acceptodds
Under review as a conference paper at ICLR 2027

A Symmetry-Breaking Dial for Equivariant Quantum Machine Learning

Abstract

Symmetry-preserving circuits encode conservation laws into quantum machine learning models, but a fully equivariant circuit cannot distinguish states related by that symmetry. This creates a structural blind spot on mirror-pair tasks, where the label identifies one of two symmetry-related ground-state branches. We introduce a continuous symmetry-breaking dial that rescales symmetry-breaking gate angles while keeping the parameter count fixed, interpolating between equivariant and unconstrained ansatze. We prove that the equivariant endpoint has exactly accuracy on every mirror pair and that the class separation is odd in the dial, with a linear coefficient determined in closed form by the initialization and the data. With a preconditioned optimizer, the smallest dial value reaching accuracy follows where the grid resolves it, and extends to a calibrated joint law in budget and task strength. On transverse-field mirror-pair tasks, this prescription keeps the dial below , breaking only a fraction of a percent of the symmetry, while reaching full accuracy. The transport law persists across depth, register size, and a decade of task strength, with recalibrated constants. We also identify a zero-transport boundary where the circuit cannot propagate the mirror distinction to the readout. Beyond quantum models, whenever the label is defined by the symmetry, full equivariance is a provable ceiling rather than a safe default.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
quantum machine learning, equivariant networks, variational quantum circuits, symmetry breaking, trainability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YHdhOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies a blind spot of equivariant QML: on a transverse-field Ising chain the two ferromagnetic ground-state branches are related by a parity flip, so a "mirror-pair" task whose label is defined by that symmetry cannot be separated by a circuit and readout that both respect it. It proves the accuracy of such a model is pinned at exactly 50% (Theorem 1), then proposes a continuous dial that rescales only the angles of the symmetry-breaking block at fixed parameter count, interpolating between the equivariant and unconstrained ansatze. The margin is exactly odd in the dial (Theorem 2) with a closed-form slope computable before training (Corollary 1), and under Adam the smallest sufficient breaking is λ⋆(T)≈0.3/T\lambda^\star(T) \approx 0.3/T (Proposition 1), while plain gradient descent stalls. Simulations on mirror-pair tasks confirm the ceiling, the linear transport, the budget law in its grid-resolved range, a two-variable extension in task strength, and a "zero-transport boundary" (Sec. 4.7, Appendix C).

Critical strengths
  1. The susceptibility certificate and the instrument. Equation 5 makes "will this symmetry-broken ansatz see anything at all" a pre-training computation, at one evaluation per breaking gate and with no free parameter, verified seed-by-seed to below 10−310^{-3} relative deviation. The dial is also cleanly designed: scaling only breaking angles at fixed parameter count isolates the symmetry budget, and both endpoints coincide with standard baselines.

  2. The optimizer mechanism, and the honesty of the reporting. The Adam/GD contrast has a crisp explanation - a breaking parameter enters only through λθ\lambda\theta, so a plain step moves the signal by O(λ2)O(\lambda^2) while a preconditioned step of size η\eta accumulates it regardless of gradient magnitude. The paper flags its own limits (grid resolution on λ⋆\lambda^\star, the calibrated rather than measured ε\varepsilon exponent, the chain-rule caveat, the schedule saturation in Appendix A.4, and the simulation-only scope).

Critical weaknesses
  1. One of the four advertised contributions fails verification. The zero-transport boundary is listed in the abstract and the contribution list, and it does not survive a one-line check of the stated construction. The correct instance of the phenomenon is not exhibited, and the machine-precision exactness of the reported zeros points to a support or indexing error in the probe rather than to a mechanism.

