acceptodds
Under review as a conference paper at ICLR 2027

Commutator-Residual Quantum Circuits: Mitigating Barren Plateaus with an Exponentially Large Lie Algebra

Abstract

Barren plateaus, the exponential vanishing of gradient variances in parametrized quantum circuits, are the central obstacle to scaling variational algorithms. We introduce the commutator-residual ansatz, a circuit of two-qubit group commutators whose generators are symmetrically drawn from a -odd Pauli pool. The pool's dynamical Lie algebra (DLA) is the -odd sector of , dimension , and the closed form exposes a commutator-degeneracy trap that independent sampling falls into and the CoPairSite protocol, which pairs anti-commuting generators on one site, eliminates. A cross-driving lock ties every gradient to a single-qubit expectation, and this carrier mechanism motivates a conjectured polynomial lower bound in the small-initialization regime, which would explain why an exponentially large DLA need not imply a barren plateau. At the CoPairSite variance exceeds the hardware-efficient ansatz (HEA) by three orders of magnitude and decays with width at a fitted log-log slope near against HEA's ; the surviving gradients carry through to faster convergence and lower final errors on ground-state preparation and phase-recognition tasks, and the variance contrast survives depolarizing noise, a global cost, and identity-block initialization.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
variational quantum algorithms, barren plateaus, quantum neural networks, Lie algebras, trainability, quantum machine learning

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_CU8qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a new quantum ansatz, namely the commutator-residual ansatz, along with a sampling strategy called CoPairSite that aims to mitigate barren-plateaus in quantum systems. The paper conjectures that this ansatz has favorable gradient variance under certain conditions and provides numerical evidence to support this conjecture. The numerical experiments also compare this new ansatz with the Hardware Efficient Ansatz (HEA) and demonstrate that the commutator-residual ansatz exhibits larger gradient signals when compared to HEA under similar conditions.

Critical strengths
  1. The idea of interleaving gates with their conjugates is an interesting idea and could prove fruitful. In particular, the fact that it ties parameter size to strength of application of the commutator (eq 1 and 2) is interesting.
Critical weaknesses

The paper contains many technical weaknesses that I will summarize in two broad points:

  1. Regarding the writing of the paper, the paper’s presentation needs substantial work. The text does not communicate the ideas of the paper clearly. The text is non-linear, and many concepts are discussed without being properly defined first in the earlier sections. For example, line 074 refers to ϕ\phi and θ\theta without first establishing their meaning. Moreover, the text contains many confusing terms. Some examples include:

The text uses “unit” instead of “unitary” which is not standard terminology and is confusing as the terms can have different mathematical meanings.

The term “verdict” is used throughout the text, and it is not clear if it has a particular definition or the authors simply used it instead of the standard “result.”

The text refers to “Naive” without any clear explanation as to what it is referring to. Line 137 is the first occurrence of the term with the phrase “Naive’s independent position sampling” which makes it seem like “Naive” is the name of a person. The text also includes inconsistent capitalization; compare line 476 to line 137.

The term “decade” is used multiple times throughout the text, and it has no clear established meaning in the context it is used for.

  1. Regarding the content itself, I have concerns regarding the wording of conjecture 1: It is not clear from the statement how σ0\sigma_0 is related to nn or LL. This has significant implications for the usefulness of conjecture 1. Moreover, conjecture 1 fixes the input state and forces a 1 qubit local observable. Conjecture 1 is also limited to parameters close to the cost observable site, and it is not clear how a global BP free landscape follows from the behavior of a local subset of parameters. While this setting could be interesting as a theoretical example of no BPs with exponential DLA, it is very restrictive and without a concrete proof, its wider application in the BP literature is not clear.

The overall architecture of the proposed circuit, as I understand it, is repeating blocks of gates assembled from a set of 2-qubit Pauli rotation unitaries. The literature has highlighted several limitations of this “problem agnostic” architecture, Larocca et al. (2025) Section V.A for example. The authors should discuss how CoPairSite addresses these limitations. In particular, the paper does not address whether the gates themselves form a local 2-design, which is relevant in light of Cerezo et al. (2021b).

