betaFeedback

WorksPhysics and Astronomy

Random-walk scaling without Bernoulli statistics: transport, persistence and resonances in the triangle map

Abstract

The triangle map is an area-preserving, piecewise-parabolic map of the torus with vanishing Lyapunov exponent. We study the symbolic process Xn=±1X_n=\pm1 given by the two halves of the torus, whose partial sums SnS_n are the transport of the lifted momentum, with an integrator free of accumulated rounding errors and exact computations on polygonal cylinder sets. We prove that the Koopman operator has no non-constant H1H^1 eigenfunctions, that the symbol is not an H1H^1 coboundary, and that runs of equal symbols are uniformly bounded. For generic parameters the walk is normally diffusive and Gaussian, but its diffusion coefficient is not that of a coin: away from the edges of the parameter square it varies between 0.20.2 and 1414 times the Bernoulli value, and for a benchmark pair, with D=3.09±0.01D=3.09\pm0.01, the excess is produced by the order of the bounded inter-event intervals, not by their distribution. The first-passage exponent is 1/21/2 with a non-Bernoulli amplitude, and worst-case Birkhoff averages decay as for a uniquely ergodic system. For β=0β=0 the Green-Kubo sum cancels, D=0D=0, yet Var(Sn)\mathrm{Var}(S_n) keeps growing slowly, sign changes are rare and the correlations of Fourier modes decay as n−0.2n^{-0.2}. We identify the zero-sum words, which become pure twists at β=0β=0, as resonances that trap orbits with a computable 1/ℓ1/\ell law, set the crossover time β−1/2β^{-1/2} at small ββ, and, in an asymmetric deformation of the symbol, produce a logarithmic term in the variance whose coefficient we compute exactly. We formulate conjectures on unique ergodicity, mixing, the central limit theorem and the cohomology of the symbol.

Lead paragraph. Diffusion in deterministic systems is usually traced to chaos: nearby orbits separate exponentially, the motion quickly forgets its past, and the dynamics behaves like a sequence of coin tosses. Many systems, however, diffuse without any exponential instability; billiards in polygons, where nearby trajectories separate only linearly in time, are the classical example. We study the triangle map, a simple model of the irrational triangular billiard, and ask what its diffusion actually reveals. Scaling alone turns out to say little: the symbolic walk spreads as the square root of time with Gaussian fluctuations, as a coin would, but its diffusion coefficient depends on the parameters and is fixed by the order in which waiting times occur, not by their statistics. For a special parameter the diffusion nearly stops. We trace this to resonant symbol patterns that trap orbits for long times, and show how a small perturbation turns these traps into measurable transport.

Introduction

Deterministic dynamical systems may display statistical properties usually associated with stochastic processes even in the absence of exponential instability and positive entropy [1, 2, 3]. Polygonal billiards and related piecewise isometric systems are the natural setting for this phenomenon [4]. Billiards in a dense GδG_\delta set of polygons are ergodic [5] and weak mixing is also generic [6]; recent preprints announce ergodicity and weak mixing for every triangle with an angle irrational relative to π\pi [7, 8], building on transverse regularity and the cohomological equation for flows on flat surfaces [9]. Numerical studies have long reported mixing-like, diffusive and anomalous transport in such systems [1, 10, 11, 3], whose dynamical origin is only partly understood.

The triangle map of Casati and Prosen [12], a model of the irrational triangular billiard, is one of the simplest systems in this class. It is the area-preserving map of the torus

pn+1=(pn+α θ(qn)+β)(mod1),qn+1=qn+pn+1(mod1),(1)\begin{aligned} p_{n+1} &= (p_n+\alpha\,\theta(q_n)+\beta)\pmod{1},\\ q_{n+1} &= q_n+p_{n+1}\pmod{1}, \tag*{(1)} \end{aligned}

where θ(q)=+1\theta(q)=+1 for q∈[0,1/2)q\in[0,1/2) and −1-1 for q∈[1/2,1)q\in[1/2,1). In a previous study [10] we found evidence of ergodic and mixing behavior in a generic, non-resonant case, where the symbolic sums

Sn=∑j=0n−1Xj,Xj=θ(qj),(2)S_n=\sum_{j=0}^{n-1}X_j,\qquad X_j=\theta(q_j), \tag*{(2)}

scale as n1/2n^{1/2}, and much slower growth, slow correlation decay and strong arithmetic sensitivity for β=0\beta=0. That work introduced the polygonal and symbolic descriptions of the map, the cubic growth of the number of cylinders and finite Markov approximations.

Here we ask to what extent the stochastic-like behavior of the triangle map is that of an uncorrelated random process. The n1/2n^{1/2} scaling cannot settle this: normal diffusion is compatible with strong temporal correlations, which enter the diffusion coefficient through the Green–Kubo relation [13], and decay of correlations does not imply a positive diffusion coefficient, coboundaries being the standard counterexample [14]. We compare the map with an unbiased Bernoulli sequence through three groups of diagnostics: the two-point quantities that control transport (Var⁡(Sn)\operatorname{Var}(S_n), Green–Kubo sums, the spectrum of XX); persistence, via first passages and the Sparre Andersen theorem [15]; and the separation of the distribution of inter-event intervals from their order [16]. Since double-precision orbits turn out to be unreliable for β=0\beta=0, all results use an integrator based on the exact lift of the dynamics, and many quantities are computed without sampling from the polygonal cylinder sets.

The results separate three situations. The Bernoulli sequence has D=1D=1. The generic triangle map is normally diffusive, with Gaussian fluctuations and first-passage exponent 1/21/2, but its diffusion coefficient is not that of a coin: D=3.09±0.01D=3.09\pm0.01 for the benchmark pair (α,β)=(1/e,(5−1)/2)(\alpha,\beta)=(1/e,(\sqrt{5}-1)/2), three times the Bernoulli value, and DD varies between 0.20.2 and 1414 away from the edges of the parameter square, with deep minima at rational β\beta. For the benchmark, DD is set by the order of the inter-event intervals, which multiplies the value implied by their distribution by 3.73.7 (2.42.4–3.73.7 for other pairs), and the persistence amplitude is not Bernoulli. For β=0\beta=0 the Green–Kubo sum cancels and D=0D=0, while Var⁡(Sn)\operatorname{Var}(S_n) keeps growing slowly, almost independently of α\alpha, sign changes of SnS_n are rare and Fourier modes decorrelate as n−0.2n^{-0.2}. A small β\beta restores diffusion after a time of order β−1/2\beta^{-1/2}.

Three analytical observations frame and partly explain the numerics: the Koopman operator has no non-constant H1H^1 eigenfunctions (Proposition 1), so the symbol is not an H1H^1 coboundary; runs of equal symbols are uniformly bounded (Lemma 1); and at β=0\beta=0 every zero-sum word is a pure twist on its cylinder, carrying segments of periodic points that trap nearby orbits for times with a 1/ℓ1/\ell law and an exactly computable coefficient. These resonances fix the crossover time β−1/2\beta^{-1/2}, and an asymmetric deformation of the symbol, which makes them displace the walk, turns them into a logarithmic term of the variance with a coefficient we compute exactly and test numerically. We do not prove ergodicity or mixing; worst-case Birkhoff averages over adversarially optimized initial conditions decay as N−1/2N^{-1/2}, as for the uniquely ergodic skew shift, and we state precisely what the data support in three conjectures (§ 8).