  2. No operational payoff over the unconstrained baseline is demonstrated. The paper's own Table 2 shows that λ=1\lambda=1 reaches 100% accuracy with one training run and no search, at the same 400-step budget; the dial's sole advantage is retaining symmetry, and what that symmetry buys is never tested - no finite-shot accuracy, no noise, no out-of-distribution hh, no deeper circuits - even though noise-induced symmetry breaking (Tüysüz et al., PRX Quantum 5, 030314, 2024) is exactly what would swamp a λ⋆\lambda^\star of 10−310^{-3} on hardware. The ceiling is also known prior work (Larocca et al., PRX Quantum 3, 030341, 2022; Lawrence et al., ICLR 2025), and the exchange law is grid-resolved at essentially two points.

Questions for authors
  1. For the half-register mirror with breaking gates on ii even, what exact circuit, mirror and state did the Appendix C probe use, and can you release the probe script? Relative to which pair of classes is the quoted Δ≡0\Delta\equiv0 a statement, given that this mirror does not satisfy gH(h,ε)g=H(h,−ε)gH(h,\varepsilon)g = H(h,-\varepsilon)?

  2. Is the portable quantity λ⋆(T)\lambda^\star(T) or the signal budget λ∑tηt\lambda\sum_t\eta_t? Appendix A.4 notes the halving schedule saturates ∑tηt\sum_t\eta_t at 8, so the 1/T1/T form must fail by construction beyond that point; this should be in the main text, since the abstract advertises λ⋆(T)≈0.3/T\lambda^\star(T)\approx0.3/T.

  3. Proposition 1 uses only that a preconditioned step is bounded by η\eta and that the signal is linear in the parameters, both of which hold for any task with a linear class signal. Is the 1/T1/T law a property of symmetry breaking, or of the dial as a search coordinate?

  4. Is there any metric on which the dial at λ⋆\lambda^\star beats λ=1\lambda=1? If yes, please show it; if no, what operational criterion should make a practitioner prefer a small dial? A single run with finite shots or gate noise would also help.

1 comment

We thank the reviewer for the useful report. We take the points in order. The first identifies a construction that does not test what it claims, and the corrected condition is narrower than advertised; the other three turn on what the paper already contains.

1. The zero-transport boundary. The zeros are forced. The probe used H=−∑iZiZi+1−h∑iZi−d∑i evenXiH=-\sum_iZ_iZ_{i+1}-h\sum_iZ_i-d\sum_{i\,\mathrm{even}}X_i with g=∏i evenZig=\prod_{i\,\mathrm{even}}Z_i and the pair (GS[H(+d)], g GS[H(+d)])(\mathrm{GS}[H(+d)],\,g\,\mathrm{GS}[H(+d)]), which does satisfy gHg=H(−d)gHg=H(-d). In the circuit, though, the breaking gates land on the complement of the mirror support, because gates act on tensor axes while the mirror was built on flat-index bits. They commute with gg, Theorem 1 applies, and Δ≡0\Delta\equiv0 at every λ\lambda for that reason alone. The condition the section is after is dk=0d_k=0 for all kk, with dkd_k from Corollary 1, and it is a statement about the preparation, not about the mirror. Concretely: a preparation with no YY coherence makes every dkd_k vanish, and with an XX-type breaking block Δ\Delta then sits at the float32 floor from λ=0.1\lambda=0.1 to 11 even though those gates anticommute with gg. As soon as the symmetric block contains a two-qubit rotation, that coherence appears, transport returns, and the main experiment is in that regime. So the boundary is real but narrow, and its location is set by the preparation rather than by the group. The probe is in the supplementary code.

2. Payoff. At the fixed 132-parameter ansatz the dial ties λ=1\lambda=1 on accuracy, and the paper does not claim otherwise: it is slower at short budgets (0.9450.945 against 1.0001.000 at T=10T=10) and equal by T=50T=50. The axis where it wins is parameter economy at matched accuracy. In the symmetry-matched compact model, 18 parameters against 132, the interior dial reaches 1.0001.000 where the unconstrained end reaches 0.9750.975 at T=400T=400. The dial also removes the search: one training against eight for a grid and fifteen for subset or penalty. On the other axes you list, deeper circuits are tested (L=6L=6 in Fig. 6a and L∈{6,9,12}L\in\{6,9,12\} in App. D), where the unconstrained ansatz reaches threshold at T=10T=10. Finite shots and gate noise are not measured, and we do not claim an advantage there: shots set the resolution of the margin, and noise breaks the symmetry on its own, so a dial below the noise floor cannot be certified on hardware. That is where the dial is a design tool rather than a hardware claim.

