{"id":"27f8780e-717e-4537-a0ff-0810baf1df3e","arxiv_id":"2505.10786","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A frequency-domain MIMO model with spatial and temporal regularization estimates the ECoG-to-EEG brain channel on simultaneous macaque recordings, claiming lower reconstruction error than LS and MMSE.","lead":"The paper models the path from brain-surface ECoG signals to scalp EEG signals as a wireless MIMO communication channel and estimates it with spatially and temporally regularized least squares. It reports roughly 15 percent lower EEG reconstruction error than least squares or MMSE baselines on macaque data, while describing a trade-off in choosing the segment length.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 15% improvement is likely an artifact of an undocumented evaluation protocol: in-sample reconstruction error cannot favor a regularized estimator over LS on the same data.","rationale":"The reader's second listed weakest assumption is exactly the in-sample reconstruction metric issue, and I agree that it is the most load-bearing part of the paper. The brain channel model's linearity and stationarity are modeling assumptions that one could debate, but the evaluation problem is decisive: the reported advantage is inconsistent with the mathematics of least squares unless a different evaluation setup was used. The paper does not report a train/test split, hyperparameters, or the relevant dimensions (K and M), so the experimental section cannot be independently reproduced. The spatial/temporal regularization framework is coherent as an optimization problem, and the U-shaped symbol-length curve is a falsifiable observation that roughly matches the stated trade-off; these are useful contributions. However, the central performance claim—that STARE beats LS and MMSE by about 15% in MSE—cannot be true under the metric as written, because LS is the exact minimizer of the data term. If the advantage was measured on held-out data, the paper must say so, and then the claim would be about generalization, not about in-sample channel reconstruction. This concern is sufficient to justify the reader's REJECT verdict; no change to that verdict is needed. The authors could address it by releasing code and a precise evaluation protocol, but absent that, the central claim remains unsupported.","tokens_in":7484,"tokens_out":7225,"duration_ms":79950,"concrete_test":"Recompute the comparison in Fig. 4 using the exact protocol described in Section III, without adding any hidden split: for each frequency f and frame k, compute the LS solution H_LS = Y_k X_k^H (X_k X_k^H)^{-1} (or the minimum-norm solution if X_k X_k^H is singular) and the STARE minimizer of Eq. (5); then evaluate Eq. (13) on the same (f, k). If STARE's in-sample MSE is lower than LS's for any configuration, the evaluation code or data-handling procedure differs from the described method. If STARE's in-sample MSE is never lower, the 15% advantage must come from a held-out evaluation, and the paper must disclose that split and report the held-out MSE. This single check determines whether the central claim reflects a real estimation advantage or an artifact of the evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central performance claim is unsupported because the evaluation metric in Eq. (13) appears to be computed as in-sample reconstruction error on the data used to fit the channel. For a fixed frame k and frequency f, the unregularized LS estimator minimizes the data term \\|Y_k - H X_k\\|_F^2; any STARE solution that also minimizes this term plus nonnegative spatial and temporal penalties cannot achieve a strictly smaller residual on the same data. If the least-squares problem is underdetermined, the minimum residual is attained by a set of solutions, and the regularized estimator can match it but not beat it. The paper reports a 15.4% improvement of STARE over LS (Fig. 4), which is mathematically impossible if the metric is genuinely in-sample. This implies either an undocumented train/test split, a different evaluation protocol (e.g., averaging over symbol lengths with different data-composition properties), or an error in the simulation pipeline. The paper does not specify the number of frames K, the number of symbols per frame M, the values of the regularization parameters (μ, ν, ρ, t_{max}), or any separation between training and test data, so the reported numbers cannot be reproduced or independently checked. The temporal regularization in Eq. (3) couples H_k and H_{k+1}, but if the objective is minimized