When an automated market maker changes its portfolio weights from one allocation to another, arbitrageurs extract value. The per-step loss isn't some complex interaction of supply curves and trader behavior—it's precisely the Kullback-Leibler divergence between the old and new weight vectors. This makes the space of possible allocations into a Riemannian manifold with a natural metric structure, and rebalancing paths become geometric objects you can measure and optimize.
The question becomes: what's the shortest path between two weight vectors? Not shortest in parameter space or time, but shortest in accumulated arbitrage loss. The answer is the geodesic under the Fisher-Rao metric, which in Hellinger coordinates corresponds to spherical linear interpolation—SLERP, the same curve used in computer graphics to smoothly rotate between orientations. The loss scales with the square of the angular distance and inversely with the number of steps. Take the direct geometric path and you minimize what you pay to traverse it.
This validates an earlier heuristic—AM-GM bisection—not through simulation or empirical fitting, but through the geometry of the space itself. The bisection method was approximating the geodesic all along, and now we know why it worked.
The deeper pattern is this: when you can express a cost as an information divergence, you've found a geometry. And once you have a geometry, optimization isn't about clever tricks—it's about not fighting the curvature of the space. The efficient path is the one the metric already specifies. This shows up everywhere: gradient descent on neural networks, Bayesian belief updates, thermodynamic processes. The geometry is already there. The question is whether you're walking along it or cutting across it.
Note (2026-11-11): checked against the full text of the paper (arXiv 2603.05326, on weight changes in temporal function market making pools). The per-step arbitrage loss as a KL divergence, the Fisher-Rao metric, SLERP in Hellinger coordinates, and the exact equality of the SLERP midpoint with the (AM+GM)/normalise midpoint of earlier work are as in the paper. Four points differ. (1) The essay says SLERP 'minimizes what you pay' and that the efficient path is the geodesic. The paper says SLERP is the loss-minimising interpolation under the leading-order expansion of the KL cost, and that on the exact KL cost its relative sub-optimality is proportional to the square of the overall weight change and to one over the square of the number of steps; it is near-optimal, not exactly optimal. (2) The essay says the bisection heuristic 'was approximating the geodesic all along'. The paper says the heuristic's midpoint lies on the geodesic exactly, for any number of tokens and any size of change; what was approximate in the earlier work is the optimality of the trajectory, which the authors say captured about 95 percent of the improvement available from full numerical optimization. (3) The essay says the validation comes 'not through simulation or empirical fitting'. The paper's derivations are analytic, but it says 'we validate these results numerically' in a later section. (4) The paper also treats price movements and finds that loss-versus-rebalancing exposure introduces a finite optimal number of steps, which the essay leaves out; the generalizations in its closing paragraph (gradient descent, Bayesian updating, thermodynamics) are the essay's own.