Section 2 sets up the symbolic walk and the exact lift. Section 3 contains the rigorous results. Sections 4 and 5 treat the generic and β=0\beta=0 regimes, Sec. 6 the correlations of Fourier modes, Sec. 7 the asymmetric deformation, and Sec. 8 the conjectures. Numerical methods are in the Appendix; further data and checks are in the supplementary material.

The map and the symbolic walk

The discontinuity set of the map is D={q=0}∪{q=1/2}\mathcal{D}=\{q=0\}\cup\{q=1/2\}. One iteration is the composition T=S∘Rβ∘KαT=S\circ R_{\beta}\circ K_{\alpha} of the discontinuous kick Kα(q,p)=(q,p+αθ(q))K_{\alpha}(q,p)=(q,p+\alpha\theta(q)), the translation Rβ(q,p)=(q,p+β)R_{\beta}(q,p)=(q,p+\beta) and the parabolic shear S(q,p)=(q+p,p)S(q,p)=(q+p,p). Away from D\mathcal{D} the tangent map is

A=(1101),An=(1n01),(3)A=\begin{pmatrix}1&1\\0&1\end{pmatrix},\qquad A^n=\begin{pmatrix}1&n\\0&1\end{pmatrix}, \tag*{(3)}

so nearby orbits that follow the same symbolic branch separate at most linearly and the Lyapunov exponent vanishes. When a cut separates two orbits, their kicks differ by 2α2\alpha, and the shear converts this momentum difference into a growing separation in qq until further cuts split them again (Figure 1). The stochastic-like behavior is generated by this interplay of parabolic shear and symbolic splitting, not by exponential instability; the recurrence of orbits near the cuts and the resulting orbit complexity were quantified in Ref. [17].

Figure 1. Mechanism of the triangle map. (a) Within a continuity domain nearby orbits separate parabolically. (b) A discontinuity gives them kicks differing by 2α2\alpha. (c) The shear converts the momentum difference into separation in qq, and further cuts generate new symbolic splittings.

The symbolic sum SnS_n is a transport observable. On the lift, iteration gives

p~n=p0+nβ+αSn,Wn=∑k=1nSk,\tilde{p}_n=p_0+n\beta+\alpha S_n,\qquad W_n=\sum_{k=1}^{n}S_k,
q~n=q0+np0+βn(n+1)2+αWn,(4)\tilde{q}_n=q_0+np_0+\beta\frac{n(n+1)}{2}+\alpha W_n, \tag*{(4)}

with (qn,pn)=(q~n,p~n) mod 1(q_n,p_n)=(\tilde{q}_n,\tilde{p}_n)\bmod1, so that Var⁡(p~n−p0−nβ)=α2Var⁡(Sn)\operatorname{Var}(\tilde{p}_n-p_0-n\beta)=\alpha^2\operatorname{Var}(S_n). Equation (4) also gives a non-accumulating integrator: SnS_n and WnW_n are integers, and qnq_n is recomputed at every step in double-double arithmetic, with an error of order 10−1610^{-16} that does not grow along the orbit (Appendix § 11). All results below use it.

Initial conditions are distributed according to Lebesgue measure, which is invariant, so XnX_n is stationary with ⟨Xn⟩=0\langle X_n\rangle=0. With C(k)=⟨X0Xk⟩C(k)=\langle X_0X_k\rangle, Dn=Var⁡(Sn)/nD_n=\operatorname{Var}(S_n)/n and GK=C(0)+2∑k=1KC(k)G_K=C(0)+2\sum_{k=1}^{K}C(k), stationarity gives

Dn=C(0)+2∑k=1n−1(1−kn)C(k)=1n∑K=0n−1GK,(5)D_n=C(0)+2\sum_{k=1}^{n-1}\left(1-\frac{k}{n}\right)C(k)=\frac{1}{n}\sum_{K=0}^{n-1}G_K, \tag*{(5)}

so DnD_n is the Cesàro mean of the Green–Kubo partial sums and differs from GnG_n at finite nn. If ∑k∣C(k)∣<∞\sum_k|C(k)|<\infty, both converge to the diffusion coefficient

D=lim⁡n→∞Dn=G∞.(6)D=\lim_{n\to\infty}D_n=G_\infty. \tag*{(6)}

Note the normalization Var⁡(Sn)∼Dn\operatorname{Var}(S_n) \sim Dn, without the factor 2 common in physics; the momentum diffusion coefficient in the convention Var⁡(Δpn)∼2Dpn\operatorname{Var}(\Delta p_n) \sim2D_p n is Dp=α2D/2D_p = \alpha^2 D/2. The two-sided spectral density SX(ω)=∑kC(k)e−ikωS_X(\omega) = \sum_k C(k)e^{-ik\omega} satisfies

Var⁡(Sn)=12π∫−ππSX(ω)sin⁡2(nω/2)sin⁡2(ω/2) dω(7)\operatorname{Var}(S_n) = \frac{1}{2\pi}\int_{-\pi}^{\pi} S_X(\omega)\frac{\sin^2(n\omega/2)}{\sin^2(\omega/2)}\,d\omega \tag*{(7)}

so SX(0)=DS_X(0) = D when SXS_X is continuous at zero. For the Bernoulli benchmark C(k)=0C(k) = 0 for k≥1k \ge1, D=1D = 1 and SX≡1S_X \equiv1.

Rigorous results

This section collects the few statements we can prove. They concern the spectrum of the Koopman operator, the cohomology of the symbol and the combinatorics of the symbolic sequence; none of them establishes ergodicity, weak mixing or mixing, but they frame the numerical results and, through the word map below, partly explain them.

No H1H^1 eigenfunctions and no H1H^1 transfer function

The recent results on triangular billiards recalled in the Introduction [7, 8] build on the framework of Chaika and Forni and of Forni and Moll [6, 9], in which transverse regularity of eigenfunctions plays a central role. This suggests the following elementary discrete observation for our torus map. It is an analogy of method, not a consequence of their statements about billiard flows, and it complements the Fourier analysis of the Koopman operator in Ref. [10], where eigenfunctions depending on pp alone were exhibited for rational α\alpha.

Proposition 1. If α\alpha is irrational and β\beta is arbitrary, every H1(T2)H^1(\mathbb{T}^2) eigenfunction of the Koopman operator Uf=f∘Tα,βUf = f \circ T_{\alpha,\beta}, for any eigenvalue ∣λ∣=1|\lambda| = 1 including λ=1\lambda= 1, is constant.

Proof. Write g=∂qfg = \partial_q f, h=∂pfh = \partial_p f and suppose Uf=λfUf = \lambda f. On the two continuity domains the derivative of Tα,βT_{\alpha,\beta} is AA, and the chain rule gives, almost everywhere, Ug=λgUg = \lambda g and U(g+h)=λhU(g + h) = \lambda h. Thus Uh=λ(h−g)Uh = \lambda(h - g) and Unh=λn(h−ng)U^n h = \lambda^n(h - ng). Since UU is unitary on L2L^2, ∥h−ng∥2=∥h∥2\|h - ng\|_2 = \|h\|_2 for all nn, so g=0g = 0 and f(q,p)=F(p)f(q,p) = F(p). The eigenvalue equation on the two half-tori gives F(p+β+α)=λF(p)=F(p+β−α)F(p + \beta+ \alpha) = \lambda F(p) = F(p + \beta- \alpha), so FF is invariant under rotation by 2α2\alpha and hence constant.

