pith. sign in

arxiv: 2504.08170 · v1 · submitted 2025-04-10 · 🪐 quant-ph · cs.LG· physics.comp-ph

Efficient measurement of neutral-atom qubits with matched filters

Robert M. Kent , Linipun Phuttitarn , Chaithanya Naik Mude , Swamit Tannu , Mark Saffman , Gregory Lafyatis , Daniel J. Gauthier This is my paper

Pith reviewed 2026-05-22 19:34 UTC · model grok-4.3

classification 🪐 quant-ph cs.LGphysics.comp-ph
keywords neutral-atom qubitsqubit readoutmatched filtersreadout crosstalkmachine learningmeasurement fidelityscalable quantum computing
0
0 comments X

The pith

Matched filters for neutral-atom qubit readout cut errors by up to 43% compared to Gaussian thresholds while using far fewer parameters than convolutional neural networks.

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

The paper develops two matched-filter algorithms for measuring states of qubits in a neutral-atom array. A local filter processes one site while an array filter adds signals from neighbors to suppress crosstalk between closely spaced atoms. Both models lower readout errors relative to a standard Gaussian threshold detector, with the array version reaching a 43% reduction. They achieve this using two orders of magnitude fewer trainable parameters and four orders of magnitude fewer multiplications than a recent neural-network method, and the learned weights remain physically interpretable. The results target the practical requirement for accurate, low-cost measurements that can scale to larger neutral-atom processors.

Core claim

The central claim is that an array matched filter trained on experimental fluorescence data incorporates neighbor-site information to mitigate readout crosstalk and produces up to 43% lower error rates than a conventional Gaussian threshold approach, while requiring only two orders of magnitude fewer parameters and four orders of magnitude fewer arithmetic operations than a convolutional neural network that achieves comparable accuracy.

What carries the argument

The array matched filter, a linear combination of fluorescence signals from a target atom and its neighbors whose weights are learned from calibration data to minimize total classification error.

If this is right

  • The array model delivers up to 43% error reduction and the local model up to 32% reduction versus a Gaussian threshold detector.
  • The array matched filter uses two orders of magnitude fewer trainable parameters and four orders of magnitude fewer multiplications than a convolutional neural network while increasing error by only 3.5%.
  • Visual inspection of the learned filter weights reveals experimental imperfections such as crosstalk sources.
  • A convolutional neural network can be pruned to 70 times fewer parameters and 4000 times fewer operations with comparable error rates.
  • Performance remains stable across a range of readout durations with only minor error increase.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The low computational cost could allow real-time retraining when experimental conditions drift, extending usable lifetime of the calibration.
  • The same matched-filter construction may transfer to other physical qubit platforms that suffer from spatially correlated readout noise.
  • Interpretable weights could directly inform hardware redesigns that reduce crosstalk at the source rather than correcting it after the fact.
  • Because the model is linear, its error statistics may combine more cleanly with standard quantum error-correction analyses than black-box neural networks.

Load-bearing premise

The training data gathered on the present apparatus remains representative of future operating conditions so that the learned filters do not overfit to transient hardware imperfections or particular atom placements.

What would settle it

Apply the trained filters to a new data set taken with altered atom spacing, different laser intensity, or a different trap configuration and verify whether the stated error reductions of 32% and 43% continue to hold.

Figures

Figures reproduced from arXiv: 2504.08170 by Chaithanya Naik Mude, Daniel J. Gauthier, Gregory Lafyatis, Linipun Phuttitarn, Mark Saffman, Robert M. Kent, Swamit Tannu.

