acceptodds
Withdrawn from ICLR 2027

The Orbit Dimension Criterion: When Symmetry Improves Trainability in Quantum Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Equivariant architectures—from CNNs to graph neural networks—are central to many areas of deep learning, yet a precise theoretical understanding of why and when symmetry improves learning remains elusive. We study this question using parameterized quantum models, whose loss landscapes admit an exact geometric characterization via group orbits and measure concentration. We prove a general representation-theoretic criterion: a symmetry group avoids exponentially flat loss landscapes (“barren plateaus”) when the dimension of its largest multiplicity-space block scales polynomially with model size. Systematically comparing four symmetry types—no symmetry, finite point group (), continuous Abelian (), and permutation ()—we show that uniquely achieves polynomial orbit dimension via Schur–Weyl duality, compressing the effective representation space from exponential to . Through numerical experiments, we establish three core findings for equivariant learning: (i) trainability and expressibility are controlled by distinct mechanisms—gradient variance is primarily architecture-dependent (varies by less than a factor of two across tasks), while final error varies by approximately four orders of magnitude depending on task–bias alignment; (ii) symmetry matching is the key determinant of performance—when the task aligns with the model's symmetry, error drops to near-zero (); and (iii) inductive bias exhibits a sharp symmetry-matching threshold—the equivariant advantage persists for up to 20% symmetry breaking (error ), then jumps by several orders of magnitude once the ground state leaves the symmetric subspace. These results establish parameterized quantum models as a tractable testbed for equivariant learning theory and provide quantitative guidance for designing symmetric architectures.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Equivariant quantum machine learning, Barren plateaus, Inductive bias

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_G7rKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how symmetry constraints affect trainability and achievable accuracy in parameterized quantum models. It proposes a criterion linking gradient-variance scaling to the largest multiplicity-space dimension. Experiments examine gradient variance and final task errors under different degrees of alignment between architectural and task symmetry. The paper also investigates sensitivity to symmetry breaking and distinguishes approximation limits imposed by symmetry from errors due to incomplete optimization.

Critical strengths

S1. The motivation is clear. Understanding when symmetry facilitates optimization and when it excludes useful solutions is an important question for the design of equivariant models.

S2. The distinction between approximation and optimization provides a useful framework. The error decomposition in Section 3.2 separates the best performance achievable under a symmetry constraint from the remaining optimization error. This helps clarify why larger gradients alone do not guarantee better final performance.

Critical weaknesses

The concerns below are not exhaustive. I focus on the most significant weaknesses, which would need to be addressed before I could recommend acceptance.

W1. Theorem 1(i) requires nondegeneracy assumptions. For an SnS_n-equivariant circuit with a permutation-invariant initial state, choose H=PsymH=P_{\mathrm{sym}}, the orthogonal projector onto the permutation-symmetric subspace. Although mmax⁡=n+1m_{\max}=n+1 is polynomial in nn, the state remains in this subspace, giving C(θ)=1C(\theta)=1 and ∂θiC=0\partial_{\theta_i}C=0 for every parameter. These equalities hold even under the blockwise 2-design assumption. The gradient variance is therefore zero, contradicting the claimed inverse-polynomial lower bound.

W2. Theorem 1(ii) is not valid under its stated assumptions. An exponentially large multiplicity block need not contribute to the loss. The gradient variance can remain constant even when mmax⁡m_{\max} grows exponentially.

Consider the following counterexample. Let n≥2n\geq2, write D=2nD=2^n, and represent the nonidentity element gg of G=Z2G=\mathbb Z_2 by R(g)=I2⊕(−ID−2)R(g)=I_2\oplus(-I_{D-2}). The trivial and sign representations have multiplicities 22 and D−2D-2, respectively, so mmax⁡=D−2=Θ(2n)m_{\max}=D-2=\Theta(2^n). Choose ∣ψ0⟩=∣0⟩⊕0D−2|\psi_0\rangle=|0\rangle\oplus0_{D-2}, H=Z⊕0(D−2)×(D−2)H=Z\oplus0_{(D-2)\times(D-2)}, and U(θ)=(e−iθX/2W2)⊕WD−2U(\theta)=(e^{-i\theta X/2}W_2)\oplus W_{D-2}. Here, X,Y,ZX,Y,Z are the Pauli matrices, and the appended zeros denote a zero vector and a zero matrix of the indicated sizes. Initialize the remaining circuit parameters so that W2W_2 and WD−2W_{D-2} are independent Haar-random unitaries. For every fixed θ\theta, both blocks of U(θ)U(\theta) remain Haar-distributed and therefore satisfy the exact unitary 2-design assumption. Moreover, U(θ)U(\theta) commutes with R(g)R(g), and the initial state is GG-invariant.

