Work

Reading the Curve, Not Counting the Points: A Two-Ring Petrov–Galerkin Curvature Force. An Abelian Control for Discrete Yang–Mills Dynamics on the Poincaré Homology Sphere

Year: 2026
Type: preprint
Abstract: A broken vessel A field on a mesh is a broken vessel. It was whole once; what is left to rebuild it from is the edges of the pieces. That is not a figure of speech. A computer never sees a gauge field whole, only its integrals along the edges of a mesh and its fluxes through small faces, and everything else has to be read back out of them. Archaeologists do this for a living. Hand one a single rim sherd and they will lay it on a rim chart and tell you the diameter of the whole pot, because the curve at the edge is enough. They are not counting points on the sherd. They are reading its curve. That is the idea of this paper, and the reason for its title. The error that refuses to go away The standard way to compute a field's force from a mesh, the Whitney reading, treats each little cell as if it were flat. Think of a robot arm driven through a list of waypoints: it moves in straight segments joined at corners. Give it more waypoints and the segments get shorter, but they never become a curve, because nothing in the controller measures one. A child fitting a peg into a hole does no arithmetic and still reads the curvature straight off the outline, from how the edge under one finger relates to the edge beside it. The difference is not how much either of them sees. It is what they measure. The flat-cell reading is the robot. On the space studied here its force error stops falling at 26% however fine the mesh gets. An error that shrinks as you refine is a resolution problem; an error that stops shrinking means the description itself is wrong, at every scale. Where it is measured: Poincaré's space Take a solid dodecahedron and glue each face to the face opposite, with a tenth of a turn before they meet. Everything closes up, with no boundary and no seams: this is the Poincaré homology sphere S3/2I, the space Poincaré built in 1904 as the counterexample to his own first guess, and the reason the Poincaré conjecture took the form Perelman finally proved. It is curved everywhere by the same amount and has no edge, so there is no flat corner for a flat-cell assumption to hide in, and no boundary to blame a bad number on. That makes it the honest place to test a curvature reading. Magnets, not clamps The fix came first as a physical intuition about how parts seat together. When you want two pieces to lock, the worst tool is a clamp, which holds whatever position you force, imperfections and all; the best is a magnet, which lets the parts find their own seat and pulls from a distance. So: stop clamping the curvature fit to the cell, and let the cell's neighbourhood draw it into place. In SU(2) this needs one more step that anyone who has lined up two sheets of paper will recognise: each neighbour is carried onto the cell's own sheet, into register, before anything is compared. Count the clues before you trust the answer A magnet does not say how far it has to reach. That has an exact answer, and it is a counting question. One ring is a clamp in disguise. A closed linear curvature on a cell has 11 unknowns. The ring of faces around it supplies exactly 11 readings. Eleven clues, eleven unknowns: the fit has no freedom at all, so it must pass through every reading exactly, noise included. It amplifies rough data by a median 125–252, and the resulting operator has 489 of 961 physical modes growing instead of decaying at n = 2. Two rings take a vote. Two rings supply 26–29 readings for the same 11 unknowns. Now the fit can weigh its witnesses instead of obeying each one: an equality-constrained least-squares vote, still exact on closed linear curvature. The amplification settles to a mesh-uniform median of 30.5–35.2. One order up, count again. Reading curvature to second order needs 26 unknowns. Two rings supply 26 readings: exactly determined again, and the median amplification jumps to about 1,490. A third ring restores the vote. That is the hybrid reading below. What the two-ring reading gives Accurate. Tested against the original Whitney variations, the two-ring force is consistent. Its error falls to 6.22%, 3.83%, 2.79% and 1.45% at mesh levels n = 1, 2, 3, 5 (observed orders 0.73, 0.80, 1.29) and passes the unchanged BW05 convergence gate. The 26% floor is gone. Stable. Every decomposed physical eigenvalue (n ≤ 3) lies in the right half-plane, the smallest at the control's true value of 4; at n = 1, 2 the symmetric part is positive definite, so the semi-discrete flow is dissipative. Nothing added. Nothing is fitted and no data are added: only the size of the neighbourhood the reading consults changes. Two to four well-chosen second-ring cells per cell already suffice; the same number chosen at random is unstable. Read one order up. Three rings, read to second order with the cell's own fluxes kept exact (the hybrid reading), bring the error to 0.158% at n = 5, against 1.45% for two rings, with observed orders 2.84, 1.88 and 1.83. It passes BW05 and the stability