{"id":"a22f60a0-9e83-4620-85b2-ea643c31d817","arxiv_id":"2610.19077","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Classical shadows recorded during quantum federated training allow unlearning any subset of quantum states via a single Newton correction, matching retrained forget accuracy within 0.5 points.","lead":"A quantum neural network trained across sensors cannot replay quantum data to forget it, because measuring a state destroys it. SHADE records a classical shadow of each training state and estimates the gradient and curvature of the forget data at any later parameters, enabling a single Newton correction that removes the data without replay.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The Newton correction assumes ∇L(θ⋆;D)=0, i.e., θ⋆ is a minimizer of the full loss. In practice, training may not converge to a stationary point; the resulting error is unaccounted for and could break the forget accuracy guarantee.","rationale":"The reader identified the strong‑convexity assumption of the retain loss as the weakest point. That concern is valid and acknowledged by the paper. However, a more fundamental assumption—that θ⋆ is a minimizer of the full loss—is not discussed in the limitations. It is used implicitly in the Newton correction derivation and is not verified experimentally. The experiments likely operate close to a local minimum, but the paper claims a general capability without testing regimes where convergence is poor. Adding a test for the initial gradient norm would directly address this risk. Therefore, I recommend a CONDITIONAL acceptance pending this additional verification. My agreement with the reader is partial because the minimizer assumption is a different (and arguably more basic) gap not captured by the convexity criticism.","tokens_in":16176,"tokens_out":3695,"duration_ms":41744,"concrete_test":"Re-run the main experiment (50 nodes, 8 qubits, sample deletion 5%) with models taken at early training epochs (e.g., rounds 10, 20, 30, 40) where ∥∇L(θ⋆;D)∥ is significant. For each early model, compute the SHADE correction and measure ∥θ_u − θ_r∥ and the forget accuracy. If the distance exceeds 0.03 (the value reported at round 60) or the FA drops more than 1 point from the retrained FA, the minimizer assumption is load-bearing. Report ∥∇L(θ⋆;D)∥ for each epoch.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central unlearning procedure (Eq. 8) derives from the identity ∇L(θ⋆;D_r) = −g_f, which holds only if ∇L(θ⋆;D)=0. The paper explicitly states (sentence b55) that θ⋆ is taken as a minimizer of L(·;D). In a realistic QFL deployment, the model is trained by noisy parameter-shift estimates on finite copies, and the global convergence of SGD is not guaranteed on the non‑convex QNN landscape. Hence θ⋆ may have a non‑zero gradient, introducing an additional error term in the Newton step that is not captured by Proposition 2. The paper does not bound ∥∇L(θ⋆;D)∥ or test how the correction degrades when training has not reached a stationary point. The experiments use a fully trained model (60 rounds), which may be near a local minimum, but the claim of ‘within 0.5 points of retraining’ is not demonstrated under less ideal convergence conditions. If the initial gradient norm is large, the uncorrected term could be comparable to the first term in Eq. (10), invalidating the error target.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes SHADE, a method for federated unlearning of quantum data when the data cannot be replayed (measured and consumed). SHADE records classical shadows of training states on the backward light cone of the observable during training, then estimates the gradient and curvature of the forget set at the current model and applies a federated Newton correction to remove their effect. The authors prove that the shadow size needed scales exponentially in the light cone size rather than the total number of qubits, and bound the parameter distance between the corrected and retrained model under strong convexity and Lipschitz Hessian assumptions. Experiments on 50 simulated nodes and on edge boards with noise show that SHADE matches the forget accuracy of full retraining within 0.5 points, takes under 0.2 s, and does not replay any states.","tokens_in":16626,"tokens_out":3146,"duration_ms":33007,"significance":"If the results are robust, SHADE solves a practically important problem for quantum federated learning: forgetting specific quantum data that cannot be remeasured. The use of classical shadows to enable deferred gradient and curvature estimation is novel, and the light-cone analysis makes the record cost manageable for shallow circuits. The method has been implemented on real edge boards with device noise, which strengthens the claim of practical applicability. The paper provides formal guarantees for the shadow estimates and a Newton correction bound, and the experimental evaluation covers three deletion types and multiple baselines.","major_comments":[{"comment":"The identity ∇L(θ⋆; D_r) = -g_f, on which the Newton correction rests, assumes ∇L(θ⋆; D)=0 (i.e., θ⋆ is a minimizer of the full loss). The