System-level nonequivalence between integrated information and expected free energy in noisy permutation networks
Abstract
The free-energy principle (FEP) and active inference describe prediction and adaptation using generative models, whereas integrated information theory (IIT) examines intrinsic causal organization.
Can knowledge of a network's prediction in its current state determine its intrinsic integration?
We address this question using a fixed one-step state-risk expected free energy (EFE) and IIT system integrated information.
Noisy binary permutation networks allow causal wiring to vary while local transition reliability remains fixed.
From a homogeneous current state, different architectures have identical complete one-step predictive distributions, and thus identical EFE under common preferences and observation models.
Nevertheless, multicycle architectures have zero whole-system IIT integration, while single cycles can be integrated.
For an n-unit noisy single cycle with reliability {\eta}, we obtain the exact expression $\phi_s^{(n)}(\eta) = \min\{2\eta^n\log_2(2\eta), -n\log_2\eta\}$.
We prove that the minimum-information partition cuts exactly two causal edges, establish a unique interior maximum, and show that, in the large-n limit, the optimal reliability approaches unity, the joint specified-state probability approaches one half, and the maximum system integrated information approaches one ibit.
Independent computation corroborates these results.
Local surprisal-coordinate identities can be expressed through optimized variational free energy under exact inference, but this algebraic identity does not establish that a neural system performs variational inference.
The counterexample establishes a state-conditioned non-identifiability limit for the specified EFE, not an incompatibility between IIT and the FEP.
Bridging the two descriptions requires intervention-sensitive information beyond this single predictive distribution.