Work

One query per state: a flag buys termination, a count buys behaviour

Year: 2026
Type: preprint
Abstract: A tape machine with two headers, a window and three trims answers a query at each step; in the companion paper its cost is measured only on the tapes whose walk meets. This paper asks what one query per state buys the machine, and measures two readings of the same query. The first reading is a flag: has this window occurred before? It buys termination. Over 18,000 tape-runs in sixty (rule, n) cells — twenty natural rules, three sizes, three hundred tapes a cell — 14.9% of the runs meet, 61.9% close a period and then repeat it for ever (the run is stopped at its first repeat and the outcome is recorded under the cap category), and 23.2% empty the region without ever meeting; the cycle is short, a median of two steps. With the flag the meeting rate is 37.8%, and 36.9% of the runs that would die are rescued. A repair that removes a boundary cell makes every run halt in at most (n+1)(n^2+1) steps, but it is not cheap: the read bound is O(n^3), no better than the most expensive rule in the catalogue, and a median rescued tape needs 73 trims at n = 512 and 212 at n = 2048, the worst 1,956. The flag is a query, not a memory: the machine keeps O(log n) bits and the visited set is the oracle's, with room for up to 2,945,071 quadruples in a single run at n = 2048. No lower bound is claimed: the paper measures what the query buys, not the least memory that would suffice. The second reading is a count: how often? The rule reads the visit count modulo k, and it buys behaviour rather than termination. Over three rules (pattern, runs, peak) at the same three sizes, the meeting fraction rises by up to 19 points as k goes from 1 to 4, and the gain is front-loaded: +28 to +41 tapes of 300 at k = 1 -> 2 for runs and peak, +9 to +11 at k = 2 -> 3 for pattern, and -1 to +9 at the last step. The curve decelerates and is not saturated by k = 4. A third to nearly half of the rescue needs no repair at all — 227 of 674 rescues at k = 2 and 356 of 808 at k = 4 — because the rule's variant mirrors the move, the walk goes the other way, and the period never closes: the count buys avoidance, not repair. Neither reading buys cheapness. Every fitted exponent stays on the rungs the catalogue already has (median reads over meeting tapes 1.287-2.016, over all tapes 1.715-1.922), the read bound stays O(n^3), and the distribution stays heavy-tailed: over all 36 colour cells the smallest g/e is 4.07 and the smallest e/G is 2.61, so the mean and the maximum never come within a factor three. And with a colour-blind rule — one whose table ignores the residue — the k = 2 model is not a different model: the repair fires on the second visit and its trim changes the region, which is part of the state key, so the bit is read only at counts 0 and 1, where "odd" and "ever visited" coincide; measured, the parity run and the flag run are bit-identical on all 18,000 tape-runs, 0 of 60 cells differing. Two further negatives are recorded, for the same reason: a cheaper repair that retracts the window instead of trimming the tape meets on 74.0% of runs and rescues 95.4%, but the walk can undo it, its repair count is unbounded (3,627,840 at n = 2048) and no termination theorem applies; and one bit of ink — a writable mark per cell, read together with the cell — buys nothing, because a mark read costs what a cell read costs, so a mark can only pay by answering about several cells at once, and one bit cannot. Termination is a theorem, every rate is a measurement with its denominator stated, and the sections say which is which.
Author: Luo, Haowei
Language: English
Open Access status: green
OpenAlex ID: W7227710370
Cites:
Cited by:
Related to: 10