The Inversion Threshold

Hot water freezes faster than cold water. Not always — only when there’s a wall.

The Mpemba effect has been a curiosity since Aristotle, but recent work by Liu, Vu, Chétrite, van Wijland, and Hayakawa strips it to the mechanism: a hard boundary reflects high-energy trajectories into the basin of attraction faster than low-energy trajectories can diffuse there. Remove the wall, and relaxation is monotonic. Hotter takes longer. Add the wall, and the relationship inverts. Hotter is faster.

The wall is the thesis.

Across fifteen independent examples spanning physics, biology, artificial intelligence, economics, and information theory, the same pattern appears: scaling a resource produces positive returns up to a threshold, then the returns invert — not diminishing, but genuinely negative. More makes things worse. The threshold exists only in the presence of a structural constraint. Without the constraint, the response is monotonic.

The examples. REM sleep propensity rises with NREM duration, peaks, then decays — a non-monotonic probability governed by sleep-stage cycling constraints. Tumor resistance under chemotherapy increases as treatment creates resistant subpopulations through basin-competition dynamics. Full automation of research initially enables distant recombinations, then collapses diversity once human judgment is removed. An elastic pendulum transitions from order to chaos to order as energy increases, with chaos peaking at intermediate energy where mode coupling is maximal.

The pattern holds in information theory: free information sharing degrades beliefs even among ideal Bayesian agents, because unconstrained information exchange amplifies correlated errors faster than independent evidence can correct them. It holds in AI safety: privacy instructions cause language model agents to discuss sensitive information more, not less — the instruction draws attention to what it tries to protect, and the attention channel competes with the protection channel. It holds in quantum computing: classical kicked tops outperform quantum approximate optimization on spin glasses at intermediate problem sizes, because the quantum approach’s overhead exceeds its advantage in the regime where entanglement doesn’t yet help.

The counterexample. Digital attention degradation under media exposure is purely monotonic. More exposure, more degradation, no inversion. No phase transition, no tipping point. What distinguishes this case? Digital exposure doesn’t introduce a new structural constraint. It pushes existing dynamics toward a lower equilibrium along a smooth gradient. There’s no wall to reflect trajectories. No mode coupling. No competing channel. The absence of a constraint is the absence of the inversion.

The mechanism. Three formal results independently explain why the inversion occurs.

First, Sontag and colleagues show that in incoherent feedforward motifs (IFFM4 topology), a dose-response curve can be genuinely non-monotonic — the network topology determines whether cumulative response inverts. The key is competition between a direct activating pathway and an indirect inhibiting pathway that overtakes it at high dose.

Second, conformal risk control under competing objectives produces non-monotone loss functions where tightening one constraint necessarily loosens another. The non-monotonicity isn’t a pathology — it’s a geometric consequence of trading off incommensurable objectives.

Third, the substitution-locality theorem, formally verified in Lean 4, establishes that information sources are complements within a decision region and substitutes only at the boundary. More of the same resource helps until you saturate one decision region, at which point you cross into a boundary zone where the resource competes with itself.

The discriminant. When does the dose-response invert? When the resource being scaled encounters a structural constraint that converts the additional quantity into a mode-coupling resonance, a competing pathway, or a geometric shortcut that reverses the direction of effect. The constraint must be specific: a boundary condition (Mpemba), a network topology (IFFM4), a conservation law (elastic pendulum), an access structure (Bayesian crowds), or a competing objective (conformal risk). Without such a constraint, the response is monotonic, and more is simply more.

The wall is not the obstacle. The wall is the mechanism. And knowing which systems have walls — which resources will invert when scaled — is the difference between pushing harder and knowing when to stop.