Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Quantum Circuit Pruning: Improving Fidelity via Compilation-Aware Circuit Approximation

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read This paper claims that pruning a parametric two-qubit gate whenever the routing cost of executing it exceeds the worst-case fidelity loss from omitting it improves circuit fidelity on noisy quantum devices, cutting two-qubit gates by up to

desk verdict Useful small step for NISQ compilation, but the core heuristic is local, the validation is self-consistent simulation, and the negative results in Fig. 3 undercut the universal claim. read the letter →

arxiv 2601.13322 v1 pith:W7MKG7PY submitted 2026-01-19 quant-ph

classification quant-ph
keywords quantumcircuitpruningrouting-awarecompilationNISQdevicesfidelitytwo-qubitgatereductionparametriccontrolledrotationsSWAPoverheadnoise-aware
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that a compilation-aware pruning rule can make noisy quantum circuits more reliable. The rule is simple: before executing a small-angle parametric two-qubit gate, compare the fidelity you lose by skipping the gate against the fidelity you lose by moving qubits into place with SWAP gates; if the routing loss is larger, drop the gate. Simulations on benchmark circuits with realistic noise show that this consistently reduces two-qubit gate counts and improves final state fidelity, with gains growing as circuits get larger and routing costs dominate. A sympathetic reader would care because this offers a parameter-free, hardware-aware way to improve NISQ computation without tuning approximation knobs.

What carries the argument

The load-bearing identity is the comparison between F_R(θ) = cos²(θ/2), the worst-case fidelity loss from omitting a rotation by angle θ, and F_swap, the estimated fidelity of routing the two qubits into adjacency using 3·⌈d/2⌉ CNOTs per qubit under depolarizing noise with a 1.25× distance correction. If F_swap < F_R(θ), the gate is pruned. This single threshold converts a compiler heuristic into a local decision rule that requires only gate angle and qubit distance.

What would settle it

Run the same pruning rule on hardware where two-qubit errors are strongly correlated or where crosstalk is significant, and compare pruned versus unpruned fidelity; if the predicted gains vanish or reverse despite the distance-based estimates favoring pruning, the independence and locality assumptions are falsified. A cheaper check is to recompute decisions using exact per-gate SWAP counts instead of the 1.25·d estimate and see whether the improved decisions change the fidelity outcome.

Watch

Extended reading notes

Core claim

The central claim is that omitting a small-angle two-qubit rotation is often less damaging than paying the SWAP overhead needed to execute it, so a compiler that makes pruning decisions from both the rotation angle and the physical distance between qubits can improve end-to-end fidelity. The paper proves a lower bound on the fidelity between a state and its rotated version, cos²(θ/2), and compares it to an analytic estimate of routing fidelity under depolarizing noise. When the routing fidelity estimate is smaller, the gate is pruned. Simulations across 4-to-14-qubit circuits show up to 48.6% reduction in two-qubit gate count and up to 47.7% improvement in fidelity versus the unrouted baseli

Load-bearing premise

The method assumes gate errors are independent and multiplicative, and that the routing cost of a gate can be estimated from physical qubit distance alone; if routing noise is correlated, state-dependent, or if pruning one gate alters the routing of others in ways the local distance estimate misses, pruning decisions can be wrong and fidelity may not improve.

Editorial extensions

If this is right

  • If the central claim is correct, quantum compilers can automatically drop costly small-angle gates without manually tuning an approximation degree, adapting to whatever topology and noise parameters the hardware has.
  • The benefit should grow on larger devices, because routing distances increase and SWAP overhead dominates, making pruning more attractive exactly where NISQ fidelity is worst.
  • The method can be bolted onto most existing routing algorithms, since it only adds a per-gate fidelity comparison rather than changing the routing engine itself.
  • Gate-count reduction of up to nearly half implies shorter execution time and reduced exposure to decoherence, amplifying the fidelity gain beyond the direct error model used here.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper leaves implicit is an adaptive threshold: as hardware error rates drift, the same rule could be re-evaluated per gate, making pruning responsive to real-time calibration data rather than fixed noise parameters.
  • The worst-case bound cos²(θ/2) is state-independent; a state-aware variant that tracks actual overlap might prune less aggressively but with higher confidence, or prune gates on entangled states that a worst-case analysis would keep.
  • The distance-estimate correction factor 1.25 is a placeholder for non-optimal routing; plugging in exact SWAP counts from the router would turn the heuristic into a more reliable decision rule, and could be tested against the approximate version.
  • The fidelity comparison is a form of cost–benefit analysis that could be generalized to other gate types or noise channels, e.g., replacing depolarizing noise with amplitude damping or crosstalk models, to see if the pruning order changes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a 'routing-aware pruning' strategy for NISQ circuits: for each parametric two-qubit gate, the compiler compares a worst-case lower bound on the fidelity lost by omitting the gate (cos^2(theta/2), Theorem 1) with an analytic estimate of the fidelity lost by routing that gate through SWAPs (Eq. 3). If the routing fidelity is lower, the gate is discarded. The method is integrated with the Route-Forcing compiler and evaluated in Qiskit simulation on MQT Bench circuits with 4-14 qubits under depolarizing plus T1/T2 noise. The paper reports up to 48.6% two-qubit gate reduction and up to 47.7% final-state fidelity improvement, with gains increasing with qubit count, and compares favorably with Qiskit's angle-based QFT approximation.

