Gauges: When There Is No Shared Frame
Published:
The problem, restated precisely
The geodesics chapter ended stuck: building convolution on a surface needs an angular reference at every point, and no intrinsic rule supplies one.
Gauge theory reframes the difficulty as a feature. Accept that the choice is arbitrary. Make it anyway, at every point, and then insist that nothing observable depends on what you chose.
A gauge is a choice of local frame at a point. A gauge transformation is a change of that choice — at one point, independently of every other. A model is gauge equivariant when transforming the frames transforms the features correspondingly, so that anything you actually predict is unaffected.
Transport, and why it does not commute
If features at \(u\) and \(v\) live in different frames, comparing them requires a map. Call it \(\mathcal{F}_{u \to v}\): the transport taking a vector expressed in \(u\)’s frame to the same vector expressed in \(v\)’s.
Now follow a closed loop. Transport a vector from \(u\) around some cycle and back to \(u\). It has returned to its starting point, expressed in its original frame. Is it the vector you started with?
Not necessarily. The composition of the transports around a loop is the holonomy, and when it is not the identity the connection is curved.
Take three transports, each a 90° rotation:
A vector \((1,0)\) sent round that loop comes back as \((0,-1)\) — same point, different vector. Whereas three 120° rotations compose to \(R(360^\circ) = I\), and every vector returns unchanged: that loop is flat.
This is a sheaf
Everything above has a discrete counterpart, and it is the object Book III is about.
| Gauge theory | Cellular sheaf |
|---|---|
| local frame at a point | stalk \(\mathcal{F}(v)\) at a node |
| transport between frames | restriction map \(\mathcal{F}_{v \trianglelefteq e}\) |
| gauge transformation | change of basis in a stalk |
| flat connection, zero holonomy | global section exists |
| curvature | obstruction measured by \(\Delta_{\mathcal{F}}\) |
A cellular sheaf on a graph is a discrete connection. The restriction maps are the transports; a global section is an assignment consistent under all of them; and the sheaf Laplacian \(\Delta_{\mathcal{F}} = \delta_0^{\top}\delta_0\) measures the failure of consistency, with kernel exactly the global sections.
The triangle holonomy example in the sheaf overview is the same calculation as the rotation composition above, with restriction maps in place of rotation matrices.
What it costs
Parameters. A scalar edge weight becomes a \(d \times d\) matrix per edge. That is the price of transport being a linear map rather than a number, and it is why the sheaf literature keeps finding that diagonal restriction maps — \(O(d)\) per edge rather than \(O(d^2)\) — perform as well as full ones. Three independent papers now report it.
Theory that does not survive contact with the architecture. The spectral guarantees hold for the linear diffusion. Once you insert nonlinearities, attention gates, or swap the Laplacian for something else, they stop applying — as DNSD does deliberately and BrainDyn does inadvertently.
A symmetry you may not need. Gauge equivariance is the right constraint when the domain genuinely lacks a global frame — meshes, molecules in local coordinates, heterogeneous graphs whose node types have incomparable feature spaces. When one global frame is fine, the machinery is pure overhead.
The blueprint, closed
That completes the five domains:
| Domain | Symmetry | Architecture |
|---|---|---|
| Grid | translation | convolution |
| Group | any group \(G\) | group / steerable convolution |
| Graph | permutation | message passing |
| Geodesic | isometry | geodesic convolution |
| Gauge | frame change | sheaf / connection Laplacian |
One procedure, five instantiations, each forced rather than invented. And running underneath, the equivalence showing rows two and three are the same operator on different graphs — with the hardware, not the mathematics, deciding which one the field adopted.
✅ Key Takeaways
- A gauge is a choice of local frame; gauge equivariance means predictions do not depend on the choice. It is the answer to the ambiguity the geodesics chapter could not remove.
- Comparing features across points requires explicit transport. Composing transports around a loop gives the holonomy.
- \(R(90°)^3 = R(270°) \neq I\) — a vector \((1,0)\) returns as \((0,-1)\). \(R(120°)^3 = I\) is flat. Curvature is the information, not a defect.
- A cellular sheaf is a discrete connection: stalks are frames, restriction maps are transports, global sections are flat assignments, and \(\Delta_{\mathcal{F}}\) measures the obstruction.
- A sheaf GNN is permutation-equivariant and gauge-equivariant — the graph and gauge rows simultaneously.
- The cost is \(d \times d\) per edge, which is why diagonal maps keep proving sufficient, and theory that applies to the linear diffusion rather than the trained network.
References
- Bronstein, M. M., Bruna, J., Cohen, T., & Veličković, P. (2021). Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. arXiv:2104.13478.
- Cohen, T., Weiler, M., Kicanaoglu, B., & Welling, M. (2019). Gauge Equivariant Convolutional Networks and the Icosahedral CNN. ICML 2019.
- Hansen, J., & Ghrist, R. (2019). Toward a Spectral Theory of Cellular Sheaves. Journal of Applied and Computational Topology, 3(4), 315–358.
- Bodnar, C., Di Giovanni, F., Chamberlain, B. P., Liò, P., & Bronstein, M. (2022). Neural Sheaf Diffusion. NeurIPS 2022.
- Singer, A., & Wu, H.-T. (2012). Vector Diffusion Maps and the Connection Laplacian. Communications on Pure and Applied Mathematics.