jointly and the metric is computed on the same frames, the sum of LS residuals remains the minimal achievable data-term value; a regularized joint solution cannot lower it. Thus the headline 15.4%/15.3% improvement rests on an evaluation protocol that is either unspecified or internally inconsistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a frequency-division MIMO framework for modeling the 'brain wireless channel' from intracranial ECoG to scalp EEG, formalized as Y = H X + N. It introduces STARE, a regularized estimator with spatial and temporal smoothness penalties solved via ADMM, and validates it on the public MDR macaque dataset. The evaluation metric is the reconstruction error between simulated EEG (estimated channel times ECoG) and recorded EEG (Eq. 13). The paper reports 15.4% and 15.3% lower MSE than LS and MMSE, respectively, and claims a U-shaped trade-off between frequency resolution and temporal stability with an optimal symbol length L = 33000 samples.","tokens_in":7834,"tokens_out":11432,"duration_ms":96586,"significance":"The conceptual analogy to MIMO communication and the coherent ADMM formulation are strengths, as is the use of a public dataset and the explicit statement of a falsifiable model. However, the central empirical claim is unsupported: the reported improvement over LS is inconsistent with the in-sample reconstruction metric, and the symbol-length trade-off is validated circularly. The mathematical derivation of the spatial penalty also contains an error. If the evaluation were corrected with proper train/test separation and a valid channel-error proxy, the framework could be of interest to the BCI and communications communities, but as presented the results do not establish the claims.","major_comments":[{"comment":"The reported 15.4% improvement of STARE over LS is incompatible with the described evaluation protocol. For a fixed frame k and frequency f, the LS estimator minimizes exactly the data term \\|Y_k - H_k X_k\\|_F^2; since Eq. (3) defines STARE's objective as that same term plus nonnegative spatial and temporal penalties, any STARE solution has a data term no smaller than the LS minimum. Because Eq. (13) computes that same data term on the data used for estimation, STARE cannot achieve a strictly lower MSE than LS. The manuscript reports the opposite, which implies an undocumented train/test split, a different evaluation rule, or an error in the simulation pipeline; none of these is described.","section":"Section III.A, Eq. (13), Fig. 4"},{"comment":"The optimal symbol length L_opt is selected as the minimum of the in-sample MSE curve (Fig. 3), and this same curve is then presented as 'the precise alignment between empirical evidence and theoretical analyses' validating the U-shaped trade-off. This is circular: the MSE curve is used both to define L_opt and to confirm that L_opt is optimal. A falsifiable test would require deriving L_opt from the stated mechanisms (frequency resolution and temporal stationarity) and evaluating it on held-out data or with an independent metric.","section":"Section III.B, Figs. 3 and 4"},{"comment":"The identity converting the edge-wise spatial penalty to a trace expression is incorrect. For an undirected graph with Laplacian L = D - A (which should be N x N, where N is the number of EEG electrodes, not P x P as stated), the correct identity is \\sum_{(i,j)\\in E} \\|H_k(i,:)-H_k(j,:)\\|_F^2 = 2 \\mathrm{Tr}(H_k^\\top L H_k). The expression in Eq. (4), 2 \\mathrm{Tr}(H_k^\\top L^\\top L H_k), penalizes squared differences of the Laplacian and does not correspond to the physiologically motivated neighborhood-similarity constraint. Consequently, the ADMM updates in Eqs. (5), (9), and (10) do not implement the penalty described in Eq. (3).","section":"Section II.B, Eq. (4)"},{"comment":"The experimental description omits values needed for reproduction: the number of frames K, symbols per frame M, the regularization parameters \\mu, \\nu, \\rho, the ADMM iteration count t_max, the initialization of H_0 for the first frame, and the exact MMSE implementation (noise covariance and channel correlation assumptions). The manuscript also does not state whether any held-out data were used. Without these, the numerical comparisons in Figs. 3 and 4 cannot be independently verified, and the reported MSE values (5.44 vs 6.44 and 6.42) are not reproducible.","section":"Sections II.C and III"}],"minor_comments":[{"comment":"The title states 'EEG-to-ECoG' but the model and experiments treat ECoG as the transmitting source and EEG