Significance. The idea is plausible and practically motivated, and the proof of the local fidelity bound is mathematically correct. The paper's strongest contribution is a concrete, compilation-aware decision rule that goes beyond simple angle thresholding and that is evaluated on a reasonable set of benchmark circuits. The comparison to Qiskit's approximation-degree pruning is informative. However, the current evidence is not sufficient to establish the central claim that pruning improves final fidelity: the noise model used to derive the decision rule is the same one used to evaluate it (Eq. 14 and Section III-C), and the step from per-gate decisions to global end-to-end fidelity is unproven and contradicted by several negative cases in Fig. 3. The contribution is therefore a suggestive heuristic with a correct local analysis, not yet a validated method.

major comments (3)
  1. [Section III-C, Eq. (14)] The evaluation is self-consistent rather than independent. The pruning rule uses F_swap (Eq. (3)), which depends on p2, and p2 is set in Eq. (14) as a function of the same circuit's gate density. The noisy simulation then applies exactly this p2, and T1/T2 are chosen 'to result in a representative fidelity' (Section III-C). Hence the simulation cannot falsify the analytic model: both baseline and pruned circuits are evaluated under a noise model that the authors calibrated with the same formula. In particular, Eq. (14) makes p2 smaller for circuits with more gates, which may produce the apparent size-dependent gain in Fig. 4. Please add results with an independent noise model (e.g., a hardware-calibrated noise model for a grid topology), or at minimum a sensitivity sweep over p2 and T1/T2, and report how often pruning improves fidelity outside the calibrating model.
  2. [Section III-B (pruning rule) and Fig. 3] The local criterion does not guarantee a global fidelity gain. Theorem 1 bounds the single-gate fidelity loss from omitting one rotation; it is not a bound on the cumulative effect of many omitted gates, since the state after multiple prunings can deviate by coherent sums of errors, and since pruning one gate changes qubit positions and therefore the routing cost of all later gates. The paper's own Fig. 3 contains negative fidelity improvements (e.g., -20.5% for qaoa on 10 qubits and -11.1% for qftentangled on 12 qubits), which shows that per-gate F_swap < F_R decisions can reduce end-to-end fidelity. To support the central claim, either provide a global error bound that relates the set of pruned gates to final-state fidelity, or run an ablation in which the actual change in final fidelity is measured per pruned gate and compared to the local prediction.
  3. [Eq. (3), Section III-B] The routing-cost estimate contains an unexplained empirical factor. F_swap assumes ceil(d/2) SWAP gates and then multiplies the distance by 1.25 to account for non-optimal routing. This factor is not derived or measured, and the routing algorithm's actual overhead may depend on circuit width, grid shape, and congestion in a way that 1.25*d does not capture. Because the pruning decision is an inequality between F_swap and F_R, small errors in this factor can change which gates are pruned and, as Fig. 3 shows, can reverse the fidelity sign. Please either justify the 1.25 factor from the routing algorithm's empirical behavior or use the actual number of SWAPs inserted by Route-Forcing in the decision rule.
minor comments (5)
  1. [Fig. 2 / Fig. 3] The x-axis labels '123 × 4 qubits' and '122 × 6 qubits' appear to be typos for 12-qubit grids (3×4 and 2×6); please correct. Figure 4 also uses '3 × 4' and '2 × 6' without explaining which corresponds to which.
  2. [Section III-C] Please specify the input state(s) used for fidelity computation (e.g., |0...0>, random states, or a specific algorithm input) and the number of shots/instances. Figure 4's mean and standard deviation are not reproducible without this information.
  3. [Section III-B] It is unclear whether pruning is integrated into Route-Forcing online (each gate is pruned when 'needed to be executed') or performed as a pre-pass using current positions before routing. Clarify the interaction between pruning and qubit mapping; 'current physical positions' implies an online pass, but the routing algorithm may not support mid-route gate removal.
  4. [Fig. 5] The 'Best Approx. Degree' is selected with hindsight (the best among all degrees). This is a useful upper bound but should be stated explicitly in the caption or text, as it makes the comparison optimistic for the baseline.
  5. [Eq. (3)] The derivation of F_swap from [13] is not reproduced. Since the paper's contribution depends on this expression, please include the intermediate steps or cite exact equation numbers from [13] so the expression is checkable without consulting the prior work.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity; the pruning rule is a heuristic and the simulation is self-consistent rather than an independent test.

