acceptodds
Under review as a conference paper at ICLR 2027

Conservation Laws in Parameterized Quantum Circuits: Complete Characterizations and Rank Separation

Abstract

Conservation laws provide a precise description of parameter combinations that cannot change under gradient flow. While such invariants have been studied extensively in classical neural network, their role in parameterized quantum circuits (PQCs) is substantially less understood because the trainable unitary, the measurement observable, and the circuit architecture jointly determine which parameter directions are visible to optimization. We develop a geometric theory of local smooth data-independent conservation laws for PQCs and use it to characterize the active training dimension of widely used Basic Entangler and Strongly Entangling architectures. Using the differential geometry of the measurement map, we expose two distinct sources of inactive directions: unitary-parametrization redundancies and nontrivial unitary motions that are invisible to the readout observable. For commuting trainable blocks, we show that all conservation laws arise from linear invariants associated with a fixed matrix nullspace, yielding a complete characterization of Basic Entangler circuits at arbitrary width and depth. For noncommuting circuits, we derive stabilizer-induced laws and establish completeness for two-layer Strongly Entangling circuits and a class of two-layer circuits with Clifford entanglers. We further construct circuits whose Lie closure rank exceeds the instantaneous Jacobian rank by an amount that grows with circuit width, showing that invisible directions at a single point can overestimate the number of conservation laws. Experiments on synthetic data and CIFAR-10 confirm the predicted invariants during training.

Primary area:
learning theory
Keywords:
conservation laws, gradient flow, parameterized quantum circuit, quantum machine learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_z1K3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors provide a first attempt to define and characterize conservation laws in parameterized quantum circuits. They take three types of quantum circuits and derive theoretical results and verify them experimentally.

Critical strengths

Characterizing PQC behaviour is a very timely topic, as it is generally not understood well. Conservation laws are commonly studied in classical ML and have not yet been transported to QML, thus, it is an interesting research question and motivation.

Critical weaknesses

I am not certain about the relevance and scope of this work. As mentioned in the questions, I am uncertain about the contribution of Theorem 4.4 (which would be the practically relevant circuits), and the extension to the broader class of circuits is the class of 2 layer Clifford circuits, which are not commonly employed in practice (otherwise I would ask the authors to provide a reference) and additionally are classically simulable. I am thus not convinced about the practical implications of these results.

Futher, I think the experimental section is quite limited. The authors state that the motivation for their approach is to reduce dimensionality and gradient measurement costs, however, for the problem instances they consider (2 and 3 qubits), this is not an issue. Since it is also easily possible to simulate 8-10 qubits on a standard laptop, I do not understand why the experiments were not scaled to reasonable problem sizes. Considering the limited scope of the experiments, I do not think the results are particularly informative.

As this is the first work on quantum conservation laws, I would also expect some more background from the classical side in the main part of the paper. I only found this afterwards in the appendix, but it is never referenced in the main text.

As a minor comment, there are several inconsistent usages of citep and cite in the paper. Further, I would encourage the authors to rewrite the statement in the experimental results in the main part that states "deeper circuits are provided in the appendix". I think this is an exaggerated statement considering that the results go from 2 to 3 qubits.

Questions for authors
  • Have you considered how the data encoding affects the conservation laws? I am aware and fully agree with the data-agnostic approach, however, the data encoding can also induce symmetries independently of the exact angles that might merit a discussion here. I also believe that this should be mentioned in the inroduction and background more prominently.
  • For the strong PQC definition, you note n+τn+\tau, I assume that this should be modulo the number of qubits? Could you clarify the notation?
  • Can you clarify the notation of C1C^1, used in Def 2.1?
  • In Remark 2.2 you mention that the conservation laws can be characterized similarly for different losses. Can you elaborate on this and provide some sentence on the applicability of the quantum setting compared to the classical one? The same holds for Remark 2.3. I think this would be very to put in a background section, as mentioned in the weaknesses above.
  • I am struggling to see how Theorem 4.4 is not trivial. For case (i) if the two commute, by definition of RSR_S the factor cancel, and the last layer has no effect on the output of j. (ii) The same goes here, if the last layer has no effect on qubit j, then evidently I can remove the gates by definition in both cases. Am I missing something? Could the authors clarify?
  • When and how is training stopped?
  • Why was training only started from one specific parameter initialization and not tested with different ones?
5 comments
AuthorsOpenReview ↗

Response to Reviewer z1K3 (Part I)

Dear Reviewer z1K3,

Thanks for your constructive and valuable feedback. Below are our responses to your questions. We hope that we have addressed your concerns regarding our paper and will ultimately convince you to raise your score.


W1: I am not certain about the relevance and scope of this work. [...]

Answer. We thank the reviewer for this comment. Our results include complete characterizations for Basic PQCs at arbitrary width and depth (Theorems 4.1, 4.2), final-layer laws for Strong PQCs at arbitrary depth (Theorem 4.4), and completeness for the stated two-layer families (Example 4.5 and Theorem 4.6). For Theorem 4.4, we address the specific role by answering Q5 below.

The circuits in Theorem 4.6 contain arbitrary-angle RYR_Y and RZR_Z rotations; the Clifford assumption concerns E1E_1, not the full trainable circuit. Strongly Entangling architectures are used in quantum learning, for example by [R1]. Our theorem covers two-layer configurations satisfying its readout and distinct-support assumptions, including the range-one CNOT ring of Corollary 4.7(i) for every n≥3n\ge3 and the three-qubit default configuration described in our response to Reviewer XA1X (Q3).

For product-state inputs, this family's observable structure permits efficient classical evaluation. Our contribution concerns which parameter combinations remain conserved and whether further independent laws exist. These questions are not resolved by the ability to simulate the circuit.

The results also reduce optimization variables and gradient-measurement settings. The linear laws allow unregularized Euclidean gradient flow and exact-gradient SGD to be represented in fewer variables. For the three-qubit default configuration, eight independent laws reduce the derivative settings from 3636 to 2020; on the regular set ΩC\Omega_C, Proposition B.9 further reduces this number to 1818. Our additional experiments for Corollary 4.7(iii), with n=3,…,8n=3,\ldots,8, give a reduction from 9696 to 3838 settings at n=8n=8. The finite-shot results reported in W2 support these measurement savings.

