{"id":"c7702729-2d5b-442f-8e25-a1de81a5e5d8","arxiv_id":"1908.06678","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Network connections must scale as b^{-z} with system size to preserve scale-free dynamics, and a Bayesian method using coarse-grained calcium imaging recovers this exponent with task-state differences across cortical regions.","lead":"This paper derives how connections in a network must rescale when the system size changes so that the dynamics stay scale free. It then uses this rule to estimate a dynamical critical exponent from mouse brain imaging, finding that the exponent shifts between rest and task in different cortical regions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The coarse-graining identity X(t)=x(b^{-α}t) (Eq. 23) is asserted, not derived; averaging and re-z-scoring 2×2 blocks can alter temporal correlation structure, so the estimated z (Eq. 33) may be biased.","rationale":"The central claim has two parts: the theoretical relation α=-z under coarse graining, and the empirical task/spontaneous difference in z. The Kepler simulation is a genuine construct validation for the DCM/PEB recovery of a global scaling exponent (α≈-1.47 vs -1.5), and I would not dismiss it. However, the transfer to imaging data requires Eq. 23, which is the only place where the coarse-graining operation is connected to the scaling theory. It is specifically asserted, not derived, and the simplification in Eq. 22 is not sufficient because block averaging is a linear spatial filter that changes correlation structure in a way that simple z-scoring cannot undo. Since Eq. 32 and Eq. 33 both follow from Eq. 23, an invalid Eq. 23 would invalidate the estimated z and therefore the task/spontaneous differences. The proposed synthetic test would settle this directly by checking whether the pipeline recovers a known z. The reader's verdict already flags Eq. 23 as the weakest assumption and gives a conditional verdict; my read does not move that, but it sharpens the required condition. The empirical issues (n=3, no error bars) are real but secondary to whether the quantity being estimated is actually z.","tokens_in":11961,"tokens_out":9702,"duration_ms":112587,"concrete_test":"Generate a ground-truth scale-free system with known z, e.g., a Gaussian field on a 64×64 grid with autocorrelation C(r,t)=r^{-a} f(t/r^z), simulate regional time courses, then apply the paper's exact coarse-graining pipeline (2×2 averaging, z-scoring, first-level DCM, PEB line search over z in 0.005 steps). Compare the recovered z to the true z. If the recovery error exceeds the 0.005 grid step or shows a systematic bias, Eq. 23/33 is not validated and the empirical z differences cannot be interpreted as dynamical critical exponents. A second, quicker check: compute sigma_r^2(t) and sigma_b^2(t) from the simulated data and directly test whether t_b = b^z t_r as required by Eqs. 30–31.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The linchpin of the coarse-graining analysis is Eq. 23: the block time course X(t), formed by averaging 2×2 region time courses and re-z-scoring (Methods 'Coarse graining', steps a–e), is asserted to equal the original region time course under a temporal rescaling, x(b^{-α}t). Eq. 22 claims z-scoring justifies dropping the amplitude factor in Eq. 21, but z-scoring only rescales a single time course; it does not guarantee that the average of four neighboring (possibly correlated) z-scored region time courses has the same temporal correlations as one region, nor that the pair (X,x) obeys the DCM scaling A'=b^{-z}A in Eq. 33. The identification α=-z (Eq. 32) is derived from Eq. 23 via the decay-time definitions in Eqs. 24–31, so any failure of Eq. 23 propagates directly into the z estimates that drive the task/spontaneous comparison. Because the Kepler example validates only the DCM/PEB estimator under an exact global scaling relation, it does not validate this local coarse-graining step. The manuscript explicitly flags the simplification in Eq. 22 as an assumption, but it does not test or derive the key identity Eq. 23 for averaged, z-scored signals.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives scale-freeness constraints for a bilinear dynamical system (the DCM recovery model) and shows that, to maintain identical equations of motion under a space-time rescaling, the intrinsic connectivity matrix must transform as A → b^{-α} A. In the orbital case this recovers Kepler's third law (α = −3/2), and the authors validate the estimation framework by recovering α ≈ −1.47 from simulated three-body orbits. The paper then extends the framework to coarse-grained neuroimaging data, defining a coarse-graining identity X(t) = x(b^{-α} t) that leads to α = −z, where z is the dynamical critical exponent. Using DCM/PEB on calcium imaging data from three mice, the paper reports that the dynamical critical exponent is higher during spontaneous activity in a task-irrelevant ROI and higher during task performance in a task-relevant ROI. The main theoretical contribution is the connection between temporal rescaling exponents and connectivity transformations in a DCM hierarchy; the main empirical claim is the state-dependent z differences in the two cortical ROIs.","tokens_in":12281,"tokens_out":7273,"duration_ms":77003,"significance":"If the central derivation and coarse-graining identity are valid, this paper provides a principled way to estimate a renormalization-group-type dynamical critical exponent from empirical neuroimaging data, using a