Probing INSIDEOUT: A Mathematical Analysis of Thermodynamic Connectomics

A rigorous audit of a popular brain-dynamics metric that finds it measures something simpler—and more useful—than advertised.

The INSIDEOUT framework claims to detect the brain's "arrow of time" by measuring thermodynamic irreversibility in fMRI data. The core idea is appealing: a healthy, conscious brain is far from equilibrium, and that non-equilibrium character should leave a statistical fingerprint in the temporal asymmetry of brain signals. The empirical results are real and clinically interesting—consciousness loss, anaesthesia, and psychiatric disorders all shift the metric. But this paper asks: what is the math actually doing, and does it justify the thermodynamic story?

The answer is that the entire INSIDEOUT computation reduces to something much simpler than it appears. You compute one matrix of lagged mutual informations between brain regions, then measure how asymmetric that matrix is—specifically, the squared Frobenius norm of its antisymmetric part. The elaborate step of time-reversing the data, which gives the framework its dramatic framing, turns out to be mathematically redundant: it just re-expresses the negative-lag content of the forward correlation using a positive-lag estimator. No new information is created. The "arrow of time" detection collapses into a familiar lead-lag analysis.

The thermodynamic interpretation is real but conditional. Under a specific model—the multivariate Ornstein-Uhlenbeck process, which is linear, Gaussian, and Markovian—asymmetric lagged correlations are provably equivalent to broken detailed balance and nonzero entropy production. The metric has the right zero: it vanishes exactly at equilibrium. But the actual entropy production rate has a different mathematical form (it requires weighting by the inverse covariance matrix), so INSIDEOUT's magnitude is not a calibrated thermodynamic quantity. More importantly, the OU assumption is never empirically tested for fMRI data, and there is a genuine alternative explanation: a fully reversible deterministic system observed through a narrow measurement window will also look irreversible, because information flows in and out of the observable subspace. The paper proposes a concrete test—vary parcellation resolution and see if the metric scales with it.

The most consequential design flaw is sign-blindness. The Fisher-Stratonovich transform squares the correlation before computing asymmetry, discarding whether region i excites or inhibits region j. This means the metric can detect that asymmetry exists but not who is on top in the directed hierarchy, and it cannot distinguish entropy production from entropy reduction. Ironically, the instantaneous variant iFLOW actually computes the signed asymmetry frame-by-frame—and then immediately throws the sign away by taking absolute values. A signed reformulation is trivially available: retain the antisymmetric matrix D_ij = r_ij(τ) - r_ji(τ), sum over pairs to get per-region source/sink scores, and track these over time. This would identify which regions lead versus follow, whether coupling is excitatory or inhibitory, and map more directly onto the entropy production rate the framework is trying to approximate.

The paper closes by situating INSIDEOUT within the broader Kolmogorov Theory program of algorithmic agency, noting that dissipation and computation are conceptually distinct—a seizure is highly dissipative without being usefully computational—and that the unsigned scalar cannot decompose what the directed flows are actually doing. Five concrete forward steps are proposed: signed reformulation, multi-resolution coarse-graining tests, HRF deconvolution controls, multi-lag spectral extension, and an inverted-U test of whether healthy brain states correspond to intermediate rather than maximal irreversibility.