acceptodds
Under review as a conference paper at ICLR 2027

Quantum-Enhanced Riemannian-Conditioned Graph Representation Learning for ALS-EEG Decoding

Abstract

Cross-patient electroencephalography (EEG) decoding for amyotrophic lateral sclerosis (ALS) requires learning from small clinical cohorts while capturing both local EEG evidence and trial-specific inter-component relations. Compact deep encoders capture local temporal–spatial patterns, whereas Riemannian and graph methods explicitly represent relational structure, but these two information sources are typically compressed or aggregated without preserving their component-wise correspondence. We introduce QRG-EEG, a quantum-enhanced Riemannian-conditioned graph model that preserves component identity across deep, graph, and quantum representations. Local evidence from component controls a single-qubit operation on qubit , while its Riemannian relation directly conditions the interaction between qubits . We analytically show how this edge conditioning becomes observable and induces a joint node–node–edge response. We evaluate QRG-EEG on four public ALS-EEG datasets comprising 30 patient-level LOSO folds, using strict source-only selection, matched classical and quantum controls, and targeted edge interventions. Extensive experiments demonstrate that QRG-EEG achieves the strongest dataset-level Macro-F1, together with secondary metrics (AUROC for P300, balanced accuracy for MI), among the representative competing methods, while removing or misaligning trial-specific edges consistently degrades performance. These results support component-aligned node–edge conditioning as a structured inductive bias for scarce-data cross-patient EEG decoding.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
EEG decoding, brain–computer interfaces, cross-subject generalization, Riemannian geometry, hybrid quantum–classical learning, graph representation learning.

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_HUc2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QRG-EEG, a quantum-enhanced Riemannian-conditioned graph representation learning framework for cross-patient ALS-EEG decoding. The method combines component-wise EEG feature extraction, Riemannian covariance representations, and graph-conditioned variational quantum circuits.

Critical strengths

The paper addresses the important and challenging problem of cross-patient EEG decoding under limited clinical data.

The component-aligned node–edge conditioning mechanism is clearly formulated and supported by theoretical analysis.

The experimental evaluation is relatively comprehensive, including patient-level LOSO validation, multiple classical and quantum baselines, targeted ablations, and paired statistical comparisons.

Critical weaknesses

The proposed framework mainly integrates existing EEG encoders, Riemannian representations, graph modeling, and variational quantum circuits. Although component-aligned conditioning is an interesting design choice, its methodological novelty over existing structured node–edge interaction mechanisms remains insufficiently established.

The theoretical analysis demonstrates that quantum measurements can capture joint node–node–edge dependencies. However, similar nonlinear interactions can also be modeled using classical architectures. It has not yet been fully explained why this correspondence must be implemented via quantum circuits. Classical graph neural networks, bilinear interactions, or explicit node-edge multiplicative interactions could also preserve the correspondence.

The current Fixed-entanglement VQC changes both the edge values and the graph topology. Could the authors provide a controlled comparison that preserves the original graph topology and quantum gates while replacing only trial-dependent edge values with source-derived constants?

The method relies on a source-fitted CSP, PCA, or xDAWN projection. Could the authors examine whether the proposed interaction mechanism remains effective under different latent representations and consistent component re-indexing, and clarify whether component identity has any interpretation beyond computational index correspondence?

Questions for authors

See the weakness.

3 comments

We sincerely thank the reviewer for the careful and constructive questions. Signed values are QRG-EEG minus comparator Macro-F1 on bigP3BCI/BNCI/EEGET/UCL (seeds averaged within patient; all 30 patient-level LOSO folds; 95% patient-bootstrap intervals in Tab. 12, all excluding 0); unsigned values are absolute scores.

