Older brains are often described as "dedifferentiated": regions that used to do separate jobs start behaving alike. One influential piece of evidence comes from O-information, a measure of whether a group of brain regions shares the same information (redundancy) or carries information only jointly (synergy). Applied to resting-state fMRI, it shows that redundancy rises with age across groups of 3 to 20 regions, concentrated in a prefrontal–motor "redundancy core" (High-order interdependencies in the aging brain). These numbers are computed by plugging finite, autocorrelated fMRI time series into Gaussian entropy formulas. Older and younger scans differ in temporal structure as well as in the brain's organization. If the estimator's bias depends on temporal structure, part of the reported age effect could come from the estimator itself.
We propose to derive the leading bias of Gaussian O-information through a second-order Taylor expansion. The expansion weights the curvature (Hessian) of the log-determinant against the sampling covariance of the estimated correlation matrix. We then correct the estimate and ask how much of the dedifferentiation signal survives. A short calculation gives the key prediction, which we will verify first. Under independent samples, the leading bias of O-information cancels exactly. It also cancels when every region shares the same autocorrelation shape. What remains is a bias driven only by regional heterogeneity in temporal autocorrelation, and it is amplified when region groups are nearly collinear. Collinear groups are exactly where redundancy is measured to be high. As a transfer test, we apply the same estimator to microbial communities perturbed by antibiotics. There the question "do ecological guilds dedifferentiate?" runs into a second problem: log-ratio coordinates blow up when a species' abundance goes to zero, much as Euler angles blow up at gimbal lock.
Core Claim
O-information for $n$ variables $X=(X_1,\dots,X_n)$ is
$$\Omega(X) = (n-2)H(X) + \sum_{i=1}^{n}\big[H(X_i) - H(X_{-i})\big],$$
where $H$ is differential entropy and $X_{-i}$ is all variables except $X_i$. Under a Gaussian model, $H$ of a $k$-dimensional subset is $\tfrac12\log\det(2\pi e\,\Sigma)$ for that subset's covariance $\Sigma$. The plug-in estimator replaces $\Sigma$ with the sample covariance $\hat\Sigma$ from $T$ time points. Its bias follows from the delta method, the same second-order Taylor argument that Generative Local Metric Learning for Nearest Neighbor Classification uses to show that finite-sample nearest-neighbor error is governed by density Hessians. Here the expanded function is $\log\det$, whose Hessian in direction $E$ is $-\mathrm{tr}(\Sigma^{-1}E\Sigma^{-1}E)$. Combining this with Bartlett's formula for the covariance of lagged sample covariances gives, to leading order,
$$\mathbb{E}\log\det\hat\Sigma - \log\det\Sigma \approx -\frac{1}{2T}\sum_{\tau}\Big[\mathrm{tr}\big(\Sigma^{-1}\Gamma(\tau)\Sigma^{-1}\Gamma(\tau)^{\top}\big) + \big(\mathrm{tr}\,\Sigma^{-1}\Gamma(\tau)\big)^2\Big],$$
where $\Gamma(\tau)$ is the lag-$\tau$ cross-covariance matrix of the subset.
For independent samples only $\tau=0$ contributes, and the bias reduces to $-k(k+1)/(2T)$. Substituting this into $\Omega$ makes the subset terms cancel exactly. The same cancellation holds if $\Gamma(\tau)=\rho(\tau)\Sigma$, meaning all regions share one autocorrelation shape. So the plug-in O-information is first-order unbiased in the textbook case. Its bias comes entirely from heterogeneous, region-specific temporal structure, and it is scaled by $\Sigma^{-1}$, which grows large when an $n$-plet is close to rank-deficient.
This yields three concrete deliverables:
- A closed-form correction that is cheap enough to apply to the millions of $n$-plets in a 2,514-region parcellation.
- A falsifiable prediction: the corrected age effect on redundancy should shrink in proportion to age differences in autocorrelation heterogeneity.
- A diagnostic that flags $n$-plets where the correction is unreliable because $\Sigma$ is near singular.
Where the Other Directions Enter
The gimbal-lock literature supplies the lens on near-singular estimation. Minimal coordinate systems create sensitivity spikes and ill-conditioned Fisher information that belong to the chart, not the physical system (Representation Issues in the ML Estimation of Camera Motion). Density models working in such charts lose roughly 40% in log-likelihood near the singularity (Are Euler angles a useful rotation parameterisation for pose estimation with Normalizing Flows?).