as the receiving array; the direction in the title should be 'ECoG-to-EEG'.","section":"Title"},{"comment":"The text contains grammatical errors, e.g., 'This work establish' and 'the complex physiology of brain challenges'; these should be corrected throughout.","section":"Abstract and Introduction"},{"comment":"References [16] and [17] are listed but never cited in the text; the macaque-similarity paragraph cites only [14], [15], [18], and [19].","section":"References"},{"comment":"The description of Fig. 3 does not specify axis labels or units; stating the time durations corresponding to the sample lengths (e.g., 13, 33, 49 seconds at 1 kHz) would improve readability.","section":"Fig. 3"},{"comment":"The symbol M is used for the number of symbols per frame and also appears in the acronyms MSE and MMSE; while context makes the meaning clear, the repeated letter in Eq. (2) and in the evaluation metric may confuse readers.","section":"Notation"}],"recommendation":"reject","confidential_remarks":"I concur with the reader's assessment. The in-sample metric issue is not a matter of interpretation: a regularized estimator cannot outperform LS on the same data in the metric being reported, so the central quantitative claim is unsound. The graph Laplacian error in Eq. (4) adds a further technical defect. The manuscript would need a fundamentally different evaluation protocol and a corrected derivation before it could be considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper is a cleanly written application of MIMO channel estimation to simultaneous ECoG/EEG recordings from the public MDR macaque dataset. The formulation is standard: frequency-domain linear model Y = H X + N, with a graph-Laplacian spatial smoothness penalty and a temporal continuity penalty, solved by ADMM. The ADMM updates are correct and the presentation is clear. That part is fine.\n\nThe problem is the evaluation. The metric in Eq. (13) is in-sample reconstruction error: they fit H on the same data they then reconstruct. For a fixed frame and frequency, the LS estimator minimizes that exact data term. Any regularized estimator that also fits the data plus nonnegative penalties cannot achieve a strictly lower residual on the same data. So STARE's reported 15.4% improvement over LS is mathematically impossible under the stated protocol. The paper gives no train/test split, no held-out frames, no hyperparameter values, and no details on how the average over symbol lengths is computed. Either the comparison uses a different protocol or there is an error in the pipeline. As written, the central quantitative claim is unsupported.\n\nThe U-shaped trade-off between symbol length and MSE is also weakened. The optimal L is picked from the same curve used to validate the U-shape, which is circular, and the same in-sample problem applies.\n\nWhat is genuinely new is the combination of spatial-temporal regularization with this dataset and the conceptual framing of the brain as a MIMO channel. That idea might be worth pursuing, but the current evidence does not establish it. The authors should redo the evaluation with a proper cross-validation split (e.g., estimate on some frames, test on others) and report regularization parameters.\n\nThis is a solid candidate for a major revision rather than an outright kill, but only if the evaluation is fixed. As it stands, the paper should not be accepted.\n\nBring it to a reading group? Maybe, as a cautionary example of in-sample evaluation. Cite it? No.\n\nThe thinking is clear and the math is honest, so I'd say serious_thinker = yes. A serious editor should send it to peer review because the flaw is subtle and the underlying framework is not obviously wrong.","headline":"Clean MIMO estimation formulation but the headline 15% improvement over LS is mathematically impossible under the stated in-sample MSE metric; the empirical claims need a proper train/test split.","tokens_in":8345,"tokens_out":3172,"would_cite":false,"duration_ms":31367,"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":"A MIMO channel model for the brain link cuts EEG reconstruction error 15 percent","keywords":["brain-computer interface","MIMO channel estimation","ECoG-to-EEG channel","wireless brain channel","spatial-temporal regularization","ADMM","frequency resolution trade-off","simultaneous EEG-ECoG recordings"],"falsifier":"Use the estimated channel from one macaque session to predict EEG from held-out ECoG of another session or another animal; if reconstruction MSE is no better than using a channel estimated from