hierarchical DCM/PEB framework. The Kepler simulation is a genuine strength: it provides a clear ground-truth validation that the DCM/PEB pipeline can recover a known scaling exponent. The theoretical result relating α and z in Eq. (32) is elegant and could be useful for characterizing scale-free dynamics in biological systems. However, the empirical z estimates depend critically on an assumed coarse-graining identity that is not derived or separately validated, and the reported differences between task and spontaneous states are based on point maxima of free-energy curves without uncertainty quantification. Thus the significance, while potentially high, is not yet fully established.","major_comments":[{"comment":"The identity X(t) = x(b^{-α} t) in Eq. (23) is asserted rather than derived for block time courses defined as the means of 2×2 z-scored region time courses. Averaging four (possibly correlated) z-scored signals changes the temporal correlation structure in ways that cannot generally be represented by a simple temporal rescaling of a single region time course; for example, if the constituent regions are independent, the block variance and autocorrelation differ from those of any individual region. Equations (28)-(33) and the subsequent neuroimaging analyses rest entirely on this identity, so any failure of Eq. (23) propagates directly into the estimated dynamical critical exponent z. The Kepler simulation does not validate Eq. (23) because the orbital scaling is an exact rescaling of the same trajectory, not a spatial average. The authors should either derive the conditions under which Eq. (23) holds for averaged z-scored signals (e.g., spatial homogeneity and independence assumptions) or provide a simulation-based validation where the ground-truth z for a coarse-graining process is known.","section":"Methods, 'Coarse graining'; Eq. (23)"},{"comment":"The empirical z estimates are the maxima of free-energy curves over the hand-picked range 0.1 to 0.4 (Methods, final paragraph of 'Coarse graining'). The paper reports no posterior confidence intervals, no free-energy differences between the maxima, and no statistical test comparing task versus spontaneous conditions. The claim of consistency across all three animals is based on visual inspection of the peak locations in Fig. 3C. Without uncertainty quantification, the magnitude and reliability of the reported differences are unknown; the authors should report model evidence differences, credible intervals or bootstrap errors, and justify why the admissible z range is restricted to 0.1–0.4 rather than, for example, allowing z ≤ 0 or z > 0.4.","section":"Results, 'Neuroimaging data'; Fig. 3C"},{"comment":"The relation α = −z in Eq. (32) is derived under the assumption that Eq. (23) holds 'on average', but the paper does not formally define what 'on average' means in the context of the DCM/PEB estimation. The second-level PEB model fits a single z across all coarse-graining scales, which presupposes that the transformation A → b^{−z} A in Eq. (33) is exact at every scale. If Eq. (23) holds only approximately (e.g., with scale-dependent deviations), the PEB model is mis-specified and the free-energy maxima may be driven by misspecification rather than by a genuine scale-free relationship. Please clarify the sense in which the identity holds and discuss the robustness of the z estimates when the assumed scaling is only approximate.","section":"Methods, 'Intrinsic connectivity and the dynamical critical exponent'; Eqs. (28)-(33)"}],"minor_comments":[{"comment":"Reference 24 is malformed: 'Mouse A, and Coordinate, C.' is not a proper citation for the Allen Mouse Common Coordinate Framework; please correct it.","section":"References"},{"comment":"The definition of the block correlation function σ_b²(t) in Eq. (26) is unclear: the prefactor b^{−?} is garbled and the averaging notation is ambiguous. Please rewrite it to make explicit how the block average is taken over blocks.","section":"Methods, Eq. (26)"},{"comment":"In the orbital derivation, the exponent algebra between Eqs. (7) and (9) contains a typographical slip (the exponents on b in the intermediate steps are not consistent with the final α = −3/2). Since the final result is correct, please check and correct the intermediate display equations.","section":"Methods, Eqs. (5)-(10)"},{"comment":"The paper does not include a data availability or code availability statement; for a computational study of this kind, such a statement would be helpful for reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea of estimating scaling exponents via DCM/PEB is attractive, and the Kepler simulation is a convincing proof of concept. However, the neuroimaging section currently rests on an unvalidated coarse-graining identity and lacks statistical rigor in the z estimates. I believe these issues are addressable within the manuscript's scope, hence major revision rather than rejection. The authors should consider strengthening the paper by validating Eq. (23) in a controlled simulation and by providing uncertainty quantification for the empirical z differences. If the empirical results remain qualitative after revision, the claims should be softened accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper has a real contribution: it derives how a DCM's connectivity matrices must transform under spatial scaling to preserve scale-free dynamics, connects that scaling exponent to the dynamical critical exponent z, and packages this as a hierarchical Bayesian estimator (DCM + PEB) that can be run on coarse-grained imaging data. The Kepler validation is the strongest part. Using orbital simulations, the estimator recovers alpha = -1.47, close to the theoretical -1.5. That gives genuine construct validity for the estimation scheme.