Pith. sign in

REVIEW 2 major objections 5 minor 2 cited by

Quantum computational sensing using quantum signal processing, quantum neural networks, and Hamiltonian engineering

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Interleaving sensing with quantum computation before measurement classifies signals more accurately than estimating them first, with simulated gains over 20 percentage points at equal sensing time.

desk verdict A coherent framework with clean bosonic analytics, but the headline static-signal advantage is an artifact of a baseline that forbids coherent integration; worth refereeing with that fixed. read the letter →

arxiv 2507.15845 v1 pith:D74GTFJO submitted 2025-07-21 quant-ph

Saeed A. Khan , Sridhar Prabhu , Logan G. Wright , Peter L. McMahon This is my paper
classification quant-ph
keywords quantumcomputationalsensingsignalprocessingneuralnetworksHamiltonianengineeringnonlinearfunctionapproximationsingle-shotclassificationspatiotemporalsamplingnoise
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 sets out to show that a quantum sensor can do better at classification and nonlinear function-approximation tasks if the quantum system computes while it senses, instead of measuring a raw signal and postprocessing it classically. The proposed protocol interleaves $L$ sensing operations with $L$ coherent computing operations before a single measurement, with the circuit specialized to the task either analytically or by supervised learning on finitely sampled measurement results. Across simulated binary and four-class discrimination of static phases, arbitrary polynomial approximation of a displacement, and binary classification of real spatiotemporal magnetoencephalography signals, the authors report a regime where their quantum computational sensor beats a conventional quantum sensor at the same number of sensing periods $N$, with accuracy advantages exceeding 20 percentage points for some tasks. The advantage appears even for small systems — a single qubit, a few qubits, or one qubit coupled to one bosonic mode — and even in the single-shot regime ($S=1$). If the claim holds, near-term quantum hardware too small for general quantum computation could still improve decision-oriented sensing tasks.

What carries the argument

The load-bearing object is the interleaved circuit $U_{\rm QCS}=U_{\rm meas}\prod_{l=1}^{L}U_{\rm sense}(u^{(l)})U_{\rm coh}^{(l)}$: $L$ signal-dependent sensing operations alternating with $L$ coherent computing operations before a single measurement. Because the signal enters several times, the circuit can realize nonlinear functions of it: for a single qubit, quantum-signal-processing rotations (the QSP construction in which interleaved non-commuting rotations implement polynomial functions of an input phase); for several qubits, trainable unitaries of the kind used in quantum neural networks; for bosonic modes, a nonlinear quantum non-demolition amplifier $\hat{H}_{\rm NL}=\sqrt{2}g\,\hat{f}\hat{P}_b$ with Hermitian $\hat{f}=\sum Q_{nm}\hat{a}^{\dagger n}\hat{a}^m$, whose readout-mode heterodyne outcome has mean $g\langle\hat{f}\rangle$ and therefore directly outputs the desired polynomial of the sensed displacement. The same design principle appears in every platform: target functions are engineered so that measurement probabilities saturate toward 0 or 1, and because single-qubit sampling variance vanishes at those extremes, this both performs the classification and suppresses the quantum sampling noise that limits the conventional baseline — the property that makes single-shot operation work.

What would settle it

For a static phase signal, compute the classification error of a single Ramsey measurement that coherently integrates over the full $N$ periods (one shot of duration $N\cdot\tau$) and compare it with the QCS error at the same $N$ for the single-qubit task of Fig. 2. If the long coherent Ramsey baseline reaches or beats the QCS error, the claimed quantum computational advantage for static single-qubit classification does not survive a matched-coherence comparison; if QCS still wins, the advantage is genuinely nonlinear processing rather than coherence accounting.