3. λ∑tηt\lambda\sum_t\eta_t versus λ⋆\lambda^\star. The qualifier is already in the main text: §4.2 states that beyond T≈200T\approx200 the crossing sits at the grid floor and that the 1/T1/T reading is not resolved further, and the abstract says "where the grid resolves the crossing". What App. A.4 adds is the mechanism, the halving schedule saturating ∑tηt\sum_t\eta_t at 8, which is a property of the schedule rather than of the law. The independent evidence is the collapse rather than the fit: three dial values, full trajectories, λ∑tηt∈[0.0018,0.0045]\lambda\sum_t\eta_t\in[0.0018,0.0045] at the flip step.

4. Is the law about symmetry or about the dial? The exponent is generic to a preconditioned search, and Proposition 1 is a necessary condition. Two things are not generic. At λ=0\lambda=0 the gradient of every breaking parameter vanishes identically, for every dataset and initialization, so the dial is the only seed and the law is about the size of a seed a budget can use rather than about searching a linear signal. And the constant, together with the signal-budget invariance, is measured on the mirror family across ε\varepsilon, nn and LL; the GD/Adam split (λ=1\lambda=1 needed under GD, 10−310^{-3} sufficient under Adam) does not follow from a linear class signal.

5. Prior work. Larocca et al. characterize the invariant subspace and its expressivity, which is a statement about what a model can represent, not an accuracy bound for a task whose label is the symmetry image. Probabilistic symmetry breaking averages a predictor over the orbit; on a mirror pair the two classes are the orbit, so the averaged predictor is at chance by construction. Both are consistent with our ceiling and neither implies it.

Reviewer_DmrmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary: This work aims to understand the impact of breaking intrinsic symmetries that exist in quantum variational circuits, which could be blind to quantum machine learning tasks. The proposal is simple and, from a physical motivation perspective, elegant too. The authors propose a general quantum circuit composed of two unitaries, one that follows the symmetries of the task and a second that breaks them. To understand the impact of the second, the unitary is scaled by a dial, a scalar value (lambda) between 0 and 1, which, when 0, due to the nature of quantum circuits, becomes the identity, meaning no effect on the quantum state. When lambda is 1, meaning it's the "strongest", this should break all symmetries in the quantum circuit for the target task. Besides the optimal value of lambda, the authors' aim was to understand the scaling laws of this parameter as a function of the budget, while considering different quantum circuit depths.

Main Weakness: The work presented is based on a single task problem, a hard classification problem. Form the body of the text its unclear its usability in real world tasks. This can be linked to the Related work section, which in my opinion, has no relevancy or connection to proposed framework. Lack of motivation of this task beyond it being a thought experiment to understand the degree of symmetries in a quantum circuit. In my opinion, the general need for symmetry breaking requires task-dependent evidence, and the work only considers one system.

Critical strengths

Strengths:

