PaperlayerView exact revision on arXiv

Source-linked AI summary

Symmetry Discovery in Quantum Learning: Observable-Level and Task-Level Inference from Finite Measurements

Zeyu Chen

arXiv:2610.00157v1quant-phcs.LG

TL;DR

Quantum-learning symmetries must match both what measurements retain and how labels transform. This paper develops finite-measurement methods to infer supported symmetries and shows how probe-span structure and task-level validation constrain which groups can be imposed.

  • Problem

    A symmetry constraint determines which distinctions a quantum model can represent, and invariant prediction can suppress an odd label distinction that an equivariant predictor would preserve.

    Source: §1 · p. 1

  • Method

    The paper combines projected-state stabilizer analysis under invariant probe spans with shadow-based finite-dictionary recovery and task-level symmetry tests.

    Source: §1 · p. 2, §4.2 · p. 13, §10 · p. 27

  • Results

    Finite-dictionary recovery has an inverse-gap measurement rate, and a commuting-qubit lower bound shows this gap dependence is necessary at fixed snapshot scale.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The measurements support imposing the closed subgroup generated by transformations validated at both the observable and task levels.

    Source: §9 · p. 27, §10 · p. 27

    A zero encoded-mean discrepancy supports symmetry only for the declared representation and predictor interface, since distinct distributions can share a mean.

    Source: §4.2 · p. 15

Abstract

from arXiv · show

Symmetry reduces the capacity of a quantum learning model, but the imposed group must match both the measured information and the label transformation. We establish a finite-measurement theory for inferring this group from candidate transformations. The central structural result identifies observable-invisible transformations with the stabilizer of a projected state whenever the probe span is invariant. It turns recovered generators into a valid subgroup and identifies the continuous invisible space with its Lie algebra. For finite dictionaries, an unbiased shadow statistic distinguishes zero from positive squared expectation discrepancies with an inverse-gap measurement rate, improving the inverse-square-gap rate of uniform discrepancy estimation. A commuting qubit lower bound proves the gap dependence optimal at fixed snapshot scale, and simultaneous intervals support data-dependent tolerances. Task validation then tests either the joint distribution through a characteristic kernel or its encoded mean through a classical--quantum discrepancy. An exact group-average identity relates the latter to joint-state asymmetry and specifies its conversion to binary task breaking mass. Projection bias quantifies the cost of excessive symmetry, while an $\ell_1$ readout bound quantifies the capacity gained by relaxing it. At an invariant pure-state backbone, retained and nontrivial breaking sectors are Fisher-orthogonal. Ising-chain calculations connect finite-shot recovery, label-dependent symmetry, and physical sector drift. These results determine which symmetry the measurements support and provide the statistical and geometric basis for a subsequent release decision.

1 Introduction

The introduction argues that a useful quantum-learning symmetry must fit both the information measured and the label transformation. It presents projected-state recovery, gap-sensitive finite-dictionary testing, and separate task validation, while distinguishing these goals from related symmetry and optimization methods.

1 Introduction

  • The symmetry constraint must match both the measurement interface and the input–label law, since observable invisibility and label invariance are distinct requirements.The introduction notes that a spin-flip-symmetric ensemble may carry an odd magnetization label, which calls for sign equivariance rather than an invariant predictor.Source: §1 · p. 1
  • When the probe span is invariant, observable-invisible transformations stabilize the projected state, and their infinitesimal null space is its Lie algebra.Exact recovered generators then generate an invisible subgroup; restricted probes can still allow second-order changes despite a bracket-closed tangent null space.Source: §1 · pp. 1–2
  • Finite-dictionary recovery uses an unbiased quadratic statistic for zero versus positive squared discrepancies, with inverse-gap measurement dependence and a commuting-qubit lower bound.Simultaneous intervals allow data-dependent tolerances, while a covariance estimator supports recovery of continuous invisible directions.Source: §1 · p. 2
  • Task validation can test the joint input–label distribution or an encoded quantum mean, connecting joint-state asymmetry to binary task-breaking mass.Projection bias and an ℓ1 readout bound quantify approximation and capacity costs, while representation theory gives Fisher orthogonality at an invariant pure-state backbone.Source: §1 · p. 2