shuffled electrode labels or a constant baseline, the MIMO channel model is not capturing real structure. An even more direct check is to compare STARE's prediction against a nonlinear reconstruction that allows cross-frequency or time-varying coupling on the same data: if the nonlinear model wins by a large margin, the linear instantaneous-mixing assumption fails.","tokens_in":1654,"feed_emoji":"🧠","tokens_out":4751,"duration_ms":100715,"temperature":0.7,"pith_summary":"This paper argues that the path from intracranial ECoG signals to scalp EEG electrodes can be treated as a frequency-division multiple-input multiple-output (MIMO) wireless channel, with ECoG arrays as transmitters and EEG electrodes as receivers. It proposes STARE, a channel estimator that adds spatial smoothness and temporal continuity constraints derived from neurophysiology to the usual data-fitting objective, and solves the resulting problem with an alternating-direction method. Using simultaneous macaque ECoG and EEG recordings, the paper reports that STARE lowers reconstruction mean-squared error by about 15 percent relative to least-squares and MMSE estimators, and that reconstruction error follows a U-shaped curve in the segmentation symbol length, with an optimum near 33,000 samples. If true, this would let communication-theoretic tools such as channel estimation, equalization, and modulation be applied to brain-computer interfaces.","feed_headline":"MIMO model for brain signals cuts EEG error 15%","feed_subtitle":"Treating cortical signals as transmitters and scalp electrodes as receivers opens BCI decoding to wireless tools.","key_machinery":"The central object is the frequency-division MIMO channel matrix $H_k^{(f)}\\in\\mathbb{C}^{N\\times P}$ mapping $P$ ECoG sources to $N$ EEG electrodes at frequency $f$ in frame $k$. STARE (Spatial-Temporal Adaptive Regularized Estimation) is the algorithm that estimates it: a single objective with three terms — the Frobenius-norm data fit $\\|Y-HX\\|_F^2$, a spatial smoothness penalty expressed through the graph Laplacian $L=D-A$ of the electrode adjacency graph, and a temporal continuity penalty $\\|H_k-H_{k-1}\\|_F^2$. The two regularizers are coupled through $H$, so the paper introduces an auxiliary variable $G$ and solves with ADMM, yielding closed-form $H$ and $G$ updates per iteration and carrying the previous frame's estimate forward as initialization. This machinery is what lets physiological priors be injected into channel estimation.","core_discovery":"The paper's central claim is that a frequency-division MIMO model describes the brain channel: after segmenting simultaneous ECoG and EEG into symbols and taking discrete Fourier transforms, each frequency bin obeys $Y_k^{(f)}=H_k^{(f)}X_k^{(f)}+N$, where $X$ is the ECoG transmit matrix, $Y$ the EEG receive matrix, and $H$ the unknown channel matrix. The contribution is the STARE estimator, which minimizes the squared reconstruction error plus two neurophysiology-informed penalties: a spatial-smoothness term over neighboring electrodes written with a graph Laplacian, and a temporal-continuity term tying successive frames. Because the constraints couple the unknowns, the paper solves the objective with ADMM and gives closed-form updates for each block. On a public simultaneous macaque EEG-ECoG dataset, STARE achieves an average MSE of 5.44 across symbol lengths, versus 6.44 for LS and 6.42 for MMSE, i.e. reductions of 15.4 percent and 15.3 percent; at the optimal symbol length of 33,000 samples the average MSE reaches 0.55. The reported U-shaped dependence of MSE on symbol length is presented as evidence of a trade-off between frequency resolution and temporal stationarity.","pith_inferences":["If the linear MIMO model holds, the estimated channel matrices encode volume-conduction geometry and might be invertible for source-localization or tissue-anomaly detection, an application the paper does not develop.","Because the reported MSE is computed on the same data used to fit the channel, an out-of-sample or cross-session validation would be the natural next test; if it passes, the channel estimate is a genuine transfer function.","The per-frequency independence assumption suggests an immediate extension: a block-diagonal or coupled-frequency model could capture cross-frequency neural interactions that the current framework treats as noise.","The macaque results justify but do not guarantee human transfer; the same pipeline could be applied directly to any future simultaneous human ECoG-EEG recording."],"forward_implications":["Brain