Watch

Extended reading notes

Core claim

The central claim is that coherently processing a signal inside the sensor before measurement can beat the standard strategy of estimating the signal and then deciding classically, whenever the task is to compute a nonlinear function of the signal. The architecture is the unitary $U_{\rm QCS}=U_{\rm meas}\prod_{l=1}^{L}U_{\rm sense}(u^{(l)})U_{\rm coh}^{(l)}$, in which $L$ sensing operations are interleaved with $L$ trainable computing operations before one measurement; for qubit sensors the computing layers are quantum-signal-processing rotations (single qubit) or trainable multi-qubit unitaries (quantum-neural-network style), while for bosonic sensors the computation is done by a nonlinear quantum non-demolition amplifier with Hamiltonian $\hat{H}_{\rm NL}=\sqrt{2}g\,\hat{f}\hat{P}_b$, where choosing $\hat{f}=\sum_{n,m}Q_{nm}\hat{a}^{\dagger n}\hat{a}^m$ makes a heterodyne readout directly return a chosen polynomial of the displacement $\alpha$. All protocols are compared with conventional quantum-sensing baselines at equal total sensing periods $N$, and for every task the paper reports a regime in which the quantum computational sensor reaches lower classification error or lower mean-squared error, with the gap growing as the task becomes harder.

Load-bearing premise

The conventional sensor is benchmarked as many separate one-period measurements, while the quantum computational sensor uses all $N$ sensing periods in one unbroken run; the paper does not compare against a conventional sensor that also integrates the whole signal coherently, so part of the reported advantage may come from that asymmetry.

Editorial extensions

If this is right

  • For a fixed budget of $N$ sensing periods, a protocol with $L$ interleaved computing layers reaches lower classification error than the conventional sensor at the same $N$, and the largest advantage occurs in the single-shot limit $L=N$, $S=1$.
  • The advantage grows with task complexity: for single-variable tasks with $R$ noncontiguous class regions, the conventional sensor's error scales approximately as $0.063(2R-1)$ while the QCS error scales approximately as $0.015(2R-1)$ at $N=64$, so harder tasks show larger gaps.
  • A single qubit suffices for binary classification, $M=\log_2 C$ qubits suffice for single-shot $C$-class discrimination, and a one-qubit-plus-one-bosonic-mode hybrid sensor classifies the Circles task at 1.7% error versus 11.2% for the conventional protocol.
  • A bosonic sensor with an engineered nonlinear amplifier estimates arbitrary polynomials of a complex displacement with lower variance than a phase-preserving linear amplifier followed by classical postprocessing; for the XOR task the QCS estimator variance is half the QS variance at large gain.
  • Training can be done end-to-end on finitely sampled measurement results — even a single shot per inference — so the protocols do not require exact outcome probabilities or expectation values.

Reading between the lines

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

  • A direct test the paper does not run: for static signals, compare against a single Ramsey measurement that coherently integrates over the full $N$ periods. If that baseline matches the QCS error, part of the reported advantage is coherence accounting rather than nonlinear processing; if QCS still wins, the advantage is genuine computation in the sensor.
  • The resource relation $N = L \times S$ implies a measurable tradeoff curve: for fixed $N$, classification error plotted against $L$ should dip to a minimum whose position shifts with task complexity, a signature a small experimental device could check in a few parameter sweeps.
  • The design principle of pushing final measurement probabilities toward 0 and 1 to suppress sampling noise is task-agnostic and may transfer beyond sensing to any few-shot quantum classifier trained on noisy hardware.
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

2 major / 5 minor

Summary. The paper proposes a general framework for quantum computational sensing (QCS) in which L sensing operations are interleaved with L coherent processing operations before a single measurement, and applies it to qubit-based, bosonic, and hybrid qubit-bosonic platforms. The tasks include binary and multiclass classification of static phases, classification of time-varying spatiotemporal MEG signals, and polynomial function approximation of complex displacements. The authors compare each QCS protocol against a conventional quantum sensing (QS) baseline that first estimates the signal and then classically postprocesses, reporting accuracy advantages that in several simulated tasks exceed 20 percentage points at equal total sensing periods N. Training is performed with finitely sampled measurement outcomes, including single-shot inference, and the bosonic protocols are supported by analytic derivations of unbiased estimators and their variances.

