acceptodds
Under review as a conference paper at ICLR 2027

Localizing SLD Geometry for Scalable Quantum Natural Gradient under Circuit Noise

Abstract

Variational quantum circuits generally produce mixed states in the presence of circuit noise. Quantum natural gradient (QNG) accounts for this mixed-state geometry using the symmetric-logarithmic-derivative (SLD) quantum Fisher information matrix, whose direct estimation can exceed practical circuit-shot budgets. Tractable QNG approximations developed for pure states do not directly provide this geometry, while mixed-state approaches often use more accessible surrogates, such as the Hilbert–Schmidt metric. We propose an approach to approximate the SLD geometry using local reductions of the noisy state. We define graph-local physical windows of radius and restrict the local SLD score space within each window to centered Pauli strings of weight at most . We assemble the local SLD metrics across fixed parameter blocks to obtain a block-local approximation. Increasing or yields nested approximation hierarchies, and the complete centered Pauli score space recovers the exact reduced-state SLD metric when no score-space regularization is used. For finite-shot implementation, we estimate the score-Gram and response matrices using local Pauli and parameter-shift measurements without reconstructing the reduced states. We derive finite-shot metric-estimation error bounds and show that, under fixed-locality assumptions, a sufficient metric-estimation cost scales as for fixed accuracy and confidence. We also establish convergence of the regularized finite-shot update to a stationary neighborhood of the noisy objective under standard stochastic-optimization assumptions. Experiments on noisy variational quantum eigensolver (VQE) instances for the Heisenberg model show improved convergence per cumulative shot compared with SGD, Adam, and approximate mixed-state QNG baselines under matched finite-shot budgets.

Primary area:
optimization
Keywords:
quantum natural gradient, symmetric logarithmic derivative, quantum Fisher information, variational quantum algorithms, noisy quantum circuits, mixed-state geometry, local quantum geometry

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_6467OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Noisy variational circuits output mixed states, whose natural geometry is the SLD quantum Fisher information matrix. Estimating it is costly, so existing mixed-state QNG methods use Hilbert–Schmidt (HS) surrogates. The paper localizes the SLD variational principle in two ways. First, it uses the reduced state on a radius-r window of the circuit interaction graph. Second, within each window it restricts the SLD score to centered Pauli strings of weight ≤ k_S. This gives F = BA⁺Bᵀ, which can be estimated from local Pauli and parameter-shift measurements without state tomography. Per-qubit parameter blocks are assembled into a block-diagonal preconditioner. The theory covers Loewner-nested hierarchies in r and k_S, an exact expression for the localization residual, finite-shot perturbation bounds, an O(n ε⁻² log(n/δ)) shot bound at fixed locality, and convergence to a stationary neighborhood. Experiments use noisy Heisenberg VQE with a hardware-efficient ansatz. At n = 6 they compare against the exact SLD-QFIM; at n = 12, depth 6, they run finite-shot optimization against SGD, Adam, Koczor HS-QNG, and WA-QNG at matched cumulative shots.

Critical strengths

None

Critical weaknesses

The main optimization result is in the regime where the paper's geometric claim does not hold. The block-local metric is closer to the SLD-QFIM only for η ≥ 0.002 (Fig. 2, Fig. 5). At η = 0.001 under non-uniform rates, Koczor HS has both lower metric error and a higher direction cosine. The headline finite-shot experiment (Fig. 4) runs at η = 0.0005. Even at η = 0.005, where the metric error is much lower (0.41 vs 0.85 for HS), the direction cosine is slightly worse (0.972 vs 0.983). So the evidence does not link "closer to SLD" to "better optimization." The finite-shot gain may come from cheap, block-diagonal, well-conditioned preconditioning rather than from SLD fidelity.

Questions for authors

Can you add a local-HS (and/or block-diagonal Fubini–Study) baseline with exactly the same windows, blocks, ridges, and shot schedule, to isolate the value of the SLD geometry itself?

2 comments

Here are our responses:

Weakness 1. We clarify that the geometric result concerns the approximation hierarchy within our construction. Increasing rr or kSk_S systematically retains more SLD geometry. It does not require the block-local metric to have lower error than every alternative metric at every noise rate.

Figures 2 and 4 use different circuit sizes and depths. Because η\eta is the Pauli-Lindblad rate per noise application, equal values of η\eta do not imply equal accumulated noise or state mixedness. Therefore, the noise threshold observed in the 66-qubit comparison cannot be transferred directly to the deeper 1212-qubit circuit.