W2. The experimental section is quite limited to small number of qubits.

Answer. We thank the Reviewer for this comment. We emphasize that our theoretical results are not restricted to small numbers of qubits. Specifically, for Basic PQC, Theorem 4.2 establishes the complete set of independent conservation laws for arbitrary numbers of qubits and layers. For Strong PQC, Theorem 4.4 identifies conservation laws associated with the parameters in the last layer, while Theorem 4.6 provides a complete characterization for the two-layer architecture with an arbitrary number of qubits.

In Figure 1 and Appendix C, we use configurations with n=L∈2,3n = L\in\\{2,3\\} to numerically illustrate these theoretical findings. Our choice of small-scale circuits is primarily motivated by visualization and interpretability rather than theoretical limitations on the number of qubits. Larger circuits would involve substantially more conserved functions, making their visualization considerably more complex without necessarily providing additional insight into the theoretical results.

To further address the reviewer's concern, we conducted additional experiments on the maximal-rank-gap family of two-layer nn-qubit Clifford circuits in Corollary 4.7(iii), extending the evaluation to n∈3,…,8n\in\\{3,\ldots,8\\}. As detailed in Appendix C.2 in the revised manuscript, we compared standard parameter-shift estimation with our gradient reconstruction method under finite-shot noise, using Nsh∈128,256N_{\mathrm{sh}}\in\\{128,256\\} shots per setting. Our estimator consistently achieves lower gradient estimation error across all circuit widths while remaining below the theoretical bound in Proposition B.9. These results extend our numerical evaluation beyond the original small-scale experiments and support the predicted measurement efficiency as the number of qubits increases.

W3. The paper need some more background from the classical side in the main part of the paper.

Answer. We have include the relevant result from the classical side in the Subsection 2.2 Conservation Laws in Gradient Flows. We suggest the reviewer refer to the revised manuscript for more detail.

W4. There are several inconsistent usages of citep and cite. Further, the paper should rewrite the statement in the experimental results that states "deeper circuits are provided in the appendix".

Answer. We have revised the manuscript to ensure consistent usage of citep and cite. We have also rephrased the statement in the experimental results section to explicitly refer to additional experiments with 3-layer circuits.


Reference.

[R1] Patrick Holzer and Ivica Turkalj. Spectral invariance and maximality properties of the frequency spectrum of quantum neural networks. Quantum Machine Intelligence, 8(1):33, 2026.

AuthorsOpenReview ↗

Response to Reviewer z1K3 (Part II)

Q1. Have you considered how the data encoding affects the conservation laws?

Answer. We agree that the choice of data encoding can introduce additional symmetries and additional laws. However, our work only focuses on data-independent conservation laws that hold for arbitrary input quantum states, allowing us to isolate invariants induced by the circuit architecture and measurement observable. We agree that this deserve more emphasis and we will revise the introduction and background accordingly.

Q2. For the strong PQC definition, you note n+τn + \tau [...] Could you clarify the notation?

Answer. Thank you the reviewer for raising this issue, the correct notation should be ((i+τℓ−1) mod n)+1((i+\tau_\ell-1)\bmod n)+1 instead of only i→i+τℓi \to i+\tau_\ell. We have corrected this in the revised manuscript.

Q3. Can you clarify the notation of C1C^1, used in Def 2.1

Answer. The notation C1C^1 denotes the space of continuously differentiable real-valued functions on Θ\Theta, as is standard in real analysis.

Q4. In Remark 2.2 you mention that the conservation laws can be characterized similarly for different losses. Can you elaborate on this and provide some sentence on the applicability of the quantum setting compared to the classical one? The same holds for Remark 2.3. I think this would be very to put in a background section, as mentioned in the weaknesses above.

Answer. We thank the reviewer for this comment. Regarding Remark 2.2, the extension to other loss functions follows from Assumption 2.4 and Corollary 2.5 in [Marcotte et al., 2025]. For a scalar-output QNN and any differentiable loss satisfying span⁡y∈Y{∂zℓ(z,y)}=R\operatorname{span}_{y\in\mathcal Y}\{\partial_z\ell(z,y)\}=\mathbb R where z=Fρ(θ)z=F_\rho(\theta), we have Wloss(θ)=span⁡ρ{∇θFρ(θ)}=WQ(θ).\mathcal W_{\mathrm{loss}}(\theta)=\operatorname{span}{\rho}\{\nabla_\theta F_\rho(\theta)\} =\mathcal W_Q(\theta). Thus, our conservation-law characterization dΦθ(∇h)=0d\Phi_\theta(\nabla h)=0 remains valid, including for squared-error and cross-entropy losses. This result applies to quantum predictors because the underlying characterization concerns general smooth parametric models.

We clarify that the framework of [Marcotte et al, 2024] applies to general smooth parametric predictors, including our QNN model Fρ(θ)=tr⁡(ρU(θ)†M0U(θ))F_\rho(\theta)=\operatorname{tr}(\rho U(\theta)^\dagger M_0U(\theta)). [Marcotte et al., 2024, Propositions 2.8-2.9] address momentum-based dynamics, while Proposition 2.12 characterizes generalized gradient flows. The additional zero-gradient-dataset assumption required by Proposition 2.9 is satisfied in our setting by choosing the maximally mixed state ρ∗=Id/d\rho_*=I_d/d which gives Fρ\*(θ)=tr⁡(M0)dF_{\rho_\*}(\theta) = \frac{\operatorname{tr}(M_0)}{d}, and therefore ∇θL(ρ∗,y)(θ)=0,∀θ,y∈Y.\nabla_\theta L_{\\{(\rho_*,y)\\}}(\theta)=0,\forall\theta,y\in\mathcal Y.

