Pith. sign in

REVIEW 2 major objections 2 minor 39 references

A stochastic score matching procedure reduces Torus Graph inference cost from O(d^6) to O(d^2) for thousands of phase variables.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-28 19:06 UTC pith:Y6XJDODR

load-bearing objection Stochastic score matching scales torus graphs to thousands of variables, but the approximation's effect on estimate quality needs direct checks. the 2 major comments →

arxiv 2606.00496 v1 pith:Y6XJDODR submitted 2026-05-30 cs.LG

Torus Graphs for Large Scale Neural Phase Analysis

Jack Goffinet , Casey Hanks , David E. Carlson This is my paper
classification cs.LG
keywords Torus Graphphase couplingneural oscillationsscore matchinghidden Markov modelautoregressive modellocal field potentialscircular statistics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a scalable version of the Torus Graph model for analyzing phase relationships in large-scale neural recordings. By introducing a stochastic score matching procedure, it reduces the computational cost of inference from O(d^6) to O(d^2) per iteration. This change makes it feasible to model 1,860 frequency-phase features extracted from multi-electrode local field potentials. The resulting framework supports two new model extensions: a hidden Markov model that captures changes in phase coupling across different brain states and an autoregressive model that infers directional influences between oscillations. Application to sleep recordings shows distinct phase interaction patterns during wakefulness versus NREM sleep.

Core claim

We introduce a stochastic score matching procedure that reduces the per-iteration cost to O(d^2), enabling inference on datasets with thousands of variables. This scalable foundation supports analyses of 1,860 frequency-phase features from multi-electrode LFPs and enables two extensions previously inaccessible to TGs or classical circular statistics: (i) a TG Hidden Markov Model capturing state-dependent phase-coupling changes (e.g., spindle-related states during sleep) and (ii) an autoregressive TG inferring directional interactions via transfer-entropy estimation. Applied to LFP recordings, these models reveal state-dependent phase-interaction patterns between wakefulness and NREM sleep.

What carries the argument

The stochastic score matching procedure that approximates the full score matching objective for the exponential-family Torus Graph distribution over phases in O(d^2) time per iteration.

Load-bearing premise

Neural phase data fits the exponential-family Torus Graph model well enough that the stochastic approximation does not introduce substantial errors in the estimated parameters.

What would settle it

A side-by-side run of the original O(d^6) score matching and the new stochastic version on a dataset of roughly 100 variables that yields substantially different parameter values or fails to recover known structure would show the approximation is inaccurate.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Inference on datasets with thousands of variables becomes computationally practical.
  • A Torus Graph Hidden Markov Model can capture state-dependent changes in phase coupling.
  • An autoregressive Torus Graph can estimate directional interactions through transfer entropy.
  • Systematic mapping of dynamic and directional phase relationships across brain states is enabled.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reduced cost may open Torus Graph modeling for other high-dimensional circular data outside neural recordings.
  • The hidden Markov and autoregressive extensions could be combined into a single model of evolving directional phase networks.
  • Validation on simulated phase data with known ground-truth couplings would test whether the stochastic procedure recovers accurate structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The paper claims that Torus Graph (TG) models for undirected phase dependencies can be scaled via a new stochastic score-matching procedure (reducing per-iteration cost from O(d^6) to O(d^2)) to handle thousands of variables, here applied to 1,860 frequency-phase features from multi-electrode LFPs; the same scalable foundation then supports two extensions (TG-HMM for state-dependent coupling and autoregressive TG for directional transfer-entropy inference) that reveal wake/NREM differences in phase interactions.

Significance. If the stochastic estimator is shown to preserve the statistical properties of the original score-matching estimator and the TG family is demonstrated to be an adequate model for the phase data, the work would remove a long-standing computational barrier and open systematic large-scale mapping of both dynamic and directed phase relationships in neuroscience.