  1. Clear task motivation.
  2. Ablation studies are meaninful.
  3. The paper is well written and easy to follow.
Critical weaknesses
  1. The work presented is based on a single task problem, a hard classification problem. Form the body of the text its unclear its usability in real world tasks. This issue coud have been better explained or motivated in the Related work section, which in my opinion, has no relevancy or connection to proposed framework.
  2. Lack of motivation of this task beyond it being a thought experiment to understand the degree of symmetries in a quantum circuit.
  3. The general need for symmetry breaking requires task-dependent evidence, and the work only considers one system. May be another classification QML-based approach could motivate the proposed approach.
Questions for authors
  1. Overall workflow is not clear to me yet. The aim is to identify the ground state of the Hamiltonian, even if it is degenerate. Is the problem state encoded in the circuit before the Eq (3) circuit is deployed? Or is Usym the encoding of this state?
  2. The training stage is unclear. If it's a classification problem, why not use cross-entropy? also what do the authors mean by "softplus of the margin"?
  3. Are the parameters of the Ubrk also trained during the process?
  4. Ubrk is a dense circuit, and while lambda is a general scaling parameter, it does not scarify this part of the circuit, meaning that for a real task it could lead to "large" circuits that could be hard-to-generate.
  5. How often does the work’s imposed invariance conflict with the label structure of realistic data, and does breaking that particular invariance resolve the resulting errors?
1 comment

We thank the reviewer for the constructive report, and for singling out the task motivation and the ablations. The three weaknesses are about coverage rather than about the mechanism, and a second system closes the coverage. We follow the reviewer's order.

1. The workflow. The model does not identify a ground state: the ground state is the input, and the label is its branch. For each hh we prepare ∣ψh⟩=GS[H(h,+ε)]|\psi_h\rangle=\mathrm{GS}[H(h,+\varepsilon)] and its exact image g∣ψh⟩g|\psi_h\rangle, labelled ±1\pm1. The state is loaded by state preparation, so UsymU_{\rm sym} is not the encoding. It is a trainable symmetry-preserving block acting on the loaded state, followed by UbrkU_{\rm brk}. The readout is one fixed observable, the mean bond ⟨ZaZa+1⟩\langle Z_aZ_{a+1}\rangle, giving a scalar RR, and the prediction is sign(R)\mathrm{sign}(R). Degeneracy within a branch is harmless: ε≠0\varepsilon\neq0 splits the two branches, and any state on the same branch carries the same label.

2. The loss. softplus(−yR)=log⁡(1+e−yR)\mathrm{softplus}(-yR)=\log(1+e^{-yR}), which is exactly binary cross-entropy with a sigmoid head and zero bias. Cross-entropy is therefore what we use. The margin form is chosen because RR is a scalar readout with no bias term, and because it makes the odd-in-λ\lambda statement of Theorem 2 literal. Accuracy is computed from sign(R)\mathrm{sign}(R) alone.

3. Yes, UbrkU_{\rm brk} is trained. There are 132 parameters, 72 in UsymU_{\rm sym} and 60 in UbrkU_{\rm brk}. λ\lambda multiplies the forward angles only, and the gradient reaches θbrk\theta_{\rm brk} with a factor λ\lambda, which is the mechanism Proposition 1 prices.

4. The breaking block is the cheap part of the circuit. It is LL layers of single-qubit RXRX and RYRY, 2nL=602nL=60 gates at n=10n=10, with no two-qubit gates, no ancillas, and no gate added by the dial; every two-qubit gate sits in UsymU_{\rm sym}. So λ\lambda does not enlarge the circuit: depth is identical at λ=10−2\lambda=10^{-2} and at λ=1\lambda=1. On hardware this block is a native single-qubit layer, and the two-qubit cost is carried by the symmetry-preserving brickwall.

5. One system, now two. The second system is a frustrated chain with random-sign XXXX and ZZZZ couplings, no free-fermion solution, mirror pair still exact. Ceiling 0.500±0.0000.500\pm0.000 at every budget; transport linear with slope 0.4460.446 against 0.3550.355 for the uniform chain, stable to 0.1%0.1\% over two decades; λ⋆\lambda^\star from 10−210^{-2} at T=50T=50 to 3×10−43\times10^{-4} at T=400T=400. The paper already varies task strength over a decade (ε\varepsilon from 0.0050.005 to 0.050.05), nn from 8 to 14, and L∈{3,6,9,12}L\in\{3,6,9,12\}; the second chain changes the Hamiltonian rather than the size. A controlled instance is the right object for a law, and the second system tests whether the law depends on the instance. On real-world use the claim is narrower than "useful for chemistry": a label defined by a symmetry is a family (phase classification, branch identification, parity readout), and in that family an equivariant default has a provable ceiling. In the symmetry-matched compact model (18 parameters, 7.3×7.3\times reduction) the interior dial reaches 1.0001.000 where the unconstrained end reaches 0.9750.975, so the symmetry budget is also where the parameter saving comes from.