Furthermore, Theorem 2.1 in [Marcotte et al., 2025] establish a structural relationship between conservation laws with and without weight decay under the same zero-gradient-dataset assumption. Thus, these general results apply to our QNN setting without quantum-specific modifications. We emphasize that applicability of these frameworks does not imply that the original conservation laws remain unchanged under alternative training dynamics. Our analysis is restricted to Euclidean gradient flow without weight decay. We have revised Remark 2.2 and 2.3 to clarify these extensions and include the relevant references.


Reference.

[Marcotte et al., 2024] Keep the momentum: Conservation laws beyond euclidean gradient flows. arXiv preprint arXiv:2405.12888, 2024

[Marcotte et al., 2025] Transformative or conservative? conservation laws for resnets and transformers. arXiv preprint arXiv:2506.06194, 2025

AuthorsOpenReview ↗

Response to Reviewer z1K3 (Part III)

Q5: I am struggling to see how Theorem 4.4 is not trivial. For case (i) if the two commute, by definition of RSR_S the factor cancel, and the last layer has no effect on the output of j. (ii) The same goes here, if the last layer has no effect on qubit j, then evidently I can remove the gates by definition in both cases. Am I missing something? Could the authors clarify?

Answer. We thank the reviewer for this question. We agree that the cancellations in cases (i) and (ii) follow directly from the stated commutation and support conditions. However, the two cases have different consequences. In case (i), commutation with ZjZ_j removes the dependence on ωL,j\omega_{L,j}, while ϕL,j\phi_{L,j} and βL,j\beta_{L,j} may still affect the output. In case (ii), the effective observable acts as the identity on qubit jj, so the entire local Euler block has no effect on the measured output.

The role of Theorem 4.4 is to give explicit, readout-dependent criteria for identifying these coordinate laws at arbitrary depth, without computing the Lie closure. Since the cancellations hold at the operator level, the corresponding derivatives vanish for every input state. Thus, the identified parameters remain fixed under unregularized Euclidean gradient flow and exact-gradient SGD, independently of the dataset. In particular, an effective observable Z(J)Z(J) gives one coordinate law for each j∈Jj\in J and three for each j∉Jj\notin J, yielding 3n−2∣J∣3n-2|J| independent coordinate laws. Their updates and derivative measurements can therefore be omitted.

The proof of Theorem 4.4 in Appendix B.3.3 also establishes a single-layer completeness result that is not stated explicitly in the main text. When L=1L=1 and Φ1S=Z(J)\Phi_1^S=Z(J), the identified coordinates form a complete set of independent local smooth conservation laws on the open dense set where sin⁡β1,j≠0\sin\beta_{1,j}\neq 0 for every j∈Jj\in J. Indeed, the derivatives of the measurement map with respect to ϕ1,j\phi_{1,j} and β1,j\beta_{1,j}, j∈Jj\in J, are nonzero and pairwise Hilbert--Schmidt orthogonal. Therefore, with p=3np=3n and q=3n−2∣J∣q=3n-2|J|, we have 2∣J∣=r0≤rLie≤p−q=2∣J∣,2|J|=r_0\le r_{\mathrm{Lie}}\le p-q=2|J|, and Theorem 3.5 gives local completeness. Thus, in this single-layer setting, the analysis also excludes additional independent local smooth laws, a conclusion that does not follow from cancellation alone. For L≥2L\ge 2, the identified laws need not be complete, since additional cross-layer laws may arise. We will make this single-layer completeness result explicit in the discussion following Theorem 4.4.

Q6. When and how is training stopped?

Answer. Thank you for your question. We agree that the stopping criterion was not clearly specified in the experimental setup. Specifically, training is conducted for a fixed budget of 3,000 stochastic gradient descent (SGD) updates, with a batch size of 32, a constant learning rate of 0.01, and no weight decay. The training process terminates upon reaching the prescribed number of updates, without employing any early stopping criteria. We will explicitly clarify this stopping criterion in the revised manuscript.

Q7. Why was training only started from one specific parameter initialization and not tested with different ones?

Answer. We thank the reviewer for this question. For each configuration in Figure 1, we use a single parameter initialization sampled from θ0∼U([−π,π]p)\theta_0\sim\mathcal{U}([-\pi,\pi]^p), independently of the ground-truth parameter θ∗\theta^* used to generate synthetic labels. We present one trajectory per configuration to clearly illustrate the evolution of individual parameters and their conserved combinations.

Importantly, the conservation laws hold for arbitrary initializations. Specifically, for each identified linear law hah_a, the exact-gradient SGD updates satisfy ha(θk)=ha(θ0)h_a(\theta_k)=h_a(\theta_0) at every iteration, as established analytically in Lemma B.5. Thus, different initializations may yield different values of the conserved quantities without affecting their preservation. The random initialization in Figure 1 is therefore chosen solely for visualization and does not restrict our theoretical guarantees.

Reviewer_XA1XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies conservation laws for parametrized quantum circuits. A conservation law is a function that remains invariant under the gradient flow, which is the idealized evolution of parameters during training. Studying these conservation laws allows to determine which parameter directions remain invisible during the optimization process. The paper focuses on two commonly used architectures, Basic Entangler and Strongly Entangling. For commuting trainable blocks (covering the Basic architecture, the paper provides a characterization of conservation laws (Theorem 4.1 and 4.2). In the non-commuting case, they also provide sufficient conditions (covering the Strong architecture). In the two-layer case, they further generalize to Clifford entanglers. The paper concludes with a numerical experiment.

