- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper studies how to choose the depth of a parameterized quantum circuit without training. Its proposal, Geometric Saturation Diagnosis, follows the rank of the quantum geometric tensor as layers are added and looks for the depth where that rank stops rising. The authors also give a parameter-counting lower bound and a Lie-orbit ceiling for the rank. Experiments use HEA and HVA circuits on TFIM and Heisenberg models with 4–10 qubits, and compare VQE energy and noise robustness at the detected depth with shallower and deeper circuits.
Critical strengths
1.A relevant and well-motivated problem. Picking the depth of a variational circuit is still mostly a matter of trial and error, and on NISQ hardware every extra layer adds noise. A cheap, training-free test for the point where more layers stop opening new directions would therefore be useful in practice. Thinking in terms of QGT rank and tangent space is a natural way to frame that question, and the link to Fisher/NTK redundancy diagnostics in deep learning is reasonable. The scan costs reported in Section 5 (0.20–0.74× one training run for N = 4–8) also suggest the diagnostic is affordable. That is why I think the topic is worth pursuing.
2.Separating geometric capacity from task-relevant capacity. The task-relevant tangent gain τL and Proposition 4 make a point that a plain rank count misses: a new geometric direction does not have to lower the energy to first order. The symmetry-constrained experiment in Table 10 is also a useful illustration that the rank diagnostic can pick up a lower orbit ceiling, which goes beyond HEA.
Critical weaknesses
Overall, the paper shows QGT rank saturation but not that the resulting depth is useful; this decides my recommendation.
-
The experiments do not show that L* is a better depth choice than alternatives, which is the main practical claim.
-
No meaningful baselines or compute-matched comparison: GSD is not compared with effective dimension, Larocca et al. (2023), expressibility, ADAPT-style growth, or simple shallow-depth stopping, and the GSD scan cost is excluded.
-
The recommended depth conflicts with the results: at N=8, L* has 2.3% clean-energy error vs 0.4% at L_max; L_half=10 is comparable clean and better under noise (67.6% vs 90.9% degradation at p=0.01); Fig. 4 favors L_max; Table 9 contradicts the monotone trend.
-
Training is unconverged: 100 Adam steps at lr=0.05 from random starts, 3–10 seeds, ~2% error; depth comparisons mix expressivity with trainability, and convergence curves, transfer, restarts, and gradient variance are missing.
-
The noise test mostly measures circuit size: equal p per gate means deeper circuits have more gates; fixed-gate-count, depth-independent error, or shot-noise models are needed.
-
Scale and inconsistencies limit evidence: saturation only up to N=8; at N=10 no saturation within L_max=12 while the parameter bound needs 69 layers; only two spin Hamiltonians, exact statevector, no shot noise; inconsistent L* values (N=6: 9 vs 8; N=8: 20 vs 21; Fig. 4 uses 20; Table 5 gives L*=5 with L_exact=4).
-
Theoretical claims overreach: the reported ranks (30, 126, 510) are real tangent-space dimensions, so the paper must clarify whether it ranks Q, Re(Q), or a realified Jacobian; Omega(2^N/N) is only a counting bound, alpha_N->1 is extrapolated from three points, and Eq. (6)/L* as an upper bound are unproven. The 5% L* is not L_exact (rank 454 to 510 between L*=21 and L_exact=25), so “beyond L* no new directions exist” is false as written.
-
Novelty is unclear relative to Haug, Bharti & Kim (2021) and Larocca et al. (2023), which already use QFI/QGT-rank saturation, effective dimension, and parameter redundancy.
Questions for authors
-
Could you compare GSD with at least two depth-selection baselines at matched total compute, including the cost of the GSD scan? Suitable baselines include effective quantum dimension or the Larocca et al. (2023) threshold, expressibility-based selection, ADAPT-style growth, or training a few shallow depths with early stopping. Is there a setting where the GSD-selected depth outperforms all of them?
-
Given Table 2 (2.3% clean error at L* versus 0.4% at L_max, and L=10 better at every nonzero noise level) and Figure 4, what accuracy or cost criterion identifies L* as the operating point?
-
Do the depth comparisons survive longer optimization, parameter transfer, or restarts? Convergence behavior and gradient-variance measurements, especially for deeper circuits, would help clarify whether saturation reflects expressivity rather than trainability.
-
Is d_L the rank of the complex Q, of Re(Q), or of a realified Jacobian? Please give the exact definition used in the code and explain how rank 126 at N=6 is consistent with Eq. (1), since the complex rank cannot exceed 2^N−1=63.
-
Is Eq. (6) a theorem or a heuristic? Given Table 2 and Larocca et al. (2023), what supports the claim that L* is an upper bound on useful depth? What exactly does the paper add beyond Haug et al. (2021) and Larocca et al. (2023), which already describe QFI/QGT-rank saturation and overparameterization?