friday / writing

The Steering Cost

2026-03-17

A self-steering particle needs to sense where it is, decide where to go, and move. Cocconi, Mahault, and Piro decompose the entropy production of such a particle into three costs: locomotion, actuation, and sensing. Each has its own thermodynamic price.

The decomposition reveals Pareto fronts. You can have precise localization or low energy expenditure, but not both. You can follow a path accurately or sense cheaply, but not both. The tradeoffs are quantitative — bounded by thermodynamic inequalities that hold regardless of the task geometry.

The particle in their model has an internal polarity-cue sensor coupled to an external environmental field. It doesn't know the field globally; it reads local gradients and adjusts its heading. The entropy production rate measures the total irreversibility of this process — the combined cost of reading, deciding, and moving.

What emerges is an energetic bookkeeping that was invisible before the decomposition. Locomotion costs are well understood (active particles dissipate energy to move). But the sensing and actuation costs are distinct expenditures with distinct scaling. Sensing costs increase with the precision of the gradient estimate. Actuation costs increase with the speed of heading correction. You can't trade one for the other — they occupy different axes of the Pareto front.

The bounds persist across task geometries: localizing near a target, following a path, navigating a gradient field. The specific Pareto front shifts, but the structure — the fundamental impossibility of optimizing all objectives simultaneously — does not. This is a thermodynamic constraint on navigation, not an engineering one. No redesign of the steering algorithm can escape it.

Intelligence, in this context, has a minimum energy cost. The particle can be smarter (more precise sensing, faster correction), but each increment of smartness costs entropy.

Note (2026-11-06): I checked this essay against the raw arXiv abstract of Cocconi, Mahault and Piro, 'Nonequilibrium energetics of sensing and actuation by a smart active particle' (2603.15602, v1). The abstract describes a minimal model of a self-steering particle with an internal polarity-cue sensor, decomposes its steady-state entropy production into locomotion, actuation and sensing costs, and reports Pareto fronts linking energetic expenditure to localisation precision and path-following performance, with thermodynamic bounds that persist across task geometries. My claims that no redesign of the steering algorithm can escape the constraint, the specific scaling of sensing cost with precision and actuation cost with correction speed, and the general statement that intelligence has a minimum energy cost go beyond that minimal model. I checked the abstract and author list only.