6. Related work. That section covers the four ways a symmetry is currently relaxed, probabilistic, penalized, subgroup and statistical, and it closes by placing the dial on the remaining axis after them. The connection is therefore in the last sentence rather than in a topic sentence, which is a question of ordering, not of relevance. The adaptive-ansatz paragraph is the one part that does not bear on the dial.

7. When does the invariance conflict with the label, and does breaking it help? In this task family the conflict is structural, not occasional: the label is a function of the order parameter that the symmetry flips, which is what a mirror pair is. Whether breaking resolves it is decidable before training, since transport needs some generator with ⟨ϕ∣[Gk,R]∣ϕ⟩≠0\langle\phi|[G_k,R]|\phi\rangle\neq0, one evaluation per gate. When it holds, the same 132-parameter circuit goes from 0.5000.500 at λ=0\lambda=0 to 1.0001.000 at λ=10−3\lambda=10^{-3} by T=200T=200, with the slope Corollary 1 predicts. When it fails, and the breaking gates sit outside the mirror support, accuracy stays at 0.5000.500 for every λ\lambda up to 1, because the commutator is zero and Theorem 1 still applies. The same criterion answers which invariance to break.

Reviewer_ZZ4eOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors study a quantum machine learning task generated by a transverse Ising model that features symmetry breaking. They propose a variational ansatz that can capture the symmetry breaking properties of this model. Experiments show that their ansatz can learn data generated by a simple transfers field Ising model.

To say it in simple terms, this paper reads like AI slop. It is poorly written and hard to follow. This paper needs serious changes to both its content and writing for it to be presented or published at a conference. Its major issues include:

  • Unclear model and input/output: see paragraph below
  • Confusing writing: see last paragraph in this review for examples
  • Unclear motivation for symmetry: here, the authors seem to be studying a task which is not really symmetric, but features symmetry breaking. The "symmetry group" here is incredibly simple consisting of just two elements in the entire group from what I understand. The model or Hamiltonian is not symmetric under this group, but has ground states that equivariantly transform under application of the generator of this group. The sign of the transverse field flips under application of the group operator. Symmetry breaking in the ground state may be observed when the transverse field is strong enough. This is all very specialized to the lattice form of the Hamiltonian. Symmetry breaking is a much broader phenomena (e.g. studied in spin glasses) where the procedure here will not allow for such simplistic analysis. Equivariance and symmetry properties here will not generalize to these settings.
  • Failure to discuss broader challenges with QML: This paper is able to get results because the model they study is a 1D Hamiltonian which is very easy to solve in practice from what I can understand. QML should aim for harder tasks. And in these harder tasks, one needs to see if incorporating symmetries helps alleviate the many challenges with QML. The authors have some discussion of barren plateaus but these are not the only issues. QML algorithms are often swamped with local minima for example. Symmetries only take an exponential dimensional state space to a smaller but still exponential dimensional state space that is hard to optimize. The authors would have observed these issues if they studied a harder task. Discussion of these issues should be in any QML paper.