W1. The problem existing node–edge mechanisms leave open. Cross-patient ALS-EEG decoding needs local temporal evidence of each latent component and trial-specific relations between components. Riemannian–deep hybrids (CRGNet, Graph-CSPNet) fuse or aggregate them after compression; EEG quantum models (QEEGNet; pyRiemann-qiskit) process flattened vectors; weighted-graph quantum circuits (Skolik et al.; Mernyei et al.) condition gates on generic graph weights not tied to signal components. QRG-EEG keeps both kinds of evidence separately addressable while making them interact by component identity: one index ii is shared by spatial filter, node feature, covariance row, graph node, and qubit, so [Hb]ij[\mathbf{H}_{b}]_{ij} drives exactly the RZZR_{ZZ} on qubits ii and jj. Controls support this: the same Riemannian information vectorized into a generic VQC (C9) gives +.083/+.148/+.066/+.106, and test-time edge misassignment (C8) +.082/+.132/+.052/+.077; consistent permutation during training (C7) keeps a clear gap on P300 (+.058/+.104) and a smaller one on MI (+.031/+.050), where the benefit comes mainly from trial-specific relations (C5, W3).

Revision: Sec. 1, Sec. 3, Tab. 2, Tab. 12, Fig. 5.

W2. Why a VQC for node–edge interaction? As stated in the Introduction, message-passing GNNs propagate relations by neighbourhood aggregation, which mixes each node's state with its neighbours' and, as layers deepen, can over-smooth node representations (Li et al., 2018), diluting component-level evidence. In the graph-conditioned circuit, node ii sets the rotation on qubit ii, relation (i,j)(i,j) generates the two-qubit gate on exactly that pair, the mixer turns the edge phase into terms multiplicative in both incident nodes and the relation (Eq. 15; App. B.8), and the Z/ZZ readout keeps per-node and per-edge coordinates. The circuit thus jointly preserves local temporal–spatial evidence and trial-specific relational geometry. Classical operators with identical aligned inputs, objective, selection, and calibration remain below QRG-EEG, with the largest gap for the GAT that aggregates the same graph:

AlternativebigP3BCIBNCIEEGETUCL
B6 GAT, same graph+.076+.229+.161+.185
B7 node–edge products+.054+.108+.086+.125
B12 trig. features+.064+.200+.088+.152

The edge-conditioned GNNs remain below QRG-EEG as well: where QRG-EEG obtains .512/.699/.524/.583 Macro-F1, NNConv/ECC obtains .508/.642/.450/.334, GINE .498/.636/.398/.287, and a pairwise MLP ϕ(vi,vj,eij)\phi(v_i,v_j,e_{ij}) .482/.641/.417/.378.

Revision: Sec. 1, Sec. 5 (RQ1, Fig. 4), Tab. 12, App. J (Tab. 18).

W3. Same topology and gates, only trial-dependent edges removed. C5 is exactly this control: topology, RZZR_{ZZ} gates, mixers, and readout are unchanged (identical parameter count), and each eb,ije_{b,ij} is replaced by its source mean. C6 changes only the topology:

bigP3BCIBNCIEEGETUCL
C5+.076+.152+.061+.082
C6+.065+.120+.033+.072

Removing trial dependence alone lowers performance on every cohort, by more than changing the topology.

Revision: Tab. 2, Tab. 5, Tab. 12, Sec. 5 (RQ2).

W4. Latent representations and component identity.

  • Representations. Source validation preferred xDAWN with ERP-residual covariance (P300) and CSP (MI) over plain covariance and PCA in every fold (16/16, 14/14). With the selected representation, QRG-EEG leads B7 by +.054/+.108/+.086/+.125; forcing the non-selected one on both models gives QRG-EEG .492/.665/.485/.502 and B7 .475/.638/.425/.462. QRG-EEG thus stays above B7 under both, so the mechanism does not hinge on one projection.
  • Identity and re-indexing. Before projection, indices refer to EEG channels. After X~b=P⊤Xb\tilde X_b=P^\top X_b (Eq. 1), index ii denotes the ii-th source-fitted spatial filter pip_i (signs fixed; App. B.2.1), shared by node, covariance row, graph node, and qubit; its physiological meaning is carried by pip_i and its scalp pattern, not by a channel. A consistent re-indexing of all of these is only a relabeling (App. B.3); inconsistent re-indexing changes the function, which C7/C8 test.