Near-collinear $n$-plets sit near the boundary of the cone of positive-definite matrices, where $\Sigma^{-1}$, and therefore our bias term and its variance, explode. We treat this as an instrument problem. We will report the condition number alongside every redundancy estimate. We will not report corrected values where the second-order expansion is untrustworthy, which we determine by simulation. We will not apply a correction blindly in that regime.
The microbiome direction supplies a second regime and a sharper version of the chart problem. Relative abundances live on the simplex, the set of proportions that sum to one. The compositional Lotka–Volterra model works in additive log-ratio coordinates, which diverge as any taxon or the reference taxon approaches zero (Compositional Lotka-Volterra describes microbial dynamics in the simplex).
O-information is unchanged when each variable is separately and invertibly transformed. Log-ratio and centered log-ratio maps are not of this kind: they mix taxa. As The Base Measure Problem and its Solution shows for densities in general, any entropy-based quantity on compositions is undefined until a base measure is fixed. The centered log-ratio covariance is singular by construction.
The microbiome question is whether modules that existing models treat as static, such as the Dirichlet-process modules of Robust and Scalable Models of Microbiome Dynamics and the keystone modules of Learning ecosystem-scale dynamics from microbiome data with MDSINE2, become redundant under perturbation. That question is only answerable once both the chart and the bias are controlled.
Research Program
Step one: derivation and verification. We derive the full leading-order bias of $\Omega$ for stationary Gaussian processes and verify the constants numerically. This includes the cancellation results and the finite-$T$ normalization. We also derive the analogous correction for shrinkage covariance estimators, since practitioners use them for large $n$.
Step two: simulation. We generate vector-autoregressive and long-memory Gaussian processes whose stationary covariance gives an analytically known O-information. We independently control three things: redundancy, regional heterogeneity of autocorrelation, and collinearity. This maps where the plug-in estimator fails, where our correction works, and where both break down near singularity.
Step three: brain data. We re-run the age-redundancy analysis on public lifespan data. These are the Human Connectome Project development, young-adult, and aging cohorts used in Analysis of functional connectivity changes from childhood to old age: A study using HCP-D, HCP-YA, and HCP-A datasets. We follow the protocol of the 164-subject study: Brain Hierarchical Atlas with 2,514 regions grouped into 20 modules, interaction orders 3 to 20, and four age bands. We then apply the correction.
Results are known to vary with parcellation and with fMRI's slow sampling (Temporal and Spatial Scales of Human Resting-State Cortical Activity across the Lifespan), so we repeat the analysis at a coarser parcellation. Regional temporal scales are themselves reported to change across the lifespan, which is what makes the heterogeneity mechanism plausible rather than hypothetical.
Step four: microbiome data. We carry the estimator to microbiome time series. One dataset is the humanized-mouse cohorts of MDSINE2, which include qPCR-measured total biomass across diet and antibiotic perturbations. The other is the C. difficile mouse data used by Gibson and Gerber: 13 commensals, 26 fecal samples per mouse over 56 days. We compute O-information on four representations: per-taxon log absolute abundance (a legitimate per-variable transform away from zeros), centered log-ratio, additive log-ratio under several reference taxa, and square-root proportions on the sphere. The square-root embedding is the simplex analog of moving from Euler angles to quaternions: it stays smooth when a species goes extinct. With so few time points per animal the bias term is large, so this step stress-tests the method rather than adding a third headline.
The contrast with closest prior work is direct. Riemann2: Learning Riemannian Submanifolds from Riemannian Data treats connectomes as points on the positive-definite manifold and corrects generative likelihoods for exponential-map distortion. It never asks whether information functionals estimated from finite autocorrelated signals are biased. Gatica et al. motivate the question but use uncorrected plug-in estimates. The microbiome module models constrain our design, because their static modules give the null hypothesis that ecological dedifferentiation must reject.
Baselines
Naive reference. The plug-in Gaussian O-information exactly as computed by Gatica et al., on identical preprocessing, parcellation and subjects. Every other method is scored by how it changes this number.
Exact independent-sample correction. This uses the closed-form Wishart log-determinant bias, together with its effective-sample-size variant that replaces $T$ by an autocorrelation-adjusted $T_{\text{eff}}$. It is the strongest analytic competitor and the one most reviewers would reach for. Our derivation predicts that both versions change O-information by essentially nothing, because their bias cancels across subsets. Confirming this in simulation is itself a result.
Moving-block bootstrap bias correction. This is the strongest general-purpose competitor. It preserves each region's autocorrelation and cross-correlation within blocks, and needs no derivation. We give it the same data and choose its block length by the standard data-driven rule, without tuning on ground truth. Because resampling every $n$-plet in a 2,514-region atlas is infeasible, we compare it with our method on a fixed random sample of $n$-plets at equal wall-clock budget, and on the full simulated systems where cost is not binding.