Related work

  • Prior work studies equivariant architectures and trainability, state-stabilizing algorithms [9], and learned transformation structure.This paper instead targets a specified quantum measurement interface and fixed candidate transformations, with separate validation of the label action.Source: Related work · p. 2
  • Prior quantum symmetry tests target full-state symmetry or hidden-subgroup recovery [16], while this work characterizes the group visible to a restricted probe span and assumptions for identifying it with a physical stabilizer.Source: Related work · p. 2

2 Setting and separation of symmetry levels

The paper separates state, observable, task, and parameter-redundancy symmetries, then defines a dictionary-relative recovery target. Observable invisibility need not form a subgroup unless the probe-span condition supplies closure.

2 Setting and separation of symmetry levels

  • The setup takes a finite or compact group continuously represented by unitaries on a finite-dimensional Hilbert space, with normalized Haar averages on closed subgroups.Distributional statements use Borel probability laws on Polish input and label spaces with continuous group actions.Source: §2 · p. 3
  • State-level symmetry preserves a density operator, observable-level symmetry preserves chosen expectations, task-level symmetry preserves the joint example law, and parameter redundancy preserves the represented state or task function.Parameter redundancy is treated separately through quotient-space quantum natural gradient.Source: §2 · p. 3
  • Finite-dictionary recovery fixes a nonempty candidate dictionary and probe maps independently of estimation samples, though independent pilot data may propose them.The target is dictionary-relative to the relevant stabilizer.Source: §2 · p. 3
  • State- and task-level stabilizers are subgroups, but observable-invisible candidates need not compose into invisible transformations without the invariant-span condition.With a subgroup guarantee, the generated subgroup's closure preserves exact invariance; approximate acceptance controls only the tested tolerance and bounded-length words, not exact invariance of the generated group.Source: §2 · p. 3

3 Observable-level symmetry discovery from finite quantum measurements

This section studies observable-level symmetry discovery from finite quantum measurements, identifying transformations invisible to probe families and developing finite-dictionary and continuous recovery guarantees. It also examines cyclic-subgroup recovery from memory-basis occupation patterns, including finite-sample threshold guarantees under added uniform noise.

3.1 Operator-algebraic discrepancy

  • With an invariant probe span, exactly invisible transformations form the stabilizer of the projected state, and generators with discrepancy at most τ have discrepancy at most ℓτ on words of length ℓ.The projected state is the Hilbert–Schmidt projection of ρ onto the probe span; without the invariant-span condition, a restricted-probe null space can miss higher-order changes. Invisibility is relative to the measured span unless the probes separate the candidate orbit.Source: §3.1 · pp. 4–5

3.2 Classical-shadow estimator and recovery guarantee

  • The finite-dictionary shadow statistic recovers zero versus positive squared discrepancies at an inverse-gap measurement rate, with gap dependence optimal at fixed snapshot scale.A commuting one-qubit lower bound establishes the inverse-gap dependence; uniform additive discrepancy estimation would instead require an inverse-square-gap rate. The result does not establish optimal dependence on probe locality or dictionary size.Source: §3.2 · pp. 5–6

3.3 Continuous discovery through a Lie-algebraic null space

  • Continuous discovery recovers first-order invisible directions as the low-eigenvalue space of a K-dimensional observable covariance matrix, with stable subspace recovery when the nonzero spectrum has a positive gap.Under the invariant-span condition, the null space is the invisible subgroup’s Lie algebra; for general restricted probes it certifies only first-order invisibility. Complete probes identify the full state stabilizer, while local probes identify the projected tangent response and require injectivity for full-stabilizer agreement.Source: §3.3 · pp. 7–10
  • For a real state, probes all in the same reality class as candidate generators can make those generators first-order invisible even when they are not symmetries.Resolving both classes requires opposite-class probe responses; in the XXZ example, adding specified probes reduces the kernel to the U(1) generator Mz.Source: §3.3 · p. 8