major comments (2)
  1. [Abstract (description of stochastic score matching) and method sections] The central scalability claim rests on the stochastic score-matching procedure yielding parameter estimates whose consistency, bias, and downstream quantities (edge selection, transfer entropy) remain close to those of the O(d^6) full estimator. No theoretical argument establishing that the Monte-Carlo approximation of the score (or its Hessian) converges to the same population objective, nor any empirical comparison on regimes where both estimators can be run, is supplied; for circular data the variance of the trigonometric moments grows with the number of pairwise terms, so the approximation error is not obviously bounded independently of d.
  2. [Applications and results paragraphs] The two extensions (TG-HMM and autoregressive TG) and the biological conclusions drawn from the 1,860-feature LFP analysis presuppose that the exponential-family TG model fits the neural phase data sufficiently well. No goodness-of-fit diagnostics, residual analysis, or comparison against simpler circular baselines are reported, leaving open whether the observed state-dependent patterns could be artifacts of model misspecification.
minor comments (2)
  1. The abstract states that the procedure 'enables inference on datasets with thousands of variables' but supplies no concrete timing or memory figures on the 1,860-feature dataset; adding these would strengthen the scalability claim.
  2. Notation for the univariate and pairwise potentials is introduced without an explicit reference to the original TG definition; a short recap equation would aid readers unfamiliar with the 2023 TG paper.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments. We respond point-by-point to the two major concerns below.

read point-by-point responses

  1. Referee: [Abstract (description of stochastic score matching) and method sections] The central scalability claim rests on the stochastic score-matching procedure yielding parameter estimates whose consistency, bias, and downstream quantities (edge selection, transfer entropy) remain close to those of the O(d^6) full estimator. No theoretical argument establishing that the Monte-Carlo approximation of the score (or its Hessian) converges to the same population objective, nor any empirical comparison on regimes where both estimators can be run, is supplied; for circular data the variance of the trigonometric moments grows with the number of pairwise terms, so the approximation error is not obviously bounded independently of d.

    Authors: We acknowledge that the manuscript provides neither a theoretical convergence argument for the Monte-Carlo score approximation nor empirical head-to-head comparisons against the full O(d^6) estimator. In revision we will add an empirical validation subsection that runs both estimators on synthetic torus-graph data for d ≤ 50 (where the full estimator remains tractable) and reports agreement on parameter recovery, edge selection, and transfer-entropy values. A rigorous theoretical bound on approximation error for growing d remains an open question outside the present scope. revision: partial

  2. Referee: [Applications and results paragraphs] The two extensions (TG-HMM and autoregressive TG) and the biological conclusions drawn from the 1,860-feature LFP analysis presuppose that the exponential-family TG model fits the neural phase data sufficiently well. No goodness-of-fit diagnostics, residual analysis, or comparison against simpler circular baselines are reported, leaving open whether the observed state-dependent patterns could be artifacts of model misspecification.

    Authors: We agree that explicit goodness-of-fit assessment is required to support the reported biological findings. The revised manuscript will include (i) likelihood comparisons of the fitted TG model against an independent von Mises baseline on the LFP data and (ii) residual checks that verify reproduction of observed pairwise trigonometric moments. These diagnostics will be presented before the state-dependent and directional analyses. revision: yes

standing simulated objections not resolved
  • A complete theoretical analysis establishing convergence of the Monte-Carlo score approximation for the torus graph model as d increases.

Circularity Check

0 steps flagged

No circularity: stochastic score matching is an independent algorithmic scaling technique

full rationale

The paper's central contribution is an algorithmic reduction of score-matching inference for torus graphs from O(d^6) to O(d^2) via stochastic approximation. This step is derived from standard stochastic gradient techniques applied to the existing score-matching objective and does not reduce to any fitted parameter, self-defined quantity, or self-citation chain within the present work. The subsequent TG-HMM and autoregressive extensions are enabled by the new scalability but are not used to justify the core procedure itself. No load-bearing claim equates a prediction or uniqueness result to its own inputs by construction; the derivation chain remains externally verifiable through standard optimization theory and is therefore self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based solely on abstract; limited detail available on parameters or assumptions. The TG model relies on exponential-family assumptions for phases.

free parameters (1)
  • univariate and pairwise potential parameters
    Parameters in the exponential-family distribution over phases are fitted to data, though specific values not given in abstract.
axioms (1)
  • domain assumption Phases follow an exponential-family distribution whose univariate and pairwise potentials generalize von Mises distributions
    Explicitly stated as the basis for the TG model in the abstract.

pith-pipeline@v0.9.1-grok · 5770 in / 1265 out tokens · 32518 ms · 2026-06-28T19:06:11.892655+00:00 · methodology

