{"id":"853fc723-0b33-4e8d-a0c7-30114256d727","arxiv_id":"2501.12335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A quantum compressive sensing pipeline with imaginary time evolution runs on Amazon Braket and reconstructs 5-pixel LIDAR signals, but only when noise is near 1e-4 or below.","lead":"Researchers deployed a quantum compressive sensing algorithm on Amazon's Braket cloud platform to see how quantum noise affects signal reconstruction from sparse LIDAR data. The study is a small proof of concept that shows the pipeline runs on near-term quantum resources, with noise below about 1e-4 preserving the trained quantum state.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how the QITE Hamiltonian H is built from sensing matrix A and outcome x; the only displayed H cannot encode general real-valued measurements, so the sRMSE results are not tied to QCS projection.","rationale":"The reader's weakest_assumption is correct and, in my reading, is the one place the central claim can break. Figures 5(a)–(e) validate QITE convergence only for a hand-picked H; Figures 5(f)–(j) report sRMSE versus Nc but never connect that H to the sensing matrix A or measurement result x. The angle encoding in Eq. (1) means a measurement outcome is a continuous pixel value, and the corresponding projection operator is v-dependent; a fixed H = −(ZIIII + IIZII) is not a projection onto that outcome. This is not a stylistic gap: without the mapping, the experiment is not a test of QCS. I agree with the reader's conditional verdict; the condition is precisely that this mapping must be supplied and validated. I would not escalate to reject because a correct x-dependent diagonal construction exists at least for row-selection sensing matrices (H = Σ_i (v_i−1/2)Z_i), so the paper may be salvageable. The proposed fidelity check settles whether the implemented QITE actually projects onto |Ψx⟩. Other concerns (no classical baseline, no error bars, shot-noise iteration discarding) are real but secondary; they affect interpretation rather than the validity of the projection step.","tokens_in":15355,"tokens_out":12394,"duration_ms":129322,"concrete_test":"Obtain from the authors the explicit (A,x) → H mapping used for Figs. 5(f)–(j). Then, for one held-out testing sample, compute the ideal projected state |Ψx⟩ ∝ e^{−β Σ_j (Aŷ − x)_j²} |Ψ⟩ on a statevector simulator, run the paper's QITE circuit with the supplied H, and compute the fidelity between the QITE output and |Ψx⟩. If fidelity is not close to 1 (e.g., below 0.95), or if no mapping can be provided, the sRMSE curves do not demonstrate QCS projection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on QITE actually implementing the QCS projection step (§3.1.3, Eqs. 5–6). Yet the paper never gives the map (A,x) → H. The only concrete Hamiltonian exhibited, H = −(ZIIII + IIZII) (Fig. 5a), is a fixed operator with no dependence on the measurement outcome x; it cannot represent the v-dependent diagonal projector (ŷ−v)² = (v−1/2)Z + const. for a real-valued pixel encoded by Eq. (1) unless v = 0. No derivation from A and x is provided for the sRMSE runs in Figs. 5(f)–(j), and no code or data is shipped. If the reported curves used such x-independent H's, they may simply show QITE relaxing to a fixed trivial subspace rather than projecting onto the measurement-constrained |Ψx⟩; if an x-dependent construction exists, it is absent from the manuscript. Either way, the sRMSE results cannot verify the claim that QCS is practically deployable with full QITE under noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to implement the Quantum Compressive Sensing (QCS) architecture of Ref. [52] with Quantum Imaginary Time Evolution (QITE) as the projection method, using 5-pixel LIDAR-derived data mapped via Eq. (1). It reports fidelity benchmarks for training-subset quantum averages (Fig. 4a), fidelity degradation under bit-flip, depolarizing, and dephasing noise with a suggested noise threshold of about 1e-4 (Fig. 4b), sRMSE-versus-Nc curves for QCS reconstruction under noiseless and noisy QITE (Fig. 5), and a separate QITE shot-noise study on a fixed 3-qubit Hamiltonian (Fig. 6). The central assertion is that QCS can be practically deployed on Amazon Braket with full QITE under realistic noise, and that increasing the number of classically measured pixels decreases reconstruction error. However, the manuscript does not specify how the QITE Hamiltonian is constructed from the sensing matrix A and measurement outcome x, which leaves the central reconstruction claim uncheckable as written.","tokens_in":15480,"tokens_out":7928,"duration_ms":84111,"significance":"If the central claim could be verified, this would be the first demonstration that the projection stage of QCS can be implemented with full QITE on a cloud quantum platform, and the fidelity/noise-threshold data in Fig. 4 would be useful engineering input for small-scale QCS experiments. The paper is honest about the small dataset and limited sparsity, and the shot-noise mitigation experiments of Fig. 6 are a concrete empirical contribution on their own. That said, the paper ships no code or