paper states this assumption (sentence b55) but does not analyze the error when training has not converged to a stationary point—a typical situation for non-convex QNN landscapes under SGD. If ∇L(θ⋆;D) ≠ 0, an additional term ∥(Ĥ_r)^{-1}∇L(θ⋆;D)∥ appears in the correction, which is not captured by Proposition 2. Please either (a) bound ∥∇L(θ⋆;D)∥ in terms of training parameters and include it in the error analysis, or (b) provide experiments with early-stopped or less-converged models to demonstrate that the method remains accurate within the claimed margin.","section":"§II.B & §III.C, Eq. (8) and Proposition 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The theoretical guarantees (Proposition 2) rely on L(·; D_r) being κ-strongly convex and having an M-Lipschitz Hessian on a ball containing θ⋆ and θ_r. These conditions are not verified for the QNN circuits used in the experiments, and QNN loss landscapes are generally non-convex. While the empirical results suggest the method still works, the paper should discuss the plausibility of these assumptions for the shallow circuits employed (8 qubits, layered brickwork) and, if possible, report estimated values of κ and M or bounds derived from the circuit structure. Without such discussion, the applicability of the parameter-distance bound to the tested setting is unclear.","section":"§IV, Proposition 2 and Corollary 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The experiments are conducted with 60 training rounds; in this regime the model is likely near a local minimum. To support the claim that SHADE works under less ideal convergence (e.g., when ∇L(θ⋆;D) is not negligible), please include a sensitivity study where the number of training rounds is varied (e.g., 10, 20, 40 rounds) and measure the distance to retrain and forget accuracy. This would directly address the skeptic's concern about the practical validity of the minimizer assumption.","section":"§V, Table I and Fig. 6","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The constant α in Eq. (9) is said to be absolute but is not specified or referenced. Please either provide its value or cite a standard median-of-means bound (e.g., from Huang et al., Nature Physics 2020) so that readers can compute the required S.","section":"§IV, Theorem 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The step 'symmetrizes Ĥ_r, raises its eigenvalues below μ to μ' should be explicitly described as projecting onto the set of symmetric matrices with eigenvalues at least μ. This projection is convex and does not increase the distance to H_r, but a brief justification would improve clarity.","section":"§III-C, Eq. (8)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The testbed uses only 3 nodes, while the simulation uses 50. It would be helpful to comment on how the testbed results (in Table I columns) are expected to generalize to a larger number of nodes, particularly regarding network latency and estimation parallelism.","section":"§V-A, Testbed description","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Time' column for SHADE reports 0.17–0.18 s in simulation and 0.18–0.19 s on testbed. Please clarify whether this includes only the server's Newton solve and the upload of estimates, or also the per-node estimation time from shadows. If the latter, mention the node-level wall-clock time.","section":"§V-B, Table I","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The description of the shadow protocol mentions calibrating for readout errors on real devices. It would be helpful to cite the calibrated-shadow literature (e.g., Chen et al., PRX Quantum 2021) for the variance bound after calibration.","section":"§III-A, Eq. (5)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The phrase 'quantum data cannot be replayed' could be slightly softened: quantum data cannot be replayed after measurement because the state is destroyed, but they could be replayed if additional copies are available. The paper already acknowledges this (the no-cloning theorem and finite copies). Consider rewording in the abstract to avoid implying absolute impossibility.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper tackles a genuine and underappreciated problem: how to remove specific quantum states from a federated QNN when the states cannot be replayed. The core idea—record classical shadows of every state at training time, then estimate the forget data's gradient and the retained data's curvature from the shadows and apply a federated Newton correction—is clever and well-motivated. The light-cone bound on snapshot storage is a nice practical addition. The experiments are thorough, covering three deletion types, multiple baselines, and real hardware under a realistic noise model. The key claim (forget accuracy within 0.5 points of retraining, in under 0.2 seconds) is convincingly demonstrated.