3.4 Cyclic finite-memory benchmark

  • In the cyclic finite-memory benchmark, discovering symmetry from local projector probes is equivalent to finding shifts that preserve the observed memory occupation pattern.The discrepancy is determined by the observed distribution p_a.Source: §3.4 · p. 10
  • With M independent memory-basis measurements, thresholding b_DO(r_ℓ) at γC/2 recovers K with probability at least 1−δ under the stated sample condition.The empirical distribution from the measurements is substituted into b_DO.Source: §3.4 · p. 10
  • The threshold γC/2 separates null values at most γC/16 from alternatives at least 9γC/16.The passage attributes this gap to the stated sample condition.Source: §3.4 · p. 10
  • For m = 12 and K = ⟨r4⟩, Figure 1 plots empirical recovery success over 2,000 simulations as uniform noise ε is added to a distribution uniform on {0, 4, 8}.Recovery uses the γC/2 threshold from Proposition 3.11.Source: §3.4 · p. 10, Figure 1 · p. 11

4 Task-level discovery for supervised quantum learning

Task-level symmetry must describe the joint transformation of inputs and labels, not just invariance of the input ensemble. The section develops classical and quantum-native tests, with the latter assessing symmetry of an encoded task mean and relating it to joint-state asymmetry.

4 Task-level discovery for supervised quantum learning

  • A task symmetry transforms inputs and labels together; an invariant input ensemble can still have an order-parameter label requiring a nontrivial output action.Source: §4 · p. 11

4.1 Maximum mean discrepancy on the joint distribution

  • With a characteristic kernel, MMD tests invariance of the joint input–label law; the covariance-commutator residual is only a candidate pre-screen.Under exact feature equivariance and an orthogonal feature action, joint-law invariance implies the covariance commutation condition; the residual depends on the features and can fail under approximate equivariance.Source: §4.1 · pp. 11–12

4.2 Quantum-native task discrepancy through classical–quantum states

  • The quantum-native discrepancy compares the encoded classical–quantum task mean with its transformed state, and zero discrepancy implies joint-law invariance when the encoding is injective.For finite alphabets encoded by orthogonal basis states and permutation actions, the mean embedding is injective on the joint law.Source: §4.2 · pp. 14–15
  • A reusable classical-shadow bank gives an unbiased CQ-discrepancy estimator without global SWAP measurements or additional quantum measurements for each dictionary candidate.The order-two statistic excludes same-example pairs to remove self-overlap bias; under the stated moment bound and margin, thresholding recovers the zero-discrepancy candidates.Source: §4.2 · pp. 14–15
  • Group-averaged CQ discrepancy equals joint-state asymmetry, while the binary breaking-mass relation depends on the label representation and class probabilities.For a spin flip, the sign-label action can pass while the invariant-label test fails; a passed sign-action test can coexist with a positive odd-sector gate.Source: §4.2 · p. 16

5 Symmetry mismatch and statistical complexity

Imposing excessive symmetry can exclude task-relevant regression functions, while relaxing symmetry enlarges the readout class and changes its estimation complexity. The section quantifies projection bias and gives generalization bounds whose dictionary-size dependence is logarithmic under specified readout constraints.

5.1 Projection bias from imposing too much symmetry

  • An overly large enforced group incurs projection bias by excluding part of the regression function.The theorem frames this as irreducible bias under excessive symmetry.Source: §5.1 · p. 17, §5 · p. 17

5.2 Commutant dimension and logarithmic generalization scale

  • For an independently selected symmetric readout basis with operator norm at most one and an ℓ1 coefficient constraint, the generalization scale depends logarithmically on the commutant dictionary size.The bound assumes bounded Lipschitz loss and independent labeled examples; choosing among J classes on the same risk sample requires a simultaneous bound.Source: §5.2 · p. 18
  • The full permutation group's commutant is identified with the collective End(C2) action and is isomorphic as a complex vector space to Sym^n(C4).The section uses this commutant dimension to express the dictionary-size change under permutation symmetry.Source: §5.2 · p. 18
  • Relaxing symmetry enlarges the ℓ1 readout dictionary and changes the capacity bound through √log m_H, while whether prediction improves depends on the approximation error removed.The comparison is conditional on fixed coefficient radius and normalized basis observables.Source: §5.2 · p. 18

6 Retaining and releasing symmetry

After discovery and task validation, a retained symmetry can serve as a backbone while selected symmetry-breaking directions are released. At an invariant pure-state backbone, retained and nontrivial breaking directions have block-orthogonal Fisher geometry under the stated conditions.