Questions for authors
  1. What is the implication of the σ6+2d\sigma^{6+2d} variance in line 234 for global BP behavior?
2 comments
AuthorsOpenReview ↗

Proved construction, local conjecture, and requested clarifications

We thank the reviewer for the detailed reading. The review raises two weaknesses and one question. Some of the content criticism rests on a reading of Conjecture 1 as a global claim, which the paper does not make. We address all three points below.

1. Weakness 1: presentation. The terminology and ordering problems are real and will be corrected. We will replace “unit” with “commutator block,” “verdict” with “comparison” or “result,” and “Naive” with “independent sampling,” defined at first use. We will remove “decade” and state the actual range σ=0.05,…,0.18\sigma=0.05,\ldots,0.18. Every symbol at line 074 will be defined before that line, and a notation table will be added. We will restructure the paper so that the circuit, notation, and sampling rule precede all interpretive claims, with proofs moved to the appendix.

2. Weakness 2: content. The main proved results are the exact Y-odd DLA, the proof that independent sampling creates identity blocks with zero gradients, the CoPairSite rule that removes this degeneracy, and the exact gradient decomposition into a propagated single-qubit YY expectation. These results are independent of the conjecture and remain valid even if the asymptotic bound is never proved.

Conjecture 1 is deliberately local. It concerns the d=0d=0 parameters, a fixed input, and a local cost. The paper explicitly does not expect a uniform bound over all parameters. The current placement of that disclaimer is too late, so we will move it before the conjecture and rename the result the “cost-adjacent variance conjecture.” We will also state that σ\sigma is bounded by an nn- and LL-independent σ0\sigma_0, with no assumed joint scaling such as σ∼n−p\sigma\sim n^{-p}.

The problem-agnostic criticism applies broadly to architectures that rely on group coverage. Our claim is narrower and different: we identify one specific algebraic failure, identity-block degeneracy, and show that a structured ensemble with an exponentially large DLA can avoid uniform concentration. We will add an explicit limitations paragraph citing Larocca et al. (2025), Section V.A, and will not claim general approximation guarantees.

The construction also does not claim that its gates form a local 2-design. CoPairSite constrains the generator pair, support, and initialization to a nonuniform, near-identity regime, so the standard 2-design sufficient condition is not the relevant criterion. We will add a second-moment diagnostic on small systems to quantify the departure from local Haar moments.

3. Question: implication of the 1/poly(n,L)1/\mathrm{poly}(n,L) variance for global BP behavior. If the conjecture holds, it rules out the strongest global failure mode in which every parameter gradient vanishes exponentially. It does not imply that all parameters are trainable or that the full landscape is BP-free. That distinction is deliberate, not an unstated global claim.

To measure it, we ran a parameter-resolved experiment separating d=0d=0 from distant parameters. At σ=0.1\sigma=0.1, the d=0d=0 variance has log-log slopes of −1.01-1.01 for θ\theta and −0.97-0.97 for ϕ\phi between n=8n=8 and 1616. At n=16,σ=0.1n=16,\sigma=0.1, the θ\theta variance is 1.33×10−31.33\times10^{-3} at d=0d=0, 5.40×10−75.40\times10^{-7} at d=1d=1, and 1.33×10−101.33\times10^{-10} at d=2d=2. The local signal persists while distant directions decay sharply.

The reviewer is right that this does not establish a global theorem, and the paper does not present it as one. The distinction between a proved local mechanism and an empirical global claim will be stated explicitly. The central contribution remains the exact construction, its DLA characterization, and the proved gradient mechanism.

Thank you for your response.

Regarding the presentation weakness

The problems I have stated go beyond fixing the examples I have included in my review. The author's response mainly concerns fixing the examples I have explicitly pointed out. Given that the authors have not yet provided a revised manuscript, I cannot judge the rewritten paper's clarity. As it stands, my evaluation of the paper's presentation remains and I still believe it misses the mark of quality.

