{"id":"1c425b26-d71a-499d-8af8-7f7f93b674b1","arxiv_id":"2504.16297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Pre-sampling error sets before state evolution and batching shot collection makes noisy quantum trajectory simulation up to 10^6x faster at generating data, demonstrated on 35-qubit and 85-qubit error-correction circuits.","lead":"This paper introduces Pre-Trajectory Sampling with Batched Execution, a way to simulate noisy quantum circuits by deciding the error patterns in advance and then measuring many shots from each prepared state. It reports large speedups in data collection: up to a millionfold for a 35-qubit statevector simulation and 16-fold for an 85-qubit tensor network simulation, producing trillion-shot datasets for training AI decoders.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PTSBE's Algorithm 2 uses fixed per-Kraus probabilities, which is only exact for unitary-mixture channels; the paper neither states this restriction nor validates the generated data against a standard reference, so the massive datasets may be systematically biased.","rationale":"The reader's weakest_assumption identified exactly the same load-bearing gap: Algorithm 2 samples Kraus operators with fixed probabilities, which is only valid for unitary-mixture channels, while the paper presents PTSBE as a general method for simulating noisy quantum circuits. I agree with the conditional verdict because the speedup mechanism (batching many shots per prepared state) is real and may well survive even if the sampling distribution needs correction or a stated restriction. However, the data-generation claim, especially for training ML decoders, depends on the generated shot distribution faithfully representing the noisy circuit. The paper provides no statistical validation against standard trajectory simulation or exact density-matrix evolution, and no explicit statement limiting Algorithm 2 to unitary-mixture channels. This is a correctness risk, not just a consensus disagreement: the internal description in Sec. 2.2 acknowledges state-dependent Kraus probabilities, while Algorithm 2 uses a state-independent pre-sampling step. The proposed concrete test would settle the issue directly by comparing distributions on a non-unitary channel. If the test confirms bias, the paper should be revised to either restrict the method or implement state-aware pre-sampling; if it shows agreement for the tested noise models, the concern is resolved. Since the evidence is currently absent, the conditional verdict remains appropriate.","tokens_in":12895,"tokens_out":5384,"duration_ms":60985,"concrete_test":"Run a small (e.g., 2-4 qubit) version of the MSD circuit or a random circuit under a non-unitary channel such as amplitude damping (Kraus operators not proportional to unitaries) with the same CUDA-Q PTSBE code. Generate e.g. 10^5 shots with PTSBE and with standard trajectory simulation (or exact density-matrix evolution for small n). Compute the total variation distance between the resulting empirical shot distributions and compare to a depolarizing/unitary-mixture channel as control. If the non-unitary case shows TV distance significantly above the control (e.g., >0.01), Algorithm 2 is biased for general CPTP channels and the paper must either restrict claims to unitary-mixture channels or replace the sampler with a state-dependent one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PTSBE generates noisy quantum datasets suitable for training QEC decoders at orders-of-magnitude lower cost. The load-bearing assumption is that Algorithm 2's pre-sampled Kraus sets, followed by many shots per set, reproduce the distribution a standard trajectory simulation would produce. This is only true for unitary-mixture channels (K_i ∝ U_i), where outcome probabilities are state-independent. For a general CPTP channel, the probability of outcome i is <ψ|K_i^†K_i|ψ> (Sec. 2.2), which depends on the state at the error location, so it cannot be drawn in advance from fixed {p}. Algorithm 2 samples each Kraus operator with fixed probability p and never conditions on the evolving state; the manuscript also never states a unitary-mixture restriction and reports no comparison between PTSBE output and a standard trajectory or density-matrix reference. Without such validation, the one-trillion-shot and one-million-shot datasets may carry a systematic bias that no amount of shot batching removes; for decoder training this is exactly the kind of error that can silently degrade model accuracy. The speedup numbers themselves may survive (batching is real), but the data-validity half of the central claim is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PTSBE (Pre-Trajectory Sampling with Batched Execution), a workflow that separates the stochastic choice of Kraus errors from state preparation and measurement. A sampling algorithm pre-selects sets of Kraus operators, and the batched-execution stage prepares each resulting state once and collects many measurement shots from it. The authors implement PTSBE in CUDA-Q on statevector and tensor-network backends, report roughly 10^6x and 16x improvements in shot-collection throughput on 35-qubit and 85-qubit magic-state-distillation circuits, and describe generating one trillion and one million shots, respectively. They motivate the method as a way to produce the large noisy-quantum datasets needed for training ML-based QEC decoders.","tokens_in":13200,"tokens_out":8151,"duration_ms":82094,"significance":"If the data produced by PTSBE are statistically equivalent to data from standard trajectory simulation, this is a valuable engineering contribution: batching shots avoids redundant state preparation, and the