A matched control in the same 1212-qubit, depth-66 setting and at the same noise rate as Figure 4 gives lower finite-shot energies for block-local SLD-QNG than for local-HS QNG (Appendix G.7, Figure 6).

Please see the Common response on SLD fidelity and finite-shot optimization for the discussion of how population SLD fidelity relates to finite-shot performance.

Question 1. We added the requested matched local-HS baseline using the same windows, parameter blocks, ridge, refresh period, and shot schedule. The result is reported in Appendix G.7 (Figure 6). After the early iterations, block-local SLD-QNG has lower mean energy over nearly all of the remaining shared budget, reaching −14.67-14.67 versus −12.42-12.42 for local-HS QNG at 70.9970.99 million shots.

We thank the reviewer for the comments. We would be grateful for any further feedback and are happy to clarify any remaining concerns.

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

This paper studies quantum natural gradient (QNG) for noisy variational circuits, where the symmetric-logarithmic-derivative (SLD) quantum Fisher information matrix is used to accoount for the mixed-state geometry. Directly estimating this is hard, so the authors propose tractable QNG apprxomiations of the SLD geometry using local reductions of noisy state. Finite-shot metric error bounds are presented with cost scale analysis. Experiments on noisy Heisenberg VQE are performed comparing other methods under finite-shot budgets.

Critical strengths
  • The paper is well-written and easy to follow. It is nice to have exact residual characterization.
  • It seems the shot accounting is carefully done and includes the metric-estimation cost. The limitations, including bounded windows and blocks are discussed honestly.
  • The block-local metric is much closer to the full SLD-QFIM than the HS-based surrogates in higher noise. In the 12-qubit experiment it reaches the lowest mean energy per shot over most of the budget.
Critical weaknesses
  • My main concerns is the results do not seem to support that a more accurate SLD approximation leads to better optimization. At η=0.005\eta=0.005, Koczor HS-QNG has a higher cosine (with the full-SLD direction) and under nonuniform rates it also gives the highest cosine. WA-QNG also performs well even though SLD error is large.
Questions for authors
  • Can you clarify my main concern in weakness?
  • Is it possible to add a baseline with the HS metric restricted to the same parameter blocks and windows?
  • For the 12-qubit setting, how close are the three metrics to the full SLD-QFIM?
4 comments

Here are our answers to the questions:

Question 1. For a fixed gradient and ridge, a more accurate SLD approximation helps keep the regularized update close to the full-SLD update. Under a finite-shot budget, however, optimization also depends on estimation cost and sampling error, so population SLD accuracy alone does not determine performance.

Koczor HS-QNG's closer population agreement at low noise is consistent with its pure-state limit, but estimating its dense metric requires approximately 16.75×16.75\times as many additional geometry shots per refresh as block-local SLD-QNG in the 1212-qubit experiment. At η=0.005\eta=0.005, block-local SLD-QNG has lower operator-norm error, while Koczor HS-QNG has a slightly higher direction cosine. This is not inconsistent because operator-norm error compares the complete metrics, while cosine tests the update direction for one gradient. The two diagnostics therefore need not give the same ordering.

WA-QNG targets a different, Hamiltonian-weighted geometry, so its performance is not determined by its distance from the full SLD-QFIM. Please see Common response on SLD fidelity and finite-shot optimization for further details.

Question 2. We added the requested matched local-HS baseline using the same windows, parameter blocks, learning rate, ridge, refresh period, and shot schedule. After the early iterations, block-local SLD-QNG has lower mean energy over nearly all of the shared budget. At the shared endpoint, the mean energies are −14.67-14.67 for block-local SLD-QNG and −12.42-12.42 for local-HS QNG. The result is reported in Appendix G.7 (Figure 6).

Question 3. Computing metric closeness at 1212 qubits requires the exact full SLD-QFIM, including a 4096×40964096\times4096 noisy state and its 144144 parameter tangents at each evaluated checkpoint. A multi-seed calculation exceeds what we can complete within the discussion period. We therefore report exact metric fidelity only for the 66-qubit comparison in Figure 2.

We thank the reviewer for the comments. We would be grateful for any further feedback and are happy to clarify any remaining concerns.