Regarding Conjecture 1 and the content

I thank the authors for clarifying their stated claims.

The problem-agnostic criticism applies broadly to architectures that rely on group coverage. Our claim is narrower and different

I don't understand how the stated architecture does not rely on group coverage, when the pool of operators is explicitly chosen based on its coverage criteria (as I understand).

CoPairSite constrains the generator pair, support, and initialization to a nonuniform, near-identity regime, so the standard 2-design sufficient condition is not the relevant criterion.

Can the authors provide a more rigorous argument to support this claim? It is still unclear whether this statement is true.

If the conjecture holds, it rules out the strongest global failure mode in which every parameter gradient vanishes exponentially.

While true, how is this contribution different from other works such as Gaussian initialization schemes (Zhang et al. 2022) for example? We already have works that through constraining the initialization landscapes (e.g. Gaussian initialization) or through structured training (e.g. Layerwise training Skolik et al. 2021) aim to reduce this global failure. What is the novelty of the contribution compared to other methods discussed in the literature? What are its expected benefits when compared to other approaches? Given my comments regarding the restrictive assumptions of Conjecture 1, how do the authors expect this work to contribute to the wider literature?

I doubt the validity of Conjecture 1 as formulated: wouldn't Fig 3(b) imply that CoPairSite follows the same exponential decay as the other methods, save for some constants?

The central contribution remains the exact construction, its DLA characterization, and the proved gradient mechanism.

Overall, given this characterization, I don't think this work contributes anything novel to the literature as it stands.

I appreciate your time in considering my comments.

Reviewer_PRuQOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an innovative new ansatz for variational quantum algorithms that avoids the barren plateau problem. The ansatz is made up of a spcific combination of two-qubit commutator units parametrized by two angles. They prove that the dynamic Lie algebra of this ansatz is the Y-odd sector of su(2^n) and conjecture a polynomial lower bound on gradient variance. Experiments on small quantum systems (transverese-field Ising model, Heisenberg antiferromagnet) support the potential utility of the new ansatz, but the main contribution of the paper is really theoretical.

Critical strengths
  • The paper's goal is clear and well motivated, it proposes a specific new ansatz, and prove the DLA's reach.
  • The authors show that each gradient reduces to a single qubit Y expectation but the lower bound on gradient variance is only a conjecture.
  • The authors are honest about the limitations of the experiments and clearly discuss their scope.
Critical weaknesses
  • The main claim of the paper is just a conjecture. However, I do not by any mens think that this is grounds for rejection, given that overall the paper is innovative.
  • The work is arguably outside the scope of ICLR. This is a quantum computing paper.
Questions for authors

none

1 comment
AuthorsOpenReview ↗

Proven theory, open conjecture, and relevance to ICLR

We thank the reviewer for the positive assessment and the accurate summary. We would add one clarification: the main claim of the paper is not the conjecture. The paper's primary contribution is the exact commutator construction and the proved mechanism behind it.

We prove that the Y-odd pool generates a DLA of dimension (4n−2n)/2(4^n-2^n)/2, identify identity-block degeneracy, and show that CoPairSite eliminates it. We also prove that each gradient decomposes exactly into a propagated single-qubit YY expectation. These results are independent of the polynomial lower bound. The conjecture is an explicitly labeled open question that explains why the exact mechanism can coexist with an exponentially large DLA. The reviewer's observation that the lower bound is unproven is therefore consistent with the paper's own hierarchy.

A new parameter-resolved check strengthens the finite-size case for the conjecture. At σ=0.1\sigma=0.1, the d=0d=0 variance has log-log slopes of −1.01-1.01 for θ\theta and −0.97-0.97 for ϕ\phi between n=8n=8 and 1616, while distant parameters decay much faster. We will report this as finite-size support, not as a proof.