channel matrices could be used as a forward model for BCI decoders, replacing ad hoc spatial filters with communication-style equalization.","The symbol-length trade-off gives a concrete rule for choosing segmentation windows: near 33,000 samples at 1 kHz in this setup, balancing spectral resolution against stationarity.","The framework turns EEG reconstruction error into a channel-estimation metric, so future BCI algorithms can be benchmarked against LS and MMSE baselines.","Spatial and temporal regularization, if validated, provides a principled way to suppress physiological artifacts without discarding data.","The same estimation pipeline could support future human ECoG-EEG recordings if simultaneous data become available."],"supporting_citations":[{"why":"Supplies the massive MIMO formalism and channel-matrix language on which the FD-MIMO model is built.","marker":"[10]"},{"why":"Provides the ADMM framework used to solve the coupled spatial-temporal regularized estimation objective.","marker":"[13]"},{"why":"Provides the only public whole-cortex simultaneous macaque EEG-ECoG recordings used for experimental validation.","marker":"[19]"},{"why":"Supports the claim that macaque skull anatomy is close enough to human to make the animal model a valid testbed.","marker":"[14]"},{"why":"Supports the claim that primate cortical-tissue electrophysiology resembles human, justifying transfer of the channel model.","marker":"[15]"}],"fun_headline_variants":["MIMO framework maps brain signals as wireless channel","EEG-to-ECoG modeled as MIMO: 15% error drop","Brain signals treated as MIMO: EEG error down 15%","Frequency-domain MIMO cuts EEG reconstruction error","STARE estimator: MIMO for EEG-ECoG, 15% better"],"cache_read_input_tokens":10496,"weakest_assumption_plain":"The load-bearing premise is that scalp EEG is a linear, frequency-by-frequency mixture of ECoG signals plus noise — $Y=HX+N$ with $H$ constant within each short segment — so that a single channel matrix per frequency and frame captures the physiology; if volume conduction is nonlinear or the two recordings are not a clean transmitter-receiver pair, the estimated $H$ is not a physical channel.","fun_headline_variants_meta":{"raw":{"variants":["MIMO framework maps brain signals as wireless channel","EEG-to-ECoG modeled as MIMO: 15% error drop","Brain signals treated as MIMO: EEG error down 15%","Frequency-domain MIMO cuts EEG reconstruction error","STARE estimator: MIMO for EEG-ECoG, 15% better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1874,"prompt_tokens":1011,"completion_tokens":863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":783}},"tokens_in":627,"tokens_out":863,"duration_ms":7032,"temperature":1.0,"reasoning_tokens":783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:04:18.314809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the estimated channel from one macaque session to predict EEG from held-out ECoG of another session or another animal; if reconstruction MSE is no better than using a channel estimated from shuffled electrode labels or a constant baseline, the MIMO channel model is not capturing real structure. An even more direct check is to compare STARE's prediction against a nonlinear reconstruction that allows cross-frequency or time-varying coupling on the same data: if the nonlinear model wins by a large margin, the linear instantaneous-mixing assumption fails.","supporting_citations":[{"cited_title":"A survey on some recent developments of alternating direction method of multipliers,","cited_arxiv_id":null,"evidence_quote":"Provides the ADMM framework used to solve the coupled spatial-temporal regularized estimation objective."},{"cited_title":"A new method for quantifying the performance of eeg blind source separation algorithms by referencing a simultaneously recorded ecog signal,","cited_arxiv_id":null,"evidence_quote":"Provides the only public whole-cortex simultaneous macaque EEG-ECoG recordings used for experimental validation."},{"cited_title":"Ct anatomy and normal radiography of the skull of the rhesus monkey (macaca mulatta),","cited_arxiv_id":null,"evidence_quote":"Supports the claim that macaque skull anatomy is close enough to human to make the animal model a valid testbed."},{"cited_title":"Resolving the mesoscopic missing link: Biophysical modeling of eeg from cortical columns in primates,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that primate cortical-tissue electrophysiology resembles human, justifying transfer of the channel model."}],"review_version":1}