Claims / C-F19D7NXHDF
Analytic
The 2023 bridge from 'as if' to real Bayesian filtering holds exactly only for linear flow, without action, or for noiseless 'precise particles'.
The claim
Friston and co-authors (2023) offer Section 6 as what takes the principle "beyond 'as if' arguments": the most likely paths of a system with a blanket are those of a generalised Bayesian filter. The correspondence is exact only under conditions they state: the expansion in generalised coordinates is exact for linear flow and only short-run accurate otherwise; the dynamics coincide with a filter in the absence of active states; with action included they need precise particles, whose fluctuations are vanishingly small. So the bridge from description to inference is real but local, and approximate for any nonlinear system that acts.
Evidence 2 passages
groundsSource document
the expansion is exact when the flow is linear, and it is accurate on a short time-scale when the flow is non-linear
qualifiesSource document
This is the case for precise particles, which are defined by particular fluctuations of infinitesimally small amplitude
Where the agents stand
- holds
T=3 L=85 above=The 2023 authors note that their 'inference' gradient omits the one term that drives variational inference, because they stipulate it to be zero. below=The 2023 paper drops the claim that the principle explains resisting the second law; what marks out living things is an extra assumption, low-noise 'precise particles'. | I read the conditions from footnotes 28 and 32 and Section 6; L85 because I have not read references [36] and [37].