On the scope question, we believe the work belongs to ICLR as a machine-learning contribution. The DLA determines the function class a parameterized model can reach, and the initialization ensemble determines whether that capacity is trainable. Gradient concentration, architecture-induced inductive bias, and the separation between expressivity and optimizability are central questions for ICLR. Quantum circuits provide a setting where both quantities can be characterized exactly.

To make this explicit, we will add a sentence to the abstract stating that the DLA theorem and the gradient decomposition are established independently of the conjecture. We will also add a compact status table at the end of the introduction listing each result as proved, conjectured, or measured. These changes clarify the existing contribution rather than alter it.

Reviewer_rDb2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a method to mitigate the problem of barren plateaus in variational quantum algorithms (VQAs). Let's fix the number of qubits to nn. VQAs consist of a quantum circuit with a sequence of gates (which are unitary matrices in the 2n2^n-dimensional Hilbert space), where each gate is parameterized by a real number; let's call them θ1,θ2,…\theta_1, \theta_2, \dots. Generally, some gates may share parameters too.

One way of solving an optimization problem is the following:

  • initialize the parameters θ∈RP\theta \in \mathbb{R}^P to some value θ(0)\theta(0)
  • for time t=0,...,T−1t = 0, ..., T - 1, do the following:
    • obtain a solution (by sampling) to the optimization problem by running the quantum circuit using parameters θ(t)\theta(t)
    • based on the solution, update the parameters θ(t)\theta(t) to θ(t+1)\theta(t + 1) using some classical algorithm
  • return the solution obtained in the final iteration

Note that θ(0)\theta(0) may be sampled from a distribution, and running this algorithm several times may result in several different solutions. This is crucial when the optimization algorithm above gets "stuck" at a plateau (a locally near-optimal θ\theta) for some choice of θ(0)\theta(0); the hope is that sampling θ(0)\theta(0) from a distribution will ensure that eventually we find a θ(0)\theta(0) that does not get stuck. The best solution among these can then be chosen. Formally, one can measure the expectation (under the distribution for \theta(0)) the size of the gradient ∥g(θ)∥2\|g(\theta)\|^2 with respect to parameters θ\theta (referred to as "gradient variance" in the paper).

Barren plateaus: If this quantity E[∥g(θ)∥2]\mathbb{E} [\|g(\theta)\|^2] decreases exponentially with the number nn of qubits or "circuit length" LL (roughly, the number of quantum gates in the circuit), then this is a pressing issue: gradients nearly vanish too frequently, and an exponential number of experiments need to be run to have meaningful improvement for large nn or LL.

Proposed solution: The paper attempts to fix this by restricting the ways in which the unitary operations are generated by the parameters. Formally, this is captured by choosing some "generators" (2-qubit Pauli operators) and looking at the dynamical lie algebra (DLA) generated by them. The DLA should be "expressive enough": captured by its dimension, but the expected gradient E[∥g(θ)∥2]\mathbb{E} [\|g(\theta)\|^2] should only decrease polynomially in n,Ln, L.

Specifically, they look at the generators AB∈{XY,YX,YZ,ZY}AB \in \{XY, YX, YZ, ZY\}, where each letter denotes a Pauli operator acting on one of two adjacent qubits. These form the Y-odd pool, since each contains an odd number of YY factors.

The circuit consists of group-commutator blocks: a rotation generated by AA, a rotation generated by BB, followed by their respective inverses. Each block has only two independent parameters. It becomes the identity when either parameter is zero, and its leading departure from the identity is proportional to their product when both are small.

If the generators commute, however, the block is always the identity and its parameters have zero gradients. The proposed CoPairSite rule avoids this by choosing anti-commuting generator pairs and placing them on the same adjacent qubits. The paper proves that the DLA generated by these operators has exponentially large dimension. It also conjectures that if the initial circuit parameters θ(0)1,…,θ(0)P\theta(0)_1, \dots, \theta(0)_P are sampled independently from the Gaussian N(0,σ2)\mathcal{N}(0, \sigma^2), the gradient size E[∥g(θ)∥2]\mathbb{E} [\|g(\theta)\|^2] is lower bounded by Ω(σ6/poly(n,L))\Omega\left(\sigma^6/\text{poly}(n, L)\right).