Since Uf=λf∈H1Uf = \lambda f \in H^1, its global weak derivatives coincide with the branchwise ones, and the jumps of TT do not contribute. The proposition does not prove weak mixing, whose spectral criterion concerns all L2L^2 eigenfunctions: obtaining transverse regularity from a measurable eigenfunction is precisely the hard part of the billiard results, whose estimates do not control the jumps of our map.

Call X=θX = \theta a coboundary if X=g−g∘TX = g - g \circ T almost everywhere for a measurable gg, so that Sn=g−g∘TnS_n = g - g \circ T^n.

Corollary 1. (i) If XX is a coboundary, then −1-1 is an eigenvalue of UU, with eigenfunction eiπge^{i\pi g}; in particular T2T^2 is not ergodic. (ii) For irrational α\alpha and any β\beta, XX is not a coboundary with g∈H1(T2)g \in H^1(\mathbb{T}^2).

Proof. Since eiπX=−1e^{i\pi X} = -1, f=eiπgf = e^{i\pi g} satisfies Uf=−fUf = -f. If g∈H1g \in H^1, then f∈H1f \in H^1, contradicting Proposition 1.

By a classical criterion for ergodic TT [18 Theorem 11.8], XX is a measurable coboundary if and only if the laws of SnS_n are tight; for L2L^2 transfer functions, if and only if Var⁡(Sn)\operatorname{Var}(S_n) is bounded [14]. Hence, if T2T^2 is ergodic, SnS_n cannot stay bounded in probability.

Since Xj=±1X_j=\pm1, Sn≡n(mod2)S_n\equiv n\pmod{2}, so (−1)n=(−1)Sn(-1)^n=(-1)^{S_n} and

(−1)nXn=(−1)Sn(Sn+1−Sn).(8)(-1)^nX_n=(-1)^{S_n}(S_{n+1}-S_n). \tag*{(8)}

The time-alternating symbol is therefore a function of the path of the walk alone: a step from level kk contributes (−1)k(-1)^k times its direction. This identity is used in § 5.2.

Bounded runs and the word map

On the half-torus {θ=s}\{\theta=s\}, s=±1s=\pm1, the map is the skew shift

Fc(q,p)=(q+p+c,p+c)(mod1),c=cs=β+sα.(9)F_c(q,p)=(q+p+c,p+c)\pmod{1},\qquad c=c_s=\beta+s\alpha. \tag*{(9)}

Lemma 1. If csc_s is irrational, there is Ls<∞L_s<\infty such that every orbit of Tα,βT_{\alpha,\beta} contains no run of more than LsL_s consecutive symbols equal to ss.

Proof. For irrational cc the skew shift FcF_c is minimal [19]: Fcn(q,p)=(q+np+cn(n+1)/2,p+nc)F_c^n(q,p)=(q+np+cn(n+1)/2,p+nc), and for (k1,k2)≠0(k_1,k_2)\ne0 the sequence k1qn+k2pnk_1q_n+k_2p_n is a polynomial in nn with irrational leading coefficient, hence equidistributed by Weyl’s criterion. Let UU be the interior of {θ=−s}\{\theta=-s\}. The open sets Fc−kUF_c^{-k}U, k≥0k\ge0, cover T2\mathbb{T}^2, and by compactness finitely many, k≤Kk\le K, already do. A run of rr symbols ss has Fcsi(xj)∉UF_{c_s}^i(x_j)\notin U for i<ri<r, so r≤Kr\le K.

The bound holds for every orbit, but it does not make every orbit dense. Even density of all orbits would not, together with the invariance of Lebesgue measure, imply ergodicity (minimal interval exchanges need not be uniquely ergodic [20]), and unique ergodicity would not imply weak mixing, as the skew shift itself (α=0\alpha=0) shows. It excludes broad (infinite-variance) waiting-time distributions, and it makes the interval statistics of § 4.2 exactly computable.

The bound is Ls=L(cs)L_s=L(c_s) with a single function L(c)=L(−c)=L(1−c)L(c)=L(-c)=L(1-c), the longest run of FcF_c in a half-torus, so L+=L−L_+=L_- at β=0\beta=0 but not in general (for the generic benchmark L+=17L_+=17, L−=7L_-=7). L(c)L(c) is small for most cc and diverges, as ∣c−a/b∣−1/2\lvert c-a/b\rvert^{-1/2}, only at the rationals a/ba/b for which Fa/bF_{a/b} has a periodic orbit inside one half-torus, such as 00, 1/31/3, 2/52/5, $1/2 (Figure S1 in the supplementary material).

Figure S1. The run bound of Lemma 1 (Experiment 21). (a) Longest run L(c)L(c) of the skew shift FcF_c in a half-torus, computed exactly for 20 000 values of c∈(0,1/2]c \in(0,1/2]; dotted lines mark rationals, dashed the law c−1/2c^{-1/2} near c=0c=0. Dots: c±c_{\pm} of the generic benchmark and c=αc=\alpha for β=0\beta=0 with α=1/e,2−1\alpha=1/e,\sqrt{2}-1. (b) max⁡(L+,L−)\max(L_{+},L_{-}) with L±=L(β±α)L_{\pm}=L(\beta\pm\alpha) in the (α,β)(\alpha,\beta) plane; the dot is the generic benchmark (β≡g−1)(\beta\equiv g-1), the squares the two β=0\beta=0 cases.

The argument extends to words. For w=(w0,…,wm−1)w=(w_0,\ldots,w_{m-1}) with partial sums sk=w0+⋯+wk−1s_k=w_0+\cdots+w_{k-1}, the lift gives on the cylinder [w][w]

Tm(q,p)=(q+mp+dw,p+cw),cw=mβ+αsm,(10)T^m(q,p)=(q+mp+d_w,p+c_w),\qquad c_w=m\beta+\alpha s_m, \tag*{(10)}

with dw=βm(m+1)/2+α∑k=1mskd_w=\beta m(m+1)/2+\alpha\sum_{k=1}^{m}s_k. If cwc_w is irrational, repetitions of ww are uniformly bounded. At β=0\beta=0, however, every zero-sum word (sm=0s_m=0) has cw=0c_w=0: on [w][w] the block map is a pure twist, and the lines mp+dw∈Zmp+d_w\in\mathbb{Z} carry sets of period-mm points. These are the horizontal segments of periodic points of Ref. [10]; for w=+−w=+- they lie at p=1−α/2p=1-\alpha/2 and p=12−α2p=\frac{1}{2}-\frac{\alpha}{2}, with lengths α/2\alpha/2 and (1−α)/2(1-\alpha)/2. A point at momentum distance δ\delta from a segment of length LL is translated along it by mδm\delta per block, i.e. by δ\delta per step, and repeats ww until it leaves the segment, for a time L/∣δ∣L/\lvert\delta\rvert if it enters at one end (Figure S2 in the supplementary material). The repetition containing a point, i.e. the maximal stretch of consecutive steps along which its symbols repeat ww periodically (about ℓ/m\ell/m copies of ww for a stretch of ℓ\ell steps), is therefore at least ℓ\ell steps long if ∣δ∣≤L/ℓ\lvert\delta\rvert\le L/\ell, i.e. in a strip of area 2L2/ℓ2L^2/\ell along the segment, and likewise along its m−1m-1 images under T,…,Tm−1T,\ldots,T^{m-1}, which are translates of it. Since Lebesgue measure is invariant, the probability that a given time belongs to a repetition of ww of length at least ℓ\ell (the fraction of time averaged over Lebesgue-distributed initial conditions) is the area of these strips,