Significance. If the comparative claims are upheld, the paper would be a substantial contribution to quantum sensing: it unifies several QCS proposals under one architecture, demonstrates finite-shot training with single-shot inference, and provides clean analytic constructions for bosonic function approximation. The availability of code and data is a strength, as is the explicit treatment of sampling noise in the training objective. However, the headline claim of a sensing-time advantage currently rests on a conventional-sensor baseline that is restricted in a way that may unfairly favor the QCS protocols, and the MEG results involve a post-selection procedure that is not fully specified. These issues need to be resolved before the central claim can be accepted.

major comments (2)
  1. [Sec. III A, Appendix B 5, Figs. 2f, 3e, 3k]
  2. [Appendix C 4, Appendix C 5, Fig. 17]
minor comments (5)
  1. [Sec. III A]
  2. [Fig. 2f inset]
  3. [Appendix C 4]
  4. [Sec. III D]
  5. [Sec. IV D]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's analytic constructions and numerical claims are self-contained; the main caveat is a benchmark-fairness question, not a circular reduction.

full rationale

All load-bearing derivations are carried out within the paper. For the qubit QSP/QNN protocols (Sec. III A-C and App. B), the QCS circuits are trained on finite-shot samples with a separate test evaluation; the reported errors are Bayes/MLP test errors, not re-statements of the training loss. The analytic QS error expression in App. B 5 is derived from a Gaussian approximation to binomial Ramsey statistics and is then compared with independently trained QCS errors; it does not assume the QCS result. For bosonic function approximation (Sec. IV and App. D), the QCS estimator FQCS = (1/2g)(X + X*) with f = sum Wnm a†^n a^m gives E[FQCS] = F* by explicit construction, and the QS coefficients Cnm are likewise solved from the unbiasedness condition; the MSE comparison is a variance computation, not an equivalence. The hybrid sensor protocol (Sec. V) is trained with a fixed resource constraint (equal average photon number) and evaluated on held-out test data. The only self-citation with overlapping authors is the companion QCSA paper [1], used to introduce terminology and prior context; none of the paper's quantitative advantage claims depend on it. A legitimate benchmark-fairness caveat exists: the conventional QS baseline is taken as S = N independent single-period Ramsey shots (Sec. III A, Fig. 2f; Sec. III C, Figs. 3e and 3k), while QCS uses all N periods coherently in a single shot; the paper does not benchmark against a conventional Ramsey sensor that coherently integrates for the full N periods. That is a potential overstatement of the reported advantage, but it is a modeling/comparison choice rather than a circular reduction of a prediction to its inputs.

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

The central claims rest on standard unitary quantum mechanics, the QSP/QNN expressivity assumptions, the specific QND amplifier model from the literature, and the modeling of the MEG data as a noise-free signal. No new physical entities are introduced. The free parameters are the trainable circuit parameters and the task-specific scaling choices; these are fitted to data but the reported test-set performance is what supports the QCSA claim.

free parameters (5)
  • trainable QCS circuit angles and unitaries = learned per task
    For qubit-based QCS, the rotation angles (φ^(l), ζ^(l)) and two-qubit unitaries V^(l) are optimized by supervised learning on the training set (Sec. III).
  • Qnm coefficients for bosonic nonlinear amplifier = Spirals coefficients in Table VI; XOR: ±i
    For known tasks Qnm = Wnm analytically; for the Spirals task, Wnm is learned from the training dataset via supervised learning (Sec. IV, Appendix D).
  • hybrid ansatz parameters β^(d), R^(d) = learned, D=16 layers
    Conditional displacement amplitudes and qubit rotations in the bosonic QSP-inspired ansatz are trained via supervised learning (Sec. V).
  • MEG signal scaling / RMS phase θ_RMS = varied 0.0 to 0.3
    The strength of the imparted phase is scaled to set the operating regime; performance is reported as a function of this scaling (Appendix C).
  • hyperparameters L, S, T = L up to 2^6, T=10, S=1
    Number of processing layers, shots, and time bins define the resource budget N = L×S; chosen by the authors for each task (Sec. III, IV).
assumptions (5)
  • domain assumption Quantum mechanics: unitary evolution and projective measurement, with ideal noiseless operation.
    All QCS/QS protocols are modeled as unitary circuits followed by measurement; experimental imperfections are excluded (Sec. VI B).
  • standard math QSP/QNN expressivity: L interleaved sensing operations can realize polynomial functions of the sensed parameter.
    The expressivity of QSP and QNN ansatze is assumed; increasing L increases function complexity (Sec. III A, III C).
  • domain assumption Nonlinear amplifier solution b_out = b_in + g f for the QND Hamiltonian H_NL = -ig f(b - b†).
    Taken from Ref. [27] and used to derive E[X_QCS] = g<f> (Sec. IV C, Appendix D4).
  • domain assumption The MEG dataset is treated as a 'quantum noise-free' ground truth, scalable to arbitrary RMS phase.
    The recorded magnetoencephalography data are assumed to contain negligible classical noise and represent the true signal to be sensed (Appendix C2).
  • domain assumption Training uses exact gradients of the quantum circuit (analytic backpropagation) although the forward loss is evaluated on finite samples.
    The paper states gradient descent assumes access to the exact gradient; parameter-shift and noisy-gradient training is only conjectured (Sec. VI B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum computational sensing using quantum signal processing, quantum neural networks, and Hamiltonian engineering." pith.science (2026). https://pith.science/paper/D74GTFJO