6.1 Hard, soft, and released models

  • A hard model enforces the retained invariant or equivariant class, whereas a released model activates generators with components outside the commutant.A soft gate controls the allowed amplitude of released branch coordinates under a fixed radius constraint.Source: §6.1 · p. 19, §6 · p. 19
  • At a stationary release point, positive κ_B gives second-order descent along the released branch, but nonlinear supervised losses require a chain-rule Hessian correction.Activation still depends on simultaneous measurement bounds, release amplitudes, and independent model validation.Source: §6.1 · p. 19

6.2 Geometric decoupling at the symmetry release point

  • At an invariant pure-state backbone, the QFIM is block diagonal between retained symmetric tangents and breaking generators with no trivial retained-group component.The multi-branch result additionally requires the breaking directions to lie in pairwise non-isomorphic nontrivial real irreducible sectors.Source: §6.2 · pp. 19–20
  • The Fisher inverse respects this decomposition at the invariant point, although the task Hessian may still couple directions within an isotypic block.For one fixed left-applied branch the retained–branch cross block stays zero for all β; general multibranch parameterizations need only decouple at the origin.Source: §6.2 · p. 20

7 Optimization and finite-size diagnostics

The section distinguishes physical symmetry breaking, local parameter distinguishability, and representation-sector redistribution as complementary diagnostics. It develops conditions for interpreting their finite-size and finite-shot behavior, including a transverse-field Ising baseline.

7.1 Observable, Fisher, and sector diagnostics

  • Persistent probe deviations, sector weights, and the Fisher spectrum respectively track observable symmetry breaking, representation-content redistribution, and local distinguishability in chosen coordinates.Consistency across runs and system sizes strengthens physical interpretation, but no universal release criterion requires all three diagnostics to change.Source: §7.1 · p. 21

7.2 Physical quantum Fisher scaling

  • Under the stated translation-invariant pure-state and correlation assumptions, the physical Fisher signal grows as L^(2−z−η) when z+η<2, logarithmically when z+η=2, and remains bounded when z+η>2.These are physical unitary-family scalings; relating them to circuit Fisher eigenvalues requires a controlled tangent map and specified coordinate normalization.Source: §7.2 · p. 22
  • The Ising exponents describe the specified magnetization-generated family and do not determine the smallest eigenvalue of a variational circuit's multi-parameter Fisher matrix.The circuit comparison needs an aligned tangent direction and fixed coordinate normalization.Source: §7.2 · p. 22, §7.5 · p. 24

7.3 Sector drift from symmetry-breaking perturbations

  • For a symmetry-breaking generator supported in one nontrivial sector, the sector weight begins quadratically in the perturbation, with coefficient tied to the physical Fisher information.The fourth-order remainder is stated at fixed system size under the single-sector support condition.Source: §7.3 · pp. 22–23

7.4 Shadow estimation of physical Fisher signals

  • For one-dimensional transverse-field Ising critical magnetization, where dlat = 1, w = 1, and γ = 7/4, the measurement requirement is N = O(L^1/2ϱ^-2 log(L/δ)).Source: §7.4 · p. 23

7.5 Transverse-field Ising scaling baseline

  • The Ising magnetization Fisher signal grows faster with L in the critical and ordered regimes than in the disordered regime.For h = 1.5, 1.0, and 0.5, respectively, the fitted exponents are approximately 1.048, 1.759, and 1.993.Source: §7.5 · p. 24, Figure 2 · p. 24

8 Worked example: the transverse-field Ising chain

The transverse-field Ising example separates observable-level recovery from task-level symmetry validation, showing that the same spin flip can receive different verdicts depending on the label action. It then connects release geometry, sector drift, and local energy curvature on specified backbones.