test. One honest cost. The force is not the gradient of an energy, and the paper shows quantitatively why every symmetrized variant fails. An accurate, stable and reciprocal force is left open. Into SU(2) For the non-Abelian group the readings must first be brought into register, as above: each neighbouring holonomy is carried onto the cell's own fibre by link transport. Without that step the force stays about 29% wrong, no better than the standard one. With it, the covariant recovery converges at second order on the Levi-Civita fields of the round and a Berger metric. On the Berger field the native SU(2) force stays 25.5% wrong at every level, while the two-ring force falls to 2.51% at n = 5 and passes BW05. Its linearised spectrum has no rough modes (the one-ring force has 207, then 1,451), and its numerical range, which bounds transient growth, has its left edge at −4.66 at n = 2, against −4.68 for the native force and −19,993 for one ring. One correction since version 2: in SU(2) the cell's own faces cannot all be held exactly, because the non-Abelian Bianchi identity is multiplicative. The fit matches them as closely as any closed field can, and the residual falls like h² (median 2.53%, 0.60% and 0.27% of the own readings at n = 1, 2, 3). The force results include it. Letting the lattice move Everything so far is a snapshot: a force measured once, on data sitting still. The harder test is a moving picture. The space itself changes shape underneath (the Berger metric (3, 1, 1) follows its Ricci flow over t ∈ [0, 0.02]), the links start exact and move under the force, and at the end they are compared with where the true field went. The control is the one a sceptic would ask for: run it again with the background refused permission to move, and see whether the lattice notices the difference. The result. The two-ring force ends 13.8%, 1.76% and 0.87% off the exact trajectory (error over the distance travelled) at n = 1, 2, 3. The native force ends 7.01%, 2.21% and 1.07%: native leads on the coarsest mesh, and two rings lead from n = 2 on. The control. Freezing the metric at its starting value leaves an error near 8% that does not shrink with the mesh, so the lattice demonstrably follows the moving background. Said plainly. Most of the field's motion over this window happens even with the metric frozen. Measured against only the part caused by the moving metric, the two-ring error is 10.6% at n = 3, down from 18–28% at n = 2: falling, and a tenth rather than a hundredth. A prediction that failed, and why. The native force was expected to stall in time, as its instantaneous error stalls near 25%. It did not. A first-order filter model, which predicts every measured trajectory error to within 6.1%, explains why: 99% of the native error at n = 2 lies in stiff directions that the dynamics damps within the window. A second forecast, that two rings would pull further ahead at n = 3, also missed: both errors halved. The step. The finest level uses the Lie implicit midpoint rule built on the Cayley map, which keeps every update exactly on SU(2). It comes from the companion paper Three Ways Across a Turn, where the Crank–Nicolson step is the tangent reading of a turn. Before use it reproduced the fourth-order runs at n = 1, 2 to within 1.5% with eight steps. An exact target: the Floor Theorem A lattice computation is only as trustworthy as the answer it can be checked against. For the Berger field, the Levi-Civita connection of the stretched sphere (3, 1, 1) read as an SU(2) gauge field, version 3 supplies one. In every sector of spin j/2 of the linearised Yang–Mills operator, (j(j+2) − 14)/3 is an exact eigenvalue on an explicit eigenvector. A Weitzenböck identity and certified interval-arithmetic inertia counts prove that it is the bottom of its sector in the four low sectors that survive on the Poincaré sphere, and bound every higher one above 142. On the Poincaré sphere the first sector above the left-invariant one begins 168/3 above the field's growth mode: one third of the sphere's first scalar eigenvalue. Both lattice forces resolve these values. The theorem is developed on its own in the companion paper The Floor Theorem for the Yang–Mills Hessian at the Berger Field (doi:10.5281/zenodo.24386614). In the engine The SU(2) computations are now GIGI engine verbs: RECOVER_CURVATURE … GROUP SU2, LINK_FORCE (which states in every answer whether the force is the gradient of an energy; only the native one is) and LINK_FLOW with the Cayley step. Each was held to a criterion fixed before it was built: the covariant reading agrees with the research code to 1.1 × 10−13, the force to 1.2 × 10−13, the linearisation to 2.5 × 10−11, and every Cayley run reproduces with identical Newton counts. Each mechanism, removed, turns its test red in the same run that passes. The two-ring force at n = 5 takes 10 s in the engine against 1,512 s in the research loop. What is new in version 3 The count one order up and the hybrid three-ring reading
Author: Davis, Bee
Institution: Brown University
Language: English
Open Access status: green
OpenAlex ID: W7235885961
Cites:
Cited by:
Related to: 10