{"id":"037f6be7-1260-45d5-9e1b-8c354bdedcec","arxiv_id":"2506.14858","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Adding a log-overhead regularizer with trainable cutting angles reduces circuit-cutting sampling cost in QML regression, at similar reported accuracy, but without an unregularized baseline.","lead":"This paper adds a penalty term to the loss function of a quantum machine learning model that discourages the sampling overhead of circuit cutting, and trains the cutting angles alongside the model. The idea could make it easier to run larger quantum circuits on small hardware, but the experiments lack a baseline comparison.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No baseline: the two 18-qubit runs both use CutReg (Eq. 5), so the claimed balance between accuracy and overhead is not demonstrated; regularized training must be compared against static cuts and unregularized adaptive training.","rationale":"I read the paper as proposing a regularizer that makes circuit-cutting angles trainable and penalizes their sampling overhead during QML training. The idea is plausible and the overhead formula is a reasonable penalty target. However, the experimental section provides no baseline without the regularizer and no static-cutting control. The two reported arms both use the regularizer, so the key observation—convergence to similar accuracy and overhead from different initializations—is consistent with the regularizer being inert or with the cost landscape itself favoring moderate angles. This missing control is the most load-bearing gap because the central claim is specifically that the regularizer enables a balance that static cutting or unregularized training would not achieve. The reader's weakest_assumption concerned expressiveness: that angles can be driven to low-overhead values without sacrificing model performance. My concern is adjacent but distinct: the paper never tests whether this expressiveness cost occurs, because it compares only regularized models against each other. A three-arm experiment with static cuts, unregularized adaptive angles, and CutReg would settle both issues at once. I also note the λ inconsistency between Eqs. (3) and (5) and the text, which is secondary but must be resolved for reproducibility. Given these gaps, the reader's CONDITIONAL verdict is appropriate; my analysis does not move it, but it sharpens the condition: the method must be shown to change the operating point relative to a no-regularizer adaptive baseline and to a static-cut baseline without sacrificing test accuracy.","tokens_in":6612,"tokens_out":16441,"duration_ms":177592,"concrete_test":"Run the same 18-qubit regression task with at least 10 random seeds under three conditions: (A) static cutting with α_l fixed at π/2 and no regularizer; (B) trainable α_l with no regularizer; (C) trainable α_l with the CutReg objective of Eq. (5) using the stated λ schedule. Report test MSE and S_total at epoch 100 for each arm. The central claim is supported if (C) matches (A) in test MSE while S_total is close to 10 and substantially below (A)'s 6561, and if (C) reduces S_total below (B). If (B) already matches (C), the regularizer is unnecessary; if (C) has materially worse test MSE than (A), the expressiveness trade-off is confirmed and the claimed practical benefit is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adding the overhead regularizer lets the optimizer find a better accuracy/overhead trade-off than static circuit cutting. The only quantitative support is Section IV-B, Figure 3, which compares two initializations of the regularized loss (π/2 vs 0.1) and shows both converge to S_total≈10 and similar test MSE. Both arms include the regularizer, so the experiment cannot distinguish the effect of CutReg from the natural behavior of trainable cutting angles. Without an unregularized run with trainable α and a static-cut run with α fixed, it is unknown whether the reduction from S_total=6561 to ≈10 is caused by the regularizer or by the optimizer independently pulling angles away from π/2. The paper also does not report whether test MSE degrades relative to a static π/2 cut, which is the key expressiveness risk: if the regularizer lowers overhead only by sacrificing accuracy, the claimed 'balance' is not a balance but a forced trade-off. Additionally, Eq. (5) omits the λ that Section III defines and Section IV describes as decaying from 0.01 to 0.0001, making the exact objective ambiguous. The verification of the central claim therefore requires controlled baseline comparisons.