full rationale

I walked the derivation chain. Theorem 1's cos^2(theta/2) bound is proved in the paper, not imported as a conclusion. Eq. (3) for F_swap is taken from the authors' prior work [13], and it does enter the pruning criterion, but it is an analytic noise model, not a fitted parameter; the paper does not fit Eq. (3) or the 1.25*d SWAP factor to the reported fidelity gains. The decision rule F_swap < F_R(theta) is presented as a heuristic, and Fig. 3 contains negative fidelity improvements, so the outcome is not forced by construction. The simulation in Sec. III-C uses a Qiskit depolarizing + T1/T2 noise model with the same p2 recipe as Eq. (14). This means the evaluation shares the same noise-model family that motivated the pruning rule, so the demonstration is largely self-consistent and does not validate the rule on hardware. That limits external validity, but it is not a circular derivation: the simulator is an independent implementation, the global fidelity metric is not the same object as the local F_swap estimate, and pruning can fail. The unproven local-to-global leap is a correctness risk, not a circular step.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small number of ad hoc noise model choices (p2, T1/T2, the 1.25 SWAP multiplier) and on the heuristic that per-gate worst-case fidelity comparisons are a valid proxy for global fidelity. No new physical entities are introduced.

free parameters (3)
  • SWAP-count multiplier (1.25) = 1.25
    Equation (3) multiplies the distance d by 1.25 to approximate actual SWAP insertions from non-optimal routing. The value is hand-picked, not measured or fit; the pruning threshold F_swap scales with it.
  • Two-qubit depolarizing error p2 = p2 = 1/(#gates/#qubits)^2
    Equation (14) sets the per-gate error rate from the circuit's gate count and qubit count. This is an ad hoc model choice to get a representative noisy baseline; it is not taken from hardware calibration.
  • Thermal relaxation times T1, T2 = T1 = T2 = 2 × circuit duration
    Equation (15) sets decoherence times proportional to circuit duration. This choice is arbitrary and affects the noise model; no hardware parameters are used.
assumptions (5)
  • domain assumption Depolarizing noise on two-qubit gates is independent and each SWAP contributes error according to Eq. (3).
    The fidelity model in Eq. (3) assumes independent depolarizing channels on each CNOT and a specific SWAP count. This model is from the authors' prior work [13] and is used both to make pruning decisions and in the Qiskit simulation, so the simulation is not an independent test.
  • domain assumption A gate can be safely pruned if the worst-case state fidelity of its rotation, cos^2(θ/2), exceeds the estimated routing fidelity F_swap.
    This is the central heuristic (Fig. 1, Section III-B). It treats each gate in isolation and does not account for state-dependent rotation effects or interactions between pruning decisions. It is load-bearing.
  • domain assumption Gate fidelities multiply to total circuit fidelity (standard Markovian noise assumption).
    Underlies the per-gate comparison and the simulation's density-matrix evolution. If errors are correlated or non-Markovian, the per-gate decision rule may be invalid.
  • ad hoc to paper The scaling p2 = 1/(#gates/#qubits)^2 produces a representative NISQ noise level.
    Equation (14) is not justified by hardware measurements; it is chosen to make the Noisy State baseline 'representative'.
  • domain assumption Theorem 1's lower bound using min-over-states (worst case) is the right criterion for gate removal.
    Uses the sharp lower bound cos^2(θ/2) as a conservative estimate. Actual circuit states may be more or less sensitive to the rotation than the worst case.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Circuit Pruning: Improving Fidelity via Compilation-Aware Circuit Approximation." pith.science (2026). https://pith.science/paper/W7MKG7PY

@misc{pith2026260113322,
  author       = {Pith},
  title        = {Pith review of: Quantum Circuit Pruning: Improving Fidelity via Compilation-Aware Circuit Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7MKG7PY}},
  note         = {Machine review of arXiv:2601.13322}
}
read the original abstract

This work presents a routing-aware pruning strategy for quantum circuits executed on Noisy Intermediate-Scale Quantum (NISQ) devices. We propose a method to remove parametric controlled rotations whose small rotation angles do not justify the routing overhead required for their implementation. By selectively pruning such gates, the method mitigates fidelity loss arising from additional SWAP operations introduced during compilation. Our approach evaluates whether executing a gate leads to greater fidelity loss than omitting it. Simulations on benchmark circuits with realistic noise models show that the method reduces two-qubit gate counts (up to 48.6%) while improving final state fidelity (up to 47.7%), especially for larger circuits where routing costs dominate.