Revision: Sec. 3 (Eq. 1), App. B.2.1, B.3, App. K (Tab. 19).

Thank you for the detailed responses. My concerns have been adequately addressed, and I will increase my rating.

Thank you for your prompt response. We sincerely appreciate your positive feedback.

Reviewer_JsQ8OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces QRG-EEG, a hybrid quantum-classical model for cross-patient ALS EEG decoding. The EEG trials are projected into a shared latent space where a component-wise EEGNet-like branch extracts local node evidence. A Riemannian branch further constructs trial-specific relations. Quantum encoded features then feed a linear classifier. The model is evaluated on four public ALS-EEG cohorts, where QRG-EEG achieves the highest dataset-level Macro-F1, and ablations then show that removing trial-specific components reduces performance. The paper also includes both classical and quantum comparison models.

Critical strengths

Good comparison with relevant baselines such as QEEGNet and GAT on the same graph.

The contribution, to my limited knowledge, is fairly novel and would be interesting for the ML community in general.

The ablations are useful as they show that both the trial-specific relations and their correct alignment are important.

Critical weaknesses

The method is evaluated on only four small ALS-EEG datasets with just 6-8 subjects per dataset; this makes it hard to see if this method can actually be used in clinical practice.

The quantum part is evaluated using exact statevector simulations this makes it difficult to know whether the same behaviour and performance would hold in a practical quantum computing setting.

Questions for authors

How computationally efficient is this algorithm? How does it scale to a large number of channels?

1 comment

We sincerely thank the reviewer for the positive assessment and for the practical questions, which we found very helpful.

Q1. Computational efficiency and scaling with the number of channels. Channels do not enter the circuit. They enter only through the source-fitted projection P∈RC×qP\in\mathbb R^{C\times q} (xDAWN for P300, CSP for MI), applied once per trial at cost O(CqT)O(CqT). Every later stage (node encoder, q×qq\times q covariance and tangent map, graph, circuit) depends on qq alone, and source validation selected q=8q=8 in all 30 folds although the cohorts have 16/8/32/19 channels (bigP3BCI/BNCI/EEGET/UCL). At q=8q=8 the circuit has 12 edges, 72 gates (24 two-qubit RZZR_{ZZ}), 52 circuit parameters, and 20 commuting Z/ZZ readouts. The full model has 695 trainable parameters, fewer than EEGNet (1,242), EEG Conformer-style (1,200), QEEGNet-style (1,378), and CRGNet-style (2,970).

Measured cost. On one NVIDIA RTX A6000, a complete final fit (all source patients, three-stage schedule) plus held-out prediction takes a median of 2,867 s for QRG-EEG and 1,926 s for the parameter-matched classical B7, with 89 MB and 88 MB peak GPU memory; the difference mainly reflects the exact 2q2^q-dimensional statevector simulation. In a controlled benchmark of the deployed architecture, inference takes 0.203 ms per trial with 132 MB peak GPU memory at q=8q=8. Increasing qq from 4 to 12 raises this from 0.107 to 0.308 ms, whereas increasing CC from 8 to 256 raises the projection cost only from 0.0011 to 0.0032 ms per trial. Exact simulation therefore scales as 2q2^q in qq but not in CC, and the gate count and readout size grow linearly in qq.

Revision: App. M (runtime, memory, scaling; Tab. 21), Sec. 5 and Tab. 7 (channels).

W2. Realistic quantum-computing settings. We agree that this distinction matters. Our scope is a quantum-enhanced learning framework: we test, at the simulation level, whether structured, data-conditioned entangling interactions improve graph representation learning for cross-patient ALS-EEG, and we make no claim of hardware readiness. Within this scope:

  1. All readout observables {Zi}∪{ZiZj}\lbrace Z_{i}\rbrace \cup \lbrace Z_{i}Z_{j}\rbrace commute, so one computational-basis setting with SS shots per trial estimates all 20 coordinates, each with variance ≤1/S\le 1/S.
  2. For the trained QRG-EEG models of all 30 folds, replacing the exact readout by SS Born-rule shots changes Macro-F1 by .000/−.006/−.005/−.005 at S=1024S=1024 and by −.001 on every cohort at S=4096S=4096.
  3. The circuit is shallow: two blocks of mutually commuting RZZR_{ZZ} gates on a 12-edge graph, read out through weight-≤2\le 2 observables.