\n\nThe math in the derivation (equations [12]-[20] and [28]-[33]) is internally consistent, and the paper is honest about the assumption in equation [22] that z-scoring removes the amplitude factor. My main concern is the step the paper does not test: equation [23], the assertion that the block time course X(t) equals the original region time course under a temporal rescaling, x(b^{-alpha}t). That identity is the linchpin for the neural application. Averaging four neighboring z-scored time courses does not automatically preserve the temporal correlation structure of a single region; correlated signals can change the effective dynamics after averaging and re-z-scoring. Since alpha = -z is derived from equation [23], any failure there propagates directly into the z estimates that drive the task/spontaneous comparison. The Kepler example does not rescue this, because it operates under exact global scaling, not local averaging.\n\nThe empirical results are plausible but only provisionally supported. The z values are point maxima of free energy curves over a narrow range (0.1 to 0.4) with no error bars, n=3 mice, and the ROI classification is post hoc. The cross-animal consistency is encouraging, but the central claim about task-relevant vs task-irrelevant cortices needs uncertainty quantification before I would take it as established.\n\nWho is this for? Anyone working on brain criticality, renormalization-group analyses of neural data, or DCM methodology. The method itself is useful, and the Kepler validation makes the framework worth engaging with. A serious referee should push for either a derivation of equation [23] or a validation on synthetic data with known scaling, plus error bars or posterior intervals on z. That is enough to make this a paper worth revising rather than rejecting.\n\nIt deserves a serious referee.","headline":"Useful estimator for the dynamical critical exponent with a clean Kepler validation, but the coarse-graining identity at the core of the neural application is asserted rather than tested.","tokens_in":12804,"tokens_out":2526,"would_cite":true,"duration_ms":24815,"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":"The paper derives and measures the network scaling constraint behind scale-free dynamics, recovering Kepler's third law as a check and finding task-dependent dynamical critical exponents in mouse cortex.","keywords":["scale freeness","dynamical critical exponent","coarse graining","dynamic causal modelling","parametric empirical Bayes","calcium imaging","Kepler's third law","renormalization group"],"falsifier":"Estimate $z$ from surrogate data built by phase-randomizing each region's time course before coarse graining; if the pipeline still returns different $z$ values between task and spontaneous states, the contrast is an artifact of the averaging scheme rather than a property of the dynamics. More directly, compute $X(t)$ and $x(t)$ from real data and test equation [23]: if the equality fails, the derived relation $\\alpha=-z$ does not apply to z-scored block averages.","tokens_in":11803,"feed_emoji":"🧠","tokens_out":16281,"duration_ms":150320,"temperature":0.7,"pith_summary":"The paper tries to establish a constraint that any network must obey if its dynamics are to look the same at every spatial scale: as the system is rescaled, its effective connectivity must shrink by a factor set by a single exponent. The paper derives this from a generic dynamic causal model, shows the time-rescaling exponent $\\alpha$ is the negative of the renormalization-group dynamical critical exponent $z$, and validates the estimation pipeline by recovering Kepler's third law from simulated orbits. Applied to wide-field calcium imaging of mouse cortex, the same pipeline finds that $z$ is higher during task performance than at rest in task-relevant cortex, and lower in task-irrelevant cortex. The result matters because it turns 'scale freeness' from a descriptive label into a quantity that can be measured from ordinary time-series data and compared across brain states.","feed_headline":"Task flips the brain's scale-free exponent in engaged cortex","feed_subtitle":"Relevant cortex shows a higher critical exponent during tasks; irrelevant cortex, the opposite.","key_machinery":"The load-bearing object is the DCM recovery model, a bilinear state-space model $d\\boldsymbol{r}/dt = A\\boldsymbol{r}$ (with $B$ and $C$ terms set to zero here) that estimates an intrinsic coupling matrix $A$ from observed time series. The paper combines it with Parametric Empirical Bayes at the second level, which tests whether the scale-dependent changes in $A$ follow $A\\to b^{-z}A$ and returns a free-energy approximation to model evidence for each candidate $z$. The link between the fitted connectivity and critical phenomena is equation [33], equivalently $\\alpha=-z$, which connects the time-rescaling exponent in equation [22] to the renormalization-group dynamical critical exponent. The coarse-graining operation that feeds this machinery is repeated $2\\times2$ spatial averaging of z-scored region time courses.","core_discovery":"On the paper's own terms, the central discovery is that scale freeness fixes how the intrinsic coupling matrix of a dynamical system must