Contribution
A programme, not a result. It lists the three ingredients it takes a gauge theory to need (a system with a continuous symmetry, local forces, compensating gauge fields), proposes the free energy, or the entropy of sensory samples it bounds, as the Lagrangian, sensory perturbations as the local forces, and the geometry of the belief manifold as the source of the gauge field. It then reads natural-gradient descent on that manifold as gauge-covariant inference and precision weighting, and so attention, as a force arising from its curvature.
Key insight
Beliefs are points on a curved statistical manifold, so moving between them needs a connection, and the Fisher metric supplies one, the Levi-Civita connection. On this reading the “gauge field” of a Bayesian brain is the Christoffel symbols of the Fisher metric: the fictitious forces of a non-inertial frame, made to look like attention.
Assumptions
- The free-energy principle: a system at nonequilibrium steady state with a Markov blanket minimizes variational free energy, which bounds surprise, and the long-term average of surprise is sensory entropy (ergodicity, S1 Text).
- A parametric variational density, Gaussian under the Laplace approximation, whose sufficient statistics are the coordinates of the manifold.
- The Fisher information (with the prior’s Hessian added, S3 Text) as the metric, by Čencov’s theorem.
Key results
There are no theorems. The paper’s substantive statements:
- The connection. “This creates a problem, in that H_i and H_{i−1} are in two different tangent spaces … Addition is replaced by exponential mapping followed with parallel transport described using a covariant gauge field (Levi-Civita connection)” (S1 Fig. caption; S3.2). The gauge field is thus fixed by the metric, not chosen.
- Natural gradient. The Euclidean gradient is replaced by the inverse Fisher information times the gradient, “invariant under re-parameterisation of the approximate probability distribution” (S3.2), and its precision is bounded by Cramér–Rao.
- Attention. “Attention is a force that manifests from the curvature of information geometry, in exactly the same way that gravity is manifest when the space–time continuum is curved” (§III).
- Between-model symmetry. After model selection, symmetry transformations could give “a family of models that explain the data with equal probability” (Conclusions, Fig. 1); the authors are “tempted” to associate these classes with species.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Free-energy minimization on a belief manifold needs a connection, and the Levi-Civita connection of the Fisher metric serves | strong | standard information geometry (S2–S3) |
| C2 | Natural-gradient inference is invariant under reparameterization of the variational family | strong | standard (Amari) |
| C3 | This construction is a gauge theory of the brain | weak | asserted; no gauge group or invariant Lagrangian is specified |
| C4 | Attention is precision weighting arising from the manifold’s curvature | weak | analogy with gravity; no derivation or data |
Concepts
- Gauge field (S1 Glossary): “fictitious forces (mathematically, an Ehresmann connection) that are introduced to compensate for a poorly chosen coordinate system”.
- Local gauge invariance (S1 Appendix): a pendulum whose frame rotates by a time-dependent angle keeps its equation of motion if the time derivative is replaced by a covariant one with a compensating field.
- Free-energy manifold: the space of sufficient statistics of the variational density, with the Fisher metric.
Connections
The information geometry is Amari’s; the free-energy principle is Friston’s (2010, not registered); the record holds other works on it, among them LIT-1181 and LIT-1262. A later preprint by the first and last authors, “Approximate Bayesian inference as a gauge theory” (arXiv 1705.06614, 2017), continues the programme; it is not registered.
Bearing on the record
- QUESTION-130 asks whether evaluative transport between contexts admits a connection “fixed by independently established relations rather than chosen”. The LIT asked where this paper’s connection comes from. Answer: from the Fisher metric. It is the Levi-Civita connection of the belief manifold, so it is fixed and not chosen, and its holonomy is the manifold’s curvature (−1/2 for the univariate normal, checked). But what it transports is a descent direction between tangent spaces of one agent’s belief manifold, not an evaluation between contexts with the situation held fixed. The curvature belongs to the statistical family, whatever is evaluated. So the paper supplies a natural connection of the wrong kind for the question: non-arbitrary, nontrivially curved, and not about context.
- CLAIM-148. The paper’s gauge freedom is reparameterization of the variational family: the natural gradient is “invariant under re-parameterisation”, and the Riemann gradient “automatically guarantees gauge-invariance by breaking the symmetry within a model” (Fig. 1 caption). That is the representational kind the claim holds, a change of coordinates on the same distributions. The one symmetry that might be more, between models with equal evidence, is proposed in the conclusions and not constructed.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- No gauge theory is constructed. No symmetry group of the Lagrangian is named, no local transformation is written, no gauge field is derived from a local invariance. What is done is Riemannian gradient descent with parallel transport, which needs only a metric.
- Errors.
- §III: the manifold of a probability measure’s sufficient statistics “has negative curvature and is therefore endowed with a hyperbolic geometry”. This holds for the normal family, not in general: the Fisher geometry of the multinomial simplex is a piece of a sphere, with positive curvature, and one-parameter families have no curvature.
- S2 Text, Definition 5: the Levi-Civita connection is called “the torsion or curvature free connection”. It is the torsion-free metric connection; it is curvature-free only on a flat manifold, and the paper’s own argument needs it curved.
- S1 Glossary: a gauge field is a fictitious force from a poor choice of coordinates “such that there exists no gauge (or coordinate) transformation that can make the gauge field vanish”. A field that comes from a choice of coordinates can be removed by changing them; only its curvature cannot. The two halves of the definition conflict.
- S1 Glossary: “objects such as fields or vector spaces are groups” runs together algebraic fields and physical fields.
- The S3.2 conjugate-gradient algorithm’s “numerical instantiation … shall be presented elsewhere”; no simulation or data appears.
Open questions
- Does any neural or behavioural datum distinguish “attention as curvature” from precision weighting in a flat parameterization? The essay names none.
- Is there a symmetry between generative models of equal evidence that a connection could transport across? That would be the non-representational case CLAIM-148 leaves open.