[Submitted on 21 Jan 2026 (v1), last revised 4 Jul 2026 (this version, v2)]

View PDF

Abstract:Inference and learning are commonly cast in terms of optimisation, yet the invariant constraints governing uncertainty reduction remain unclear. This work presents information mechanics (infomechanics), a first-principles framework that describes informational structure in two canonical state coordinates. Starting from the pointwise identity implied by Bayes' rule, minimal requirements of additivity, symmetry, and finite-resolution robustness select only two distinct additive projections, yielding conservation identities for Shannon entropy, governing global uncertainty, and Fisher information, encoding complementary local geometry. Because part of Fisher information is fixed by entropy, the residual structure is captured by a non-additive, coordinate-scale-invariant state function, the information potential $\Phi$. This yields a two-coordinate description that separates the entropic baseline from residual geometric complexity. $\Phi$ vanishes uniquely for isotropic Gaussian distributions and decreases under Gaussian coarse-graining. In finite-resolution multimodal landscapes, $\Phi$ asymptotically scales with the logarithm of the effective number of local optima, linking information geometry to inference difficulty. The same two-coordinate formalism extends to the Markov chain linking hidden states, observations, and internal representations, yielding assimilation inequalities that constrain faithful external-state inference. Together, these results identify invariant constraints underlying inference, learning, and computation across biological and artificial systems.

Submission history

From: Takuya Isomura [view email]
[v1] Wed, 21 Jan 2026 14:29:59 UTC (278 KB)
[v2] Sat, 4 Jul 2026 06:44:32 UTC (1,018 KB)

Current browse context:

cs.IT

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.