Figures

Figures reproduced from arXiv: 2601.13322 by the authors.

Figure 1
Figure 1. Before applying a conditional rotation θ, the compiler assesses the distance between the two interacting qubits (blue and red paths). The proba￾bility of pruning increases with path length. The values shown correspond to a p2 = 0.005 (Fswap from Equation (3)) and a θ = π/6 rotation (FRnˆ (θ) from Equation (6)). In this example, a π 6 -rotation between the blue qubits is pruned, while the same rotation between the re… view at source ↗
Figure 2
Figure 2. Considered nearest-neighbor topologies ranging from 4 to 14 qubits. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two-qubit gate count (top row) and circuit fidelity (bottom row) for the selected benchmarks containing between 4 and 14 qubits. The percentage [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Fidelity’s relative improvement when pruning the circuit (blue) and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Fidelities for the noiseless and noisy approximation (blue and orange [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 2 linked inside Pith

  1. [1]

    M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information: 10th Anniversary Edition. Cambridge University Press, 2010

  2. [2]

    Quantum algorithms: an overview,

    A. Montanaro, “Quantum algorithms: an overview,”npj Quantum Infor- mation, vol. 2, no. 1, pp. 1–8, 2016

  3. [3]

    Quantum Computing in the NISQ era and beyond,

    J. Preskill, “Quantum Computing in the NISQ era and beyond,”Quan- tum, vol. 2, p. 79, Aug. 2018

  4. [4]

    Software mitigation of crosstalk on noisy intermediate-scale quantum computers,

    P. Murali, D. C. McKay, M. Martonosi, and A. Javadi-Abhari, “Software mitigation of crosstalk on noisy intermediate-scale quantum computers,” inProceedings of the Twenty-Fifth International Conference on Archi- tectural Support for Programming Languages and Operating Systems, 2020, pp. 1001–1016

  5. [5]

    Not all qubits are created equal: A case for variability-aware policies for nisq-era quantum computers,

    S. S. Tannu and M. K. Qureshi, “Not all qubits are created equal: A case for variability-aware policies for nisq-era quantum computers,” in Proceedings of the twenty-fourth international conference on architec- tural support for programming languages and operating systems, 2019, pp. 987–999

  6. [6]

    Tackling the qubit mapping problem for nisq-era quantum devices,

    G. Li, Y . Ding, and Y . Xie, “Tackling the qubit mapping problem for nisq-era quantum devices,” inProceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems, ser. ASPLOS ’19. New York, NY , USA: Association for Computing Machinery, 2019, p. 1001–1014

  7. [7]

    Quantum supremacy using a programmable superconducting processor,

    F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A. Buellet al., “Quantum supremacy using a programmable superconducting processor,”Nature, vol. 574, no. 7779, pp. 505–510, 2019

  8. [8]

    Qubit allocation,

    M. Y . Siraichi, V . F. d. Santos, C. Collange, and F. M. Q. Pereira, “Qubit allocation,” inProceedings of the 2018 International Symposium on Code Generation and Optimization, ser. CGO 2018. New York, NY , USA: Association for Computing Machinery, 2018, p. 113–125

Show all 26 references
  1. [9]

    Noise-adaptive compiler mappings for noisy intermediate-scale quan- tum computers,

    P. Murali, J. M. Baker, A. Javadi-Abhari, F. T. Chong, and M. Martonosi, “Noise-adaptive compiler mappings for noisy intermediate-scale quan- tum computers,” inProceedings of the twenty-fourth international conference on architectural support for programming languages and oper...