Critical strengths
  • Understanding training dynamics for PQC is a challenging problem. Conservation laws for PQCs have not been comprehensively studied previously, the invariance results proved in the paper thus bring us a step further towards understanding PQC training dynamics.
  • Matehmatical results are clearly formulated and the proofs appear to be correct
  • Overall, this is a well-written paper
Critical weaknesses
  1. Experiment findings remain unclear.
  • As stated in the text, We have proved that if the conserved function is linear w.r.t to parameter θ, its invariant is preserved exactly under discrete update, i.e., Lemma B.5 shows that linear laws are preserved exactly under SGD. The experiments then simply measure that this is indeed the case. They conclude with As illustrated in Figure 1, these quantities remain conserved throughout discrete training., but this was already proved in Lemma B.5 anyway.
  • Then, the authors state Moreover, we observe no additional parameter relations that yield independent conserved quantities, such as further linear combinations involving sums or differences of parameters. This empirical evidence supports that the identified laws constitute a maximal independent set and hence characterize all conservation laws. However, there seems to be no evidence for this from the experiments.
  • There is an entire Section 4.4 on Training Benefits, but these are not tested in the experiments.
  1. The methodology is substantially building on (Marcotte et al., 2023), which was formulated for general predictors. The paper here applies this theory to the PQC case. This comes with challenges of course, but nevertheless the relation to (Marcotte et al., 2023) should be made more clear. For example, the introduction describes Marcotte et al. (2023) as being about "linearized models", but the paper treats general nonlinear models.
Questions for authors
  1. Experiments are carried out for n,L∈{2,3}n, L \in \lbrace 2,3 \rbrace. What happends for larger number of qubits and more layers?
  2. Could you provide more details regarding the weight decay case? In Remark 2.3 you claim as shown in (Marcotte et al., 2024), the methodology for identifying conservation laws extends to these settings in a largely analogous manner, but given that (Marcotte et al., 2024) is for classical neural networks, it is not at all clear why this works here.
  3. In the abstract and introduction, you claim to cover completeness for two-layer Strongly Entangling circuits. However, the results only cover the case n=2n=2 (Example 4.5) and n≥3n\geq3, under the conditions of Theorem 4.6; can you clarify if these conditions are satisfied for PennyLane's default entangling ranges and the readouts used in your n=3n=3 experiments?
  4. Can you test the described training benefits in experiments? Do they still hold in the presence of shot noise?
  5. Can you comment on how your findings relate to earlier studies that examined parameter redundancies, such as
  • Fontana et al., "Non-trivial symmetries in quantum landscapes and their resilience to quantum noise" (Quantum 2022)
  • van de Wetering, Yeung, Laakkonen & Kissinger, "Optimal compilation of parametrised quantum circuits" (Quantum 2025).
6 comments
AuthorsOpenReview ↗

Response to Reviewer XA1X (Part I)

Dear Reviewer XA1X,

Thanks for your constructive feedback and thoughtful review. In the sequel, we will provide our response to your questions regarding the paper and hope that it will completely address your concerns.


W1. [...] We have proved that if the conserved function is linear w.r.t to parameter θ\theta [...] The experiments then simply measure that this is indeed the case [...]

Answer. We thank the Reviewer for raising this concern. We agree that Lemma B.5 already establishes the exact preservation of linear conservation laws under discrete-time SGD with exact gradients. Figure 1 merely illustrates these theoretical results for concrete quantum circuits, including frozen parameters and conserved linear combinations across layers, and serves as a consistency check rather than independent empirical evidence.

W2. [...] Moreover, we observe no additional parameter relations that yield independent conserved quantities [...]. However, there seems to be no evidence for this from the experiments.

Answer. We thank the Reviewer for this constructive feedback. We agree that claiming the absence of additional parameter relations yielding independent conserved quantities based solely on the parameter updates in Figure 1 may be confusing. The primary purpose of Figure 1 is to verify the conservation laws rigorously established in Examples 4.3 and 4.5. Since this above statement is unnecessary for supporting our theoretical results and may cause confusion, we will remove it in the revised manuscript.

W3 + Q4. There is an entire Section 4.4 on Training Benefits, but these are not tested in the experiments. Can you test the described training benefits in experiments? Do they still hold in the presence of shot noise?

Answer. Thank you for your question. Section 4.4 and Appendix B.3.7 establish two theoretical training benefits:

  • Dimension reduction. Given qq independent linear conservation laws, Proposition B.6 provides an equivalent reparameterization θ=θ0+Qz,z∈Rp−q,\theta=\theta_0+Qz, z\in\mathbb R^{p-q}, reducing the optimization dimension from pp to p−qp-q while preserving the exact gradient-flow and SGD dynamics.

  • Reduced measurement cost. For two-layer nn-qubit Clifford circuits, where p=6np=6n, rLie=p−qr_{\mathrm{Lie}}=p-q, and Δ=rLie−r0\Delta=r_{\mathrm{Lie}}-r_0, Propositions B.7 and B.9 reduce the number of parameter-shift settings as 2p→2rLie→2r0.2p\to 2r_{\mathrm{Lie}}\to 2r_0. Moreover, Proposition B.9 constructs an unbiased gradient estimator satisfying E[∥∇Fρ^−∇Fρ∥2]≤r0(r0+2Δ)Sd,\mathbb E\left[\left\|\widehat{\nabla F_\rho}-\nabla F_\rho\right\|^2\right] \leq \frac{r_0(r_0+2\Delta)}{S_d}, where SdS_d is the total derivative-measurement shot budget.

Notably, the reconstructed gradients preserve the linear conservation laws exactly, even under finite-shot noise. Indeed, letting ha(θ)=va⊤θh_a(\theta)=v_a^\top\theta for 1≤a≤q1\le a \le q and K=spanv1,…,vqK=span\\{v_1,\ldots,v_q\\}, Proposition B.9 ensures im(Tθ)⊆K⊥im(T_\theta)\subseteq K^\perp. Thus, for g^θt=Tθtd^Jθt\widehat g_{\theta_t}=T_{\theta_t}\widehat d_{J_{\theta_t}}, each SGD update satisfies ha(θt+1)−ha(θt)−ˉηva⊤Tθtd^Jθt0ˉ,h_a(\theta_{t+1})-h_a(\theta_t)\=-\eta v_a^\top T_{\theta_t}\widehat d_{J_{\theta_t}}\=0, for every finite-shot measurement outcome, rather than only in expectation.

Additional experiments under finite-shot noise. To assess the benefits of our theoretical results under finite-shot noise, we conducted additional Monte Carlo experiments on two-layer Clifford circuits with 3≤n≤83 \le n \le 8 and Nsh∈128,256N_{sh}\in\\{128,256\\}. As shown in Fig. 4 of the revised manuscript, our reconstructed estimator achieves lower MSE with fewer measurement settings and total shots than standard parameter-shift estimation. The empirical MSE remains below the theoretical bound and approximately halves when the shot count doubles, supporting the estimator's measurement efficiency and finite-shot robustness. Please refer to Appendix C.2 of the revised manuscript for further details.