Apologies for the late reply, and thank you for the additional experiments and clarifications. The matched local-HS control alleviates my concern about the role of the shared local block structure. My remaining concern is still the significance/utility of this work. The favorable scaling concerns a fixed localized approximation, and there is no guarantee that useful geometry is retained as circuit size or depth grows. The convergence result does not establish an advantage over SGD, and the empirical evidence is limited to small VQE instances. For these reasons I retain my score of 2. My score is provisional, and I will engage in further discussion with the AC and other reviewers before finalizing my assessment.

Thank you for clarifying your remaining concern and acknowledging the matched local-HS control.

The radius rr and score cutoff kSk_S provide explicit accuracy–cost controls. Larger windows and richer score spaces retain more SLD geometry within fixed parameter blocks. The fixed-locality shot bound applies when window and block sizes remain bounded.

To provide further evidence of practical utility, we have added a 16-qubit Heisenberg experiment conducted during the rebuttal period (Appendix G.10, Figure 10). Block-local SLD-QNG achieves lower mean energy than SGD, Adam, and WA-QNG over most of the shared cumulative-shot range with all energy, gradient and metric-estimation costs included.

Together with the noise-strength and TFIM controls, this larger-instance experiment strengthens the empirical evidence that the localized geometry provides useful preconditioning. The shot-efficiency advantage persists beyond the main 12-qubit instance after accounting for metric-estimation costs.

Reviewer_UDjWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets natural-gradient optimization of variational circuits under noise, where the relevant geometry is the SLD quantum Fisher matrix of a mixed state. Estimating the full QFIM is too expensive, and existing surrogates (Hilbert–Schmidt, Hamiltonian-weighted HS) are not the SLD metric. The authors propose to restrict the SLD variational problem in two ways. The theory covers three things. First, two Loewner-order hierarchies, in r and k_S, with an exact localization residual. Second, a finite-shot concentration bound with an O(n log n) shot count under fixed-locality assumptions. Third, a standard nonconvex SGD-style convergence result to a stationary neighborhood. Experiments use noisy Heisenberg VQE with Pauli–Lindblad noise. At 6 qubits the authors compare against the exact full SLD-QFIM. At 12 qubits they run finite-shot optimization against SGD, Adam, HS-QNG and WA-QNG at matched cumulative shots.

Critical strengths
  • The setup is interesting and clear. By framing localization as a restricted SLD variational principle (Lemma 3.1), we get a principled one-sided approximation (0⪯Flocal⪯FSLD0 \preceq F_{\text{local}} \preceq F^{\text{SLD}}) with two tunable knobs. Plus, the exact residual formula (Thm 4.2) and the separation of EcrossE_{\text{cross}} from Eloc(r)E_{\text{loc}}^{(r)} make it easy to track down where errors come from.

  • The estimator looks practical.

  • The authors are refreshingly upfront about their limitations: increasing rr doesn't capture cross-block couplings, and the convergence theorem only covers the independent-gradient-batch variant. They do use non-uniform noise and a second Hamiltonian as helpful controls, though.

Critical weaknesses
  • The empirical results don't fully back up the main claim. The core premise is that staying closer to the SLD-QFIM improves optimization, but in the finite-shot test (η=0.0005\eta = 0.0005), HS-QNG is actually closer to full SLD in both operator norm and update direction than block-local (Fig. 2a/b). Block-local only wins on metric error at higher learning rates (η≥0.002\eta \ge 0.002), where its edge pretty much disappears anyway. Meanwhile, WA-QNG performs well despite a high relative metric error (~0.9). The authors admit metric error isn't everything, but that directly undercuts the "SLD fidelity" argument. The actual gains likely come from unisolated factors like estimator variance, ridge scaling, or lower staleness.

  • Some of of the key baselines and oracles are missing, such as block-diagonal Fubini–Study/pure-state QNG, exact, shot-free full-SLD-QNG or block-diagonal-SLD-QNG

  • The theorem assumes gradient noise is conditionally unbiased, but the experiments reuse response measurements at refresh iterations, which correlates with gradients. The authors acknowledge this, but it means the theory doesn't actually cover their implementation and a clarification/justification on gap looks necessary.

  • The technical novelty is limited and the paper weights more on empirical side (not favoring any, but just understanding the scope)

Questions for authors
  1. In Fig. 2, is the block-local F computed with γ=0.01\gamma = 0.01 and the PSD projection, or with the pseudoinverse? Is the full-SLD reference computed with a pseudoinverse? How sensitive are the errors and cosines to γ\gamma, especially at small η\eta where reduced states are nearly pure and A has small eigenvalues.