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CutReg, a regularizer that augments a QML loss with the logarithm of the circuit-cutting sampling overhead, and treats the cutting angles as trainable parameters. In numerical regression experiments with hardware-efficient ansätze of 18, 24, 30, and 50 qubits, the authors report that 18-qubit models initialized with maximally entangling and partially entangling cuts converge to similar test errors and a total sampling overhead near 10, and that the approach permits training larger VQCs on a single CPU until barren-plateau gradients appear.","tokens_in":6815,"tokens_out":10688,"duration_ms":102677,"significance":"If the central claim is established, the idea of making circuit-cutting overhead part of the QML objective is a useful and timely contribution: it connects resource estimation with variational training and offers a concrete way to adapt cutting angles during optimization rather than fixing them in preprocessing. The paper's formulation is simple and the use of standard overhead formulas makes the proposal easy to adopt. However, the current experimental evidence does not yet isolate the effect of the regularizer from the behavior of trainable cutting angles, so the significance rests on a claim that is not yet demonstrated.","major_comments":[{"comment":"The central claim that CutReg enables a balance between accuracy and sampling overhead is not directly tested, because both 18-qubit arms in Fig. 3 use the regularized objective. The observed reduction in S_total from 6561 to about 10 in the pi/2-initialized arm is a direct consequence of adding a penalty proportional to log S_total to the loss, so it does not by itself show that the regularizer is needed or beneficial. Please add controlled comparisons: (i) unregularized training with trainable cutting angles (lambda = 0), and (ii) static cuts with alpha fixed at pi/2 and at 0.1, and report test MSE and S_total for each. Without these baselines, the paper cannot distinguish the effect of CutReg from the natural evolution of trainable angles, nor can it support the claim that the accuracy/overhead point found is a balance rather than a forced trade-off.","section":"IV-B, Fig. 3"},{"comment":"The objective is stated inconsistently. Eq. (3) defines L_reg = L_QML + lambda * R_overhead, and Eq. (4) defines R_overhead as the log of the product, while Eq. (5) in the experimental setup writes L_reg = MSE + log(prod(...)) with no lambda. The text then says lambda was initialized to 0.01 and stepped down to 0.0001 after 10 epochs. This makes the exact optimized function ambiguous and hampers reproducibility. Please correct Eq. (5) to include lambda, and specify whether 'stepped down to 0.0001' is a one-time change at epoch 10 or a continuous annealing schedule.","section":"III, Eq. (3); IV-A3, Eq. (5)"},{"comment":"The paper states that the 18-qubit results are averages over ten training runs, but Fig. 3 shows no error bars, confidence intervals, or per-run spread. Since the central observation is that the two initializations converge to 'similar' test error and 'similar' S_total near 10, the reader cannot judge whether the differences are statistically meaningful. Please report means with standard deviations or interquartile ranges, and specify whether the plotted quantity is test MSE or another metric.","section":"IV-A4 and IV-B1, Fig. 3"}],"minor_comments":[{"comment":"The 30- and 50-qubit results are described only by the statement that gradients were on the order of 10^-7, which is presented as evidence of barren plateaus. No learning curves, final losses, or overhead values are shown for these runs, so the claim that the method 'enabled training' of these sizes is not yet substantiated; either add supporting data or weaken the claim.","section":"IV-B1"},{"comment":"Eq. (1) is cited to reference [9], but the exact formula and its regime of applicability should be stated explicitly, since the entire overhead computation and the regularizer depend on it. A short derivation or a precise pointer to the equation in Piveteau and Sutter would remove ambiguity.","section":"III, Eq. (1)"},{"comment":"The Meyer-Wallach entanglement measure is attributed to reference [13] (Brennen), which is a related but distinct reference; please cite the original Meyer-Wallach work or clarify the provenance.","section":"II, [13]"},{"comment":"There is a typo in 'Quasiprobability simiulation' in Section III, and the notation for the total overhead switches between s(alpha) in Eq. (1), s(alpha-hat) in Eq. (2), and S_total in Section IV-A4; please unify the notation.","section":"III"},{"comment":"Fig. 2 is labeled as a decomposition of a CX gate, while the text says the cut CZ gates are replaced by trainable Rzz gates; please clarify how the CX decomposition relates to the Rzz(alpha) cuts used in the experiments.","section":"IV-A2, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and timely, but the experimental section currently supports only a qualitative demonstration that adding an overhead penalty changes the training dynamics; it does not yet support the paper's central claim of a beneficial balance relative to existing static-cutting practice. The requested baseline experiments and the objective-function correction are within the scope of a revision, so I recommend major revision rather than rejection. I also note that the paper would be strengthened by making the code and data available, since the numerical claims rest on a single synthetic regression task."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Periyasamy et al. propose making circuit-cutting angles trainable and penalizing the log sampling overhead in the QML loss. That specific combination is new as far as the cited literature goes, and it is a clean idea worth engaging. The overhead formula for Rzz cuts is standard, the regularization term is simple, and they are honest that larger circuits hit barren plateaus.