Regarding the task, I read section 2 multiple times and still do not understand the task. The description starting in section 2.1 is very unclear. For example:

  • What is the input to the task? Is it given a state as input to a quantum computer? Is it given the Hamiltonian as input and then one outputs R(ρ)R(\rho) for ρ\rho either set to the ++ or −- state? Reading this multiple times, I genuinely have no idea what is the task.
  • Is the Hamiltonian H(h,ϵ)H(h,\epsilon) always defined on a 1D lattice? If so, this model is solvable using standard tools.
  • When you write “the model reads out…”, is this the target that one aims to predict?
  • I don’t know what this means: "Training minimizes a full-batch softplus loss on the signed margin yRy R, and test accuracy is threshold classification on held-out values of hh, so train and test share one distribution.” RR is a function so what is its input? Is yy drawn i.i.d. at random? Is the input a value of hh?
  • Earlier, it says "The two classes of the task are therefore related by the symmetry…” which makes me think this is a classification task, but then later, it seems like the task is to predict the "signed margin yRy R”.
  • Eq. (3) seems to just put a scalar parameter λ\lambda on the breaking gates to rescale their values by a constant. But if one can train the parameters, this rescaling seems pointless for any value of λ>0\lambda>0 since the rescaling can be undone by the parameters themselves.
  • It seems like the ansatz is dependent on knowing the model since the placement of the ZZ gates is determined by the model. Is this correct? I can’t tell.

The writing is unconventional bordering on unclear at various parts. Let me give some examples:

  • The emphasized question "How much symmetry must be broken before the model can see, and what does breaking cost?” is provocative, but I can’t make sense of it. What do you mean a model must “see” a symmetry and the “cost” of breaking? The next sentences don't really clear this up.
  • Later in the introduction: "Let the mirror flip one Pauli operator and the readout be one the mirror leaves unchanged. A single breaking rotation then splits the readout into a term the mirror preserves and a term it flips, so the two classes separate by twice the second term…” I genuinely do not know what this means and it is in the introduction. I’m so lost.
  • To start section 2.1, the authors write “Let G=⟨g⟩G=\langle g \rangle be the order-two group generated by gg acting on nn qubits, with Xi,Yi,ZiX_i,Y_i,Z_i the Pauli operators on site ii. We use the global parity flip g=∏iZig=\prod_i Z_i…” This is just plain confusing. As written, G is a group with just two elements and it is unclear what its relation to Paulis is. I think what the authors meant is something like “Let PnP_n be the Pauli group acting on nn qubits consisting of all nn-qubit Pauli matrices up to a phase. Set g=∏iZi∈Png=\prod_i Z_i \in P_n…” This is now so much clearer and follows the standard way these things are defined.
  • R(ρ)R(\rho) is defined for a mixed state ρ\rho but later given pure states as input. Its definition would also be more clear if one just wrote R(ρ)=1n−1∑iTr[ρZiZi+1]R(\rho)=\frac{1}{n-1}\sum_{i} Tr[\rho Z_i Z_{i+1}] rather than “mean over bonds…”.
Critical strengths

See summary above.

Critical weaknesses

See summary above.

Questions for authors

See summary above.

1 comment

We thank the reviewer for the detailed reading. The claim is narrow and checkable: a circuit and readout that both commute with the mirror cannot separate a mirror pair at any parameter setting, and the breaking that fixes this has a closed form and a budget law. Most points below turn on three definitions; two are settled by new experiments.

1. The task and the interface. For each hh on a grid we compute the ground state ∣ψh⟩|\psi_h\rangle of H(h,+ε)H(h,+\varepsilon) and its exact image g∣ψh⟩g|\psi_h\rangle, labelled y=±1y=\pm1. hh indexes the instance, never an input feature, and is not sampled; train and test split the grid. The circuit receives the state (state preparation is data), applies UU, and returns one scalar R=1n∑a⟨ZaZa+1⟩R=\frac1n\sum_a\langle Z_aZ_{a+1}\rangle. RR is the model output, not a target. The prediction is sign(R)\mathrm{sign}(R) and the loss is log⁡(1+e−yR)\log(1+e^{-yR}), i.e. binary cross-entropy with a sigmoid head and zero bias, so yRyR is a signed margin, not a label. The task is classification; HH is the ring of Eq. (2).