Q1.What happens for larger number of qubits and more layers?

Answer. Thank you for your question. We study two widely used PQC architectures: Basic PQC and Strong PQC. Theorem 4.2 completely characterizes the conservation laws for Basic PQC with arbitrary numbers of qubits and layers, while Theorem 4.4 identifies conservation laws for the last layer of Strong PQC. Due to the non-commutativity of trainable blocks, a complete characterization for Strong PQC with L≥3L\geq 3 remains open. However, Theorem 4.6 provides a complete characterization for two-layer Strong PQC with arbitrary numbers of qubits by exploiting its Clifford structure.

In Figure 1 and Appendix C, we numerically verify our theoretical findings using configurations with n,L∈2,3n,L\in\\{2,3\\}, which cover the relevant theoretical cases. Increasing the number of qubits or layers would introduce substantially more conserved functions, making the figures unnecessarily complex without providing additional insight into the validity of our theoretical results.

AuthorsOpenReview ↗

Response to Reviewer XA1X (Part II)

W4. The methodology is substantially building on (Marcotte et al., 2023), which was formulated for general predictors. The paper here applies this theory to the PQC case. This comes with challenges of course, but nevertheless the relation to (Marcotte et al., 2023) should be made more clear.

Answer. We thank the reviewer for this comment. We agree that the framework of [Marcotte et al., 2023] applies to general parametric predictors, including nonlinear models. Our definition of data-independent conservation laws and the use of gradient orthogonality and Lie-closure rank follow their framework. In particular, Theorem 3.5 follows from their Theorem 3.3, as acknowledged in Section 3, while Appendix A explains how we adapt their Lie-algebraic characterization to PQCs.

For PQCs, we formulate the conservation condition through the measurement map Φ(θ)=U(θ)†M0U(θ)\Phi(\theta)=U(\theta)^\dagger M_0U(\theta). Since Fρ(θ)=Tr⁡(ρΦ(θ))F_\rho(\theta)=\operatorname{Tr}(\rho\Phi(\theta)) is linear in ρ\rho and density matrices span the Hermitian operators, the data-independent conservation condition under squared loss reduces to dΦθ(∇h(θ))=0d\Phi_\theta(\nabla h(\theta))=0 (Propositions 3.1, 3.2 and Theorem 3.3). We further characterize the kernel of this differential through the unitary parametrization and observable stabilizers (Theorem 3.6), thereby identifying quantum-specific mechanisms of conservation that are not explicitly addressed in [Marcotte et al. ,2023].

Unlike [Marcotte et al. ,2023], which provides a general characterization, our main contribution is to derive explicit conservation laws and prove their completeness for concrete PQC architectures by exploiting Pauli propagation, circuit symmetries, and Lie-bracket calculations. Theorems 4.1 and 4.2 cover commuting blocks, including Basic PQCs of arbitrary width and depth. For the noncommuting families in Example 4.5 and Theorem 4.6, we identify explicit laws, establish completeness, and compute the generic ranks of the measurement differential and Lie closure under the stated assumptions. While rank separation was previously observed in two-layer linear networks [Marcotte et al., 2023], Corollary 4.7 establishes a sharp upper bound of n−1n-1 for the class in Theorem 4.6 and constructs a family attaining it, demonstrating that the separation can grow linearly with the number of qubits.

Q2: Could you provide more details regarding the weight decay case?

Answer. We thank the reviewer for this question. Please let us elaborate more on this.

(i) Effect of weight decay. Let LDL_D denote the unregularized loss and hv(θ)=v⊤θh_v(\theta)=v^\top\theta be a linear conservation law. Proposition 3.1 ensures v⊤∇LD(θ)=0v^\top\nabla L_D(\theta)=0 for every dataset. Under the weight-decay flow θ˙=−η(∇LD(θ)+λθ),\dot\theta=-\eta\bigl(\nabla L_D(\theta)+\lambda\theta\bigr), where η>0\eta>0 and λ≥0\lambda\geq 0, we obtain ddthv(θ(t))=−ηλhv(θ(t)),hv(θ(t))=e−ηλthv(θ(0)).\frac{d}{dt}h_v(\theta(t))=-\eta\lambda h_v(\theta(t)),\quad h_v(\theta(t))=e^{-\eta\lambda t}h_v(\theta(0)). Thus, for λ>0\lambda>0, both frozen parameters and conserved linear combinations decay exponentially, while eηλthv(θ(t))e^{\eta\lambda t}h_v(\theta(t)) remains equal to hv(θ(0))h_v(\theta(0)). Similarly, SGD with weight decay and exact minibatch gradients satisfies hv(θk+1)=(1−ηkλ)hv(θk),h_v(\theta_{k+1})=(1-\eta_k\lambda)h_v(\theta_k), where ηk\eta_k is the learning rate at iteration kk.

(ii) Why these results apply to PQCs. First, PQC training follows Euclidean gradient flow on the real parameter space θ\theta, so the same parameter-space arguments apply. Second, the results of [Marcotte et al., 2024, 2025] apply to smooth parametric predictors without architecture-specific restrictions. Our predictor Fρ(θ)=tr⁡(ρU(θ)†M0U(θ))F_\rho(\theta) =\operatorname{tr}\bigl(\rho U(\theta)^\dagger M_0U(\theta)\bigr) and its squared loss are smooth in θ\theta, while Proposition 3.1 establishes the required gradient orthogonality.

Third, the assumption of [Marcotte et al., 2025, Theorem 2.1], requiring a dataset with zero loss gradient at each θ\theta, is satisfied. Indeed, choosing ρ∗=I/d\rho_*=I/d yields Fρ∗(θ)=tr⁡(M0)d.F_{\rho_*}(\theta)=\frac{\operatorname{tr}(M_0)}{d}. Hence, the prediction and squared loss are independent of θ\theta, giving a zero loss gradient. The theorem therefore applies, showing that every data-independent conservation law under weight decay takes the form H(t,θ)=H0(eηλtθ)H(t,\theta)=H_0(e^{\eta\lambda t}\theta), where H0H_0 is a conservation law without weight decay. For our linear laws, this recovers (i), implying that weight decay introduces no new laws and our completeness results carry over.