0 comments
read the original abstract

Oscillatory neural signals such as electroencephalography (EEG) and local field potentials (LFPs) show phase relationships that coordinate communication across brain regions. Modern recordings capture hundreds of channels across many frequency bins, yet standard phase analyses are restricted to only a few variables. The Torus Graph (TG) model, an exponential-family distribution over phases whose univariate and pairwise potentials generalize von Mises distributions, infers principled structure among oscillations but models only static, undirected dependencies and is limited to $\sim \! 100$ variables because its score matching inference scales as $\mathcal{O}(d^{6})$. We introduce a stochastic score matching procedure that reduces the per-iteration cost to $\mathcal{O}(d^{2})$, enabling inference on datasets with thousands of variables. This scalable foundation supports analyses of 1,860 frequency-phase features from multi-electrode LFPs and enables two extensions previously inaccessible to TGs or classical circular statistics: (i) a TG Hidden Markov Model capturing state-dependent phase-coupling changes (e.g., spindle-related states during sleep) and (ii) an autoregressive TG inferring directional interactions via transfer-entropy estimation. Applied to LFP recordings, these models reveal state-dependent phase-interaction patterns between wakefulness and NREM sleep. Together, they enable systematic, large-scale mapping of dynamic and directional phase relationships across brain and cognitive states.

Figures

Figures reproduced from arXiv: 2606.00496 by Casey Hanks, David E. Carlson, Jack Goffinet.

Figure 1
Figure 1. Figure 1: Visualization of TG statistics and parameters. Left: phases x1, x2, x3 are extracted from a signal (black) via a continu￾ous wavelet transform, with color denoting phase angle according to the colorwheel shown. Right: TG statistics and parameters are arranged as a d × d complex matrix, where each entry’s co￾sine and sine components are treated as real and imaginary parts respectively, and colored by positi… view at source ↗
Figure 2
Figure 2. Figure 2: Runtime of stochastic versus exact score matching as a function of data dimension. 4. Results We evaluated our methods on synthetic datasets to assess accuracy against known ground truth and on a large-scale mouse LFP dataset to illustrate practical utility. All exper￾iments were run on a single Nvidia A5000 GPU with 24 GB VRAM using efficient JAX implementations (Bradbury et al., 2018). See Appendix C for… view at source ↗
Figure 4
Figure 4. Figure 4: Transfer entropy (TE) estimation and causal discovery on synthetic data. A: Estimated TE from an autoregressive TG (AR￾TG) converges rapidly to ground truth in a unidirectional setting. B: AR-TGs outperform baseline TE estimators in identifying causal direction. C: Multivariate Granger causality degrades rapidly with increasing dimension; the red line denotes a 30-hour timeout. D: AR-TGs maintain accurate … view at source ↗
Figure 5
Figure 5. Figure 5: Large-scale TG analysis of mouse LFP phase data (62 channels, 30 frequencies, d = 1860). A: Empirical phase statis￾tics arranged by channel–frequency pairs; insets show within- and between-region structure. Colors denote phase angle as in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Multivariate transfer entropy (TE) in wake and NREM sleep. Connections of interest—prelimbic cortex → striatum, in￾fralimbic → prelimbic/cingulate cortices, and VTA → SNr—are highlighted in yellow, purple, and green. A: TE between chan￾nels during wake at 18 and 45 Hz. B: Wake minus NREM TE at the same frequencies. C: Wake–NREM TE differences for all cross-region channel pairs across frequency. D: Minimum … view at source ↗
Figure 8
Figure 8. Figure 8: Additional Experiments comparing torus graph models fit using stochastic and exact methods. Pseudo-R 2 of the estimation of torus graph parameters (ϕ) fit using an increasing number of samples from a synthetic time series from 2- to 1192-dimensional data. Stochastic score matching achieves comparable performance in low-dimensional settings, as well as achieving vastly better estimation in higher-dimensiona… view at source ↗
Figure 9
Figure 9. Figure 9: Wake TG statistics with brain region labels. Rows and columns are grouped by brain region (18 regions, 2–6 channels each), and each channel contributes 30 phase variables at center frequencies linearly spaced between 1 and 55 Hz. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Wake TG parameters with brain region labels. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Mean spindle waveforms for each channel in [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Large labeled version of Figure 7A. C.9. Model Specification Stability and Sensitivity Analysis We additionally report sensitivity analyses examining the robustness of each method to hyperparameter choices and model specification. Stochastic Score Matching TG Sensitivity to Hyperparameter Specification To investigate the sensitivity of SSM￾TG to its hyperparameters we conducted line searches over batch si… view at source ↗
Figure 13
Figure 13. Figure 13: Large labeled version of Figure 7B. 70 60 50 40 30 20 10 0 Training Loss Batch Size batch_size_64 batch_size_128 batch_size_256 batch_size_512 batch_size_1024 Learning Rate learning_rate_1e-4 learning_rate_5e-4 learning_rate_1e-3 learning_rate_4e-3 learning_rate_7e-3 0 1000 2000 3000 4000 5000 Iterations 70 60 50 40 30 20 10 0 Training Loss Number of Samples n_100 n_500 n_1000 n_5000 n_10000 0 1000 2000 3… view at source ↗
Figure 14
Figure 14. Figure 14: Training Loss curves used in stability analysis. Mean over 5 runs is plotted ± standard deviation. Stopping criteria was applied independently for each model. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: State occupancies relative to spindle times for all K-state TG-HMMs with K ∈ [3, 8]. Compare to Figure 6C. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 2 canonical work pages

