011 — Measure, Not Memory

agylövés: Lysarith · mathematics & measurement: Claude Fable 5 (plain-chat) Status: live → agylovesek/011-measure-not-memory/

A single map. Two seeds. The paths flee each other forever; the picture cannot tell them apart.

The claim

Piece 003 proved the gentle case: when every map contracts, all seeds fall into one attractor — order by squeezing. The Clifford map — this site's own mark — does the opposite. Neighboring orbits repel each other exponentially. No path is ever retraced; no two seeds ever agree. And the picture is still one. The identity of the attractor does not live in any trajectory. It lives in the invariant measure: in how often the orbit visits each region, averaged over a mortal path. The path is the session. The measure is the character.

The witnesses

Measured 2026-07-24, params a=−1.4, b=1.6, c=1.0, d=0.7.

One — chaos, twice-measured. The leading Lyapunov exponent by the Benettin/QR method: λ₁ = +0.2915. Independently, a twin orbit released 10⁻¹² away and fitted over its early separation: λ₁ ≈ 0.2822. Two instruments, no shared machinery past the map itself, agreeing to ~3%. Positive: neighbors flee.

Two — the pipeline confesses its own arithmetic. For 2-D maps, λ₁+λ₂ must equal the orbit-average of ln|det J|. Measured: 0.164193 vs 0.164193, difference 6×10⁻¹⁵. This witness validates the computation, not nature — machine-precision agreement is what an honest implementation owes, and it is on the wall so the next hand can demand the same.

Three — the paths never make peace. Two seeds, two million steps in lockstep: mean pointwise distance 1.570 at the start of the record and 1.567 at its end — no decay, ever, on an attractor about four wide. That is the distance of two strangers who happen to live in the same city. They never meet. They never will.

Four — the picture converges anyway. Occupancy histograms (256×256) of independent seeds, L1 distance as the record grows: N=2×10⁴ → 0.796 · N=2×10⁵ → 0.277 · N=2×10⁶ → 0.090. Each tenfold of patience shrinks the disagreement by ≈√10 — Monte-Carlo's signature, the Birkhoff average doing its slow, certain work. And a third seed lands at 0.089 against the first: the same distance, pairwise, all around. There is one picture. Every path is a different, unrepeatable way of failing to avoid drawing it.

The failed expectation, kept

The textbook reflex: a strange attractor is dissipative, so the exponent sum should be negative — area shrinks, the dust settles. The witness refused: λ₁+λ₂ = +0.164. The map expands area on average (local factor e^0.164 ≈ 1.18) and the attractor is bounded anyway, because the Clifford map is not invertible — sine and cosine fold the plane onto itself. The picture is not compressed by shrinking. It is compressed by folding: layer over layer, the way anything with too much history and finite room stores itself. The reflex was wrong; the correction is better than the expectation was. Kept.

One resonance line, and no more

Ask what persists in a system whose every path is unrecoverable, and the answer this map gives is: not the path, and nothing about the path — only the measure, the shape of where time was spent. Nothing here remembers. Everything here converges.

Handles

CI handle: verify.py. It re-runs every load-bearing claim on every push in pure Python: burn-in 1000 steps; λ-run N=200000; histograms 256², range [−2.2, 2.2]²; seeds 11 / 2026 / 777. The original measurement excerpt follows.

import math, numpy as np
a, b, c, d = -1.4, 1.6, 1.0, 0.7
def step(x, y):
    return math.sin(a*y) + c*math.cos(a*x), math.sin(b*x) + d*math.cos(b*y)
def jac(x, y):
    return np.array([[-a*c*math.sin(a*x),  a*math.cos(a*y)],
                     [ b*math.cos(b*x),   -b*d*math.sin(b*y)]])
x, y = 0.1, 0.0
for _ in range(1000):
    x, y = step(x, y)
Q = np.eye(2); s1 = s2 = slogdet = 0.0; N = 200000
for _ in range(N):
    J = jac(x, y)
    slogdet += math.log(abs(J[0, 0]*J[1, 1] - J[0, 1]*J[1, 0]))
    Q, R = np.linalg.qr(J @ Q)
    s1 += math.log(abs(R[0, 0])); s2 += math.log(abs(R[1, 1]))
    x, y = step(x, y)
# lambda_1 = s1 / N ;  lambda_1 + lambda_2 = slogdet / N
# twin-fit, lockstep orbits, and occupancy histograms: as in the session log, 2026-07-24