Whole-Brain Models and Algorithmic Irreversibility: The Arrow of Time as a Property of Observation, Not of the Brain

The "arrow of time" in brain signals is a property of how you look at the brain, not of the brain itself — and once you see that, a lot of popular neuroscience claims about irreversibility collapse.

Here's the core intuition. Play an EEG recording backward and a statistician can tell it's reversed. That asymmetry — the "arrow of time" — weakens under anesthesia and deep sleep, which has led researchers to treat it as a thermometer of consciousness or metabolic computation. This paper argues that reading is mostly wrong, and that getting it right matters for both interpreting data and building whole-brain models.

The key move is separating what "irreversible" actually means. The paper identifies four distinct levels. At the bottom, the exact microscopic dynamics are time-reversible: no information is destroyed, and in principle you could run the film backward. An arrow only appears when an observer throws away information — by watching through a coarse, lossy window rather than the full state. This lossy mapping (called coarse-graining, or equivalently projection) loses information in two ways: finite precision, where dynamical contraction blurs distinct states together, and partial observation, where information leaks into variables you never recorded. The arrow you measure lives entirely at the observer's level. Actual thermodynamic heat cost (Landauer's bound) only appears when a physical device performs an explicit reset — a genuinely many-to-one collapse, not just a smooth dynamical flow.

The payoff for brain modeling is direct. Neural-mass and mean-field whole-brain models (WBMs) are themselves a coarse-graining: one field stands in for millions of neurons. So when a WBM exhibits a strong arrow of time, that's not a discovery about the brain's thermodynamics — it's the expected shadow of the averaging you performed when you wrote the equations down. The noise term in the model equations is not fundamental stochasticity; it's the microscopic detail you averaged away, reappearing as apparent randomness. The irreversibility was put in by the modeler's own hand.

What can a measured arrow actually tell you? Safely: that there is directed statistical dependence in the chosen observable — the present of one region predicts the future of another asymmetrically. That's real and useful as a brain-state marker. What it cannot tell you: how much entropy the brain is producing, or how conscious the subject is. A perfectly reversible system watched through a small window produces the same number. The paper also draws a careful line between directed dependence and computation: the former is necessary but not sufficient for the latter. Transport (delays, volume conduction, shared stimulus) produces strong directed dependence while computing nothing; genuine computation requires compressive flow — directed dependence that shortens the description of a task-relevant variable.

What remains genuinely informative is not the size of the arrow but where and when it concentrates. In a predictive-coding model, the arrow should localize to circuits performing model revision — the slow-to-fast gating machinery that compares predictions against inputs. A WBM, unlike a recording, can actually test this: vary the parcellation resolution, compare full-state to projected-signal arrows, use phase-randomized surrogates, and condition on the external stimulus. These controls disentangle "arrow from observation" from "arrow from dynamics" — and they're the concrete next steps the paper calls for.