@misc{pith2026250715845,
  author       = {Pith},
  title        = {Pith review of: Quantum computational sensing using quantum signal processing, quantum neural networks, and Hamiltonian engineering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D74GTFJO}},
  note         = {Machine review of arXiv:2507.15845}
}
abstract

Combining quantum sensing with quantum computing can lead to quantum computational sensors that are able to more efficiently extract task-specific information from physical signals than is possible otherwise. Early examples of quantum computational sensing (QCS) have largely focused on protocols where only a single sensing operation appears before measurement -- with an exception being the recent application of Grover's algorithm to signal detection. In this paper we present, in theory and numerical simulations, the application of two quantum algorithms -- quantum signal processing and quantum neural networks -- to various binary and multiclass machine-learning classification tasks in sensing. Here sensing operations are interleaved with computing operations, giving rise to nonlinear functions of the sensed signals. We have evaluated tasks based on static and time-varying signals, including spatiotemporal signals. Our approach to optimizing the circuit parameters in a QCS protocol takes into account quantum sampling noise and allows us to engineer protocols that can yield accurate results with as few as just a single measurement shot. In all cases, we have been able to show a regime of operation where a quantum computational sensor can achieve higher accuracy than a conventional quantum sensor, with a simulated accuracy advantage of $>$20 percentage points for some tasks. We also present protocols for performing nonlinear tasks using Hamiltonian-engineered bosonic systems and quantum signal processing with hybrid qubit-bosonic systems. Overall, we have shown that substantial quantum computational-sensing advantages can be obtained even if the quantum system is small, including few-qubit systems, systems comprising a single qubit and a single bosonic mode, and even just a single qubit alone -- raising the prospects for experimental proof-of-principle and practical realizations.

Figures

Figures reproduced from arXiv: 2507.15845 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9 [PITH_FULL_IMAGE:figures/full_fig_p029_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10 [PITH_FULL_IMAGE:figures/full_fig_p031_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11 [PITH_FULL_IMAGE:figures/full_fig_p032_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12 [PITH_FULL_IMAGE:figures/full_fig_p035_12.png]
Figure 13
Figure 13. Figure 13: FIG. 13 [PITH_FULL_IMAGE:figures/full_fig_p037_13.png]
Figure 14
Figure 14. Figure 14: FIG. 14 [PITH_FULL_IMAGE:figures/full_fig_p038_14.png]
Figure 15
Figure 15. Figure 15: FIG. 15 [PITH_FULL_IMAGE:figures/full_fig_p039_15.png]
Figure 16
Figure 16. Figure 16: FIG. 16 [PITH_FULL_IMAGE:figures/full_fig_p041_16.png]
Figure 17
Figure 17. Figure 17: FIG. 17 [PITH_FULL_IMAGE:figures/full_fig_p043_17.png]
Figure 18
Figure 18. Figure 18: FIG. 18 [PITH_FULL_IMAGE:figures/full_fig_p053_18.png]
Figure 19
Figure 19. Figure 19: FIG. 19 [PITH_FULL_IMAGE:figures/full_fig_p055_19.png]
Figure 20
Figure 20. Figure 20: FIG. 20 [PITH_FULL_IMAGE:figures/full_fig_p057_20.png]
Figure 21
Figure 21. Figure 21: FIG. 21 [PITH_FULL_IMAGE:figures/full_fig_p059_21.png]
Figure 22
Figure 22. Figure 22: FIG. 22 [PITH_FULL_IMAGE:figures/full_fig_p061_22.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Simple broadband signal detection at the fundamental limit