Figure S2. The period-two twist at β=0\beta=0, α=2−1\alpha=\sqrt{2}-1. (a) The cylinder [+−][+-] (grey), the period-two segments (solid) and their images under one step (dotted). (b) qnq_n at even nn for orbits at momentum distance δ\delta from the segment p=12−α2p=\frac{1}{2}-\frac{\alpha}{2} (shaded): they drift by 2δ2\delta per block and repeat +−+- until they leave it; the stay length is compared with the remaining length of the segment divided by ∣δ∣|\delta|. (c) β=10−5\beta=10^{-5}: the momentum drift bends the motion into a parabola; the longest stay fits it into the wider segment.

Fw(≥ℓ)≃Kwℓ,Kw=2m∑jLj2.(11)F_w(\ge\ell)\simeq\frac{K_w}{\ell},\qquad K_w=2m\sum_jL_j^2. \tag*{(11)}

the sum running over the segments of periodic points of ww itself, the maximal intervals of [w][w] on the resonant lines p=(j−αW)/mp=(j-\alpha W)/m, W=∑k=1mskW=\sum_{k=1}^{m}s_k. The images of these segments are the segments of the cyclic rotations of ww, so KwK_w is the same for all rotations, and totals over words count each cyclic class once. The law holds for ℓ≫m\ell\gg m; at finite ℓ\ell the ends of the segments, where the strip meets the boundary of [w][w], give corrections (§ 5.2). For w=+−w=+-, K+−=4[(α/2)2+((1−α)/2)2]=α2+(1−α)2K_{+-}=4[(\alpha/2)^2+((1-\alpha)/2)^2]=\alpha^2+(1-\alpha)^2. For small β>0\beta>0 the drift c+−=2βc_{+-}=2\beta makes δ\delta grow linearly in time, so that after kk steps the position along the segment is δ0k+βk2/2\delta_0 k+\beta k^2/2; the longest repetition is the one whose parabola just fits in the longer segment, of length Lmax⁡=max⁡(α,1−α)/2L_{\max}=\max(\alpha,1-\alpha)/2, which gives the maximal length 22Lmax⁡/β2\sqrt{2L_{\max}/\beta}. The lift also yields exact word probabilities; for instance, at β=0\beta=0, P(−+−+)=1/12P(-+-+)=1/12 for every α\alpha (supplementary material, Sec. S1).

Generic parameters and a correction

We call (α,β)(\alpha,\beta) generic if 11, α\alpha, β\beta are linearly independent over Q\mathbb{Q}. Then every cwc_w is irrational, every word has uniformly bounded repetitions, and the map has no periodic orbits, since a period-mm orbit with word ww requires cw∈Zc_w\in\mathbb{Z}. All generic pairs used below satisfy this condition, by the transcendence of ee and the linear independence over Q\mathbb{Q} of square roots of distinct square-free integers, except (π−3,1/e)(\pi-3,1/e): the rational independence of 11, π\pi, 1/e1/e would follow from the algebraic independence of π\pi and ee, which is open.

Correction to Ref. [10]. There the generic case was defined by α,β\alpha,\beta irrational with β∉Z+αZ\beta\notin\mathbb{Z}+\alpha\mathbb{Z}, and Lemma 1 states that the map then has no periodic orbits. The necessary condition tβ+αSt≡0(mod1)t\beta+\alpha S_t\equiv0\pmod1 for a period-tt orbit excludes β∈Z+αZ\beta\in\mathbb{Z}+\alpha\mathbb{Z} only for t=1t=1; in general it gives β∈Q+Qα\beta\in\mathbb{Q}+\mathbb{Q}\alpha, so the hypothesis must be strengthened to rational independence. For example, α=1/e\alpha=1/e, β=(1−α)/2\beta=(1-\alpha)/2 satisfy the original condition, yet w=+++−w=+++- has cw=4β+2α=2c_w=4\beta+2\alpha=2, and the line p≃0.97409p\simeq0.97409 carries a segment of length 0.3420.342 of period-four points, which we verified in 50-digit arithmetic. The other rigorous statements of Ref. [10] used here are not affected.

Why ergodicity is hard. On each half-torus the map is a uniquely ergodic skew shift [19], and Veech proved strict ergodicity of Z2\mathbb{Z}_2-extensions of rotations by step cocycles [21]; both arguments need a factor. Here the lifted momentum is a rotation only modulo the dense subgroup Z+αZ\mathbb{Z}+\alpha\mathbb{Z}: the symbol feeds back into the base, and there is no factor. The billiard theorems [5, 6, 7, 8] rest on Teichmüller dynamics and do not transfer to a model map. Boshernitzan-type criteria [22] need word complexity growing linearly along a subsequence, whereas ours grows as ≃0.66n3\simeq0.66n^3 (supplementary material, Sec. S1), and the renormalization schemes for kicked oscillators [23], of which the triangle map is the parabolic member qn+1−2qn+qn−1=β+αθ(qn)q_{n+1}-2q_n+q_{n-1}=\beta+\alpha\theta(q_n), do not cover generic parameters. To our knowledge, minimality, unique ergodicity and ergodicity of the generic map remain open; we state the expected behavior as Conjecture 1 in Sec. 8.

Generic regime: correlated normal diffusion

We use the benchmark α=e−1\alpha=e^{-1}, β=(5−1)/2\beta=(\sqrt{5}-1)/2 of our earlier work.