Figure 1
Figure 1. Figure 1: FIG. 1. Matched filter designs for a single qubit with bound [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. The measurement infidelity from different analysis [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Classification infidelity for each qubit (numbered in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. False-3D images of the learned weights for each qubit [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Weights of neighbor qubits (numbered in the top left [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. (a) Infidelity comparison between CNN-Site (Baselin [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Quantum computers require high-fidelity measurement of many qubits to achieve a quantum advantage. Traditional approaches suffer from readout crosstalk for a neutral-atom quantum processor with a tightly spaced array. Although classical machine learning algorithms based on convolutional neural networks can improve fidelity, they are computationally expensive, making it difficult to scale them to large qubit counts. We present two simpler and scalable machine learning algorithms that realize matched filters for the readout problem. One is a local model that focuses on a single qubit, and the other uses information from neighboring qubits in the array to prevent crosstalk among the qubits. We demonstrate error reductions of up to 32% and 43% for the site and array models, respectively, compared to a conventional Gaussian threshold approach. Additionally, our array model uses two orders of magnitude fewer trainable parameters and four orders of magnitude fewer multiplications and nonlinear function evaluations than a recent convolutional neural network approach, with only a minor (3.5%) increase in error across different readout times. Another strength of our approach is its physical interpretability: the learned filter can be visualized to provide insights into experimental imperfections. We also show that a convolutional neural network model for improved can be pruned to have 70x and 4000x fewer parameters, respectively, while maintaining similar errors. Our work shows that simple machine learning approaches can achieve high-fidelity qubit measurements while remaining scalable to systems with larger qubit counts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces two matched-filter machine learning models for readout of neutral-atom qubits: a local site model and an array model that incorporates neighbor information to mitigate crosstalk. On experimental fluorescence traces, the models achieve error reductions of up to 32% and 43% relative to a conventional Gaussian threshold, while the array model uses two orders of magnitude fewer trainable parameters and four orders of magnitude fewer multiplications than a recent CNN approach, with only a 3.5% error increase across readout times. The learned filters are presented as physically interpretable.

Significance. If the reported gains hold, the work supplies a computationally lightweight and interpretable alternative to neural-network readout for neutral-atom arrays. The large reduction in parameter count and arithmetic operations directly addresses scalability concerns for larger qubit numbers, and the interpretability of the filters offers a route to diagnosing experimental imperfections without sacrificing performance.

major comments (1)
  1. [§4 (Experimental Results)] The 43% error reduction and efficiency claims for the array model are measured on held-out shots from the same experimental runs and apparatus. No cross-run or cross-calibration validation is reported to test whether the learned neighbor weights remain effective when laser intensity, position jitter, or loading statistics vary, which is load-bearing for the claim that the approach scales robustly to future operating conditions.
minor comments (2)
  1. [Abstract] The abstract sentence beginning 'We also show that a convolutional neural network model for improved can be pruned...' is incomplete; the intended phrase should be restored for clarity.
  2. [Figures 2-4] Figure captions and axis labels should explicitly state the number of shots and the precise definition of 'error' (e.g., assignment error per site) to allow direct comparison with the tabulated results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their positive summary and for identifying this important aspect of our validation strategy. We address the major comment point by point below.

read point-by-point responses

  1. Referee: [§4 (Experimental Results)] The 43% error reduction and efficiency claims for the array model are measured on held-out shots from the same experimental runs and apparatus. No cross-run or cross-calibration validation is reported to test whether the learned neighbor weights remain effective when laser intensity, position jitter, or loading statistics vary, which is load-bearing for the claim that the approach scales robustly to future operating conditions.

    Authors: We agree that cross-run validation would provide stronger evidence for robustness under varying conditions. Our reported results use held-out shots drawn from the same experimental runs to evaluate generalization to unseen data while remaining within the statistics and apparatus of the collected dataset; this tests performance against shot-to-shot fluctuations and any intra-run drifts that were present. The array model learns a small number of neighbor weights that primarily capture the fixed optical crosstalk geometry of the array, which is expected to remain stable for a given experimental setup. Because the model contains only a few dozen trainable parameters, it can be retrained or fine-tuned rapidly if calibration changes occur. We nevertheless acknowledge that the absence of explicit cross-run or cross-calibration tests limits the strength of the scalability claim. In the revised manuscript we will add a paragraph in §4 discussing this limitation and noting that the low parameter count makes periodic re-calibration on new data computationally inexpensive. We believe this clarification addresses the referee’s concern without altering the core experimental claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity: performance metrics are measured outcomes on held-out experimental data

full rationale

The paper trains local and array matched-filter models on experimental fluorescence traces collected from the neutral-atom apparatus and reports error rates, parameter counts, and multiplication counts directly from evaluation on held-out shots. These quantities are not defined by the fitted weights themselves nor obtained by renaming a fit as a prediction. No self-definitional equations, load-bearing self-citations, uniqueness theorems, or ansatzes smuggled via prior work appear in the central claims. The derivation chain consists of standard supervised learning on real hardware data followed by empirical benchmarking, which remains self-contained and falsifiable outside the fitted values.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard supervised learning assumptions plus the domain assumption that fluorescence traces contain sufficient information to separate signal from crosstalk. No new physical entities are postulated. The filter coefficients themselves are learned parameters but are not free parameters in the sense of being hand-tuned to produce the headline result; they are outputs of training.