    quant-ph 2026-01 conditional novelty 6.0 of 10

    Broadband AC-field detection at the Grover-like limit can be achieved by a single analog experiment using a randomized SSH control Hamiltonian and a GHZ probe, with the lower bound derived from an integrated-quantum-F...

  2. Quantum Computational-Sensing Advantage

    quant-ph 2025-07 conditional novelty 4.0 of 10

    A perspective defines quantum computational sensing (QCS) and its advantage (QCSA), and organizes many recent sensing-plus-computing protocols into a single taxonomy.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · cited by 2 Pith papers

  1. [1]

    The dataset is a time- varying, 306-channel magnetoencephalography (MEG) signal, along with a 3-axis accelerometer signals

    Dataset and classification task The dataset we use for these simulations is obtained from an open-source experiment [21]. The dataset is a time- varying, 306-channel magnetoencephalography (MEG) signal, along with a 3-axis accelerometer signals. Each channel consists of data generated by a sensor. There are 3 sensors at 102 spatial locations on a headset....

    work page

  2. [2]

    quantum noise-free

    Dataset preprocessing We now discuss the procedure we implement to allow the dataset to be compatible with our simulations. The main goal of this procedure is to reduce the complexity of the simulation of the quantum system, so that they maybe simulated in a reasonable time on a classical computer. We first perform Principal Component Analysis on the data...

    work page

  3. [3]

    % 1 M · T T emporally-coherent QCS (T)

    Quantum Neural Network architecture for QCS TABLE II. Properties of protocols used for classification of spatiotemporal MEG data. The protocols are depicted in Fig. 16. Quantum System Trainable single-qubit Trainable multi-qubit Layers Number of Architecture operations R operations V L measurements Conventional QS (R) " % 1 M · T T emporally-coherent QCS ...

    work page

  4. [4]

    Straight-Through

    T raining details The simulations are performed with PyTorch [37], which allows us to use the conventional method of backpropa- gation to train the system. All matrices, representing the sensing and programmable unitaries, and the state of the system, are stored as PyTorch tensors (with a dimension set by the Hilbert space size of the system), allowing fo...

    work page 2000

  5. [5]

    Additional results Ramsey interferometry Temporally-coherent Spatially-coherent Spatiotemporally-coherent 0.0 0.1 0.2 0.3 Ramsey interferometry Temporally-coherent Spatially-coherent Spatiotemporally-coherent 40 30 20 0.0 0.1 0.2 0.3 50Classification error (%) Signal strength (average imparted phase ) Ramsey interferometry Temporally-coherent Spatially-co...

    work page

  6. [6]

    QCSA vs. task complexity for more general tasks To emphasize that the performance difference between QS and QCS protocols for increasing task difficulty is not special to the tasks considered in Appendix B 5, we now compare their performance for more general tasks. In particular, while we still consider our general binary classification tasks with R nonco...

    work page

  7. [7]

    Multi-variate sensing: comparing quantum computational sensors against multiple Ramsey interferometers In Sec. III C of the main text, for the conventional QS benchmark we consider a protocol for which Uprobe and Umeas are entangling operations and L = 0, namely with a structure similar to quantum sensor networks [2]. In particular, this quantum sensor is...

    work page

  8. [8]

    III C of the main text, we showed how binary multi-variable classification tasks can be performed using quantum computational sensors comprising two qubits

    QCSA for multi-variable binary discrimination tasks using a single qubit In Sec. III C of the main text, we showed how binary multi-variable classification tasks can be performed using quantum computational sensors comprising two qubits. However, in principle a single-shot measurement of even a single qubit can provide sufficient information for binary cl...

    work page

Show all 21 references
  1. [9]

    (1) of the main text, and consists of a pair of bosonic modes in each case

    Architecture: Umeas and quadrature measurements The architecture of both the QS and the QCS protocols is defined by our general Eq. (1) of the main text, and consists of a pair of bosonic modes in each case. We assume the initial state prior to either protocol to be |ψ0⟩ = |0⟩...

