README.md
7.3 kB · markdown
band-return-times
- The band automaton of a direction
(z1, z2)withz1 + z2 = wand3 | z1: states are the integersjin[-(z2-1)//2, (z1-1)//2], the moves arej -> (j + a)/3over the incrementsain{0, z1, -z2}whose quotient is integral, and the band is invariant. - Out-degree is 2 on
j = 0 mod 3, 1 onj = z2 mod 3and 0 on the third class, so the mean out-degree is1 + (n0 - n2)/Nfor the counts of those two classes among theNband states, and never 1 as an identity. Nconsecutive integers split the three classes mod 3 to within one, so|mean - 1| <= 1/Nalways and the mean is exactly 1 whenevern0 = n2, which3 | Ngives and does not exhaust:3 | Nis sufficient and not necessary.- The census the verb prints over
win{13, 40, 100, 101, 121, 257, 364, 1093}: 591 coprime directions, 465 of them exactly critical, 60 of those 465 with3not dividingN. - The two smallest witnesses, also printed by the verb:
(z1, z2) = (3, 1)has states{0, 1}, degrees2, 1and mean3/2, the smallest non-critical direction;(3, 2)has degrees2, 0and mean 1 atN = 2with3not dividingN. - A walk leaves 0 by the forced increment
z1and a return is a walk back to 0; a return of lengthnspells a multipliermwithm z1andm z2binary in base 3 on disjoint supports, som wis binary of lengthn. - The first return time
d(z)is the base-3 length of the shortest binary liftm w, and forw = R_k = (3^k - 1)/2the depthceil(d(z) / k)is the block depth of the pincer. - At the horizon
n = bk, for a block depthb, the slot profile of the column transfer is uniform, so the transfer is one fixed matrix inkper residuejand the return countL(k, bk)obeys a constant-coefficient linear recurrence inkwhose dominant root is the block ratelam_b. - Every column sum of every block deviates from
2^b/3by one of two values fixed by the parity ofb,-1/3or+2/3at evenband-2/3or+1/3at oddb, solam_blies in2^b/3 + [-1/3, 2/3]at evenband in2^b/3 + [-2/3, 1/3]at oddb. - The lift half of the same pincer is the submask count
N_K(m), the submasks of a binaryKdivisible by a divisormof it; its solution set is closed under complement inK, under disjoint union and under nested difference, so it is always even and counts the distinct unions of pairwise disjoint irreducible submasks. - The decomposition into irreducibles is not unique, so
N_K(m) <= #packings <= 2^(iota)for the irreducible countiota, the right equality holding exactly when the irreducibles are pairwise disjoint; a power-of-two count does not force disjointness and the two ladders differ. - A column transfer with a state set free of
kis a linear representation ofNas a series over the column word, so its state count is at least the Hankel rank; that rank reads3, 7, 14, 31, 62, 126, 253against the full3, 7, 15, 31, 63, 127, 255at word length1..7on each side, so no such transfer with fewer than 253 states exists atk <= 15and its unboundedness past there is observed and not proved. - The reversed reading of that matrix is its transpose at equal Hankel side lengths
p = qand is no second check; the irreducible countiotahas the same full-rank Hankel, so the bounded object is not the computable one. - The headline is two-sided at every
b,|3 lam_b / 2^b - 1| <= 2^(1 - b): the parity refines which edge is which, not the rate; the printed excess3 lam_b - 2^bis positive at everybthe verb reaches, so the upper edge is the live one.
WHAT IT COMPUTES
automaton: the state count, the edge count, the out-degree profile and the mean out-degree of the band automaton over the coprime directions of the given weights, with the band invariance and the identitymean = 1 + (n0 - n2) / Nasserted, the count of exactly critical directions and of those among them with3not dividingN, and the two smallest witnesses(3, 1)and(3, 2).returns:L(k, n), the number of primitive returns of the weightR_kinside horizonn, by a transfer matrix on the column carry; the per-block growthlam_bofL(k, bk); the excessrhoover the free model2^(n-1) / R_k; and the lengths at which no return exists.hist: the first-return histogram of every coprime direction of weightR_k, its bulk against the multiplier modelSum_{max T = t} 2^k / m_T, the depth histogram and the survival of the deep tail, and the deepest return againstsqrt(R_k).critical: the deepest and the median first return over a ladder of weights prime to 3, against the scalesqrt(w)a critical walk on a band ofw/2states predicts, with the leave-one-out range of thed_maxslope, the median exponent under cuts on the occupied countZ, the two-predictor fit onlog wandlog Z, and the least-squares diagnostics.model:D(k, N) = Sum over the primitive lifts of length at most N of 2^L / mat lift lengthL, the equidistribution count of return pairs, summed by the same transfer matrix and bracketed by the length of the lift.ladder: the fixed matrix atn = bk, its head lengthsr0(j), the exact characteristic polynomial ofL(k, bk)ink, the minimal polynomial oflam_bwith its radical form where the degree allows,lam_bas a certified interval by exact bisection, the parity column-sum bracket that holds it, andmax_j rho(B(b, j))againstlam_b.lift: the depth-2 lift census,M_k = Sum_T N_K(m)by meet in the middle with the Hankel rank of the sequence itself, the irreducible supplySum_T iota_Tand the largest one, the share ofTwhose count is trivial, whose count is a power of two and whose irreducibles are pairwise disjoint, the residue classesm_T mod 9of the ladder with the two-line argument that class 4 is empty, and the Hankel rank ofNand ofiotaover the column word.check: the transfer against a brute enumeration of binary multiples, the block-ladder identities, and the breadth-first return time against the shortest lift found by brute force.
HOW TO RUN
uv run python research/lab/py/band-return-times/returns.py checkuv run python research/lab/py/band-return-times/returns.py automaton --weights 13 40 100 101 121 257 364 1093uv run python research/lab/py/band-return-times/returns.py returns --kmin 3 --bmax 9 --kbig 160uv run python research/lab/py/band-return-times/returns.py hist --kmax 14uv run python research/lab/py/band-return-times/returns.py critical --wmin 2000 --wmax 400000 --step 1.15uv run python research/lab/py/band-return-times/returns.py model --kmin 8 --kmax 16 --bmax 8uv run python research/lab/py/band-return-times/returns.py ladder --bmax 14uv run python research/lab/py/band-return-times/returns.py lift --kmax 12 --kbig 17 --nmax 7liftcosts2^(1.5k)on theM_kladder and2^(2k-1)on the irreducible census, 170 s at the defaults, 80 s of it the Hankel rank at word length 7.histcostsw^1.5, 43 s atk = 14;laddercosts a second a step tob = 14and a minute tob = 20; every other verb is under 30 s.
WITNESSES
- coprime.md, THE WINDOW AT DIMENSION ONE: the lift half as a union count over irreducible submasks, and the return count of the critical band automaton, the block ladder at every depth, the block rate
lam_bas an exact algebraic number, the return-time support and the depth reading of the deep tail.