Diffusion coefficient

For 4096 orbits followed to n=108n=10^8, the local log-log slopes of Var⁡(Sn)\operatorname{Var}(S_n) and of the mean running maximum ⟨Smax⁡(n)⟩=⟨max⁡k≤nSk⟩\langle S_{\max}(n)\rangle=\langle\max_{k\le n}S_k\rangle are 0.996 and 0.502, and Var⁡(Sn)/n=3.079\operatorname{Var}(S_n)/n=3.079 at the largest checkpoint; the amplitude ⟨Smax⁡(n)⟩/n=1.408\langle S_{\max}(n)\rangle/\sqrt{n}=1.408 matches the Brownian value 2D/π=1.400\sqrt{2D/\pi}=1.400. Correlations from 64 trajectories of length 2262^{26} agree with the values computed without sampling from the polygonal cylinders (App. § 11) for k<80k < 80: the mean squared deviation in units of the bootstrap error is 0.88 per lag, as expected for purely statistical scatter. One has C(1)=C(2)=0C(1) = C(2) = 0 exactly for all parameters (for fixed q0q_0, the uniform p0p_0 makes q1q_1 uniform, and q2=2q1−q0+αθ(q1)+βq_2 = 2q_1 - q_0 + \alpha\theta(q_1) + \beta is the image of q1q_1 under a two-to-one map that preserves Lebesgue measure), but the next correlations are large, C(4)≃0.081C(4) \simeq0.081, C(7)≃0.167C(7) \simeq0.167 (supplementary material, Figure S3), and GKG_K reaches 2.04 at K=10K = 10 and 2.66 at K=30K = 30 (Figure 3, middle left). The rest of the excess is built by a band of weaker, sign-oscillating correlations at 102≲k≲2×10310^2 \lesssim k \lesssim2 \times10^3, after which GKG_K is flat: G103=3.064±0.003G_{10^3} = 3.064 \pm0.003, G104=3.091±0.009G_{10^4} = 3.091 \pm0.009. By Eq. (5), Dn=D−2M/n+o(1/n)D_n = D - 2M/n + o(1/n) with M=∑kkC(k)≃100M = \sum_k kC(k) \simeq100, which explains the slow approach of the finite-time variance. A non-oscillating tail C(k)≃ak−3/2C(k) \simeq ak^{-3/2}, as reported for the smooth observable cos⁡πy\cos\pi y [12], is not seen (a=0.08±0.16a = 0.08 \pm0.16 over 2×103≤k≤1042 \times10^3 \leq k \leq10^4). The averaged periodogram is flat below ω∼10−3\omega\sim10^{-3} at 3.11±0.053.11 \pm0.05 (bottom left). All estimators agree (Figure 3, left):

Figure 3. Two transport regimes. Left: generic α=1/e\alpha= 1/e, β=(5−1)/2\beta= (\sqrt{5}-1)/2. Right: α=2−1\alpha= \sqrt{2}-1, β=0\beta= 0. Top: Var⁡(Sn)/n\operatorname{Var}(S_n)/n (2048 orbits). Middle: Green–Kubo partial sums GKG_K (64 trajectories of length 2262^{26}, generic; 1024 of length 2222^{22}, β=0\beta= 0). Bottom: log-binned spectral density (32 trajectories of length 2252^{25}). Bands are one standard error; dashed lines mark D=3.09D = 3.09, 0 and SX∝ωS_X \propto\omega; dotted lines the Bernoulli value 1. All data from the exact-phase integrator.

Figure S3. Symbol correlations, generic regime (α=1/e\alpha= 1/e, β=(5−1)/2\beta= (\sqrt{5} - 1)/2; 64 exact-phase trajectories of length 2262^{26}). (a) C(k)C(k) for k<80k < 80: measured (bars) and computed without sampling from the polygonal cylinders (circles); C(1)=C(2)=0C(1) = C(2) = 0. (b) Deviations in units of the standard error. (c) RMS of C(k)C(k) in logarithmic bins of kk with the noise level subtracted (solid) and the noise level (dashed); grey: the band 102≤k≤2×10310^2 \le k \le2 \times10^3.

Var⁡(Sn)∼Dn,SX(ω)→D,D=3.09±0.01.(12)\operatorname{Var}(S_n) \sim Dn,\qquad S_X(\omega) \to D,\qquad D = 3.09 \pm0.01. \tag*{(12)}

The value is not universal: away from the edges of the parameter square DD ranges from 0.2 to 14 (median 2.3), with dark lenses at rational β\beta and bright bands near low-order resonances (supplementary material, Figure S4); four other generic pairs give D=2.15D = 2.15–2.942.94 (supplementary material, § 14), and the pair A of § 7 gives D=1.19D = 1.19, close to the Bernoulli value. The generic symbolic walk is normally diffusive but correlated.

Figure S4. Diffusion coefficient over the parameter square (Experiment 24, exact lift). Dn=Var⁡(Sn)/nD_n = \operatorname{Var}(S_n)/n at n=214n = 2^{14}, 1024 orbits per point, on a 200×200200 \times200 grid; computed for α+β≤12\alpha+ \beta\le\frac{1}{2} and reflected through (14,14)\left(\frac{1}{4}, \frac{1}{4}\right) (dotted line), by the exact symmetry D(α,β)=D(12−α,12−β)D(\alpha,\beta) = D\left(\frac{1}{2} - \alpha,\frac{1}{2} - \beta\right). (a) The fundamental square [0,12]2\left[0,\frac{1}{2}\right]^2, with the benchmark pair and the pairs A, B of Sec. 7 folded into it. (b) Zoom near the skew-shift limit α=0\alpha= 0; labels mark rational β\beta.

The fluctuations are Gaussian on the explored scales: the Kolmogorov distance from the standard Gaussian of the normalized sum Sn/Var⁡(Sn)S_n/\sqrt{\operatorname{Var}(S_n)} is 0.013–0.017 for 103≤n≤3.4×10710^3 \leq n \leq3.4 \times10^7, at the level of sampling noise, the skewness decays as n−1/2n^{-1/2}, and the correlation between SnS_n and S2n−SnS_{2n} - S_n is below 0.02 for n≥104n \geq10^4. This supports, without proving, the central limit theorem of Conjecture 1 (§ 8), with asymptotically independent increments [24].

Distribution versus order of inter-event intervals

Let τi\tau_i be the intervals between consecutive occurrences of +1+1, observed from an occurrence at time 0, and let tJ=τ0+⋯+τJ−1t_J = \tau_0 + \cdots+ \tau_{J-1} be the time of the JJ-th return. By Kac’s lemma the mean interval is 2, and since [0,tJ)[0,t_J) contains JJ symbols +1+1 and tJ−Jt_J-J symbols −1-1, ∑i<J(τi−2)=−StJ\sum_{i<J}(\tau_i-2) = -S_{t_J} exactly. Hence D=12G(∞)D = \frac{1}{2}G(\infty) with G(K)=v+2∑k=1KCov⁡(τ0,τk)G(K) = v + 2\sum_{k=1}^{K}\operatorname{Cov}(\tau_0,\tau_k), v=Var⁡(τ0)v = \operatorname{Var}(\tau_0), whereas an independent renewal process with the same interval law has DR=v/2D_R = v/2 [16]. By Lemma 1 the intervals are bounded, and the set of initial conditions producing a given run is a finite union of polygons, so the interval law is known exactly and no shuffled surrogate is needed:

longest runs +/-max τ\tauvvDR=v/2D_R = v/2
generic (benchmark pair)17/781.70850.854
β=0\beta= 0, α=1/e\alpha= 1/e11/11122.14931.075
β=0\beta= 0, α=2−1\alpha= \sqrt{2} - 112/12132.12541.063
Bernoulli∞\infty∞\infty21

Table 1.

