CALLIOPE · BCOM

The Bayesian brain and Free Energy Principle propose that perception and action minimise prediction error, but these frameworks lack an explicit complexity constraint. Unconstrained Bayesian updating accumulates contextual detail indefinitely, incompatible with neural metabolic budgets. We propose that brain function is additionally governed by a Kolmogorovcomplexity prior formalised as the minimum description length (MDL) principle: the brain approximates the shortest generative program of "me in the world" that adequately explains sensory input. MDL is the algorithmic implementation of Occam's razor, partly implicit in the Bayesian Occam factor but requiring an explicit algorithmic bound to remain tractable.

Within this framework, chronic pain and tinnitus emerge as energetically adaptive yet clinically maladaptive states in which persistent percepts are integrated into the self-model to eliminate ongoing prediction error and reduce metabolic cost, explaining why percepts persist even as distress diminishes in most sufferers. Suffering arises through failure to compress the worldmodel, driven by intolerance of uncertainty and neuroticism. Large-scale longitudinal data confirm a precise dissociation: peripheral factors predict whether a phantom percept is experienced, while mood, neuroticism, sleep, and life stressors exclusively predict whether it produces suffering. This dissociation maps onto two distinct levels of Kolmogorov compression,

self-model and world-model, and is shared across tinnitus and chronic pain, implying a domaingeneral computational mechanism.

Sleep is mechanistically central: slow-wave ripples consolidate salience-weighted traces into neocortical generative models, while REM implements dropout-like regularisation by extracting context-invariant structure. Disruption creates a self-reinforcing loop in which an overfitted selfmodel is re-consolidated nightly, making sleep disturbance a mechanistic driver of suffering. Effective treatment must simultaneously target self-referential integration, world-model recalibration, and sleep architecture restoration, goals addressed by a high-definition transcranial infraslow pink-noise stimulation protocol (HD-tIPNS) paired with sleep optimisation.

The brain doesn't just predict — it compresses, and when that compression goes wrong in a specific way, you get chronic pain and tinnitus.

The standard story about the brain is that it's a prediction machine: it builds a model of the world, compares incoming sensory signals against that model, and updates when things don't match. This paper argues that story is incomplete. Prediction alone, left unconstrained, would cause the brain's internal model to bloat indefinitely — accumulating more and more detail to explain every sensory quirk. That's metabolically unsustainable. The authors propose the missing ingredient is a compression drive: the brain is also trying to maintain the shortest generative program that adequately explains "me in the world." This is Kolmogorov complexity, or operationally, the Minimum Description Length (MDL) principle — prefer the simplest model that fits the data. It's Occam's razor with a metabolic price tag attached.

With that framing, chronic pain and tinnitus snap into focus as a specific kind of compression failure. Both conditions arise from sensory deafferentation — the peripheral input goes quiet, but the brain keeps generating prediction errors because it expects a signal that isn't arriving. The brain's solution is elegant and brutal: it stops treating the phantom signal as an error to be resolved, and instead integrates it into the self-model. The ringing or the pain becomes part of "who I am," encoded as a background constant. This actually works energetically — acute pain costs roughly 60% more metabolic energy than chronic pain, which costs only 15% more than baseline. The compression is real. The cost is that the percept is now identity-level and nearly impossible to dislodge. EEG microstate data support this: both tinnitus and chronic pain patients spend about 50% of resting time in a brain state anchored to the left anterior middle temporal gyrus — a self-referential processing hub — rather than the posterior cingulate state seen in healthy controls.

But here's the key clinical puzzle: most people with tinnitus or chronic pain don't suffer severely. The percept persists, but it fades into the background. Only about 20% develop serious distress. Large-scale UK Biobank data (roughly 500,000 participants) reveal a clean dissociation: peripheral factors (hearing loss, obesity) predict whether you have the phantom percept at all, while mood, neuroticism, sleep quality, and life stress exclusively predict whether you suffer from it. The paper maps this onto two distinct compression levels. The self-model compression (integrating the percept into identity) happens in nearly everyone. What goes wrong in the suffering minority is world-model compression: the broader model of the environment becomes overfitted — everything feels potentially threatening, nothing gets explained away, salience stays chronically elevated. Neuroticism and intolerance of uncertainty are the psychological signatures of this failure.

Sleep is not a side note here — it's mechanistically central. Slow-wave sleep consolidates salience-weighted memories into the generative model; REM sleep acts like dropout regularization in machine learning, replaying experiences across shuffled contexts to extract only the invariant structure and prevent overfitting. In people who suffer, disrupted sleep means the overfitted, phantom-incorporating self-model gets re-consolidated every night, undoing any daytime progress. Stress raises noradrenergic tone, which suppresses REM, which prevents regularization, which deepens suffering — a closed vicious loop. The paper predicts that tinnitus appears in lucid dreams (where the self-model is fully online) but not ordinary dreams (where it's suppressed), and cites a recent survey confirming exactly this.

The therapeutic implications are specific: you can't fix this with a single-target intervention. The paper proposes a multi-node brain stimulation protocol (HD-tIPNS) using grey noise nested on an infraslow carrier wave, simultaneously targeting thalamocortical dysrhythmia, the noise-cancelling pathway, the salience network, and the self-model node — paired with deliberate sleep architecture restoration. The goal isn't to eliminate the percept but to demote it from a core self-model primitive to a low-salience background feature. Recovery means editing the compressed model, not suppressing its outputs.