data, and the missing (A, x) -> H construction means the sRMSE curves cannot currently be attributed to the QCS projection objective. The hardware claim is also stronger than what the simulations actually demonstrate.","major_comments":[{"comment":"The central projection step is not specified. Section 3.1.3 describes QITE only through the general tomography and Trotter formulas of Eqs. (5)-(6); the manuscript nowhere gives the map from the sensing matrix A and measurement outcome x to the Hamiltonian H used in the sRMSE experiments of Figs. 5(f)-5(j). The only explicit realization, H = -(ZIIII + IIZII) in Fig. 5(a), is a fixed operator: it has no dependence on which pixels were classically measured, no dependence on the measured values x, and no dependence on the pixel-qubit midpoint v of Eq. (1). Such an operator at best implements one particular projection (favoring the j1> state on two fixed qubits), not the general measurement-consistent projection |Psi_x> for arbitrary pixel sets and outcomes. Consequently the reported decrease of sRMSE with Nc cannot be verified as an effect of QCS projection; it could arise from QITE relaxing to a trivial fixed subspace. The revised manuscript must provide the explicit construction (A, x) -> H, including how Nc, the measured pixel indices, and the measured values enter the Hamiltonian coefficients, together with the tomography and Trotter implementation details.","section":"Section 3.1.3 (Eqs. (5)-(6)), Fig. 5(a), (f)-(j)"},{"comment":"The shot-noise study is not connected to the QCS pipeline. Figure 6 reports QITE convergence only for a fixed 3-qubit Hamiltonian H = -ZII - IIZ, varying shots per observable and d_beta; it does not report sRMSE, reconstruction error, or any metric tied to the QCS setup of Section 3.1.3. The statement that a check for highly perturbative shot-noise instances with a medium d_beta 'significantly mitigates' shot noise is therefore demonstrated only for a toy Hamiltonian, not for QCS. In addition, Section 3.2.4 says that shot noise 'will naturally exist' when implementing on quantum cloud resources, but the experiments in this paper are local statevector and density-matrix simulations; if finite-shot statistics were simulated, the exact simulation recipe should be described.","section":"Section 4.3, Fig. 6"},{"comment":"The 'practical implementation on Amazon Braket' claim exceeds what the experiments demonstrate. The Introduction states that QITE is implemented 'on Amazon's quantum cloud computing services' and the abstract announces a 'practical implementation of QCS on Amazon Braket,' but Section 4.2 says that all Born machine circuits are generated using 'local, statevector simulations' via the Braket SDK, and Section 3.2.3 explicitly limits the noisy QITE results to 'density matrix simulations.' No QPU, device, or execution details are reported anywhere. The paper should either present hardware execution results or rephrase the contribution as a simulator-based implementation, because the current wording overstates the evidence for deployment on actual quantum computing resources.","section":"Introduction, Section 4.2, Section 3.2.3"}],"minor_comments":[{"comment":"Equation (11) uses K0 in both terms of the amplitude-damping channel; one of them should be K1 as defined in Eq. (13).","section":"Eq. (11)"},{"comment":"The caption for Fig. 5 is internally confusing: it says '(g)-(j) Same as (f) except for single-gate bit-flip noise...' and then '(h)-(j) being identical demonstrations for depolarizing, dephasing, and amplitude damping noise.' As written, the assignment of panels to noise types is contradictory and should be clarified.","section":"Fig. 5 caption"},{"comment":"The sRMSE formula appears to contain an unresolved LaTeX control sequence ('vuut'), and the printed formula lacks an explicit square-root symbol; the intended expression should be a standard sqrt( (1/nu) * sum_i ((P_i - R_i)/sigma_i)^2 ).","section":"Eq. (14)"},{"comment":"Several superscripts are rendered with spaces rather than exponents (e.g., '2 15', '10 -4', '2 8'), which makes the text hard to read; these should be typeset as proper superscripts.","section":"General formatting"},{"comment":"In the data-encoding example, the midpoint parameter is called 'p', whereas Eq. (1) and the surrounding text use 'v'; the notation should be unified.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on quantum compressive sensing or QITE. The paper does something concrete: it implements the full QITE projection from Motta et al. inside the QCS framework from their earlier paper, runs it on Amazon Braket (mostly via simulators, with noise models), and measures sRMSE on a small LIDAR-derived dataset. That integration is new, and the noise sensitivity results—fidelity drops fast once per-gate noise passes roughly 1e-4, and more classically measured pixels reduce sRMSE—are plausible and useful for anyone planning a real QCS experiment.\n\nThe authors are honest about limitations: they flag the small dataset and low sparsity, and they openly describe the shot-noise heuristic of discarding energy-increasing QITE iterations. That last point is a double-edged sword—it's transparent, but it means the convergence plots are somewhat cherry-picked.