References.

[Marcotte et al., 2023] Abide by the law and follow the flow: Conservation laws for gradient flows. Advances in neural information processing systems, 2023.

[Marcotte et al., 2024] Keep the momentum: Conservation laws beyond euclidean gradient flows. arXiv preprint arXiv:2405.12888, 2024

[Marcotte et al., 2025] Transformative or conservative? conservation laws for resnets and transformers. arXiv preprint arXiv:2506.06194, 2025

AuthorsOpenReview ↗

Response to Reviewer XA1X (Part III)

Q3. In the abstract and introduction, you claim to cover completeness for two-layer Strongly Entangling circuits. [...] Can you clarify if these conditions are satisfied for PennyLane's default entangling ranges and the readouts used in your experiments?

Answer. We thank the reviewer for this question. These conditions hold for PennyLane's default two-layer ranges and the global-ZZ readout M0=Z⊗nM_0=Z^{\otimes n} for n=2,3n=2,3.

For n=2n=2, the default ranges (1,1)(1,1) satisfy Example 4.5. For n=3n=3, the default ranges (1,2)(1,2) satisfy Theorem 4.6, since E2†M0E2=Z1E_2^\dagger M_0E_2=Z_1 and

P_Y = E_1^\dagger Y_1E_1 = X_1Y_2Z_3,\quad P_Z = E_1^\dagger Z_1E_1 = Z_2Z_3,$$ have pairwise distinct supports. Thus, Theorem 4.6 establishes completeness with $q=8$ independent laws, $r_0=9$, and $r_{\mathrm{Lie}}=10$ on $\Omega_C$. Additionally, Corollary 4.7(i) establishes completeness for ranges $(1,1)$ and readout $Z_1Z_n$ for arbitrary $n\geq 3$. We clarify that the three-qubit experiments in Appendix C use three layers and only verify the final-layer conservation laws of Theorem 4.4, without claiming completeness. For other two-layer configurations, completeness holds whenever the assumptions of Theorem 4.6 are satisfied. > **Q5. Can you comment on how your findings relate to earlier studies that examined parameter redundancies, such as [R1] and [R2]** **Answer.** We will incorporate the comparison with those related literature in our revised manuscript. Fontana et al. study symmetries of parameterized unitaries, where distinct parameter configurations satisfy $U(\theta)=U(\widetilde{\theta})$. Such equivalence implies $\Phi(\theta)=\Phi(\widetilde{\theta})$, but the converse need not hold, since distinct unitaries may produce the same evolved observable. This distinction allows our framework to identify observable-dependent redundancies and associated conservation laws even when parameter variations change the underlying unitary. Moreover, while [R1] construct symmetry-related parameter configurations using $\sigma$-Pulse transformations, we characterize quantities that remain conserved along Euclidean gradient-flow trajectories. Meanwhile, [R2] study parameter redundancy from a circuit-compilation perspective, proving optimal parameter reduction for a restricted class of parametrized Clifford circuits. Their approach eliminates redundant parameters while preserving the implemented unitary, and such affine parameter reductions also induce linear conservation laws under Euclidean gradient flow. However, our work differs in two fundamental aspects. First, rather than seeking a reduced circuit representation, we characterize parameter combinations that remain invariant under gradient-flow training of the original circuit and establish completeness results for these conservation laws. Second, while their reduction preserves the implemented unitary, our framework captures observable-dependent redundancies, where parameter variations change the underlying unitary but leave the measurement map unchanged. Consequently, our analysis identifies conservation laws beyond those arising from unitary-preserving parameter reductions. --- **References.** [R1] Fontana et al., "Non-trivial symmetries in quantum landscapes and their resilience to quantum noise" (Quantum 2022) [R2] van de Wetering et al. "Optimal compilation of parametrised quantum circuits" (Quantum 2025).
Reviewer_e6qgOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper develops a geometric theory of data-independent conservation laws—parameter combinations that remain unchanged under Euclidean gradient-flow training—in parameterized quantum circuits. It distinguishes redundancies in the unitary parameterization from changes invisible to the measurement observable, and uses the Lie closure of measurement-gradient fields to determine the number of independent conserved quantities. Its main contributions are a complete characterization for Basic Entangler circuits of arbitrary width and depth through a fixed matrix-nullspace computation, explicit final-layer invariants for Strongly Entangling circuits, and completeness results for specified two-layer Strong and Clifford-entangler circuits. It also constructs examples where the Lie-closure rank exceeds the instantaneous Jacobian rank by an amount linear in circuit width, demonstrating why pointwise Jacobian analysis can overcount conservation laws. The authors derive corresponding reductions in optimization dimension and gradient-measurement settings. Experiments on synthetic regression and binary CIFAR-10 classification, using small two- and three-qubit circuits, confirm that the predicted invariants persist during SGD; empirical validation focuses on these invariants, while completeness and computational savings are established theoretically.

Critical strengths

The paper develops a geometric framework for characterizing data-independent conservation laws in parameterized quantum circuits and determining the number of functionally independent conserved quantities through the Lie-closure rank. This characterization is valuable because it provides a criterion for establishing whether an identified family of conservation laws is complete. I find the distinction between the instantaneous Jacobian rank and the Lie-closure rank particularly interesting: directions that are locally invisible to the measurement map do not necessarily correspond to quantities conserved throughout training. The explicit rank calculations and separation examples clarify this distinction and provide insight into the effective dimension of the training dynamics.

The paper illustrates its framework through concrete circuit families, including Basic PQCs and specified two-layer Strong PQCs, and numerically verifies the predicted conservation laws during training. These examples make the abstract geometric results easier to interpret by exhibiting both individually frozen parameters and conserved combinations of parameters. The numerical experiments further demonstrate that the predicted invariants persist under discrete SGD updates, connecting the continuous-time analysis to an implemented training procedure. This supports the practical relevance of the conservation phenomena, although the experiments remain small-scale and primarily validate invariance rather than improvements in optimization efficiency or predictive performance.