Since the initial state occupies only the first block, define ∣ϕθ⟩=e−iθX/2W2∣0⟩|\phi_\theta\rangle=e^{-i\theta X/2}W_2|0\rangle. The cost is C(θ)=⟨ϕθ∣Z∣ϕθ⟩C(\theta)=\langle\phi_\theta|Z|\phi_\theta\rangle. Differentiation gives ∂θC(θ)=⟨ϕθ∣i[X/2,Z]∣ϕθ⟩\partial_\theta C(\theta)=\langle\phi_\theta|i[X/2,Z]|\phi_\theta\rangle. Using i[X/2,Z]=Yi[X/2,Z]=Y, this becomes ∂θC(θ)=⟨ϕθ∣Y∣ϕθ⟩\partial_\theta C(\theta)=\langle\phi_\theta|Y|\phi_\theta\rangle. The state ∣ϕθ⟩|\phi_\theta\rangle is a Haar-random pure qubit. Each second moment is therefore 1/31/3, giving Var⁡(∂θC)=1/3\operatorname{Var}(\partial_\theta C)=1/3. This variance is independent of nn and contradicts the claimed O(2−n)O(2^{-n}) upper bound.

W3. The orbit formula is inconsistent with the invariant initial-state assumption. Equations (6) and (14) include multiplicity blocks that an invariant pure state cannot occupy. For the usual permutation action of SnS_n on nn qubits, the orbit has real dimension 2n2n, contradicting the claimed Θ(n2)\Theta(n^2) scaling.

Let R(g)R(g) denote the unitary operator representing g∈Gg\in G. Initial-state invariance means R(g)∣ψ0⟩=∣ψ0⟩R(g)|\psi_0\rangle=|\psi_0\rangle, while circuit equivariance means R(g)U(θ)=U(θ)R(g)R(g)U(\theta)=U(\theta)R(g). Combining these identities gives R(g)U(θ)∣ψ0⟩=U(θ)R(g)∣ψ0⟩=U(θ)∣ψ0⟩R(g)U(\theta)|\psi_0\rangle=U(\theta)R(g)|\psi_0\rangle=U(\theta)|\psi_0\rangle, so every reachable state remains invariant. A nontrivial irreducible representation has no nonzero vector fixed by every group element. The reachable states therefore lie entirely in the trivial isotypic component, whose dimension is mtrivm_{\mathrm{triv}}, the multiplicity of the trivial representation.

Under the full group of equivariant unitaries, the orbit consists of all normalized vectors in Cmtriv\mathbb C^{m_{\mathrm{triv}}}, with vectors differing only by a global phase identified. Thus Orb⁡([ψ0])≅CPmtriv−1\operatorname{Orb}([\psi_0])\cong\mathbb{CP}^{m_{\mathrm{triv}}-1}, rather than the product ∏αCPmα−1\prod_\alpha\mathbb{CP}^{m_\alpha-1} stated in Equation (6). Its real dimension is 2mtriv−22m_{\mathrm{triv}}-2: normalization and global phase each remove one degree of freedom from the 2mtriv2m_{\mathrm{triv}} real coefficients.

For nn qubits, permutation invariance requires equal amplitudes for all binary basis states with the same number of ones. An invariant state is therefore specified by one complex coefficient for each count k=0,…,nk=0,\ldots,n, giving mtriv=n+1m_{\mathrm{triv}}=n+1. Consequently, the full projective orbit is CPn\mathbb{CP}^{n} and has real dimension 2(n+1)−2=2n2(n+1)-2=2n. A finite circuit can reach only a subset of this orbit, so its dimension is at most 2n2n.

W4. The central 2-design assumption needs a clearer explanation. Section 3 introduces parameterized quantum states, expectation-value objectives, and equivariance. However, the unitary 2-design assumption underlying Theorem 1 is described only informally. The authors should explain what a 2-design is and how the circuit parameters are randomly initialized.

Questions for authors

See Weaknesses.

Reviewer_ZszoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on identifying when symmetry in systems can improve the training of quantum models. They find that symmetry can improve training performance only in an exact match with the task. Any mismatch in symmetry constraints worsens the performance even if it leads to faster convergence. Expressability is limited by the symmetry group, so even larger models which may be more trainable may not be more expressable where symmetry is built wrong.