transform with system size: $A \\to b^{-z}A$, where $z$ is the dynamical critical exponent, and that the time-rescaling exponent $\\alpha$ in $x(t)\\to x(b^{\\alpha}t)$ is exactly $-z$. The paper derives this from the requirement that the scaled equation of motion have the same form as the original, demonstrates the estimation machinery on orbital simulations where the recovered exponent $\\alpha=-1.47$ matches Kepler's prediction $-3/2$, and then applies the method to calcium-imaging data. In three mice, the dynamical critical exponent that maximizes model evidence is higher in a task-relevant cortical region during task performance than during spontaneous activity, and higher in a task-irrelevant region during spontaneous activity than during task performance. The authors interpret higher $z$ as greater temporal renormalization across spatial scales, i.e., more cross-scale information transmission.","pith_inferences":["A natural next test, not run by the paper, is whether $z$ responds continuously to task difficulty or attention: the reported state contrast predicts a monotonic increase in task-relevant $z$ with engagement and a decrease in task-irrelevant $z$.","Because the derivation ignores external inputs and models only intrinsic coupling, applying the pipeline to fMRI or EEG would require recalibrating $z$ under hemodynamic or volume-conduction smoothing; the paper does not address those confounds.","The inverse relation $\\alpha=-z$ makes a concrete cross-species prediction the paper does not test: if larger brains have slower intrinsic fluctuations (positive $z$), then their activity time courses should be time-rescaled by $b^{-z}$, meaning an internal clock that slows with brain size; developmental or comparative imaging could check this directly."],"forward_implications":["A single pair of coarse-graining levels suffices in principle to estimate the dynamical critical exponent from empirical connectivity matrices, avoiding indirect box-counting or avalanche statistics.","The same hierarchical Bayesian procedure can be transferred to any imaging modality that yields region time courses, provided the coarse-graining assumption holds.","The ROI contrast implies that task engagement increases cross-scale temporal coupling in task-relevant cortex and decreases it in task-irrelevant cortex, giving a quantitative, state-dependent readout of cortical criticality.","Positive $z$ values across all animals and ROIs mean larger cortical structures exhibit slower decay of fluctuations, consistent with scale-free organization at the mesoscale."],"supporting_citations":[{"why":"It supplies the Parametric Empirical Bayes second-level model used to compare scaling exponents via free energy.","marker":"15"},{"why":"It defines the dynamical critical exponent $z$ from renormalization group theory, the target quantity in equation [31].","marker":"16"},{"why":"It provides the Lagrangian definition of scale freeness from which the equation-of-motion scaling condition follows.","marker":"17"},{"why":"It introduces the dynamic causal modelling framework whose bilinear form is the recovery model in equation [11].","marker":"18"},{"why":"It provides the DEM variational inversion algorithm used to estimate intrinsic coupling matrices at each scale.","marker":"21"},{"why":"It describes the transgenic GCaMP6f mouse line whose cortical calcium signals provide the empirical time series.","marker":"22"},{"why":"It supplies the intact-skull wide-field calcium imaging preparation used to record cortical dynamics.","marker":"23"},{"why":"It supplies the anatomical coordinate framework used to define the two regions of interest and assign task relevance.","marker":"24"},{"why":"It supports the designation of posterior parietal cortex as task-relevant for auditory decision-making.","marker":"30"}],"fun_headline_variants":["Scale-free dynamics tie task state to cortex critical exponent","Task flips critical exponent in task-relevant cortex regions","Cortex scale-free exponent shifts with task relevance","Dynamical critical exponent tracks task in scale-free networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after averaging neighboring regions into blocks and re-standardizing, a block's activity trace is just a time-rescaled copy of a single region's trace; averaging can change the signal's variance and correlation structure in ways that would distort the measured critical exponent.","fun_headline_variants_meta":{"raw":{"variants":["Scale-free dynamics tie task state to cortex critical exponent","Task flips critical exponent in task-relevant cortex regions","Cortex scale-free exponent shifts with task relevance","Dynamical critical exponent tracks task in scale-free networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1364,"prompt_tokens":992,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":608,"tokens_out":372,"duration_ms":4843,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:37:42.442715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate $z$ from surrogate data built by phase-randomizing each region's time course before coarse graining; if the pipeline still returns different $z$ values between task and spontaneous states, the contrast is an artifact of the averaging scheme rather than a property of the dynamics. More directly, compute $X(t)$ and $x(t)$ from real data and test equation [23]: if the equality fails, the derived relation $\\alpha=-z$ does not apply to z-scored block averages.","supporting_citations":[],"review_version":1}