\n\nThere is one soft spot, and the stress-test note gets it right. The Newton correction in Eq. (8) depends on the identity ∇L(θ^⋆; D_r) = −g_f, which holds only if ∇L(θ^⋆; D) = 0—i.e., the trained model is a stationary point of the full loss. The paper states this assumption (b55), but in a realistic QFL deployment with noisy parameter-shift estimates and non-convex loss landscapes, convergence to a stationary point is not assured. The error introduced by a non-zero gradient is not bounded, and the paper does not test the sensitivity of the method to training convergence. The theory (Proposition 2) also assumes κ-strong convexity with M-Lipschitz Hessian—conditions unlikely to hold for QNNs and not verified for the circuits used. The experiments are done with a well-trained model (60 rounds), which may be near a local minimum, but the claim of “within 0.5 points” is not demonstrated under less ideal convergence.\n\nThat said, the paper does not hide these assumptions, and the experimental results suggest the method still works in practice despite them. The flaw is real but not fatal; it limits the theoretical rigor that can be claimed. A responsible revision would either bound ∥∇L(θ^⋆; D)∥ from training data or include experiments with fewer rounds to test robustness.\n\nThis paper is for the quantum ML and federated learning communities. It deserves a serious referee and likely acceptance after minor revisions addressing the convergence assumptions. I would definitely bring it to reading group.","headline":"Clever and well-tested solution to a real problem, but the theoretical guarantee relies on a zero-gradient condition that is unlikely to hold in practice and is not bounded or tested.","tokens_in":17041,"tokens_out":1364,"duration_ms":14942},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Classical shadows enable unlearning of quantum data without replay, achieving retraining-level accuracy in under 0.2 seconds.","keywords":["quantum federated learning","machine unlearning","classical shadows","quantum neural networks","quantum sensor networks","data deletion","Newton correction","shadow-based data erasure"],"falsifier":"If a quantum federated learning node with a shallow QNN (light cone size ≤6) records S=$10^{5}$ classical snapshots per state and a window deletion is requested, SHADE should achieve a parameter distance ∥θ_u - θ_r∥ within the theoretical bound (Eq. 10) and forget accuracy within 0.5 points of retraining on a 50-node simulation. If the observed distance exceeds the bound by more than a factor of 2, the estimation or correction procedure must fail for this regime.","tokens_in":17714,"feed_emoji":"⚛️","tokens_out":4367,"duration_ms":48514,"temperature":0.7,"pith_summary":"Quantum federated learning trains a quantum neural network across nodes that produce quantum states, such as sensor outputs. When data must be forgotten—because a node withdraws or a sensor drifted—existing unlearning methods require replaying the consumed states, which is impossible due to no-cloning and the one-shot nature of quantum measurement. This paper proposes SHADE, which records classical shadows of each training state, enabling estimation of gradients and curvatures for the forget data without re-measuring the state. A single federated Newton correction then removes the forget data's effect. The authors prove the shadow record grows with the backward light cone of the measured observable, not the total qubit count, and demonstrate that SHADE matches retraining accuracy within 0.5 points in under 0.2 seconds on a 50-node simulation and edge hardware.","feed_headline":"Classical shadows enable fast quantum unlearning","feed_subtitle":"Federated learning nodes forget samples without replay, matching retraining accuracy within 0.5 points in under 0.2 seconds.","key_machinery":"The key machinery is the classical shadow formalism combined with backward-light-cone restriction, applied to the gradient and Hessian operators of a parameterized quantum circuit. The classical shadow snapshot (Eq. 5) provides unbiased estimates of expectation values at any parameters when the circuit depth is fixed. By keeping snapshots only on the light cone Λ of the output observable O, the required number of snapshots per state grows exponentially in |Λ| but not in the number of qubits (Theorem 1). The federated Newton correction (Eq. 8) uses the estimated gradient of the forget data and the estimated Hessian of the retained data to update the model in one step, without iterative optimization or replay.","core_discovery":"The central discovery is that a quantum neural network's gradient and Hessian on a quantum state are expectation values of observables, and these can be estimated at any later parameter value from classical shadows recorded during training, without accessing the original state again. By storing only the shadow data on the backward light cone of the output observable, the storage cost depends only on the depth of the circuit, not on the total qubit count. This enables a federated Newton correction that matches the unlearning accuracy of full retraining while being orders of magnitude faster and consuming no fresh quantum copies.","pith_inferences":["The SHADE framework could be adapted to other tasks beyond classification, such as quantum state certification or quantum metrology, where forgetting past calibration data is