Finite-shot readout thus shows measurement-budget robustness at the simulator level; hardware noise and hardware execution remain future work.

Revision: App. L (Tab. 20), Sec. 5, Sec. 6, Limitations.

W1. Small cohorts and clinical relevance. Small cohorts are intrinsic to ALS rather than a choice of this study. ALS is a rare, progressive disease; recording sessions are limited by fatigue and motor decline, and longitudinal follow-up becomes difficult as the disease advances. Each public ALS-EEG cohort therefore contains 6–8 patients. Generalizing to an unseen patient under this scarcity is the methodological problem we address, and it remains under-explored. Within that setting, we treated the patient as the unit of evidence: every score comes from a patient excluded from all fitting, selection, and calibration, across 30 patients, four cohorts, and two paradigms.

  • Against the strongest non-QRG model per cohort, QRG-EEG leads by +.040 (BNCI), +.064 (EEGET), and +.125 (UCL) Macro-F1 and wins every patient (8/8, 6/6, 8/8); on bigP3BCI the XdawnCov+TS+LR pipeline is stronger.
  • Against B7 (0.6% parameter gap), the margin is positive on all four cohorts (+.054/+.108/+.086/+.125), with 95% patient-bootstrap CIs excluding 0.
  • The source-only configuration was identical in every fold of a cohort.

We therefore see the contribution as a methodological advance for data-limited cross-patient ALS-EEG decoding, not a clinically validated system. The Limitations position it this way, and online BCI use, information-transfer rate, and prospective multicenter validation are listed as next steps.

Revision: Tab. 1, Tab. 11, Tab. 12, Sec. 5, Limitations.

We simplified the Figs. 1–2 and the notation in Sec. 4. Thank you again for the encouraging and constructive review.

Reviewer_JEAxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes QRG-EEG for cross-patient ALS-EEG decoding. The method combines component-wise EEGNet features with Riemannian covariance representations to construct an EEG graph, where node features and pairwise relations are mapped to corresponding single-qubit and two-qubit operations in a variational quantum circuit. The authors also analyze how the edge-conditioned interactions affect the circuit measurements. Experiments are conducted on four ALS-EEG datasets using patient-level leave-one-subject-out evaluation, with comparisons against classical and quantum baselines and several targeted ablation studies.

Critical strengths

The component-aligned node–edge conditioning is an interesting design that explicitly incorporates Riemannian relations into the quantum circuit, rather than simply using a quantum classifier on flattened EEG features.

The experimental evaluation is reasonably comprehensive. The patient-level LOSO protocol, source-only model selection, and controlled ablations provide useful evidence for the proposed approach. The method also shows consistent improvements over the main reported baselines across all four datasets.

Critical weaknesses

The methodological novelty and the benefit of quantum interactions over classical alternatives are not fully established. The main classical interaction baseline (B7) uses a relatively restricted form of node–edge interaction, while the additional nonlinear classical controls in Appendix G are evaluated on only two folds per dataset. It remains unclear whether a more expressive classical model using the same aligned node and edge information could achieve comparable performance.

The comparison with stronger modern EEG models is limited. In Appendix I, EEG Conformer-style achieves a Macro-F1 of 0.501 compared with 0.516 for QRG-EEG on EEGET-ALS-MI. This relatively small difference, together with the limited patient cohorts, makes the practical advantage of the proposed method less convincing. A broader evaluation against competitive classical architectures would strengthen the empirical contribution.

Questions for authors

Could the authors consider comparing QRG-EEG with more expressive classical edge-conditioned models, such as Edge-Conditioned Convolution (ECC/NNConv), GINE, or an MLP with explicit pairwise node–edge interactions? Although B6 and B7 provide useful classical controls, these additional comparisons using the same node/edge inputs and training protocol would help clarify whether the improvements arise from the proposed interaction structure or specifically from its quantum implementation.