lightweight error metadata provides a useful supervised-learning label structure that is not available from physical devices. The raw throughput measurements in Figs. 4 and 5 are credible because repeated shot sampling from a prepared state is genuinely cheaper than re-preparing that state, and the scale of the demonstrated datasets is impressive. However, the statistical-equivalence claim is the load-bearing part of the paper, and it is not established. The paper's own Section 2.2 distinguishes state-independent unitary-mixture channels from general Kraus channels, yet Algorithm 2 samples with fixed probabilities in a way that does not reproduce either a general CPTP channel or even a unitary-mixture channel's joint outcome distribution. The paper also reports no comparison of PTSBE output against a standard trajectory or density-matrix reference. For a method whose stated purpose is training data generation, that validation is essential.","major_comments":[{"comment":"The pre-sampling loop treats each Kraus operator as an independent Bernoulli random variable, whereas the trajectory sampling in Algorithm 1 draws one normalized Kraus outcome per noise channel. For a single unitary-mixture channel with outcomes {sqrt(p_i) U_i}, Algorithm 2 can select no operator, one operator, or several operators; when several are selected, the 'compatible' check discards later operators in iteration order, introducing an order-dependent bias. The resulting error-sequence distribution is not the channel distribution, and the effective no-error probability becomes the product of (1 - p_i) over the listed operators rather than the channel's no-error weight. The authors must either replace this step with one categorical draw per channel, or prove and validate an equivalence that the manuscript currently does not provide.","section":"Sec. 3.1, Algorithm 2"},{"comment":"The paper itself states that for general Kraus channels the outcome probabilities are state-dependent, given by <psi|K_i^dagger K_i|psi>, and CUDA-Q computes them at runtime (Algorithm 1, lines 8-11). PTSBE's pre-sampling from fixed probabilities {p} can be exact only for unitary-mixture channels (K_i = sqrt(p_i) U_i). The manuscript never states this restriction; instead, Sec. 1 and Fig. 1 present the method as applicable to arbitrary noisy circuits and universal simulation. The authors should state the restriction explicitly and either limit all claims to unitary-mixture (e.g., Pauli) noise models or provide a correct procedure for general channels with controlled error.","section":"Sec. 2.2 and Sec. 3.1"},{"comment":"Figures 4 and 5 report only shots-per-second and the fraction of unique shots; no experiment compares the PTSBE output distribution with a standard trajectory or density-matrix simulation. For a method whose stated purpose is to generate training data, distributional equivalence is the relevant benchmark. I recommend adding a quantitative comparison, such as bitstring histograms, low-order statistical moments, or decoder accuracy when trained on PTSBE data versus standard-trajectory data for the same circuit. Without this, the suitability of the trillion-shot dataset for decoder training is unverified.","section":"Sec. 4, Figs. 4 and 5"}],"minor_comments":[{"comment":"The complexity statement for Algorithm 2, 'just ~O(|{K}|^2(p)^2) operations', appears meaningless as written; the sampling loop is O(nsamples * |{K}|). Please correct or define the expression.","section":"Sec. 3.1"},{"comment":"The proportional-sampling variant defines p'_alpha = p_alpha / sum_i p_i; if p_alpha is already a normalized joint probability, this division is incorrect, and the summation index is undefined. Please clarify the intended normalization.","section":"Sec. 3.1"},{"comment":"The word 'providence' should be 'provenance' when referring to error metadata, as in 'error providence' and the caption of Fig. 1.","section":"Sec. 2.3 and Sec. 3"},{"comment":"In the text describing Fig. 4, 'samples of 10^6 total shots are comprised of more than a 0.5 fraction of unique results' should be reworded to 'more than half' or 'a fraction greater than 0.5'.","section":"Sec. 4"},{"comment":"'deliterious effects' is a typo for 'deleterious effects'.","section":"Sec. 2.3"}],"recommendation":"major_revision","confidential_remarks":"The throughput result is likely solid, but the statistical-correctness gap is central and cannot be waved away with more benchmarks. The good news is that the fix appears local: for unitary-mixture channels, replace the independent per-operator trials with a single categorical draw per channel, state the restriction, and add a distributional validation against a standard trajectory reference. If the authors can do that, the paper could be acceptable; as written, the main application claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe PTSBE paper has a real, useful speedup at its core, but the statistical correctness argument has a load-bearing gap. The paper never limits its sampling method to unitary-mixture channels, and without that restriction the algorithm is not sampling the right distribution.\n\nWhat's actually new: pre-committing to a Kraus operator set before state preparation, then drawing many shots from the fixed prepared state. That eliminates redundant state preparation and gives error-provenance metadata as a byproduct. The 10^6x and 16x speedups are credible as batching effects – prepare once, sample 10^6 times. The engineering looks solid, and the datasets (10^12 and 10^6 shots) are real. For Pauli or other unitary-mixture noise, this is a legitimate way to generate large training corpora for ML decoders.