free parameters (1)
  • filter coefficients
    Learned weights of the matched filters; their values are determined by training rather than chosen by hand to force the reported error reductions.
axioms (1)
  • domain assumption Fluorescence intensity traces from each site are statistically independent of distant sites once nearest-neighbor crosstalk is modeled.
    Invoked when the array model incorporates neighbor information to suppress crosstalk.

pith-pipeline@v0.9.0 · 5810 in / 1403 out tokens · 50300 ms · 2026-05-22T19:34:43.554325+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 2 internal anchors

  1. [1]

    For each measurement time, there are 6,002 images, randomly shuffled and split into 60% for training, 20% for testing, and 20% for validation

    Preprocessing The first step is to crop the full 442 × 62-pixel image to obtain the 28 × 28-pixel image corresponding to the secondary path. For each measurement time, there are 6,002 images, randomly shuffled and split into 60% for training, 20% for testing, and 20% for validation. Next, we find the mean pixel intensity across all training images and subtrac...

    work page

  2. [2]

    Estimating Statistical Uncertainty in Infidelity As mentioned in the previous section, the images are split and randomly shuffled into training, testing, and validation sets. For Figures 2 and 3, and Tables I and II, the dataset is split using 10 independent random shuffles, the infidelities at each measurement time are computed, and the infidelities for each ra...

    work page

  3. [3]

    Training In the training step, the feature vectors for each im- age x ∈ Rd× 1 (see Fig. 1), where d is the number of elements in the feature vector, are placed into a matrix X ∈ Rd× M , where M is the number of images in the training dataset, and the corresponding ground truth la- bels from the primary path y ∈ [0, 1] are placed into a matrix Y ∈ R1× M . ...

    work page

  4. [4]

    For each boundary size, the weights learned in the training step are used to predict the qubit states for all feature vectors x in the validation set by computing ˆ y = ˆWx

    Optimizing model metaparameters For each qubit and measurement time, we simultane- ously optimize the threshold and boundary size. For each boundary size, the weights learned in the training step are used to predict the qubit states for all feature vectors x in the validation set by computing ˆ y = ˆWx. Next, ˆy′ is computed for all thresholds ranging fro...

    work page

  5. [5]

    Testing Lastly, the optimal boundary size, weights, and thresh- olds for each qubit are used to obtain ˆy′ for each image in the testing image set and compared to the ground truth values y to obtain the classification fidelity

    work page

  6. [6]

    Cross-Fidelity Calculation To obtain a more representative cross-fidelity calcu- lation, we apply the trained models to a larger dataset 9 a) b) c) FIG. 6. (a) Infidelity comparison between CNN-Site (Baselin e-355k Model) and Pruned CNN-Site (5k Model) (b) Infidelity comparison between CNN-Array (Baseline-74M Model) and Pru ned CNN-Array (18k Model) (c) Over...

    work page

  7. [7]

    When scaled di- rectly to multi-qubit arrays, the model complexity grows to over 74 million parameters, rendering real-time infer- ence infeasible

    Baseline CNN Model The original single-qubit CNN model [ 16] has approx- imately 355,000 trainable parameters. When scaled di- rectly to multi-qubit arrays, the model complexity grows to over 74 million parameters, rendering real-time infer- ence infeasible

    work page

  8. [8]

    Pruned Model Our optimized single-qubit model eliminates unnec- essary convolutional and feed-forward layers through strategic pruning, reducing the parameter count from 355k to approximately 5k—a 70 × reduction—while pre- serving high readout fidelity. Key insights driving this reduction include the removal of fully connected layers, which contributed min...

    work page

  9. [9]

    Higher-level layers are added selectively to account for crosstalk between adjacent qubits

    Hierarchical Multi-Qubit Model To scale effectively, we introduce a hierarchical CNN architecture that reuses our lightweight single-qubit CNN as a modular component. Higher-level layers are added selectively to account for crosstalk between adjacent qubits. This hierarchical design enables high-fidelity multi-qubit readout with just 18k parameters—a 4,000 ...

    work page

  10. [10]

    Shor, Algorithms for quantum computation: discrete logarithms and factoring, in Proceedings 35th Annual Symposium on Foundations of Computer Science (1994) pp

    P. Shor, Algorithms for quantum computation: discrete logarithms and factoring, in Proceedings 35th Annual Symposium on Foundations of Computer Science (1994) pp. 124–134

    work page 1994

  11. [11]

    D. J. Bernstein and T. Lange, Post-quantum cryptogra- phy, Nature 549, 188 (2017)

    work page 2017

  12. [12]

    Gidney and M

    C. Gidney and M. Eker ˚ a, How to factor 2048 bit rsa in- tegers in 8 hours using 20 million noisy qubits, Quantum 5, 433 (2021)

    work page 2048

  13. [13]