Critical weaknesses

A principal limitation is the computational difficulty of extending the framework to general variational quantum algorithms. Evaluating and certifying the Lie-closure rank can require constructing increasingly high-order iterated Lie brackets, differentiating complicated parameter-dependent vector fields, and determining their functional dependencies. Although the pointwise rank is finite, this does not automatically provide an inexpensive calculation or a straightforward stopping criterion. Consequently, the general characterization does not by itself yield an efficient method for counting or explicitly constructing conservation laws in arbitrary circuits. The tractability of the paper’s examples relies substantially on their algebraic structure, leaving the scalability of the approach beyond these structured settings an important open question.

Questions for authors

Do you have an efficient algorithm to upper bound the Lie-closure rank?

4 comments
AuthorsOpenReview ↗

Response to Reviewer e6qg (Part I)

Dear Reviewer e6qg,

Thanks for your constructive feedback and thoughtful review. In the sequel, we will provide our response to your questions regarding the paper and hope that it will completely address your concerns.


W1. A principal limitation is the computational difficulty of extending the framework to general variational quantum algorithms. [...]

Answer. We thank the reviewer for the useful comments. Firstly, we agree that the general characterization does not by itself give an efficient method for arbitrary PQCs. Indeed, computing the Lie closure requires higher derivatives and the identification of dependencies among parameter-dependent fields, and evaluating these fields can itself be costly.

However, our results make this characterization explicit for structured families. Theorems 4.1, 4.2 give complete descriptions for commuting blocks, including Basic PQCs at arbitrary depth, while Example 4.5 and Theorem 4.6 establish completeness for the stated noncommuting families. In each case, we identify explicit laws and prove that no additional independent local smooth laws exist. In addition, we hope that the following observations will clarify the computational issues raised by the reviewer.

(a) Bracket order and termination. A local stopping criterion is available: if two consecutive bracket-generated distributions have the same constant rank on an open neighborhood, then the Lie closure has stabilized locally [Marcotte et al., 2023, Lemma 3.2]. Moreover, for real-analytic fields on a connected parameter domain, at most p−r0genp-r_0^{\mathrm{gen}} extensions suffice to determine the generic Lie-closure rank, where r0genr_0^{\mathrm{gen}} is the generic rank of WQW_Q, provided that the procedure stops when consecutive generic ranks agree or when rank pp is reached. This result bounds the bracket depth, but not the computational cost. For our classified families, no extension is needed in Theorems 4.1, 4.2, and at most one is needed in Example 4.5 and Theorem 4.6 on the stated regular sets.

(b) Evaluating derivatives and brackets. Closed-form expressions for the fields are not required for pointwise evaluation. For fixed probes FB(θ)=tr⁡(BΦ(θ))F_B(\theta)=\operatorname{tr}(B\Phi(\theta)), we have

$

∇FB,∇FC\nabla F_B,\nabla F_C

(\theta)=\nabla^2F_C(\theta)\nabla F_B(\theta)-\nabla^2F_B(\theta)\nabla F_C(\theta).$

These derivatives can be evaluated by automatic differentiation in simulation. In addition, when each parameter occurs in a single Pauli rotation, they admit finite parameter-shift formulas. These identities are exact for exact expectation values; however, they do not by themselves give a polynomial cost for arbitrary bracket orders.

(c) Certifying completeness. Suppose that qq independent conservation laws have been established on a neighborhood of θ0\theta_0, and let A(θ)A(\theta) contain finitely many smooth measurement-gradient fields and their brackets as columns. Then

\le r_{\mathrm{Lie}}(\theta) \le p-q.$$ If $\operatorname{rank}A(\theta_0)=p-q$, a nonzero minor remains nonzero near $\theta_0$. Hence the two bounds agree locally, and Theorem 3.5 gives exactly $q$ independent local conservation laws. Thus, once the upper bound is established, a single exact rank computation certifies local completeness. For general circuits, the remaining costs include constructing brackets, evaluating their derivatives, and certifying ranks. While the above results provide bounds on bracket depth and criteria for certifying local completeness, the computational efficiency of these procedures may vary across circuit architectures. --- **Reference** [Marcotte et al., 2023] Abide by the law and follow the flow: Conservation laws for gradient flows. Advances in neural information processing systems, 2023.
AuthorsOpenReview ↗

Response to Reviewer e6qg (Part II)

Q1. Do you have an efficient algorithm to upper bound the Lie-closure rank?

Answer. Yes, in the following sense. Our characterization gives a deterministic structural bound and a randomized procedure that assumes exact access to expectation gradients. Any qq functionally independent conservation laws imply rLie(θ)≤p−qr_{\mathrm{Lie}}(\theta)\le p-q, because each law has zero directional derivative along every measurement-gradient field, and this property is preserved under Lie brackets. Thus, conservation laws provide upper bounds without computing the full Lie closure.

(i) Causal cone. Consider a circuit given as a list of local gates and a fixed readout whose support is known. We trace this support backwards through the circuit. At each step, we retain any gate that meets the current support and add all its qubits to the support. Let pconep_{\mathrm{cone}} be the number of distinct parameters appearing in the retained gates. Since Φ\Phi is independent of every other parameter, each corresponding coordinate is conserved by Theorem 3.3. Therefore, rLie(θ)≤pcone.r_{\mathrm{Lie}}(\theta)\le p_{\mathrm{cone}}. If each gate acts on a bounded number of qubits and depends on a bounded number of parameters, this procedure takes O(g+n+p)O(g+n+p) time, where gg is the number of gates. However, the bound can be loose and may equal pp.