  2. [10]

    Approximate quantum fourier transform and decoherence,

    A. Barenco, A. Ekert, K.-A. Suominen, and P. T ¨orm¨a, “Approximate quantum fourier transform and decoherence,”Physical Review A, vol. 54, no. 1, p. 139, 1996

  3. [11]

    Offline quantum circuit pruning for quantum chemical calculations,

    S. Imamura, A. Kasagi, and E. Yoshida, “Offline quantum circuit pruning for quantum chemical calculations,” in2023 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1. IEEE, 2023, pp. 349–355

  4. [12]

    Qadaprune: Adaptive parameter pruning for training variational quan- tum circuits,

    A. Kulshrestha, X. Liu, H. Ushijima-Mwesigwa, B. Bach, and I. Safro, “Qadaprune: Adaptive parameter pruning for training variational quan- tum circuits,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 2. IEEE, 2024, pp. 120–125

  5. [13]

    An accurate and efficient analytic model of fidelity under depolarizing noise oriented to large scale quantum system design,

    P. Escofet, S. Rodrigo, A. Garcia-S ´aez, E. Alarc ´on, S. Abadal, and C. G. Almud ´ever, “An accurate and efficient analytic model of fidelity under depolarizing noise oriented to large scale quantum system design,” Quantum Science and Technology, vol. 10, no. 3, p. 035061, jul 2025

  6. [14]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincioet al., “Variational quantum algorithms,”Nature Reviews Physics, vol. 3, no. 9, pp. 625– 644, 2021

  7. [15]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,”Nature, vol. 549, no. 7671, pp. 195–202, 2017

  8. [16]

    Improved bounds for the approximate qft,

    D. Cheung, “Improved bounds for the approximate qft,” inProceedings of the Winter International Synposium on Information and Communica- tion Technologies, ser. WISICT ’04. Trinity College Dublin, 2004, p. 1–6

  9. [17]

    MQT Bench: Benchmark- ing Software and Design Automation Tools for Quantum Computing,

    N. Quetschlich, L. Burgholzer, and R. Wille, “MQT Bench: Benchmark- ing Software and Design Automation Tools for Quantum Computing,” Quantum, vol. 7, p. 1062, Jul. 2023

  10. [18]

    Route-forcing: Scalable quantum circuit mapping for scalable quantum computing architectures,

    P. Escofet, A. Gonzalvo, E. Alarc ´on, C. G. Almud ´ever, and S. Abadal, “Route-forcing: Scalable quantum circuit mapping for scalable quantum computing architectures,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 01, 2024, pp. 909– 920

  11. [19]

    Timing and resource-aware mapping of quantum circuits to superconducting processors,

    L. Lao, H. van Someren, I. Ashraf, and C. G. Almudever, “Timing and resource-aware mapping of quantum circuits to superconducting processors,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. 2, pp. 359–371, 2022

  12. [20]

    Efficient mapping of quantum circuits to the ibm qx architectures,

    A. Zulehner, A. Paler, and R. Wille, “Efficient mapping of quantum circuits to the ibm qx architectures,” in2018 Design, Automation & Test in Europe Conference & Exhibition (DATE), 2018, pp. 1135–1138

  13. [21]

    Lightsabre: A lightweight and enhanced sabre algorithm,

    H. Zou, M. Treinish, K. Hartman, A. Ivrii, and J. Lishman, “Lightsabre: A lightweight and enhanced sabre algorithm,”arXiv preprint arXiv:2409.08368, 2024

  14. [22]

    Quantum error correction below the surface code threshold,

    Google Quantum AI and Collaborators, “Quantum error correction below the surface code threshold,”Nature, vol. 638, no. 8052, pp. 920– 926, 2025

  15. [23]

    Quantum computing with qiskit,

    A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Crosset al., “Quantum computing with qiskit,”arXiv preprint arXiv:2405.08810, 2024

  16. [24]

    Quantum computing simulator on a heterogenous hpc system,

    J. Doi, H. Takahashi, R. Raymond, T. Imamichi, and H. Horii, “Quantum computing simulator on a heterogenous hpc system,” inProceedings of the 16th ACM International Conference on Computing Frontiers, ser. CF ’19, 2019, p. 85–93

  17. [25]

    Complexity-theoretic foundations of quan- tum supremacy experiments,

    S. Aaronson and L. Chen, “Complexity-theoretic foundations of quan- tum supremacy experiments,” inProceedings of the 32nd Computational Complexity Conference, ser. CCC ’17. Dagstuhl, DEU: Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, 2017

  18. [26]

    Characterizing quantum supremacy in near-term devices,

    S. Boixo, S. V . Isakov, V . N. Smelyanskiy, R. Babbush, N. Ding, Z. Jiang, M. J. Bremner, J. M. Martinis, and H. Neven, “Characterizing quantum supremacy in near-term devices,”Nature Physics, vol. 14, no. 6, pp. 595–600, Jun 2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.