If the intervals were independent with the same law, the walk would have DR=v/2D_R = v/2: 0.85 in the generic regime and 1.07 at β=0\beta= 0, both close to the Bernoulli value. The measured transport is very different in both regimes, so it is set by the order of the intervals, i.e. by the covariances in G(K)G(K) [Figure 2(c)]. For the benchmark pair they are positive (long intervals tend to follow long ones) and 12G(K)\frac{1}{2}G(K) grows from 0.85 to 3.15±0.033.15 \pm0.03 at K=4096K = 4096, a factor of about 3.7, consistent with the more precise D=3.09±0.01D = 3.09 \pm0.01 obtained from the symbol sum; 80% of the increase occurs within twenty events. At β=0\beta= 0 they alternate in sign, starting with Cov⁡(τ0,τ1)≃−0.43\operatorname{Cov}(\tau_0,\tau_1) \simeq-0.43 (a long interval is followed by a short one that compensates it), and 12G(K)\frac{1}{2}G(K) falls from 1.07 to 0.05–0.15 within about ten events; the remainder, 12G≃0.005\frac{1}{2}G \simeq0.005 at K∼103K \sim10^3, decays slowly. Bounded intervals and their order thus produce both enhanced and suppressed transport, whereas anomalous transport is usually attributed to broad waiting-time distributions.

Figure 2. Persistence and inter-event intervals. (a) Survival probabilities relative to the lattice Bernoulli reference QBQ_B, Eq. (13): generic Q+Q_+ (solid) and Q−Q_- (dashed) with Wilson 68% bands, and (Q++Q−)/2(Q_+ + Q_-)/2 for β=0\beta= 0. (b) Effective persistence exponent over factor-four windows; the horizontal line is the Sparre Andersen value 1/21/2. (c) Event-time Green–Kubo sums G(K)/2G(K)/2 of the intervals between ++ events; dotted lines mark the renewal values v/2v/2 computed exactly from the interval law.

Persistence

Let Q±(n)Q_\pm(n) be the probability that the walk has not entered the positive (negative) half-line up to time nn, i.e. that ±Sk≤0\pm S_k \le0 for all 1≤k≤n1 \le k \le n. For a Bernoulli walk (independent symmetric steps ±1\pm1, hence the subscript) the two coincide and the Sparre Andersen theory [15, 25] gives

QB(n)=(n⌊n/2⌋)2−n∼2πn,(13)Q_B(n) = \binom{n}{\lfloor n/2\rfloor}2^{-n} \sim\sqrt{\frac{2}{\pi n}}, \tag*{(13)}

the lattice counterpart of the universal n−1/2n^{-1/2} law for continuous symmetric steps. The involution (q,p)↦(−q,−p)(q,p) \mapsto(-q,-p) preserves Lebesgue measure, reverses θ\theta and conjugates Tα,βT_{\alpha,\beta} to Tα,−βT_{\alpha,-\beta}, so Q+(n;α,β)=Q−(n;α,−β)Q_+(n;\alpha,\beta) = Q_-(n;\alpha,-\beta) exactly. For the generic parameters (10610^6 orbits to n=222n = 2^{22}) the effective exponent θeff=−dln⁡Q/dln⁡n\theta_{\mathrm{eff}} = -\mathrm{d}\ln Q/\mathrm{d}\ln n is compatible with 1/21/2 for n≳102n \gtrsim10^2, but the amplitudes are not Bernoulli and differ between the sides, Q+/QB≃1.75Q_+/Q_B \simeq1.75 and Q−/QB≃1.32Q_-/Q_B \simeq1.32 [Figure 2(a,Figure 2b)]. They are not fixed by the one-time marginals either. For independent steps the survival probability is determined by them through the Sparre Andersen identity ∑n≥0Q+(n)sn=exp⁡(∑k≥1skP(Sk≤0)/k)\sum_{n\ge0} Q_+(n)s^n = \exp(\sum_{k\ge1}s^kP(S_k \le0)/k) [25]; fed with the measured P(Sk≤0)P(S_k \le0) (and P(Sk≥0)P(S_k \ge0) for Q−Q_-), it reproduces simulated Bernoulli walks within 1%, but for the map the measured Q+Q_+ and Q−Q_- exceed its prediction by factors 1.38 and 1.77 at n=4096n = 4096: the excess persistence comes from temporal correlations, not from the distribution of SkS_k.

At β=0\beta= 0 persistence is anomalously strong: with 10610^6 orbits to n=1.7×107n = 1.7 \times10^7, Q/QBQ/Q_B reaches 28, and θeff\theta_{\mathrm{eff}} decreases from 0.29 at n∼102n \sim10^2 to 0.205±0.0140.205 \pm0.014 for 106≤n≤1.7×10710^6 \le n \le1.7 \times10^7, identically for α=1/e\alpha= 1/e and 2−1\sqrt{2} - 1; a logarithmic decay is excluded, and no limiting exponent is claimed. The long survivors are selected early by thin bands of initial momentum (supplementary material, Sec. S3).

The exceptional =0 regime

Variance, correlations and spectrum

At β=0\beta= 0 double-precision orbits are not adequate: their symbol sequences differ from the exact ones from n∼106n \sim10^{6}, and for an α\alpha with large partial quotients the double-precision variance is three times too large at n≳5×106n \gtrsim5 \times10^{6} (supplementary material, § 15). We followed 2048 orbits to n=225≃3.4×107n = 2^{25} \simeq3.4 \times10^{7} for α∈{2−1,3−1,(5−1)/2,1/e}\alpha\in\{\sqrt{2} - 1, \sqrt{3} - 1, (\sqrt{5} - 1)/2, 1/e\} and for two values with large partial quotients, π−3\pi- 3 and [0;2,30,2,300,2,3000,1,1,…][0; 2, 30, 2, 300, 2, 3000, 1, 1, \ldots] (continued-fraction expansion), with replicate ensembles and one run to 1.3×1081.3 \times10^{8}. The variance curves are nearly independent of α\alpha: Var⁡(Sn)≃13\operatorname{Var}(S_n) \simeq13, 33–38, 65–73 and 95–103 at n=102n = 10^{2}, 10410^{4}, 10610^{6} and 3.4×1073.4 \times10^{7} (Figure 3, right), with no feature at the convergent denominators. The local exponent dln⁡Var⁡/dln⁡nd \ln\operatorname{Var}/d \ln n decreases from 0.3 to 0.1 without a plateau, and dVar⁡/dln⁡nd \operatorname{Var}/d \ln n increases, so the growth is slower than any fixed accessible power and faster than ln⁡n\ln n. Beyond this the law is not determined: a+bln⁡2na + b\ln^{2}n, a+cln⁡n+bln⁡2na + c\ln n + b\ln^{2}n, a+b(ln⁡n)pa + b(\ln n)^{p} and a+cnγa + cn^{\gamma} fit equally well, with model-dependent parameters (supplementary material, § 15).