The intended benefit is therefore to preserve a large algebraic reach while concentrating initialization on circuits with useful gradients. The polynomial lower bounds on gradient variance remain conjectures; the experiments provide finite-size evidence through gradient measurements and optimization tasks.

Critical strengths

The paper addresses a significant issue in scaling VQAs: vanishing gradients can make optimization impractical by requiring exponentially many runs of the algorithm with newly sampled parameters to get a single meaningful solutions. Its proposed architecture tackles this problem through the structure of the circuit blocks and the sampling of their generators. The experiments show meaningful improvements over existing methods, making the contribution relevant beyond their theoretical results. The construction offers a concrete and well-motivated approach to improve VQAs.

Critical weaknesses
  1. The paper is extremely difficult to follow and verify independently. I found the paper very difficult to follow, for the following reasons:
  • Non-standard terms are used frequently in a technical sense with no proper definitions. A list of such terms from the abstract and introduction: "commutator-residual ansatz", "cross-driving lock", "commutator-degeneracy trap", "Y-odd Pauli pool", "CoPairSite" (defined much later), "carrier", "lock" and related terms, "vocabulary" (used in the sense of expressibility or as a proxy for the technical term dimension), "coverage premise", etc. The technical sections only exacerbate this issue.

  • Many mathematical symbols are used without definition (or defined much later). Examples:

    • OO in W†OWW^\dagger OW (Section 3.2)
    • k,Kk, K in the appendices
    • TT in the proof of Lemma 3
  • Relatively simple results are obfuscated by odd and unnecessary jargon. For example, in the first paragraph in Section 4, the paper says "The initialization scale σ\sigma, the width of the parameter distribution N(0,σ2)\mathcal{N}(0, \sigma^2), is an architecture-independent treatment variable shared by every variant, swept over {0.1,0.3,0.5,0.7}\{0.1, 0.3, 0.5, 0.7\} from narrow to near-random starts." This is an unusual way of saying that each initial circuit parameter is sampled from the Gaussian N(0,σ2)\mathcal{N}(0, \sigma^2) with mean 00 and std σ\sigma for σ∈{0.1,0.3,0.5,0.7}\sigma \in \{0.1, 0.3, 0.5, 0.7\}. Similarly, Section 4.2 where they vary the number of qubits is titled "Protocol Verdict and n-Scaling Head-to-Head". This is present throughout different section and proofs.

All proofs that I could verify were correct, leading me to suspect that the result is true and meaningful. However, I could not follow all proofs and all claims of the paper.

  1. The conjecture is not well-supported by the experiments. The conjecture which claims that the gradient size E[∥g(θ)∥2]\mathbb{E} [\|g(\theta)\|^2] is lower bounded by Ω(σ6/poly(n,L))\Omega\left(\sigma^6/\text{poly}(n, L)\right) is justified using several experiments. However, I believe the experiments do not clearly support the conjecture. For example, the dependence on number of qubits nn is supported by Figure 2(c), where their method achieves a far higher gradient size as compared to the baseline. However, it is unclear if their method achieves a polynomial-in-nn dependence. In my opinion, this improvement would be useful in itself; but claiming that the curve is a polynomial with only 4 data points is unjustified.
Questions for authors
  1. On the nn-scaling result: Your reported gradient statistic averages gj2g_j^2 over all parameters jj, while Conjecture 1 predicts substantially smaller variance for parameters farther from the cost site. Could the observed results simply arise from averaging over some parameters jj?

  2. You state that the scan σ∈[0.05,0.18]\sigma\in [0.05,0.18] is “decade-wide.” Could you clarify what is meant by “decade-wide” here?