\n\nThe problem is in Algorithm 2. It samples each Kraus operator with a fixed probability p, independent of the state. For a general CPTP channel, the correct Kraus outcome probability is <psi|K_i^dagger K_i|psi>, which depends on the state at that point. The paper itself states that formula in Sec. 2.2, but then presents Algorithm 2 as generic without noting that it only works when those probabilities are state-independent, i.e., for unitary-mixture channels. There is also no comparison against standard trajectory or density-matrix output. Since the goal is to train decoders, a biased training distribution is exactly the failure mode that matters.\n\nMinor: the batched-shot trick breaks with mid-circuit measurements and feed-forward, and the paper doesn't mention that limitation.\n\nMy take: the speedup is real, but the data-validity half of the claim is unsubstantiated. The authors can fix this by stating the unitary-mixture restriction (and confirming their experiments use such noise) or by giving a state-dependent pre-sampling algorithm, and by validating on a small example against exact density-matrix simulation. I'd send this to peer review – the idea is important and the flaw is fixable – but I would not recommend acceptance until the validation is in.\n\nIt's a good paper for a reading group; there's a clean conceptual issue worth dissecting. I wouldn't cite it in its current form.","headline":"Batched-shot speedup is real, but PTSBE's pre-sampling is only exact for unitary-mixture noise, and the paper neither says so nor validates it.","tokens_in":13682,"tokens_out":5899,"would_cite":false,"duration_ms":57913,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a","03.67.Pp"],"model":"deepseek-v4-flash","headline":"A new two-phase method, Pre-Trajectory Sampling with Batched Execution, claims to turn quantum trajectory simulation into a programmable noisy-data factory with speedups up to a millionfold.","keywords":["quantum trajectory simulation","pre-trajectory sampling","batched execution","noisy quantum circuits","quantum error correction","statevector simulation","tensor networks","machine learning decoders"],"falsifier":"Take a small circuit whose noise includes amplitude damping, where the Kraus probabilities depend on the state, and compare the empirical measurement distribution produced by the standard trajectory algorithm against the distribution produced by PTSBE with the same total number of shots. If the two distributions differ beyond statistical error, the central equivalence behind the method is false.","tokens_in":12696,"feed_emoji":"⚛️","tokens_out":9066,"duration_ms":84362,"temperature":0.7,"pith_summary":"Quantum trajectory simulation approximates a noisy quantum computer by following many separate pure-state histories, each subjected to randomly sampled errors, instead of evolving an exponentially large density matrix. The paper's central claim is that this process wastes almost all of its effort: standard implementations resample errors and re-prepare the quantum state for every measurement shot, while the Pre-Trajectory Sampling with Batched Execution (PTSBE) method fixes an error set first, prepares the state once, and then draws a very large number of shots from that single state. The authors report that this turns trajectory simulation into a programmable data-collection engine, producing a 35-qubit quantum error-correction dataset of one trillion shots with about a millionfold speedup over conventional trajectory methods, and a preliminary 85-qubit tensor-network run with a sixteenfold speedup. The payoff would be large labeled noisy datasets for training machine-learning decoders, at a scale previously limited to Clifford-only simulators.","feed_headline":"Pre-sampling makes noisy quantum data collection a million times faster","feed_subtitle":"Batching many shots per fixed error set turns trajectory simulation into a factory for error-decoder training data.","key_machinery":"The central mechanism is a two-phase decomposition of a noisy trajectory computation. In the Pre-Trajectory Sampling phase, a user-selected algorithm produces a list of Kraus operator sets $\\{K_{\\alpha_0},\\dots,K_{\\alpha_i}\\}$—Kraus operators being the matrices describing the individual errors a noise channel can apply—together with a shot count $m_\\alpha$ for each set. In the Batched Execution phase, the simulator prepares the pure state resulting from the coherent gates and the fixed error set once, then samples all $m_\\alpha$ measurement outcomes from that state without re-preparation or re-sampling. Algorithm 2 is the canonical sampler: for each noise location it includes a Kraus operator with probability $p$, discards physically incompatible or duplicate combinations, and assigns a large uniform shot budget to each retained set, making the trajectory ensemble a programmable sampling distribution.","core_discovery":"The paper's discovery is that the expensive part of trajectory simulation—state preparation—can be made to happen only once per carefully chosen error set rather than once per shot, and that the resulting efficiency gain grows almost linearly with the number of shots collected per batch. For unitary-mixture noise channels such as depolarizing channels, the probabilities of Kraus operators are independent of the quantum state, so replaying a fixed error set for many shots is exact. For general, state-dependent Kraus channels, the paper presents the same procedure without proving that one prepared state followed by many shots reproduces the original trajectory ensemble; it reports the speedups and dataset sizes as the main evidence. Concretely, it claims a mature 35-qubit statevector simulation of the 5-to-1 magic state distillation protocol