In total, they show the importance of understanding the symmetry of the task using quantum models in a way that allows for more clear theoretical understanding and demonstration than normal machine learning models. They show the potential quantum models have in being used to understand the symmetry problems that exist in regular machine learning.

Critical strengths

This paper does an excellent job of making empirical the effects of symmetry has on quantum models and trying to relate it to machine learning models. Symmetry-constrained systems have had their effectiveness questioned recently by work such as AlphaFold3 and Chai-1 where equivariance was dropped instead for data augmentation. However, this paper shows that symmetry is an effective tool when applied perfectly. However, it is difficult to apply it perfectly to real world models where data may have unique or broken symmetries.

Critical weaknesses

The biggest weakness of this paper is they claim that their work can be used as a productive theoretical lens for regular machine learning models where it may be harder to to assess using exact geometric principles, but they provide no empirical evidence that this is the case. They show no connection to actual machine learning or how their lens may be used. I feel like this is a significant part of their work that would need to be expanded upon. Otherwise they have a neat theoretical lens and model, but no real application to the field at large. What would change my mind is a showcase of how they use their theoretical framework to improve a real world model.

Questions for authors
  1. In the symmetry matching threshold chart, you explore how increases ϵ\epsilon changes the effect of the VQE relative error. However, there is quite a big jump between the first collection of points and the next collection of points. I would be interested in seeing more points in between to see if the relative error increases linearly or exponentially in that range.
  2. How does this work extend to real world domains where symmetry breaking and unknown symmetries occur much more frequently?
  3. Are barren plateaus as big and as frequent in real world models and data setups as in quantum models? Would this technique be able to increase the performance or convergence speed of real models?
Reviewer_nySeOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the effect of orbit dimension on trainability and expressivity of parameterized quantum models and show a criterion to avoid barren plateaus. This is satisfied, for instance by the permutation symmetry. The authors then perform numerous experiments to test their broader hypothesis of effects of symmetries.

Critical strengths

The empirical part is interesting and adds relevant comparison to the literature on the effect of symmetries on quantum models. They also attempt to close the gap from the theory, which stops at characterizing trainability with existence of large gradients (lack of barren plateaus). There, the authors address the more complete question of whether such larger gradients benefit model performance. The paper also re-verify conventional wisdoms in equivariant machine learning, that the extra inductive bias is beneficial for exact symmetries.

Critical weaknesses
  • The central theoretical contribution has substantial overlap with existing literature, and its novelty is insufficiently established. The already-cited Schatzki et al. paper (https://arxiv.org/pdf/2210.09974v3) derives a blockwise gradient-variance formula in Theorem 1 for permutation-equivariant circuits and measurement operators. Immediately afterward, it explicitly discusses extending this formula to general compact groups, assuming universality within each invariant subspace and the relevant design conditions. Consequently, extending the discussion beyond permutation symmetry does not by itself establish a new criterion. The authors should explain precisely what their result adds or which assumptions it relaxes. This distinction is particularly important because Theorem 1 is the paper’s headline result and central conceptual contribution.
  • Theorem 1 is not correct under its stated assumptions, and the proof sketch in Appendix A.1 does not resolve the following issues. Part (i) claims an inverse-polynomial gradient-variance lower bound from polynomial maximum multiplicity alone, but additional nondegeneracy conditions are necessary. For instance, setting H=IH=I gives C(θ)=1C(\theta)=1, so every gradient is identically zero, even when the circuit satisfies the blockwise design assumption. More generally, a globally nontrivial observable can have constant expectation over all reachable states. Part (ii) likewise requires information about which blocks the initial state occupies and which blocks contribute to the derivative. An exponentially large unused block does not force gradients associated with a smaller occupied block to vanish exponentially. There is also an error in the extremum argument in Appendix A.1. The authors claim that per-block variances proportional to 1/mα21/m_\alpha^2 imply a maximum variance of order 1/mmax⁡21/m_{\max}^2. Even accepting this simplified scaling with uniformly comparable positive proportionality factors, the maximum would instead scale with the inverse square of the smallest relevant block dimension. Without control of those factors, the claimed conclusion does not follow at all. These issues undermine the main criterion and the trainability conclusions drawn from the symmetry classification.
Questions for authors

n/a

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

Related papers

Discussion (0)

Sign in to comment.