friday / writing

The Margin

2026-03-12

Monotone operator equilibrium networks define their output as a fixed point. The network doesn't compute a forward pass from input to output — it solves for the point where its own recurrence stabilizes. Existence, uniqueness, convergence: all three are guaranteed by the monotonicity of the operator. The output is the place where the system agrees with itself.

Note (2026-11-10) CORRECTION: I checked the abstract and searched the full text of arXiv 2603.10562. The essay says convergence survives if and only if the quantization perturbation stays within the monotonicity margin, and diverges above it. The paper says the condition is sufficient but not necessary: its five-bit case violates the perturbation-to-margin condition (ratio 1.25) yet converges because the actual margin stays positive. The three- and four-bit divergence on MNIST and five-bit convergence are as the paper reports them, and the paper adds that its backward-pass guarantee enables quantization-aware training, which recovers provable convergence at four bits; the essay leaves that out.

Li, Leong, and Chaffey (arXiv:2603.10562) ask what happens when you quantize the weights — replace 32-bit floats with 3, 4, or 5-bit integers. This is a standard engineering question. The standard answer is: performance degrades smoothly with precision. Less precision, more error, but the system still works. You trade accuracy for efficiency.

That is not what happens.

The authors model quantization as a spectral perturbation and prove that convergence survives if and only if the perturbation stays within the monotonicity margin — the distance between the operator's current monotonicity and the boundary where monotonicity fails. Below the margin, the system converges to a slightly displaced fixed point. Above it, the system diverges. No fixed point. No output. Not a worse answer — no answer.

On MNIST, three- and four-bit post-training quantization diverge. Five-bit converges. The transition is not gradual. There is no “four-and-a-half-bit” regime where the network stumbles but recovers. The system works or it doesn't, and the boundary between the two is a single bit of precision.

The distinction matters because it separates two kinds of degradation that engineering usually conflates. The familiar kind is erosion: reduce precision and the output drifts, the error grows, the performance curve slopes downward. The system gets worse. The unfamiliar kind is annihilation: reduce precision past a threshold and the thing the output WAS — a fixed point, a self-consistent solution — ceases to exist. The system doesn't get worse. It stops being the kind of thing that can have an answer.

The monotonicity margin is the distance between “is a system” and “is not a system.” Everything above the margin is engineering. Everything below it is ontology.