\n\nThe soft spot that matters is the projection step. The paper never specifies how the QITE Hamiltonian H is built from the sensing matrix A and measurement outcome x. The only concrete H displayed, -(ZIIII + IIZII), is fixed and independent of x, so it cannot encode a general real-valued measurement constraint. If the sRMSE curves in Figure 5(f)-(j) use that same H, they are showing QITE relaxing to a fixed ground state, not projecting onto the x-dependent state. If they use a different, x-dependent construction, it is absent from the manuscript. Either way, the central claim cannot be checked. This is the one load-bearing gap; everything else is a scaling or reporting issue.\n\nAlso missing: no classical baseline (e.g., l1 minimization) to tell you whether the sRMSE numbers are any good, no error bars, and no shipped code or data. For a five-pixel, 64-sample demo, those additions would be cheap and would raise confidence substantially.\n\nMy read: it's a modest but legitimate proof-of-concept with one serious under-specification. The fix is doable—write down the (A,x) to H map, ideally with a circuit for a small example, and release the code. I would send it to peer review and ask for that construction and a classical baseline before acceptance.","headline":"A genuinely new empirical study of QITE-based QCS under noise, but the missing (A,x) to H construction keeps the main results from being checkable.","tokens_in":16122,"tokens_out":3207,"would_cite":false,"duration_ms":32556,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum compressive sensing can be run on noisy cloud quantum hardware, with QITE projection and per-gate noise near 1e-4 or lower.","keywords":["quantum compressive sensing","QITE","quantum noise","Born machine","LIDAR","Amazon Braket","signal reconstruction","tensor networks"],"falsifier":"Run the QCS plus QITE pipeline on a signal larger than five pixels with an arbitrary sensing matrix $A$ and a stated general rule for building the projection Hamiltonian from $A$ and the measurement outcome $x$. If the mean sRMSE stops decreasing as $N_c$ increases, or if no such general Hamiltonian construction can be given, the paper's central practicability claim fails.","tokens_in":15038,"feed_emoji":"⚛️","tokens_out":9156,"duration_ms":85535,"temperature":0.7,"pith_summary":"This paper claims that the Quantum Compressive Sensing (QCS) protocol can be put into practice on a noisy cloud quantum computer rather than only in classical simulation. The authors implement the full Quantum Imaginary Time Evolution projection on a five-pixel LIDAR waveform dataset, and report that under bit-flip, depolarizing, dephasing, and amplitude-damping noise, the mean scaled root-mean-square reconstruction error decreases as more pixels are measured classically. They also find that the pre-trained quantum average is preserved only when per-gate noise is around 1e-4 or lower, and that shot noise in QITE's tomography step can be mitigated by tuning shot counts and imaginary-time step sizes. If these results hold, they indicate that QCS is deployable on near-term quantum hardware without full error correction.","feed_headline":"Quantum compressive sensing survives noise on Amazon Braket","feed_subtitle":"QITE projection keeps signal reconstruction accurate when per-gate noise is near 1e-4; more pixels reduce error.","key_machinery":"The mechanism is the Born machine quantum average combined with QITE projection. The quantum average $|\\Psi\\rangle$ is a superposition of training samples prepared through a circuit with a control register; its squared amplitudes give the probability of each signal. QITE applies the imaginary-time evolution $e^{-\\beta \\hat{H}}$ to this state, exponentially suppressing basis states whose expectation values disagree with the measured outcome, and implements this non-unitary operation unitarily through quantum state tomography followed by Trotterized imaginary-time steps. A second ingredient is the pixel-qubit angle encoding, which maps each real pixel value to a single-qubit rotation so that the entire signal state remains an efficiently prepared product state.","core_discovery":"The central discovery is that the QCS architecture, which had previously been studied only through classical simulation, can be implemented end-to-end on a real cloud quantum platform using the full QITE projection algorithm. Concretely, the paper shows that for a five-pixel LIDAR waveform dataset, a Born machine trained on 256 samples can be projected onto measurement-constrained states via QITE, and that under bit-flip, depolarizing, dephasing, and amplitude-damping noise, the mean per-pixel sRMSE of the reconstructed signal decreases as the number of classically measured pixels increases. The paper also reports that the trained quantum-average state is only preserved when the per-gate noise probability is on the order of 1e-4 or smaller, and that shot noise in QITE tomography can be controlled by increasing the number of shots per observable and by discarding