  2. Can you add an exact (shot-free) full-SLD-QNG and block-diagonal-SLD-QNG run on the 12-qubit instance? This would separate geometry quality from finite-shot effects and show how much E_cross costs.

  3. Can you report circuit/setting counts and wall-clock estimates alongside shots, and give the exact ground-state energy and noisy-optimum reference in Figs. 4 and 6?

  4. To be consistent with experiments, can the convergence analysis be extended to the measurement-reuse variant actually used, and can you comment on whether the bound can show any advantage over plain SGD?

4 comments

W1 — SLD fidelity and finite-shot performance. Block-local SLD-QNG retains useful SLD geometry at substantially lower estimation cost than the dense Koczor baseline and its gain is not explained solely by staleness, ridge, or the shared local architecture. All metric-based methods use the same refresh period, so metric staleness is matched. All finite-shot comparisons evaluate the complete estimators under matched cumulative-shot budgets, so their sampling effects are included. The matched local-HS control additionally holds locality, learning rate, parameter-space ridge, and shot budget fixed. Block-local SLD-QNG still reaches a lower mean energy of −14.67-14.67, compared with −12.42-12.42 for local Hilbert-Schmidt QNG (Appendix G.7, Figure 6).

The cited fidelity ordering and finite-shot result come from two different experiments. The Koczor fidelity ordering comes from Figure 2, which evaluates exact population metrics on the 66-qubit circuit, whereas the finite-shot optimization at η=5×10−4\eta=5\times10^{-4} is Figure 4 on the deeper 1212-qubit circuit. Because η\eta is the Pauli-Lindblad noise rate per application, the circuits experience different accumulated noise and mixedness. The six-qubit population ordering therefore cannot be transferred directly to Figure 4. Appendix G.8 (Figure 7) provides the corresponding finite-shot sweep for the same 66-qubit setting. Please see the Common response on SLD fidelity and finite-shot optimization for the full Koczor and WA-QNG comparisons.

W2 — Missing baselines and oracles. The block-diagonal QNG of Stokes et al. targets the pure-state Fubini-Study geometry rather than the mixed-state geometry produced by the noisy circuit. To provide a matched control in the noisy setting, we added the matched local-HS control on the 1212-qubit instance. With the local architecture and measurement budget fixed, block-local SLD-QNG reaches lower finite-shot energies than local-HS QNG (Appendix G.7, Figure 6). The exact full-SLD QNG and global block-diagonal SLD QNG oracles are addressed in our response to Question 2 and in Appendix G.9 (Figure 9).

W3 — Measurement reuse. We extended the convergence analysis to the measurement-reuse variant used in the experiments. Reuse adds a conditional-bias term whose contribution is amortized over the refresh period, while convergence to a stationary neighborhood is preserved under the stated assumptions. This extension is given in Appendix F.4 (Theorem F.3) and discussed further in our response to Question 4.

W4 — Technical novelty. While related ingredients exist individually, our contribution is their integration into a localized SLD framework for noisy variational circuits with new structural and finite-shot guarantees. Our structural results establish Loewner-ordered hierarchies in physical-window and score-space locality, an exact localization residual, and exact recovery of the reduced-state SLD metric with complete local scores. Our finite-shot analysis provides a tomography-free estimator with concentration and fixed-locality scaling guarantees and proves convergence of the regularized update to a stationary neighborhood. The matched cumulative shot experiments then evaluate the practical utility of this theoretically characterized construction under finite-shot budgets. We view the work as a theory-driven framework with finite-shot empirical validation, rather than as primarily empirical.

Here are our answers to the questions:

Question 1 Figure 2 uses exact population metrics and does not use the finite-shot score ridge γ\gamma. Both the block-local and full-SLD metrics are computed from the spectral SLD formula with support cutoff 10−1210^{-12}. Since kS=3k_S=3 spans the complete Pauli score space on these windows, the block local metric is equivalent to BA+B⊤BA^+B^\top. No PSD projection is applied when computing the metric error. For the direction cosine, numerical negative eigenvalues are clipped before adding the common ridge μ=0.1\mu=0.1.

Although γ\gamma is used only in the finite-shot estimator, we isolate its population level effect by replacing A+A^+ with (A+γI)−1(A+\gamma I)^{-1} at η=5×10−4\eta=5\times10^{-4} for the same ten initializations.