The Green–Kubo sums explain the suppression. From 1024 trajectories of length 2222^{22}, GKG_K falls from 1 to 0.45 at K=10K = 10 and 0.09 at K=100K = 100, then oscillates around zero with decreasing amplitude (−0.029-0.029, 0.022, −0.014-0.014 at K=103K = 10^{3}, 10410^{4}, 10510^{5}, errors 0.002–0.004): the positive instantaneous variance is almost exactly cancelled by integrated anticorrelations, ∑k≥1C(k)≃−1/2\sum_{k \ge1} C(k) \simeq-1/2. The spectral density vanishes at low frequency with local exponent η(ω)\eta(\omega) close to one and arithmetic oscillations; scale by scale 1−η(2π/n)1 - \eta(2\pi/n) matches the local variance exponent, as Eq. (7) requires, and a log-corrected linear spectrum SX∼∣ω∣ln⁡(1/∣ω∣)S_X \sim|\omega|\ln(1/|\omega|), which gives ln⁡2n\ln^{2}n, is consistent with both drifts without being singled out. Exact Markov approximations of order JJ, built without sampling on the return map to the starts of ++ runs (supplementary material, Sec. S4), show the same structure on short scales (n≲102n \lesssim10^{2}): at β=0\beta= 0 their diffusion coefficient falls as DJ≃0.78/JD_J \simeq0.78/J (14≤J≤2814 \le J \le28), against DJ≃2.9D_J \simeq2.9 in the generic case, and their transfer function, the solution of the Poisson equation, grows only logarithmically in JJ instead of as J\sqrt{J}, so that at every finite order XX is close to a coboundary, as in Conjecture 2.

Thus, on the explored scales (n≤3.4×107n \le3.4 \times10^{7}, K≤1.3×105K \le1.3 \times10^{5}, ω≳3×10−6\omega\gtrsim3 \times10^{-6}), Var⁡(Sn)/n→0\operatorname{Var}(S_n)/n \to0, D=0D = 0 and SX(ω)→0S_X(\omega) \to0, while Var⁡(Sn)\operatorname{Var}(S_n) grows without saturation. If the growth persists, XX is not an L2L^{2} coboundary (§ 3: an L2L^{2} transfer function would keep Var⁡(Sn)\operatorname{Var}(S_n) bounded), although its Green–Kubo sum vanishes as for a coboundary. The bulk of the distribution widens (interquartile range 4 →\to 12 from n=10n = 10 to 3.4×1073.4 \times10^{7}), and Sn/σnS_n/\sigma_n has a stable, α\alpha-independent, non-Gaussian law with kurtosis ≃5\simeq5.

Parity and zero-sum-word resonances

The walk SnS_n moves on the integer lattice, and each step crosses one edge [k,k+1][k,k+1] of it (k∈Zk \in\mathbb{Z} a level, not a symbol). By the parity identity (8), the alternating sum An=∑j<n(−1)jXjA_n = \sum_{j<n}(-1)^jX_j adds (−1)k(-1)^k for every crossing of the edge [k,k+1][k,k+1], in either direction: +1+1 for edges with an even lower end and −1-1 for edges with an odd one. A walk that oscillates over few edges, as in a repetition of +−+-, therefore has an almost ballistic AnA_n, whereas a walk spreading over many levels crosses both kinds equally often. Since An/n=1n∑j<n(−U)jXA_n/n = \frac{1}{n}\sum_{j<n}(-U)^jX is an ergodic average for the unitary operator −U-U, the mean ergodic theorem gives An/n→P−1XA_n/n \to P_{-1}X in L2L^2, the projection of XX on the eigenspace {Uf=−f}\{Uf = -f\}; equivalently, E∣An/n∣2→μX({π})\mathbb{E}|A_n/n|^2 \to\mu_X(\{\pi\}), the mass at the frequency π\pi, that of the period-two oscillation (−1)j=eiπj(-1)^j = e^{i\pi j}, of the spectral measure of XX, defined by C(k)=∫eikω dμX(ω)C(k) = \int e^{ik\omega}\,d\mu_X(\omega). A positive limit would give an eigenfunction with Uf=−fUf = -f, hence a non-constant T2T^2-invariant function: T2T^2 would not be ergodic, the same obstruction that a coboundary would create (Corollary 1). Conjecture 2 asserts that neither occurs. In the generic case Var⁡(An)/n≃1.7\operatorname{Var}(A_n)/n \simeq1.7 is constant. At β=0\beta= 0, Var⁡(An)/n2\operatorname{Var}(A_n)/n^2 decreases slowly, from 0.051 to 0.0091 (α=2−1\alpha= \sqrt{2} - 1) between n=210n = 2^{10} and 2222^{22}, as n−0.2n^{-0.2}, and SXS_X has a peak at ω=π\omega= \pi of height ≃900\simeq900 against a median of 0.8, with weaker peaks at 2π/32\pi/3, π/2\pi/2, π/3\pi/3, π/4\pi/4. At β=0\beta= 0 the momentum stays on the lattice p0+αZp_0+\alpha\mathbb{Z}, and by Eq. (8) the sign of each + ⁣−+\!- burst is fixed by the level on which it occurs, so successive bursts add coherently at ω=π\omega=\pi.

These resonances are seen directly in the orbits. Let Fm(≥ℓ)F_m(\geq\ell) be the fraction of time spent inside repetitions, at least ℓ\ell steps long, of zero-sum words of length mm; Eq. (11) predicts ℓFm(≥ℓ)=Km=∑wKw\ell F_m(\geq\ell)=K_m=\sum_w K_w, independent of ℓ\ell. At β=0\beta=0 each orbit keeps its momentum lattice p0+αZp_0+\alpha\mathbb{Z} and samples a single distance δ\delta from each segment, so we use β=10−7\beta=10^{-7}, which lets the momentum drift while cutting repetitions only beyond 22Lmax⁡/β≃5×1032\sqrt{2L_{\max}/\beta}\simeq5\times10^3. Over 10m≤ℓ≤100m10m\leq\ell\leq100m, for α=2−1\alpha=\sqrt{2}-1 and m=2,4,6,8m=2,4,6,8, we find ℓFm=0.52\ell F_m=0.52–0.590.59, 1.481.48–1.551.55, 0.380.38–0.430.43, 0.530.53–0.580.58 against the exact Km=0.515K_m=0.515, 1.3731.373, 0.3730.373, 0.5470.547; ℓFm\ell F_m decreases with ℓ\ell towards KmK_m, the excess at ℓ≃10m\ell\simeq10m being a finite-ℓ\ell correction [the same law for the deformed map is shown in Figure 6(b)]. For β=10−3,…,10−6\beta=10^{-3},\ldots,10^{-6} the longest + ⁣−+\!- repetition agrees with the cutoff 22Lmax⁡/β2\sqrt{2L_{\max}/\beta} within 1–5% [supplementary material, Figure S2]. The KmK_m do not decrease with mm: resonant twists exist at all lengths.

Figure 6. Asymmetric deformation, α=2−1\alpha= \sqrt{2} - 1. (a) ε=1/4\varepsilon= 1/4: Var⁡(Xn)/n\operatorname{Var}(X_n)/n for generic β=3−1\beta= \sqrt{3} - 1, for β=−α(1−2ε)\beta= -\alpha(1 - 2\varepsilon) (zero mean drift), and for β=0\beta= 0 (two seeds of 10510^5 orbits; interquartile scale σIQR2/n\sigma_{\mathrm{IQR}}^2/n from n=210n = 2^{10}). (b) ε=1/4\varepsilon= 1/4, β=0\beta= 0: ℓF(≥ℓ)\ell F(\ge\ell) for stays of period two and four, with the exact K2=0.4216K_2 = 0.4216 and K4=0.5K_4 = 0.5 (dashed). (c) ε=0.2\varepsilon= 0.2, β=0\beta= 0: Var⁡(Xn)/n\operatorname{Var}(X_n)/n split into the parts accumulated outside and inside the stays; dashed, the slope c(ε)=0.230c(\varepsilon) = 0.230 of Eq. (16) anchored at n=212n = 2^{12}.