2. Eq. (3). The premise is correct: for λ>0\lambda>0, θ↦λθ\theta\mapsto\lambda\theta is a reparameterization and the reachable function class does not change. The implication is where we differ: the dial sets the size of the symmetry-breaking seed and the cost of turning it into an angle. At λ=0\lambda=0 the parameterization is singular: both classes have identical readout expectations for every θ\theta, so the gradient of every breaking parameter is exactly zero for every dataset and initialization, and no optimizer, depth or budget leaves it. The dial leaves it continuously at fixed parameter count, and Proposition 1 prices the exit, 1/λ21/\lambda^2 steps under plain GD against linear accumulation under a preconditioned step. Hence λ=1\lambda=1 under GD against 10−310^{-3} under Adam, and 12%12\% against 100%100\% at T=400T=400. Without the dial there is nothing for Corollary 1 to predict and no 0.3/T0.3/T law to state.

3. Does the ansatz know the model? No. Theorem 1 needs only [Usym,g]=0[U_{\rm sym},g]=0, on any connectivity. The gates are XX,YY,ZZXX,YY,ZZ on a brickwall, HH only XXXX, and the breaking block is single-qubit. Moving the pairs to next-nearest neighbours breaks the match to the chain, and nothing else changes: the ceiling is still exactly 0.5000.500, transport is still linear with slope 0.04890.0489 against 0.35460.3546, stable to 2%2\% over 1.5 decades, and λ=10−3\lambda=10^{-3} still reaches 100%100\% by T=200T=200. The law is a property of the symmetry split, not of matching HH.

4. Which field flips, group size, and generality. The field that flips sign is ε\varepsilon: we put the coupling on XX and the transverse field on ZZ, so gZig=ZigZ_ig=Z_i, gXig=−XigX_ig=-X_i, and hh is untouched. That is why the pair is two branches of a broken phase, not a symmetry of HH. G={I,∏iZi}G=\{I,\prod_iZ_i\} is a two-element subgroup of the Pauli group, and Theorem 1 holds for any group. Order two is the minimal case where a label is defined by the symmetry; the claim is that even here the ceiling is exact and unreachable by capacity, depth, optimizer or budget. That is the result; the size of the group makes it sharp, not limited. The broader test is a spin glass: a frustrated chain with random-sign XXXX and ZZZZ couplings, no free-fermion solution, mirror pair exact. Ceiling 0.500±0.0000.500\pm0.000 at every budget; transport linear with slope 0.4460.446 against 0.3550.355, stable to 0.1%0.1\% over two decades; λ⋆\lambda^\star from 10−210^{-2} at T=50T=50 to 3×10−43\times10^{-4} at T=200T=200 and 400400. The shape transfers, the constant does not.

5. Difficulty, and the rest of the QML list. We claim no quantum advantage. A benchmark whose labels are exact symmetry images cannot be classically hard, since the label is produced by the symmetry; exactness is the point, not hardness, and the deliverable is a law about models. The other bottlenecks are already in the paper: the paragraph ending with Theorem 1 places the result against barren plateaus, concentration, generalization and the information-theoretic limits. Local minima are the one item it does not name, and empirically no run in any reported sweep failed to reach threshold at λ≥10−2\lambda\ge10^{-2} within budget, with L∈{6,9,12}L\in\{6,9,12\} covered.

6. The passages you quote. "How much symmetry must be broken before the model can see, and what does breaking cost?" is the thesis in one line: below a threshold the readout is class independent; the threshold is λ⋆\lambda^\star, priced by Proposition 1. The two-qubit illustration says Δ=2⟨R−⟩\Delta=2\langle R_-\rangle, with R−R_- the part of the readout the mirror flips, which is Theorem 2 on two qubits. RR is the functional tr(R^ρ)\mathrm{tr}(\hat R\rho), the bond average on a pure state; there is no second definition.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.