  2. [10]

    (D7) Recall that Wnm = W ∗ mn

    F unction approximation task: target polynomials F ⋆ and expected mean-squared-error as a metric For completeness, we rewrite the target polynomials F ⋆ considered in the main text, F ⋆ = DX n,m=0 Wnmα∗nαm. (D7) Recall that Wnm = W ∗ mn. The above general form can also be used...

  3. [11]

    Nonlinear function approximation using a linear phase-preserving amplifier a. Quantum dynamics of a linear phase-preserving amplifier We now describe the specific Umeas interaction that defines the conventional QS protocol for an all-bosonic sensor, which takes the form Umeas ...

  4. [12]

    Nonlinear function approximation using a nonlinear amplifier a. Quantum dynamics of a nonlinear amplifier The Umeas interaction that defines the QCS protocol for a bosonic nonlinear amplifier is given by Umeas = exp{−i ˆHNL}, (D28) where ˆHNL takes the form ˆHNL = −ig ˆf (ˆb −...

  5. [13]

    Writing the target polynomial in the form defined by Eq

    W orked example: XOR task For the XOR task we wish to construct the target function F ⋆: F ⋆ = −iα2 + iα∗2 (D40) In this subsection, we will work through an example of constructing an unbiased estimator of the above target function for both the conventional QS approach using a...

  6. [14]

    QS and QCS protocols: 1D polynomial approximation As we have considered a large number of 1D polynomial approximation tasks, listing all the required coefficients is slightly cumbersome; these have instead been provided in the form ofPython code, which can also be used to repr...

  7. [15]

    The classical postprocessing coefficients for a linear phase-preserving amplifier for the same task are included in Table VII

    QS and QCS protocols: Higher-order Spirals task In Table VI we provide the quantum nonlinear processing coefficients Qnm required for the Higher-order Spirals classification task in the main text. The classical postprocessing coefficients for a linear phase-preserving amplifie...

  8. [16]

    19, we present the structure of the hybrid qubit-bosonic mode quantum computational sensor analyzed in Sec

    Architecture for hybrid qubit-cavity quantum computational sensors In Fig. 19, we present the structure of the hybrid qubit-bosonic mode quantum computational sensor analyzed in Sec. V of the main text. Classical neural network QCS layers layers FIG. 18. Hybrid qubit-cavity qu...

  9. [17]

    bosonic modes) is to guard against numerical artefacts that may arise due to Hilbert space truncation effects

    Hilbert space truncation An important consideration specific to the training of quantum computational sensors comprising modes with infinite dimensional Hilbert spaces (e.g. bosonic modes) is to guard against numerical artefacts that may arise due to Hilbert space truncation e...

  10. [18]

    V of the main text to a simpler conven- tional QS protocol for displacement sensing, again starting with the Circles classification task depicted in Fig

    Comparing hybrid quantum computational sensors against phase-preserving amplifiers In this subsection, we compare the hybrid QCS protocols analyzed in Sec. V of the main text to a simpler conven- tional QS protocol for displacement sensing, again starting with the Circles clas...

  11. [19]

    Straight-through

    T rainable models, classical postprocessing, and loss functions a. Differentiable models and measurement results Differentiable models for all the trainable quantum computational sensors used in this work are written in PyTorch [37], and have been made publicly available (see ...

  12. [20]

    probabilities for qubit state measurement

    Performance of training using finitely-sampled measurements In this section we provide a comparison of training using finitely-sampled measurements against training using exact expectation values, e.g. probabilities for qubit state measurement. Of course, the performance of tr...

  13. [21]

    III A of the main text, which have been trained using end-to-end training and cross-entropy-loss minimization, Eq

    Number of measurement samples for training versus inference Our use of finitely-sampled measurement results for training raises a question that does not present itself when training using exact probabilities: how many samples should be used to calculate measurement results ¯XQ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.