\n\nThe main problem is that the experiments don't evidence the central claim. The key figure compares two initializations of the regularized objective; both arms contain CutReg. Without an unregularized run with trainable angles and a static cut (alpha fixed at pi/2), you cannot attribute the drop in overhead from 6561 to about 10 to the regularizer. The optimizer might independently pull angles away from pi/2 even without the penalty. You also don't know whether test MSE degrades relative to a static cut, which is the expressiveness risk the paper itself flags. As written, the overhead reduction is by construction, not a prediction; the only genuinely empirical question is the accuracy/overhead trade-off, and that question is not directly tested.\n\nThere are also smaller consistency issues. Eq. (5) drops the lambda that Eq. (3) and the training description say is there (decaying from 0.01 to 0.0001), so the exact objective is ambiguous. The ten-run averages are plotted without error bars. The 30- and 50-qubit experiments contribute only an observation that gradients scale to 10^-7, which tells us about barren plateaus, not about CutReg.\n\nWhat is good: the paper is well-scoped, the literature placement is fair (Gentinetta et al. and Marshall et al. are correctly distinguished), and the idea is experimentally easy to adopt. If the accuracy/overhead balance survives proper baseline comparisons, it would be a useful practical tool for cut-based QML on small hardware. But that evidence is not here yet.\n\nRecommendation: deserve peer review after major revision. An editor can send it to a reviewer with a request to test baseline comparisons and fix the equation. I would not cite it yet.","headline":"A clean new regularizer idea for adaptive circuit cutting, but the experiments lack the baseline comparisons needed to prove the central claim.","tokens_in":7366,"tokens_out":1948,"would_cite":false,"duration_ms":20731,"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":"Cutting angles in quantum circuits can be trained as part of the loss function, letting the optimizer reduce sampling overhead while keeping test error low.","keywords":["quantum machine learning","circuit cutting","circuit knitting","sampling overhead","regularization","variational quantum circuits","barren plateaus"],"falsifier":"Run the same regularized training on a regression target whose features on opposite sides of the cut are strongly correlated, so the optimal model needs the entanglement generated near $\\alpha=\\pi/2$; if test error rises steeply whenever the penalty forces the angles toward $0$ or $\\pi$, the claimed balance does not generalize beyond the tested dataset.","tokens_in":6402,"feed_emoji":"✂️","tokens_out":12304,"duration_ms":114200,"temperature":0.7,"pith_summary":"Circuit cutting lets a larger quantum circuit run on smaller hardware, but the classical post-processing that reassembles the pieces demands extra samples, and that overhead depends sharply on the angles of the gates being cut. This paper proposes making those cutting angles trainable by adding a regularizer $R_{\\text{overhead}}(\\hat{\\alpha}) = \\log \\prod_{l=1}^{L} (1+2|\\sin\\alpha_l|)^2$ to the machine-learning loss. The optimizer can then move angles toward $0$ or $\\pi$ when the learning task allows it, lowering the sampling overhead, or keep them near $\\pi/2$ when the task needs the entanglement. The paper's numerical study on regression with up to 50 qubits indicates that models initialized with high- and low-entanglement cuts converge to similar accuracy and sampling overhead, so maximal entanglement at the cuts is not necessary for good performance. This matters because it turns circuit cutting from a static preprocessing step into an adaptive part of training, a step toward studying larger quantum models on limited hardware.","feed_headline":"Circuit-cut angles become trainable, lowering sampling overhead","feed_subtitle":"A log-penalty in the QML loss lets the optimizer balance regression accuracy against sampling cost","key_machinery":"The load-bearing object is the angle-dependent sampling-overhead formula $s(\\alpha)=(1+2|\\sin\\alpha|)^2$ for a cut rotation gate, together with its product form $s(\\hat\\alpha)=\\prod_l s(\\alpha_l)$ for multiple cuts. The regularizer is the logarithm of this product, which is non-negative, differentiable, and monotone in the overhead; because the cutting angles are trainable parameters, its gradients push each angle toward $0$ or $\\pi$ when the learning term permits. This is what converts circuit cutting from a fixed preprocessing decision into an adaptive part of the optimization.","core_discovery":"The central claim is that the sampling overhead of circuit cutting is not a fixed post-processing cost but a differentiable function of the cut-gate angles, so it can be optimized together with the model. For a cut rotation gate with angle $\\alpha$, the overhead is $s(\\alpha)=(1+2|\\sin\\alpha|)^2$, and for $L$ independent cuts it multiplies; the paper's regularized objective is $L_{\\text{reg}}(\\hat\\theta)=L_{\\text{QML}}(\\langle O\\rangle_{\\hat\\theta})+\\lambda \\log\\prod_{l=1}^L (1+2|\\sin\\alpha_l|)^2$, with the cutting angles $\\hat\\alpha$ part of the trainable parameters $\\hat\\theta$. Minimizing this objective lets the optimizer balance regression accuracy against sampling cost, and the paper reports that 18-qubit regression runs starting from $\\alpha_l=\\pi/2$ (overhead 6561) and from $\\alpha_l=0.1$ (overhead 4.29) both end near overhead $\\approx 10$ with comparable test error. The paper also reports training 30- and 50-qubit circuits on a single CPU, with gradient magnitudes around $10^{-7}$, the barren-plateau regime where gradients vanish, indicating that this obstacle persists independently of cutting overhead.","pith_inferences":["The regularizer should transfer to other two-qubit rotation gates such as $R_{xx}$ or $R_{yy}$ and to classification losses, since the qualitative angle-overhead curve is the same; the paper only demonstrates regression with $R_{zz}$ cuts.","The final trained angles carry interpretable signal: angles that remain near $\\pi/2$ identify cut locations whose entanglement is load-bearing for the target function, so the regularizer can double as a diagnostic for where cutting hurts.","Replacing the logarithmic form with an additive penalty $\\sum_l (s(\\alpha_l)-1)$ is a natural variant; comparing the two would reveal whether the choice of aggregation matters when cuts have very different overheads."],"forward_implications":["Any parameterized cut gate with a known angle-dependent overhead can be trained instead of being fixed before training.","Starting cuts with low entangling capability is not a disadvantage: the 18-qubit experiments suggest the optimizer reaches the same accuracy and overhead as high-entanglement starts.","The regularized objective allows 30- and 50-qubit circuits to be trained on a single CPU, making barren-plateau studies more accessible.","The learning-rate schedule for $\\lambda$ is part of the method: a strong early penalty guides angles down, then relaxing it lets the model recover overhead if accuracy demands it."],"supporting_citations":[{"why":"Supplies the negative-probability decomposition that underlies circuit cutting.","marker":"[1]"},{"why":"Quantifies the sampling overhead of cutting and formalizes the quasiprobability decomposition the regularizer builds on.","marker":"[2]"},{"why":"Formulates the sampling cost of implementing a virtual two-qubit gate from single-qubit operations, making the overhead angle-dependent.","marker":"[3]"},{"why":"Demonstrates overhead-constrained circuit knitting for variational dynamics, the closest prior use of overhead control that this method generalizes.","marker":"[6]"},{"why":"Provides the explicit angle-dependent overhead formula $s(\\alpha)=(1+2|\\sin\\alpha|)^2$ used in the regularizer.","marker":"[9]"},{"why":"Provides the circuit-cutting simulation software used to run the numerical experiments.","marker":"[10]"},{"why":"Provides the guided-SPSA gradient estimator used to train the quantum circuit parameters.","marker":"[11]"},{"why":"Defines the multipartite entanglement measure used to track entanglement during training.","marker":"[13]"}],"fun_headline_variants":["Trainable cut angles reduce sampling overhead in QML","Optimizing circuit cutting overhead as a loss term","Balancing cutting cost and accuracy via trainable angles","QML loss now includes the cost of circuit cutting","Making sampling overhead from cuts a trainable objective"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the optimizer can push the cut-gate angles toward low-overhead values (near $0$ or $\\pi$, which also reduces entanglement) without seriously hurting the model's ability to fit the target function; the paper's evidence for this is a single synthetic regression task compared only across different initializations.","fun_headline_variants_meta":{"raw":{"variants":["Trainable cut angles reduce sampling overhead in QML","Optimizing circuit cutting overhead as a loss term","Balancing cutting cost and accuracy via trainable angles","QML loss now includes the cost of circuit cutting","Making sampling overhead from cuts a trainable objective"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1456,"prompt_tokens":975,"completion_tokens":481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":591,"tokens_out":481,"duration_ms":5542,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:20.411434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same regularized training on a regression target whose features on opposite sides of the cut are strongly correlated, so the optimal model needs the entanglement generated near $\\alpha=\\pi/2$; if test error rises steeply whenever the penalty forces the angles toward $0$ or $\\pi$, the claimed balance does not generalize beyond the tested dataset.","supporting_citations":[{"cited_title":"Constructing a virtual two-qubit gate by sampling single-qubit operations,","cited_arxiv_id":null,"evidence_quote":"Formulates the sampling cost of implementing a virtual two-qubit gate from single-qubit operations, making the overhead angle-dependent."},{"cited_title":"Overhead-constrained circuit knitting for variational quantum dynamics,","cited_arxiv_id":null,"evidence_quote":"Demonstrates overhead-constrained circuit knitting for variational dynamics, the closest prior use of overhead control that this method generalizes."},{"cited_title":"Guided-spsa: Simultaneous perturbation stochastic approximation as- sisted by the parameter shift rule,","cited_arxiv_id":null,"evidence_quote":"Provides the guided-SPSA gradient estimator used to train the quantum circuit parameters."}],"review_version":1}