    A Quantum Approximate Optimization Algorithm

    E. Farhi, J. Goldstone, and S. Gutmann, A quan- tum approximate optimization algorithm (2014), arXiv:1411.4028

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  14. [14]

    Zhou, S.-T

    L. Zhou, S.-T. Wang, S. Choi, H. Pichler, and M. D. Lukin, Quantum approximate optimization algorithm: Performance, mechanism, and implementation on near- term devices, Physical Review X 10, 021067 (2020)

    work page 2020

  15. [15]

    Abbas et al., Challenges and opportunities in quantum optimization, Nature Reviews Physics 6, 718 (2024)

    A. Abbas et al., Challenges and opportunities in quantum optimization, Nature Reviews Physics 6, 718 (2024)

    work page 2024

  16. [16]

    Y. Cao, J. Romero, and A. Aspuru-Guzik, Potential of quantum computing for drug discovery, IBM Journal of Research and Development 62, 6 (2018)

    work page 2018

  17. [17]

    N. S. Blunt, J. Camps, O. Crawford, R. Izs´ ak, S. Leon- tica, A. Mirani, A. E. Moylett, S. A. Scivier, C. Sunder- hauf, P. Schopf, et al. , Perspective on the current state- of-the-art of quantum computing for drug discovery ap- plications, J. Chem. Theory Comput. 18, 7001 (2022)

    work page 2022

  18. [18]

    P. W. Shor, Scheme for reducing decoherence in quantum computer memory, Physical review A 52, R2493 (1995)

    work page 1995

  19. [19]

    A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Surface codes: Towards practical large-scale quantum computation, Phys. Rev. A 86, 032324 (2012)

    work page 2012

  20. [20]

    Acharya et al

    R. Acharya et al. , Quantum error correction below the surface code threshold, Nature 638, 920 (2025)

    work page 2025

  21. [21]

    Saffman, T

    M. Saffman, T. G. Walker, and K. Mølmer, Quantum information with rydberg atoms, Rev. Mod. Phys. 82, 2313 (2010)

    work page 2010

  22. [22]

    Henriet, L

    L. Henriet, L. Beguin, A. Signoles, T. Lahaye, A. Browaeys, G.-O. Reymond, and C. Jurczak, Quantum computing with neutral atoms, Quantum 4, 327 (2020)

    work page 2020

  23. [23]

    Graham, Y

    T. Graham, Y. Song, J. Scott, C. Poole, L. Phuttitarn, K. Jooya, P. Eichler, X. Jiang, A. Marra, B. Grinke- meyer, et al. , Multi-qubit entanglement and algorithms on a neutral-atom quantum computer, Nature 604, 457 (2022)

    work page 2022

  24. [24]

    H. J. Manetsch, G. Nomura, E. Bataille, K. H. Leung, X. Lv, and M. Endres, A tweezer array with 6100 highly coherent atomic qubits (2024), arXiv:2403.12021

    work page internal anchor Pith review arXiv 2024

  25. [25]

    Phuttitarn, B

    L. Phuttitarn, B. M. Becker, R. Chinnarasu, T. M. Graham, and M. Saffman, Enhanced measurement of neutral-atom qubits with machine learning, Phys. Rev. Appl. 22, 024011 (2024)

    work page 2024

  26. [26]

    Maurya, C

    S. Maurya, C. N. Mude, W. D. Oliver, B. Lienhard, and S. Tannu, Scaling qubit readout with hardware efficient machine learning architectures, in Proceedings of the 50th Annual International Symposium on Computer Architec- ture , ISCA ’23 (ACM, 2023) p. 1–13

    work page 2023

  27. [27]

    C. N. Mude, S. Maurya, B. Lienhard, and S. Tannu, Effi- cient and scalable architectures for multi-level supercon - ducting qubit readout (2025), arXiv:2405.08982

    work page arXiv 2025

  28. [28]

    Google Quantum AI, Exponential suppression of bit or phase errors with cyclic error correction, Nature 595, 383 (2021)

    work page 2021

  29. [29]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep Learn- ing (MIT Press, 2016)

    work page 2016

  30. [30]

    Hertz, Sequential ridge regression, IEEE Trans

    D. Hertz, Sequential ridge regression, IEEE Trans. Aerosp. Electron. Syst. 27, 571 (1991)

    work page 1991

  31. [31]

    C. R. Vogel, Computational methods for inverse problems (SIAM, 2002)

    work page 2002