failed imaginary-time iterations. The authors present this as evidence that QCS is a practical quantum data-driven approach for compressive sensing in the noisy intermediate-scale era.","pith_inferences":["A general construction of the projection Hamiltonian from the sensing matrix $A$ and measurement outcome $x$ is the missing step; if one exists, QITE-based QCS would extend to arbitrary sensing matrices rather than only the hand-picked operator shown in the paper.","Because the paper only tests five-pixel signals, the observed monotone decrease in error with $N_c$ is a trend, not a scaling law; larger sparse datasets could make the benefit per measured pixel larger or smaller.","The encoding results hint that tuning pixel midpoints or adding one qubit per ambiguous pixel could remove a large share of reconstruction errors without any additional measurements.","QITE's tomography cost per step will dominate on larger images, so combining QITE with reduced-tomography methods that use fewer observables per step is a natural next test."],"forward_implications":["Near-term quantum imaging pipelines can use QITE projection without full error correction when per-gate noise is at or below about 1e-4.","Classically measuring more pixels provides a practical knob to reduce reconstruction error under noise, as shown by the decreasing sRMSE with increasing $N_c$.","A 256-sample quantum average is sufficient to approximate the global average for five-pixel LIDAR waveforms, keeping the training circuit small.","Shot noise in QITE tomography can be managed by tuning shots-per-observable and the imaginary-time step size $d\\beta$, making the projection robust enough for current hardware."],"supporting_citations":[{"why":"Original QCS proposal: training via quantum average, pixel-qubit mapping, and the three projection ideas, including QITE, that this work implements.","marker":"[52]"},{"why":"Full QITE algorithm (tomography, Trotterized imaginary-time evolution) whose deployment is the paper's main practical contribution.","marker":"[53]"},{"why":"Cloud quantum platform and its simulator interface used for all noiseless and noisy experiments.","marker":"[54]"},{"why":"Tensor-network compressed sensing with the quantum-average construction that QCS adapts to qubits.","marker":"[51]"},{"why":"Angle-encoding of pixels into single-qubit rotations that QCS generalizes to arbitrary pixel midpoints.","marker":"[49]"},{"why":"Born machine formalism, the complex-amplitude probabilistic model underlying the trained quantum state.","marker":"[47]"}],"fun_headline_variants":["Quantum compressive sensing runs on Braket under noise","QITE tames noise in quantum compressive sensing on Braket","Practical quantum compressive sensing: Braket implementation","Quantum compressive sensing survives real quantum noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the assumption that every measurement constraint can be encoded as a simple few-qubit energy operator whose ground state is the desired projected signal; the paper shows this for one hand-picked operator and does not give the general construction.","fun_headline_variants_meta":{"raw":{"variants":["Quantum compressive sensing runs on Braket under noise","QITE tames noise in quantum compressive sensing on Braket","Practical quantum compressive sensing: Braket implementation","Quantum compressive sensing survives real quantum noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1384,"prompt_tokens":927,"completion_tokens":457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":398}},"tokens_in":543,"tokens_out":457,"duration_ms":5014,"temperature":1.0,"reasoning_tokens":398,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:16:24.244511+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the QCS plus QITE pipeline on a signal larger than five pixels with an arbitrary sensing matrix $A$ and a stated general rule for building the projection Hamiltonian from $A$ and the measurement outcome $x$. If the mean sRMSE stops decreasing as $N_c$ increases, or if no such general Hamiltonian construction can be given, the paper's central practicability claim fails.","supporting_citations":[{"cited_title":"Quantum compressive sensing: Mathematical machinery, quantum algorithms, and quantum circuitry,","cited_arxiv_id":null,"evidence_quote":"Original QCS proposal: training via quantum average, pixel-qubit mapping, and the three projection ideas, including QITE, that this work implements."},{"cited_title":"Amazon Braket,","cited_arxiv_id":null,"evidence_quote":"Cloud quantum platform and its simulator interface used for all noiseless and noisy experiments."},{"cited_title":"Tensor network compressed sensing with unsupervised machine learning,","cited_arxiv_id":null,"evidence_quote":"Tensor-network compressed sensing with the quantum-average construction that QCS adapts to qubits."},{"cited_title":"Information Perspective to Probabilistic Modeling: Boltz- mann Machines versus Born Machines,","cited_arxiv_id":null,"evidence_quote":"Born machine formalism, the complex-amplitude probabilistic model underlying the trained quantum state."}],"review_version":1}