Ablation. To isolate the key ingredient, we drop the cross-lagged heterogeneity terms by forcing $\Gamma(\tau)=\rho(\tau)\Sigma$ with a pooled autocorrelation. If this ablation performs as well as the full correction, our mechanism is wrong.
What beating them means. Lower root-mean-square error against the analytic O-information in simulation. A false-positive rate for a group difference near nominal when two simulated groups share $\Sigma$ but differ in autocorrelation heterogeneity. On real data, corrected estimates that are stable when a scan is truncated to half its length, while the plug-in drifts.
Reported reference numbers. For the brain task there is no reported result found on Lacuna for O-information magnitudes or age-effect sizes. The file reports only that older subjects show higher redundancy at orders 3 to 20 and that synergy peaks at 6 to 8 regions, so we reproduce these values ourselves. For microbiome forecasting context, the file reports that MDSINE2 outperformed regularized gLV baselines and identified 17 modules in the healthy cohort. It also reports that cLV beat additive-log-ratio linear models on the mouse diet data (p = 0.008). Neither is an O-information number.
Why not just…?
Why not just correct with the effective sample size? Because the analysis predicts it is the wrong shape of correction. A scalar $T_{\text{eff}}$ rescales every subset's bias by the same factor, and the cancellation in $\Omega$ removes it entirely. If simulations show that a $T_{\text{eff}}$-corrected estimator fixes the false positives produced by heterogeneous autocorrelation, our central derivation is refuted. That outcome would be informative in its own right.
Why not just block-bootstrap? This is the real threat. In principle it captures the same bias without any Hessian algebra. Our case rests on two things. The first is cost: a closed form per $n$-plet versus thousands of resamples per $n$-plet. The second is variance near singular $\Sigma$, where resampled covariances can cross into rank deficiency. If the bootstrap matches our accuracy at feasible scale, we will say so, and the contribution narrows to a fast, interpretable approximation plus the identification of autocorrelation heterogeneity as the driver.
Why not just use shrinkage covariance or a coarser parcellation to avoid near-singular groups? Shrinkage pulls $\hat\Sigma$ toward a diagonal target and so shrinks redundancy by an amount that depends on $\Sigma$. That introduces a structured bias, which our derivation extends to cover. Coarsening changes the scientific question, because the redundancy core was found at fine resolution. We run both, and test whether the corrected fine-scale result agrees with the coarse-scale result, which is a consistency check neither alternative supplies.
Why not just match or regress out head motion? Motion differs by age and can induce autocorrelation, so it is a confound we must handle regardless. We include framewise motion as a covariate and test the corrected age effect within motion-matched subgroups. Motion matching alone does not address heterogeneous intrinsic autocorrelation, which remains even in still subjects.
Decisive Checks
The project succeeds or fails on four checks.
The first is analytic: in simulation, the plug-in O-information must be unbiased under independent and separable autocorrelated sampling, and biased in proportion to heterogeneity otherwise. If not, the mechanism is wrong.
The second concerns brain data. Across subjects, the per-subject correction term should correlate with measured autocorrelation heterogeneity. The fraction of the age-redundancy effect that survives correction is the headline number, whether it turns out to be most of the effect or little of it. Survival or disappearance of the middle-frontal and precentral redundancy core is the anatomical test.
The third is internal consistency: within subjects, truncating scans should move the plug-in but not the corrected estimate.
The fourth concerns the microbiome. A rise in redundancy after antibiotics counts as ecological dedifferentiation only if it holds on per-taxon absolute-abundance and square-root representations. It must also hold across choices of reference taxon, rather than appearing in one log-ratio chart alone. We additionally validate on synthetic modular gLV communities in which modules are merged on a known schedule, so the onset of dedifferentiation is ground truth.
References
- High-Order Interdependencies in the Aging Brain
- Generative Local Metric Learning for Nearest Neighbor Classification
- Representation Issues in the ML Estimation of Camera Motion
- Are Euler angles a useful rotation parameterisation for pose estimation with Normalizing Flows?
- Compositional Lotka-Volterra describes microbial dynamics in the simplex
- The Base Measure Problem and its Solution
- Robust and Scalable Models of Microbiome Dynamics
- Learning ecosystem-scale dynamics from microbiome data with MDSINE2
- Analysis of functional connectivity changes from childhood to old age: A study using HCP-D, HCP-YA, and HCP-A datasets
- Temporal and Spatial Scales of Human Resting-State Cortical Activity across the Lifespan
- Riemann2: Learning Riemannian Submanifolds from Riemannian Data