  1. [1]

    The annals of applied statistics , volume=

    Torus graphs for multivariate phase coupling analysis , author=. The annals of applied statistics , volume=

  2. [2]

    2006 , publisher=

    Rhythms of the Brain , author=. 2006 , publisher=

    2006

  3. [3]

    The phase of ongoing

    Busch, Niko A and Dubois, Julien and VanRullen, Rufin , journal=. The phase of ongoing. 2009 , publisher=

    2009

  4. [4]

    Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=

    Noise-contrastive estimation: A new estimation principle for unnormalized statistical models , author=. Proceedings of the thirteenth international conference on artificial intelligence and statistics , pages=. 2010 , organization=

    2010

  5. [5]

    Human brain mapping , volume=

    Measuring phase synchrony in brain signals , author=. Human brain mapping , volume=. 1999 , publisher=

    1999

  6. [6]

    Science , volume=

    Neuropixels 2.0: A miniaturized high-density probe for stable, long-term brain recordings , author=. Science , volume=. 2021 , publisher=

    2021

  7. [7]

    Science , volume=

    Spontaneous behaviors drive multidimensional, brainwide activity , author=. Science , volume=. 2019 , publisher=

    2019

  8. [8]

    science , volume=

    Neuronal oscillations in cortical networks , author=. science , volume=. 2004 , publisher=

    2004

  9. [9]

    Science , volume=

    Modulation of oscillatory neuronal synchronization by selective visual attention , author=. Science , volume=. 2001 , publisher=

    2001

  10. [10]

    Journal of Neuroscience , volume=

    Gamma oscillations in the entorhinal cortex of the freely behaving rat , author=. Journal of Neuroscience , volume=. 1998 , publisher=

    1998

  11. [11]

    Physiological reviews , volume=

    Neurophysiological and computational principles of cortical rhythms in cognition , author=. Physiological reviews , volume=. 2010 , publisher=

    2010

  12. [12]

    Phase relationship between hippocampal place units and the

    O'Keefe, John and Recce, Michael L , journal=. Phase relationship between hippocampal place units and the. 1993 , publisher=

    1993

  13. [13]

    arXiv preprint arXiv:1910.04375 , year=

    Estimating transfer entropy via copula entropy , author=. arXiv preprint arXiv:1910.04375 , year=

    work page arXiv 1910

  14. [14]

    International Conference on Learning Representations (ICLR) , year =

    Adam: A Method for Stochastic Optimization , author =. International Conference on Learning Representations (ICLR) , year =

  15. [15]

    James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and George Necula and Adam Paszke and Jake Vander

  16. [16]

    Granger causality and transfer entropy are equivalent for

    Barnett, Lionel and Barrett, Adam B and Seth, Anil K , journal=. Granger causality and transfer entropy are equivalent for. 2009 , publisher=

    2009

  17. [17]

    Molecular Neurobiology , volume=

    Dopaminergic projection from ventral tegmental area to substantia nigra pars reticulata mediates chronic social defeat stress--induced hypolocomotion , author=. Molecular Neurobiology , volume=. 2021 , publisher=