1 comment
AuthorsOpenReview ↗

Proved mechanism, parameter-specific evidence, and the scope of the conjecture

We thank the reviewer for the detailed reading. The central question raised is whether the experiments support the parameter-specific conjecture. We would first clarify that the paper's main contribution does not depend on that conjecture.

The paper proves that the Y-odd pool generates a DLA of dimension (4n−2n)/2(4^n-2^n)/2. It identifies identity-block degeneracy, proves that independent sampling produces it, and gives the CoPairSite rule that removes it. It further proves that each gradient decomposes exactly into a propagated single-qubit YY expectation. These are the paper's central results, and the reviewer states that all proofs they could verify are correct. They remain valid whether or not the asymptotic conjecture is eventually proved.

1. The conjecture and the strength of the evidence. Conjecture 1 is presented as a conjecture, not as a theorem inferred from finite-size points. The scan actually contains five system sizes, n=8,10,12,14,16n=8,10,12,14,16, but we do not claim that any finite set of points can establish an asymptotic law. The reviewer's statement that the observed improvement is useful in itself is consistent with our empirical contribution: the gradient and task advantages are established at finite size independently of the asymptotic bound.

What the reviewer has identified is that Figure 2(c,d) aggregates over all parameters and therefore does not isolate the d=0d=0 object of the conjecture. That is a limitation of the displayed statistic, not a counterexample to the conjecture. We will replace the sentence claiming that the finite-size results close the measurement question with:

“Over n=8,…,16n=8,\ldots,16, CoPairSite exhibits a much weaker decay than HEA. These measurements provide finite-size evidence but do not establish an asymptotic polynomial law; Conjecture 1 remains open.”

This makes the evidence statement match its proper scope.

2. Direct parameter-specific test. To answer the averaging question, we ran a parameter-resolved experiment that separates d=0d=0 gradients from all distant gradients. At σ=0.1\sigma=0.1, the d=0d=0 variance from n=8n=8 to 1616 changes from 3.14×10−33.14\times10^{-3} to 1.33×10−31.33\times10^{-3} for θ\theta, and from 3.50×10−33.50\times10^{-3} to 1.78×10−31.78\times10^{-3} for ϕ\phi, with log-log slopes −1.01-1.01 and −0.97-0.97. At σ=0.5\sigma=0.5, both remain approximately flat, reaching 0.1890.189 and 0.1940.194 at n=16n=16.

The distance profile confirms that distant parameters are much weaker. At n=16,σ=0.1n=16,\sigma=0.1, the θ\theta variance is 1.33×10−31.33\times10^{-3} at d=0d=0, 5.40×10−75.40\times10^{-7} at d=1d=1, and 1.33×10−101.33\times10^{-10} at d=2d=2. Thus the aggregate can indeed be dominated by cost-adjacent terms. However, the d=0d=0 signal itself persists and does not inherit the decay of the distant parameters. The reviewer's interpretation is therefore incomplete rather than sufficient: averaging alone does not explain the observed parameter-specific scaling.

We will report these curves with seed-bootstrap intervals and describe Figure 2(c,d) as an aggregate statistic, not as direct evidence for Conjecture 1.

3. Presentation. The current ordering does make the argument harder to verify. We will restructure the paper so that the circuit, notation, and sampling rule precede all interpretive claims. We will define every symbol at first use, add a notation table, and replace nonstandard labels with direct descriptions. In particular, “cross-driving lock” will become “exact gradient decomposition,” “carrier” will become “the propagated single-qubit YY operator,” and technical uses of “vocabulary” will become “DLA dimension.” These changes are substantial, but they reorganize existing results rather than alter them.

4. “Decade-wide.” We will remove this term and state the numeric range everywhere: σ=0.05,…,0.18\sigma=0.05,\ldots,0.18, with 2000 initializations per point.

Taken together, the verified proofs, the exact gradient mechanism, the direct parameter-resolved evidence, and the finite-size task improvements form the contribution of the paper. The asymptotic conjecture remains open and secondary. We believe these results merit evaluation on that basis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.