with up to a millionfold speedup and one trillion collected shots, plus a preliminary 85-qubit tensor-network simulation with a sixteenfold speedup and one million collected shots.","pith_inferences":["An immediate test the authors do not run: compare PTSBE's output distribution to exact density-matrix evolution on a few qubits with amplitude damping and other state-dependent channels; the present evidence for exactness is limited to unitary-mixture noise, and no statistical validation is reported.","The same batching idea could be transferred to approximate density-matrix or matrix-product-state simulators, where state preparation is also expensive relative to sampling, though the equivalence question would need revisiting per formalism.","Shot budgets could be made adaptive: for highly mixed or high-entropy states almost every shot is unique, but for low-entropy states duplicate outcomes dominate; a data-quality-aware scheduler would allocate shots according to measured uniqueness.","Decoder training on pre-sampled data may inadvertently train on the simulator's noise model rather than hardware noise; comparing decoder performance on experimental magic-state data against PTSBE-trained decoders would test the transferability of the generated labels."],"forward_implications":["Large datasets from universal (non-Clifford) noisy simulators become practical, filling a gap that Clifford-only stabilizer simulators cannot cover.","Error provenance tags attached to each trajectory turn simulated shots into supervised training data for machine-learning decoders, with known error labels unavailable from hardware.","Users can target sampling toward rare or specific error combinations—probability bands, gate types, spatial correlations—for fault-tolerance studies instead of accepting black-box sampling.","Because the fixed-error-set state is prepared once and then sampled many times, the marginal cost per shot becomes nearly linear in shot count until the state's distinct measurement outcomes are exhausted.","The method parallelizes naturally both within a state preparation and across independent trajectories, so speedups from the batch mechanism multiply with GPU scaling."],"supporting_citations":[{"why":"Supplies the GPU-accelerated statevector and tensor-network simulator on which PTSBE is implemented and benchmarked.","marker":"[22]"},{"why":"Defines the Clifford-only simulation baseline that PTSBE is designed to supplement for universal noisy circuits.","marker":"[29]"},{"why":"Provides the 5-to-1 magic state distillation protocol whose compiled circuits are simulated in the scaling experiments.","marker":"[10]"},{"why":"Supplies the recently demonstrated logical magic state distillation circuits, including the [[7,1,3]] and [[17,1,5]] color codes, used in the benchmarks.","marker":"[43]"},{"why":"Motivates the data-generation goal by showing that learned decoders need large simulated noisy datasets.","marker":"[7]"},{"why":"Foundational reference for quantum trajectory methods, the statistical technique PTSBE modifies.","marker":"[20]"},{"why":"Foundational Monte Carlo wave-function method whose per-shot error resampling defines the baseline that PTSBE accelerates.","marker":"[21]"},{"why":"Gives the CPTP/Kraus formalism and the state-dependent probability rule that underlies the paper's weakest equivalence assumption.","marker":"[11]"}],"fun_headline_variants":["Batch error sets to turn one state prep into a trillion noisy shots","Pre-trajectory sampling: replaying error sets yields 10^6x speedup","One prepared state, many shots: trajectory sims become shot factories","Reuse error sets to collect a trillion shots from 35-qubit simulations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that drawing many measurement shots from one fixed, pre-selected set of errors gives the same noisy measurement distribution as drawing a fresh set of errors for every shot; for noise channels whose Kraus probabilities depend on the state at each error location, the paper does not establish this equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Batch error sets to turn one state prep into a trillion noisy shots","Pre-trajectory sampling: replaying error sets yields 10^6x speedup","One prepared state, many shots: trajectory sims become shot factories","Reuse error sets to collect a trillion shots from 35-qubit simulations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000863,"raw_usage":{"total_tokens":3738,"prompt_tokens":935,"completion_tokens":2803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2721}},"tokens_in":551,"tokens_out":2803,"duration_ms":18907,"temperature":1.0,"reasoning_tokens":2721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:07:07.655239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small circuit whose noise includes amplitude damping, where the Kraus probabilities depend on the state, and compare the empirical measurement distribution produced by the standard trajectory algorithm against the distribution produced by PTSBE with the same total number of shots. If the two distributions differ beyond statistical error, the central equivalence behind the method is false.","supporting_citations":[{"cited_title":"et al Bausch","cited_arxiv_id":null,"evidence_quote":"Motivates the data-generation goal by showing that learned decoders need large simulated noisy datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Foundational reference for quantum trajectory methods, the statistical technique PTSBE modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the CPTP/Kraus formalism and the state-dependent probability rule that underlies the paper's weakest equivalence assumption."}],"review_version":1}