(ii) Global linear laws. By Theorem 3.3, the function hv(θ)=v⊤θh_v(\theta)=v^\top\theta is conserved if and only if v∈K:=⋂θ∈Θker⁡dΦθ.v\in K:=\bigcap_{\theta\in\Theta}\ker d\Phi_\theta. Consequently, rLie(θ)≤p−dim⁡K≤pcone,r_{\mathrm{Lie}}(\theta)\le p-\dim K\le p_{\mathrm{cone}}, where the second inequality follows because KK contains every inactive coordinate direction.

The space KK can be recovered by random sampling under the following assumptions. Suppose that Φ\Phi is real-analytic on a connected open parameter domain, as for Pauli-rotation circuits with affine parameter dependence. Draw pp independent pairs (θi,ρi)(\theta_i,\rho_i). For each pair, sample θi\theta_i from a distribution with a density and, independently, take ρi\rho_i to be a product of independent Haar-random single-qubit pure states. Define Fρ(θ)=tr⁡(ρΦ(θ)),Gi,:=∇Fρi(θi)⊤.F_\rho(\theta)=\operatorname{tr}\bigl(\rho\Phi(\theta)\bigr), \qquad G_{i,:}=\nabla F_{\rho_i}(\theta_i)^\top. Then ker⁡G=K\ker G=K with probability one in exact arithmetic. Indeed, product-state projectors span the Hermitian operators, so the collection of all probe gradients spans K⊥K^\perp. Analyticity then ensures that pp independent samples span this space with probability one.

This procedure requires pp gradient evaluations and O(p3)O(p^3) additional arithmetic operations. Its total cost is therefore O(pCgrad+p3)O(pC_{\mathrm{grad}}+p^3), where CgradC_{\mathrm{grad}} is the cost of one gradient evaluation. This gives polynomial oracle complexity, but does not imply polynomial total cost for arbitrary circuits, since gradient evaluation can itself be costly. Moreover, the guarantee requires exact gradients and exact rank decisions.

The second bound includes all global linear laws, including cross-layer ones. However, it can remain loose when additional independent local laws exist. Thus, these procedures provide a general deterministic structural bound and a potentially sharper randomized bound under the stated assumptions. We do not claim an efficient method for obtaining tight bounds for arbitrary PQCs.

Dear Reviewer e6qg,

Thank you for your careful and constructive review. We sincerely appreciate the time and consideration you have devoted to evaluating our work. We hope that our responses have addressed your comments and further clarified the motivation, methodology, and contributions of the paper.

As the discussion period is drawing to a close, we would welcome any additional questions or feedback you may have. We would be glad to address any remaining concerns and provide further clarification if needed.

We look forward to hearing your thoughts.

Best regards,

Authors

I would like to thank the authors to address my concerns. This improves my understanding on the estimation of the Lie-closure rank.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

Dear Area Chair and Reviewers,

As the discussion period approaches its conclusion, we sincerely thank the Area Chair and all reviewers for their careful assessments and constructive feedback. In response to the concerns raised, we have revised the manuscript and conducted additional experiments to strengthen our findings. All modifications and new results are highlighted in blue in the revised manuscript. Below, we summarize the main concerns addressed and the corresponding improvements.

Strengths

The reviewers acknowledge that studying conservation laws in PQCs is an important and timely research direction that remains largely unexplored. They recognize that our results advance the understanding of PQC training dynamics and appreciate the theoretical contributions, particularly the geometric characterization of conservation laws and the distinction between the Jacobian and Lie-closure ranks. They also highlight the concrete circuit examples and numerical verification of the predicted invariants under SGD.

Questions and Weaknesses

  • Connection to classical work (XA1X, z1K3): Clarified that our characterization of conservation laws for structured PQCs follows a similar approach to prior classical work by Marcotte et al. (2023, 2025), which developed a general conservation-law framework and studied its application to specific architectures, including two-layer linear networks, ResNets, and Transformers.

  • Contributions (XA1X, z1K3): Clarified our contributions relative to prior PQC redundancy studies, emphasizing the identification of exact parameter invariants under gradient flow and their implications for reducing optimization dimensions and quantum training costs. We further explained the significance of Theorem 4.4, which provides explicit, readout-dependent criteria for identifying frozen parameters at arbitrary circuit depth without computing the Lie closure.

  • Computational complexity (e6qg): Addressed the computational challenges of Lie-closure evaluation by providing local stopping criteria, completeness certificates, and two methods for upper-bounding its rank: a deterministic causal-cone approach and a randomized procedure for identifying global linear laws. We also analyzed their computational costs and clarified the limitations for general PQCs.

  • Practical benefits (XA1X, z1K3): Added finite-shot experiments on 3–8 qubits to evaluate the measurement efficiency and robustness of our gradient reconstruction method. For 8 qubits, the method reduces measurement settings from 96 to 38 and MSE from 0.173 to 0.085 at 128 shots per setting, and from 0.086 to 0.042 at 256 shots per setting.

  • Completeness and applicability (XA1X, z1K3): Clarified the scope of our theoretical guarantees, including complete characterizations for Basic PQCs at arbitrary width and depth and for the specified two-layer Clifford-entangler circuits at arbitrary width. We also verified that the completeness conditions hold for PennyLane's default two-layer ranges and the global-ZZ readout.

  • Extensions to other training settings (XA1X, z1K3): Clarified why the general frameworks of Marcotte et al. (2023, 2024, 2025) apply to general parametric predictors, including nonlinear models. We also derived the evolution of linear laws under weight decay and explained the applicability of existing results to other loss functions and momentum-based dynamics.

  • Experimental interpretation (XA1X, z1K3): Clarified that the experiments are intended to illustrate the theoretically established conservation laws rather than provide independent evidence of completeness, which is rigorously proved in our theoretical analysis for the specified circuit families. We also provided additional details on parameter initialization, training duration, and stopping criteria.

  • Presentation and notation (z1K3): Expanded the classical background, corrected notation and citation inconsistencies, and improved the description of experimental configurations.

We believe these responses and additional experiments substantially strengthen the manuscript, particularly by clarifying the novelty of our theoretical contributions, demonstrating measurement-efficiency benefits under finite-shot noise, and making the scope of our results more explicit.

We sincerely thank the Area Chair and reviewers again for their valuable feedback and consideration.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.