In the fixed-entanglement ablation (Appendix B.2.4), both the graph topology and the trial-specific edge conditioning may be changed. This makes it difficult to determine which factor contributes to the performance degradation. Could the authors include an additional control that preserves the original graph topology while using fixed edge values? This would help isolate the contribution of trial-specific edge conditioning.

1 comment

We sincerely thank the reviewer for the thoughtful assessment and the concrete suggestions. Signed values are QRG-EEG minus comparator Macro-F1 on bigP3BCI/BNCI/EEGET/UCL (seeds averaged within patient); unsigned values are absolute scores. Every result uses all 30 patient-level LOSO folds, and all 95% patient-bootstrap intervals reported in Tab. 11–12 exclude 0.

W1 / Q1. Classical models with the same aligned node–edge inputs. Under exact simulation the circuit is a classical function of (v,e)(v,e) (App. B.3); our claim is the inductive bias of the aligned interaction, which we test against classical parameterizations with identical inputs, objective, source-only selection, and calibration.

Classical controls. B6 (edge-aware GAT, identical graph), B7 (node–edge products, 0.6% parameter gap), and B12 (circuit-derived trigonometric surrogate, more parameters than QRG-EEG):

bigP3BCIBNCIEEGETUCL
B6+.076+.229+.161+.185
B7+.054+.108+.086+.125
B12+.064+.200+.088+.152

On the same frozen node–edge features (App. G), finite-RFF and SVM-RBF controls also remain below QRG-EEG on every cohort.

The suggested edge-conditioned models. NNConv/ECC (Gilmer et al., 2017; Simonovsky & Komodakis, 2017), GINE (Hu et al., 2020), and a pairwise MLP ϕ(vi,vj,eij)\phi(v_i,v_j,e_{ij}) shared across all aligned edges (Graph-Network edge model; Battaglia et al., 2018) use the same viv_i, trial-specific eije_{ij}, 8-node/12-edge graph, objective, schedule, search budget, and calibration, parameter-matched to QRG-EEG within 10% (Macro-F1):

bigP3BCIBNCIEEGETUCL
QRG-EEG.512.699.524.583
NNConv/ECC.508.642.450.334
GINE.498.636.398.287
Pairwise MLP.482.641.417.378

QRG-EEG obtains the highest Macro-F1 and AUROC on every cohort against all three models; margins are +.004 to +.030 on bigP3BCI and +.057 to +.296 on the other cohorts.

Where the structure differs. Each ⟨Zi⟩\langle Z_i\rangle contains products of sin⁡βij\sin\beta_{ij}/cos⁡βik\cos\beta_{ik} over the whole neighbourhood N(i)\mathcal N(i), coupled to the node angles (App. B.8, Eq. 56) and generated by 52 circuit parameters; B7 and B12 enumerate only a truncated subset.

Same Riemannian information without pair alignment. A generic VQC on the vectorized tangent matrix (C9) gives +.083/+.148/+.066/+.106, and test-time edge permutation (C8) +.082/+.132/+.052/+.077. Retraining on consistently permuted edges (C7) keeps a clear gap on P300 (+.058/+.104) and a smaller one on MI (+.031/+.050), where the benefit comes mainly from trial-specific relations (C5).

Revision: Tab. 12 (App. D.2), App. G (Tab. 15), App. J (Tab. 18), App. B.8, Sec. 5 (RQ1).