\hline \text{Score-space inverse} & \text{Metric error} & \text{Direction cosine}\\\\ \hline A^+ & 0.447\pm0.027 & 0.852\pm0.013\\\\ \gamma=0.01 & 0.449\pm0.031 & 0.856\pm0.015\\\\ \gamma=0.3 & 0.465\pm0.029 & 0.876\pm0.012\\\\ \gamma=1 & 0.525\pm0.017 & 0.890\pm0.010\\\\ \hline \end{array}$$ Values are means $\pm$ standard errors over the ten initializations. Thus $\gamma=0.01$ introduces little population-level bias, while larger values modify the metric more. **Question 2** The requested shot-free $12$-qubit comparison requires a $4096\times4096$ noisy density matrix, $144$ parameter tangents, and a global SLD eigendecomposition at each metric refresh, together with exact gradients at every iteration. This exceeds what we can complete within the discussion period. We instead added the same controlled comparison on the $6$-qubit, depth-$4$ instance in Appendix G.9 (Figure 9). Both methods use exact noisy objectives, gradients, and population metrics with identical initializations and optimization settings. Since the oracles have no circuit-shot budget, Figure 9 reports energy against iteration. Full-SLD QNG reaches lower mean energies later in the optimization, indicating a modest effect from $E_{\rm cross}$ in this setting. Figure 4 continues to evaluate finite-shot performance per cumulative circuit shot. **Question 3** We expanded Appendix G.6 to report circuit settings and representative serial Qiskit Aer wall-clock times alongside circuit shots. The SGD and Adam rows report the ordinary-iteration cost shared by all methods. The metric-based rows report their refresh-iteration costs. For WA-QNG, each independently sampled one-shot shadow basis is counted as one setting. $$\begin{array}{lrrr} \hline \text{Method} & \text{Settings} & \text{Shots} & \text{Time (s)}\\\\ \hline \text{SGD} & 867 & 443392 & 427\\\\ \text{Adam} & 867 & 443392 & 430\\\\ \text{Block-local SLD-QNG} & 7104 & 762496 & 734\\\\ \text{WA-QNG} & 867+288\times 576 & 609280 & 686\\\\ \text{Koczor HS-QNG} & 42627 & 5788672 & 5569\\\\ \hline \end{array}$$ The exact noiseless ground-state energies are $-20.568363$ and $-9.974309$ for the $12$- and $6$-qubit instances, respectively. We do not have an exact noisy optimum. We report the exact noiseless ground-state energy, which is a lower bound on the noisy objective.

Question 4. We have extended the analysis to the measurement-reuse variant used in the experiments. Reuse adds a conditional-bias term amortized by the refresh period, while preserving convergence to a stationary neighborhood (Appendix F.4, Theorem F.3). The bound does not show an advantage over plain SGD, so the per-shot advantage reported in the paper is empirical.

At refresh iterations, the response measurements are also used to estimate the gradient. We therefore write

\qquad b_t= \mathbb{E}\left[ \widehat g_\eta^t-g_\eta^t \mid \boldsymbol\theta^t,M_t \right],$$ where $$\mathbb{E}[\zeta_t \mid \boldsymbol{\theta}^t,M_t]=0, \qquad \mathbb{E}[\Vert\zeta_t\Vert_2^2 \mid \boldsymbol{\theta}^t,M_t] \le \sigma_{g_\eta}^2.$$ We assume $\mathbb{E}[\Vert b_t\Vert_2^2]\le\beta_{\mathrm{reuse}}^2$ at refresh iterations. At all other iterations, an independent gradient batch is used and $b_t=0$. Under $L_C$-smoothness, $\mu I\preceq M_t\preceq\Lambda_{\max}I$, and $\alpha\le\mu/(4L_C)$, the extended analysis gives $$\begin{aligned} \min_{0\le t<t_{\max}} \mathbb{E}[\Vert\nabla C_\eta(\boldsymbol{\theta}^t)\Vert_2^2] &\le \frac{4\Lambda_{\max} \left(C_\eta(\boldsymbol{\theta}^0)-C_{\inf}\right)} {\alpha t_{\max}} + \frac{2L_C\alpha\Lambda_{\max}\sigma_{g_\eta}^2}{\mu^2} + \frac{3\Lambda_{\max}\beta_{\mathrm{reuse}}^2}{\mu} \left( \frac{1}{R_{\mathrm{ref}}} + \frac{1}{t_{\max}} \right). \end{aligned}$$

We thank the reviewer for the comments. We would be grateful for any further feedback and are happy to clarify any remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.