required.","The theoretical dependence on strong convexity (Proposition 2) may be relaxed by using a quasi-Newton update or a smaller step size, potentially extending the method to deeper circuits where the Hessian is not guaranteed positive definite.","The record cost (72 kB per state for 8 qubits) might be reduced further by adaptive shadow protocols that focus on the most informative measurement bases, or by compressing the shadow bits through error-correcting codes."],"forward_implications":["SHADE enables fine-grained unlearning (per-sample, per-window, per-node) in quantum federated learning without any additional quantum resource consumption after training.","Unlearning can be completed in under 0.2 seconds regardless of the number of nodes, enabling real-time compliance with deletion requests.","The shadow record size scales exponentially with the backward light cone of the observable, not with the number of qubits, allowing large shallow circuits to be forgotten efficiently.","The method maintains model utility (test accuracy above 95.7%) and privacy (membership inference near 50%) across all deletion types.","SHADE does not require any node to re-prepare or measure quantum states after the training phase, making it compatible with strict no-cloning constraints."],"supporting_citations":[{"why":"Provides the classical shadow formalism for predicting expectation values of observables from randomized single-qubit measurements.","marker":"[38]"},{"why":"Establishes that observables of a shallow QNN act only on the backward light cone of the measured qubits, which limits the required shadow size.","marker":"[39]"},{"why":"Introduces the parameter-shift rule that defines the gradient and Hessian operators whose expectations SHADE estimates from shadows.","marker":"[37]"},{"why":"Describes the certified removal approach (exact Newton correction) that SHADE approximates and against which it is benchmarked.","marker":"[33]"}],"fun_headline_variants":["Shadows let quantum models forget without replay","Classical shadows erase quantum data in 0.2 seconds","Federated quantum unlearning matches retraining accuracy","Forget quantum sensor data without destroying states","Shadow-based quantum unlearning: fast, accurate, no replay"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The error bound in Proposition 2 assumes the retain loss landscape is κ-strongly convex with an M-Lipschitz Hessian on a ball containing the trained and retrained models, an assumption that is known to fail for typical quantum neural networks outside of overparameterized regimes.","fun_headline_variants_meta":{"raw":{"variants":["Shadows let quantum models forget without replay","Classical shadows erase quantum data in 0.2 seconds","Federated quantum unlearning matches retraining accuracy","Forget quantum sensor data without destroying states","Shadow-based quantum unlearning: fast, accurate, no replay"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":7.2e-05,"raw_usage":{"cost":7.1587432e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":7.1587432e-05,"upstream_inference_prompt_cost":2.5248552e-05,"upstream_inference_completions_cost":4.633888e-05},"total_tokens":1467,"prompt_tokens":863,"completion_tokens":604,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":527}},"tokens_in":863,"tokens_out":604,"cost_source":"openrouter_usage","duration_ms":8020,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":527,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-15T16:54:50.504293+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If a quantum federated learning node with a shallow QNN (light cone size ≤6) records S=$10^{5}$ classical snapshots per state and a window deletion is requested, SHADE should achieve a parameter distance ∥θ_u - θ_r∥ within the theoretical bound (Eq. 10) and forget accuracy within 0.5 points of retraining on a 50-node simulation. If the observed distance exceeds the bound by more than a factor of 2, the estimation or correction procedure must fail for this regime.","supporting_citations":[{"cited_title":"Predicting many properties of a quantum system from very few measurements,","cited_arxiv_id":null,"evidence_quote":"Provides the classical shadow formalism for predicting expectation values of observables from randomized single-qubit measurements."},{"cited_title":"Cost function dependent barren plateaus in shallow parametrized quantum circuits,","cited_arxiv_id":null,"evidence_quote":"Establishes that observables of a shallow QNN act only on the backward light cone of the measured qubits, which limits the required shadow size."},{"cited_title":"Eval- uating analytic gradients on quantum hardware,","cited_arxiv_id":null,"evidence_quote":"Introduces the parameter-shift rule that defines the gradient and Hessian operators whose expectations SHADE estimates from shadows."},{"cited_title":"Certified Data Removal from Machine Learning Models,","cited_arxiv_id":null,"evidence_quote":"Describes the certified removal approach (exact Newton correction) that SHADE approximates and against which it is benchmarked."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}