8.1 Observable level

  • At h = 1 on the symmetric ground state, the observable-invisible set contracts to {T, R} when the spin flip changes sign on the magnetization probes.The resulting margin is γO = 0.0532.Source: §8.1 · p. 25
  • At M = 300 on the symmetric ground state, exact-recovery success is 0.92 for the unbiased U-statistic and 0.658 for the plug-in estimator.The plug-in estimator's positive finite-sample bias pushes symmetric elements above the threshold in this example.Source: §8.1 · p. 25, Figure 3 · p. 25

8.2 Task level

  • With an invariant scalar label the task rejects the spin flip at every tested field, whereas with a sign label it is exactly Z2-equivariant.For the invariant label, DCQ(P; triv) is 0.5000, 0.4595, and 0.0833 at h = 0.5, 1.0, and 1.5, respectively.Source: §8.2 · pp. 25–26

8.3 Release point

  • At the reported release point, retained–breaking Fisher cross entries vanish below 10^-14 and the branch block is 13.108.The cross entries vanish at every θ by parity; the zero within the retained block at this point is specific to the chosen backbone.Source: §8.3 · p. 26

8.4 Diagnosis

  • At the product backbone, odd-sector weight follows the predicted initial quadratic drift, with observed values 0.00319, 0.01964, and 0.07439 at β = 0.02, 0.05, and 0.1.The quadratic predictions are 0.0032, 0.0200, and 0.0800; deviation grows with β as the quartic term appears.Source: §8.4 · p. 26

8.5 Loss curvature and the release decision

  • Along M_y, release curvature is positive at the product state but negative on the exact symmetric ground state at every reported field.At h = 0.5, 1.0, and 1.5, the product-state values are 48.0, 32.0, and 16.0; the ground-state curvature ranges from −57.4 to −10.2.Source: §8.5 · p. 26
  • A finite-shot release decision must additionally control estimation error, the step remainder, and the statistical cost of enlarging the predictor.The curvature calculations identify the local task signal, while discovery and release geometry provide distinct inputs to the decision.Source: §8.5 · p. 26

9 Permutation and lattice symmetries

Permutation and lattice symmetries provide candidate transformations whose observable spans and label actions must both be validated. The retained constraint is the closed subgroup supported at both levels, with qualifications for approximate and continuous recovery.

Permutation symmetries

  • Adjacent transpositions form a candidate dictionary, and complete fixed-weight Pauli probe spans are invariant under these swaps.Local probe invisibility must still be distinguished from full many-body invariance unless candidate state variations are separated.Source: §9 · p. 27

Lattice symmetries

  • On finite periodic lattices, translations and point-group generators form a compact dictionary, while selective geometric probes require checking their orbit span.Density-wave, nematic, and magnetization observables specify different possible label representations and breaking directions.Source: §9 · p. 27
  • In quantum convolutional models, equivariant pooling must intertwine fine- and coarse-lattice actions for architectural changes to have a controlled sector interpretation.Site permutations preserve complete fixed-weight probe spans.Source: §9 · p. 27

Recovery and validation

  • The retained constraint is the closed subgroup generated by transformations supported at both the declared observable and task levels.A joint confidence statement allocates failure probability across simultaneous discovery and task events.Source: §9 · p. 27
  • For continuous candidates, spectral estimates give a nearby infinitesimal subspace, but noisy basis vectors need not form an exact Lie algebra.Exact fitted groups require structural constraints or additional validation of their finite action.Source: §9 · p. 27

10 Conclusion

The paper connects finite-measurement symmetry recovery to task-level label actions and to model choices across function, statistical, and geometric levels. It also identifies remaining work in calibrating tolerances to task loss and converting noisy continuous estimates into certified group actions.

Conclusion

  • Inverse-gap finite-dictionary recovery has a matching qubit lower bound, while projection turns observable invisibility into a subgroup condition and its infinitesimal version into a Lie algebra.Approximate acceptance propagates to finite compositions.Source: §10 · p. 27
  • Characteristic kernels test the full joint law, while a classical–quantum discrepancy tests its encoded mean and connects it exactly to joint-state asymmetry.Projection bias, norm-constrained readout capacity, and retained–breaking Fisher orthogonality describe model choice at function, statistical, and geometric levels.Source: §10 · p. 27
  • Further work includes calibrating probe-discrepancy tolerances to task loss, finding smaller separating families for non-invariant spans, and certifying finite actions from noisy tangent estimates.The paper identifies these as questions about the information and geometry needed to impose a constraint.Source: §10 · p. 27
Loading 2610.00157v1…