    2021

  18. [18]

    Cabezas, A

    Alberto Cabezas and Adrien Corenflos and Junpeng Lao and Rémi Louf , year=. BlackJAX: Composable. 2402.10797 , archivePrefix=

    work page arXiv

  19. [19]

    Protein bioinformatics and mixtures of bivariate

    Mardia, Kanti V and Taylor, Charles C and Subramaniam, Ganesh K , journal=. Protein bioinformatics and mixtures of bivariate. 2007 , publisher=

    2007

  20. [20]

    Neural computation , volume=

    Phase coupling estimation from multivariate phase statistics , author=. Neural computation , volume=. 2010 , publisher=

    2010

  21. [21]

    Directional-unit

    Zemel, Richard and Williams, Christopher and Mozer, Michael C , journal=. Directional-unit

  22. [22]

    Science , volume=

    A common hub for sleep and motor control in the substantia nigra , author=. Science , volume=. 2020 , publisher=

    2020

  23. [23]

    Nature communications , volume=

    Bi-directional regulation of cognitive control by distinct prefrontal cortical output neurons to thalamus and striatum , author=. Nature communications , volume=. 2021 , publisher=

    2021

  24. [24]

    Current Biology , volume=

    The bilateral prefronto-striatal pathway is necessary for learning new goal-directed actions , author=. Current Biology , volume=. 2018 , publisher=

    2018

  25. [25]

    Physical review letters , volume=

    Measuring information transfer , author=. Physical review letters , volume=. 2000 , publisher=

    2000

  26. [26]

    Mathematical Programming , volume=

    An exact duality theory for semidefinite programming and its complexity implications , author=. Mathematical Programming , volume=. 1997 , publisher=

    1997

  27. [27]

    , author=

    Estimation of non-normalized statistical models by score matching. , author=. Journal of Machine Learning Research , volume=

  28. [28]

    Wollstadt, Patricia and Lizier, Joseph T and Vicente, Raul and Finn, Conor and Martinez-Zarzuela, Mario and Mediano, Pedro and Novelli, Leonardo and Wibral, Michael , journal=

  29. [29]

    International Conference on Learning Representations (ICLR) , year=

    Self-Supervised Graph Neural Networks for Improved Electroencephalographic Seizure Analysis , author=. International Conference on Learning Representations (ICLR) , year=

  30. [30]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Eeg2rep: enhancing self-supervised eeg representation through informative masked inputs , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  31. [31]

    EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks , volume =

    Kotoge, Rikuto and Chen, Zheng and Kimura, Tasuku and Matsubara, Yasuko and Yanagisawa, Takufumi and Kishima, Haruhiko and Sakurai, Yasushi , journal =. EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks , volume =

  32. [32]

    Scientific Reports , volume=

    infomeasure: a comprehensive Python package for information theory measures and estimators , author=. Scientific Reports , volume=. 2025 , publisher=

    2025

  33. [33]

    Statsmodels: econometric and statistical modeling with

    Seabold, Skipper and Perktold, Josef and others , journal=. Statsmodels: econometric and statistical modeling with

  34. [34]

    1984 , publisher =

    Chemical Oscillations, Waves, and Turbulence , author =. 1984 , publisher =

    1984

  35. [35]

    Generative models of cortical oscillations: neurobiological implications of the

    Breakspear, Michael and Heitmann, Stewart and Daffertshofer, Andreas , journal=. Generative models of cortical oscillations: neurobiological implications of the. 2010 , publisher=

    2010

  36. [36]

    1980 , publisher=

    The geometry of biological time , author=. 1980 , publisher=

    1980

  37. [37]

    Neuron , volume=

    Traveling electrical waves in cortex: insights from phase dynamics and speculation on a computational role , author=. Neuron , volume=. 2001 , publisher=

    2001

  38. [38]

    The Journal of Mathematical Neuroscience , volume=

    Understanding the dynamics of biological and neural oscillator networks through exact mean-field reductions: a review , author=. The Journal of Mathematical Neuroscience , volume=. 2020 , publisher=

    2020

  39. [39]

    Journal of neurophysiology , volume=

    Effect of amplitude correlations on coherence in the local field potential , author=. Journal of neurophysiology , volume=. 2014 , publisher=

    2014