNN across even and odd sizes using multiple randomly initialized training runs.
  • Broader scope: The paper presents EQNNs as a general framework for equivariant channels, with potential applications beyond supervised learning to unsupervised, generative, and reinforcement learning.The framework extends prior unitary-model results to general equivariant channels and is described as applicable to multiple quantum-learning settings.
  • Framework: The authors present three construction methods for EQNN layers: nullspace solving, group twirling, and Choi-operator techniques.These methods target unitary and non-unitary equivariant layers and address optimization and implementation considerations.

Appendix A: Equivariance in existing quantum algorithms

The appendix interprets several quantum algorithms and QCNN constructions through equivariance, showing how their operations respect task symmetries. It also identifies the exact Haldane-model QCNN as equivariant while describing classical-shadow and purity-related procedures using equivariant maps or observables.

  • Equivariance provides a perspective for analyzing non-variational quantum algorithms beyond variational quantum machine learning.
  • Purity estimation uses unitary-equivariant transformations because purity is invariant under unitary transformations.The Bell-basis procedure is equivalent to measuring SWAP^⊗n, an observable in the commutant of the two-copy unitary representation.
  • A one-copy randomized procedure appends zero-initialized qubits, applies a random larger-system unitary, and estimates repeated measurement probabilities.
  • The exact Haldane-model QCNN has equivariant convolution, pooling, and final measurement layers, making the overall model Z2 × Z2-invariant.The convolution commutes with the symmetry action, pooling maps between corresponding reduced representations, and the final ZXZ observable lies in the output commutant.
  • The classical-shadow protocol can be interpreted as a composition of equivariant layers when its averaged observable commutes with the relevant representation.For the Clifford ensemble, the resulting observable commutes with the two-copy representation of U(2^n).

Appendix C: Deferred proofs

These deferred proofs establish parameterization and representation properties of equivariant quantum channels. They use Choi-operator decompositions to count free parameters and show that equivalent intermediate representations preserve EQNN expressibility.

  • Equivariant channels are parameterized through the irreducible-representation decomposition of the combined input-output representation.
  • Complete positivity is imposed blockwise on Choi-operator components, while trace preservation is enforced through a partial-trace constraint.
  • The rank of the trace-preservation constraint matrix equals C(Rin, Rout), reducing the number of independent real parameters accordingly.
  • In the non-equivariant case, the trace-preservation constraint removes dim(HA)^2 free parameters from a fully parameterized channel.
  • Replacing an intermediate representation with a unitarily equivalent representation does not change EQNN expressibility.Equivariant channels before and after the representation change correspond bijectively, and the composed EQNN map remains unchanged.

2. Deferred proofs from Section V

The deferred proofs reduce equivariance checks for finite and Lie groups to generators, and clarify how twirling projects maps into the commutant. They also describe algebraic and numerical routes for implementing these projections.

  • For finite groups, equivariance with respect to a generating set implies equivariance with respect to every group element.
  • For compact Lie groups with surjective exponentiation, checking commutation with generators of the Lie algebra is equivalent to checking group equivariance.
  • When exponentiation covers only a connected component, equivariance can be enforced using Lie-algebra generators together with generators of a subgroup connecting the remaining components.
  • For SO(3), exponentiation is surjective, whereas O(3) additionally requires a subgroup element to connect its two determinant components.
  • Twirling projects an operator or map onto the commutant, allowing its coefficients to be obtained by solving a linear system built from commutant basis elements.The Gram matrix contains traces of products of commutant basis elements, and its inverse is the Weingarten matrix.

2. In-circuit twirling with ancillas or classical randomness

This section describes practical implementations for twirling and equivariant channels, using ancillas, classical randomness, or approximate procedures. It also covers channel dilation, parameter constraints, optimization, and an implementation benefit for irreducible output representations.

  • In-circuit twirling: Small finite-group twirling can be implemented in-circuit with log2|G| ancillas prepared in a uniform superposition.The circuit can twirl any parametrized channel native to the hardware platform.
  • In-circuit twirling: Classical sampling of group elements can replace ancillas, and O(log |G|) samples suffice for a good approximate twirled channel by Hoeffding’s bound.
  • In-circuit twirling: In-circuit twirling preserves equivariance but does not provide the parameter-count reduction obtained by analytically parameterizing equivariant channels.
  • Channel implementation: A general quantum channel can be implemented through a unitary dilation whose environment dimension is determined by its Kraus rank.For an n-to-m-qubit channel, the maximum Kraus rank is 2^(m+n), potentially requiring up to m+n ancilla qubits.
  • Optimization: Equivariant channels can be optimized with projected gradient descent in either circuit parameters θ or classical variables x, subject to complete positivity and trace preservation.
  • Channel implementation: Equivariant channels with irreducible output representations are unital, enabling ancilla recycling through measurement and controlled unitaries.

Appendix F: SU(2)-equivariant 2-to-1-qubit and 1-to-2-qubit channels

The section characterizes SU(2)-equivariant maps between two and one qubits, deriving their basis, trace-preserving constraints, and feasible channel region. It also shows that complementary channels access different components of the input state.

  • Five SU(2)-equivariant linear maps are introduced as a basis for constructing the relevant 2-to-1-qubit channels.
  • The second basis map is removed because it increases the trace on certain inputs, while the other maps cannot cancel that increase.
  • Modified basis elements are chosen so that all maps except the trace-preserving channel annihilate trace, simplifying characterization of the feasible channel set.
  • Nonnegative Choi-operator eigenvalues determine the feasible region for the channel coefficients, including constraints on y and z.
  • The channel analyses show that different equivariant channels combine different pieces of the input density matrix and are therefore complementary.
  • The resulting channel expressions distinguish diagonal and off-diagonal Bell-basis information while preserving the input-trace condition.

3. Cross-product channel

The cross-product channel is an SU(2)-equivariant 2-to-1-qubit map that detects coherence between symmetric and antisymmetric sectors. Its implementation and physical realization require careful treatment of trace preservation, complete positivity, and the channel’s real–imaginary asymmetry.

  • The cross-product channel maps two qubits to one and combines antisymmetric Pauli commutators into an output vector.
  • The vanilla cross-product map is traceless but is not itself trace preserving or completely positive; adding a trace-preserving map can yield a physical channel for an allowed coefficient range.
  • The channel detects superpositions between symmetric and antisymmetric Bell states through specific off-diagonal density-matrix entries.
  • It outputs zero for states block diagonal in the symmetric and antisymmetric subspaces, so it specifically probes coherence between those sectors.
  • The vanilla construction captures only either real or imaginary matrix components, motivating an alternative equivariant version with a Hermitian output.
  • Combining the two cross-product versions can recover off-diagonal terms between symmetric and antisymmetric subspaces after adding sufficient depolarizing noise to ensure physicality.
  • A circuit realizes the original cross-product channel in expectation by random Pauli rotations that encode Tr[σ_iσ_jρ] into ancilla Pauli expectations.
  • The same circuit does not implement the alternative version because the relevant operator is non-Hermitian; a Hadamard test can implement it in expectation instead.
Loading 2210.08566v2…