W2. Stronger modern EEG models. Thank you for pointing to App. I. The revised Tab. 1 and App. I report the final runs of all models, including QRG-EEG; in these runs EEG Conformer-style obtains .429 against .524 for QRG-EEG on EEGET (Tab. 17). EEG Conformer-style (B10), CRGNet-style (B11), and Riemannian pipelines (B8: XdawnCov+TS+LR for P300, FBCSP+LDA for MI) were evaluated on all 30 folds with the same source-only calibration. QRG-EEG minus EEGNet (B2) is +.074/+.161/+.121/+.171, minus Conformer +.069/+.230/+.095/+.150, minus CRGNet +.051/+.155/+.064/+.174, and minus B8 −.049/+.040/+.162/+.199. QRG-EEG has the highest Macro-F1 and paradigm-appropriate secondary metric (AUROC for P300, BAcc for MI) on BNCI, EEGET, and UCL, where it exceeds the strongest non-QRG model by +.040 (B8), +.064 (B11), and +.125 (B7) for every patient (8/8, 6/6, 8/8; Tab. 11); on bigP3BCI, XdawnCov+TS+LR is stronger.

Revision: Tab. 1, Fig. 3, Tab. 11, Fig. 6, App. I.

Q2. Same topology with fixed edge values. Tab. 2 separates node-only, edge-only, fixed-entanglement, shuffled-edge, and vectorized-geometry settings. The requested control is C5: the source-fitted topology and gates are kept and each trial's edge values are replaced by their source means; C6 changes only the topology. Both keep QRG-EEG's architecture and parameter count:

bigP3BCIBNCIEEGETUCL
C5+.076+.152+.061+.082
C6+.065+.120+.033+.072

Removing trial dependence alone lowers performance on all four cohorts, by more than changing the topology, so trial-specific edge conditioning contributes independently of topology.

Revision: Tab. 2, Tab. 5, Tab. 12, Fig. 5, Sec. 5 (RQ2).

Reviewer_VsS4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QRG EEG which is a quantum learning framework for EEG classification in ALS patients. The model combines local EEG features extracted by a neural network with Riemannian relations between EEG components. These features are encoded into single qubit rotations and two qubit interactions in a variational quantum circuit. The authors also provide a theoretical analysis of how node and edge information jointly affects quantum measurements. The method is evaluated on four ALS EEG datasets using 30 patient based LOSO folds. The results show improved classification performance compared with several classical and quantum baselines, supported by ablation experiments. All quantum experiments are performed using exact statevector simulation.

Critical strengths

1.Strong experimental evaluation. The paper tests four ALS EEG datasets with 30 LOSO patient folds. It includes several classical and quantum baselines and uses source patient data for model selection. This makes the results more convincing.

  1. Clear model design and theoretical explanation. The proposed model connects EEG features, graph relations, and quantum operations in a structured way. The theoretical analysis also explains how node and edge information can jointly affect quantum measurements.
Critical weaknesses
  1. Limitations in the ablation experiments. The fixed entanglement ablation changes both the graph topology and the input dependent edge angles. Therefore, it is unclear which change causes the performance drop. Also, the effects of selecting different spatial projections and qubit counts across LOSO folds are not fully reported. These issues make it difficult to identify the main source of the performance improvement.

  2. Lack of finite shot estimation. The shot budget is an important indicator of the measurement efficiency of a quantum algorithm. However, all reported results rely on exact expectation values from simulation. We suggest that the authors provide a finite shot analysis to evaluate the measurement cost and determine whether the reported performance gains can be maintained with a limited number of shots.

Questions for authors
  1. Spatial projection and qubit selection Could the authors report how often CSP, PCA, and each qubit count were selected across the LOSO folds? Would the performance gains remain if the same spatial projection and qubit count were used across models?
1 comment

We sincerely thank the reviewer for the careful reading and for two questions that helped us make the controls cleaner. Signed values are QRG-EEG minus comparator Macro-F1 on bigP3BCI/BNCI/EEGET/UCL (seeds averaged within patient; all 30 patient-level LOSO folds; 95% patient-bootstrap intervals in Tab. 12, all excluding 0); unsigned values are absolute scores.

W1(a). Entanglement schemes. We provide additional single-factor ablations that separate topology from trial dependence, comparing entanglement schemes on the same projection PP, q=8q=8, and readout (C5 and C6 keep QRG-EEG's architecture and parameter count; C3 removes the two-qubit gates):

Entanglement schemebigP3BCIBNCIEEGETUCL
No two-qubit gates (C3)+.070+.151+.055+.089
Same graph and gates, source-mean edges (C5)+.076+.152+.061+.082
Random graph, trial-specific edges (C6)+.065+.120+.033+.072
QEEGNet-style ring-CNOT VQC (B4)+.058+.129+.106+.180

Replacing each trial's edge values by their source means (C5) lowers performance on every cohort, so the data-conditioned RZZR_{ZZ} gates contribute beyond entanglement itself; changing only the topology (C6) also lowers it on every cohort, by less than C5.

Revision: Tab. 2 and 5 (C3–C6), Tab. 12, Fig. 5, Sec. 5 (RQ2).

W1(b) and Q1. Projection method and qubit count. CSP/PCA is a choice only for MI; for P300 the projection is xDAWN, and source validation chooses its covariance input. Over the 30 outer folds the selections were identical within each cohort: ERP-residual 16/16 (plain 0/16), CSP 14/14 (PCA 0/14), q=8q=8 30/30. The reported results are therefore not a mixture of per-patient configurations.

Same projection and qubit count across models. Within each fold, C3, B6, B7, and B12 use QRG-EEG's fold object (PP, q=8q=8, graph), C9 the same PP and qq, and B4 the same qq. C9 (vectorized tangent geometry → generic VQC) gives +.083/+.148/+.066/+.106, B6 (GAT on the same graph) +.076/+.229/+.161/+.185, B7 (node–edge products, 0.6% parameter gap) +.054/+.108/+.086/+.125, and B12 (circuit-derived trigonometric features) +.064/+.200/+.088/+.152. C2 applies EEGNet to the same P⊤XP^\top X; QRG-EEG remains ahead on all four cohorts (+.055/+.103/+.040/+.060).

Forcing the non-selected option (same folds, seeds, and calibration; Macro-F1):

  • q=6q=6 lowers QRG-EEG on every cohort, in line with q=8q=8 being selected in every fold.
  • With the selected representation, QRG-EEG leads B7 by +.054/+.108/+.086/+.125. Forcing the non-selected one (plain covariance for P300, PCA for MI) on both models gives QRG-EEG .492/.665/.485/.502 and B7 .475/.638/.425/.462. QRG-EEG thus remains above B7 under both representations, so its advantage does not hinge on one projection, while source-only validation identifies the stronger one for QRG-EEG.

Revision: Sec. 5 (Data and protocol), App. F (Tab. 14), App. K (Tab. 19), Tab. 12.

W2. Finite-shot measurement. All readout observables {Zi}∪{ZiZj}(i,j)∈E\lbrace Z_{i}\rbrace \cup \lbrace Z_{i}Z_{j}\rbrace_{(i,j)\in\mathcal{E}} commute and are diagonal in the computational basis, so SS shots per trial estimate all 20 coordinates from the same bitstrings, each with variance ≤1/S\le 1/S; for the linear head, σ(ok)≤∥Wc,:k∥1/S\sigma(o_k)\le\lVert W_{c,:k}\rVert_1/\sqrt S. For the trained QRG-EEG models (all folds, three seeds), we replaced the exact readout by SS Born-rule samples (5 sampling repeats). Change relative to the exact readout of the same model (ΔMacro-F1 / ΔAUROC):

Shots SSbigP3BCIBNCIEEGETUCL
64+.009 / −.011−.041 / −.045−.035 / −.039−.038 / −.042
256+.003 / −.002−.019 / −.021−.016 / −.019−.018 / −.020
1024.000 / .000−.006 / −.007−.005 / −.006−.005 / −.007
4096−.001 / .000−.001 / −.002−.001 / −.001−.001 / −.002

From S=1024S=1024, Macro-F1 changes by at most .006 on every cohort (.001 at S=4096S=4096); S=256S=256 stays within .019, whereas S=64S=64 shows noticeable shot noise (up to .041 on BNCI). This is measurement-budget robustness at the simulator level; hardware noise and hardware execution remain future work.

Revision: App. L (Tab. 20), Sec. 5, Limitations.

We are grateful for these suggestions, which made the ablations more precise.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.