Crossover at small

For α=1/e\alpha=1/e and 2−1\sqrt{2}-1 and seven values 6×10−5≤β≤0.066\times10^{-5}\leq\beta\leq0.06 (1024 orbits to n=225n=2^{25} each), the variance follows the β=0\beta=0 curve up to a crossover time and then grows linearly. The ratio Var⁡β(Sn)/Var⁡0(Sn)\operatorname{Var}_{\beta}(S_n)/\operatorname{Var}_{0}(S_n) is approximately a function of nβn\sqrt{\beta} (scatter ±30%\pm30\%, not monotone in β\beta, larger than the statistical errors), with crossover time nc≃(2n_c\simeq(2–3)β−1/23)\beta^{-1/2}. The relevant time is not 1/β1/\beta, at which the drift nβn\beta is of order one, but the time at which the quadratic phase βn(n+1)/2\beta n(n+1)/2 of Eq. (4) is. It is the lifetime of the resonances: for β>0\beta>0 the block map of a zero-sum word acquires the drift cw=mβc_w=m\beta, and the longest + ⁣−+\!- repetition is 22Lmax⁡/β≃1.5β−1/22\sqrt{2L_{\max}/\beta}\simeq1.5\beta^{-1/2}. At the smallest β\beta,

D(β)≃βVar⁡0(Sβ−1/2),(14)D(\beta)\simeq\sqrt{\beta}\operatorname{Var}_{0}(S_{\beta^{-1/2}}), \tag*{(14)}

with ratio of the two sides 1.00–1.07 (α=1/e\alpha=1/e) and 0.8–1.25 (2−1\sqrt{2}-1) for β≲6×10−4\beta\lesssim6\times10^{-4}: the variance accumulated by the β=0\beta=0 dynamics up to the crossover is then spread diffusively.

Correlations of Fourier modes and mixing

Mixing is equivalent to the decay of ⟨f,Unf⟩\langle f,U^n f\rangle for all zero-mean f∈L2f\in L^2, and it suffices to check the Fourier modes fk,l=e2πi(kq+lp)f_{k,l}=e^{2\pi i(kq+lp)}. Their correlations were measured in double precision in [12, 10], with decay close to n−3/2n^{-3/2} (generic) and n−0.2n^{-0.2} (β=0\beta=0). We repeated the measurement with the exact lift, recomputing qnq_n and pn=p0+nβ+αSn(mod1)p_n=p_0+n\beta+\alpha S_n\pmod1 at every step, for the modes (1,0)(1,0), (0,1)(0,1), (1,1)(1,1), (2,0)(2,0), (0,2)(0,2), with 33–4×1094\times10^9 steps per parameter pair (the mode (1,−1)(1,-1) coincides with (1,0)(1,0) because qn−pn=qn−1q_n-p_n=q_{n-1}).

In the generic case the correlations decay quickly (Figure 4): the rms ∣C1,0∣|C_{1,0}| is 7×10−37\times10^{-3} for 30≤n<10030\leq n<100 and 5×10−45\times10^{-4} for 300≤n<1000300\leq n<1000, with effective exponents 1.4–2, and the signal is below the noise floor 6×10−56\times10^{-5} beyond n≃103n\simeq10^3; this is consistent with n−3/2n^{-3/2} over about one decade. At β=0\beta=0 all modes decay slowly but steadily up to n=2.6×105n=2.6\times10^5, for α=1/e\alpha=1/e and 2−1\sqrt{2}-1, with effective exponent 0.19–0.21 for (1,0)(1,0), (1,1)(1,1), (0,1)(0,1), and no plateau. The correlations are real: the symmetry (q,p)↦(−q,−p)(q,p)\mapsto(-q,-p) conjugates the map with β\beta to the map with −β-\beta, so at β=0\beta=0 every Ck,lC_{k,l} is real. (For any β\beta the modes (k,0)(k,0) are real as well, because the map is invariant with the qq-preserving reversor (q,p)↦(q,−p−αθ(q)−β)(q,p)\mapsto(q,-p-\alpha\theta(q)-\beta); the other modes are complex for β≠0.)\beta\neq0.) They carry the parity of the walk: the qq-modes nearly vanish at odd lags, and the pp-modes alternate in sign [Figure 4(c)].

Figure 4. Correlations of Fourier modes (exact lift). (a), (b) Root-mean-square ∣Ck,l(n)∣|C_{k,l}(n)| over logarithmic lag bins, after subtraction of the statistical noise, for the modes (1,0)(1,0) and (0,1)(0,1), shown where the signal exceeds three standard errors (generic noise floor ≃6×10−5\simeq6 \times10^{-5}). Grey lines: n−3/2n^{-3/2} (dashed) and n−0.2n^{-0.2} (dash-dotted). (c) Parity structure at β=0\beta= 0, α=2−1\alpha= \sqrt{2} - 1: real part of C1,0C_{1,0} at even lags and its modulus at odd lags, and C0,1C_{0,1} at even lags compared with −C0,1-C_{0,1} at odd lags.

The pp-modes have a direct meaning for the cocycle: at β=0\beta=0,

C0,l(n)=E e2πilα(Sj+n−Sj)(15)C_{0,l}(n)=\mathbb{E}\,e^{2\pi i l\alpha(S_{j+n}-S_j)} \tag*{(15)}

is the characteristic function of the increments of SS. If XX were a coboundary and TT mixing, it would tend to ∣Ee2πilαg∣2\lvert\mathbb{E} e^{2\pi i l \alpha g}\rvert^2, generically non-zero. The steady decay of C0,1C_{0,1}, from 0.39 at n≃10n \simeq10 to 0.094 at n≃5×104n \simeq5 \times10^4, is evidence against tightness. Finally, worst-case Birkhoff averages of the Fourier modes over random, near-cut and adversarially optimized initial conditions decay as N−1/2N^{-1/2} for two generic pairs, like the uniquely ergodic skew shift, while the same search recovers the periodic segments at β=0\beta= 0 and the period-four orbits of the counterexample above (supplementary material, Sec. S5).

An asymmetric deformation: resonances made visible

At β=0\beta= 0 the triangle map has two features that cannot be separated: the zero-sum-word resonances and the zero mean drift of the lifted momentum. Replacing θ\theta by

θε(q)={−1,q∈[0,ε),+1,q∈[ε,1),\theta_{\varepsilon}(q) = \begin{cases} -1, & q \in[0,\varepsilon),\\ +1, & q \in[\varepsilon,1), \end{cases}

in the same map interpolates between the triangle map (ε=1/2\varepsilon= 1/2, up to α→−α\alpha\to-\alpha and q→q+1/2q \to q+1/2) and the skew shift (ε=0\varepsilon= 0); the lift (4) and the exact integrator are unchanged. Now E Sn=n(1−2ε)\mathbb{E}\,S_n = n(1-2\varepsilon), and we consider the centered walk Xn=Sn−n(1−2ε)=−2(Nn−nε)X_n = S_n - n(1-2\varepsilon) = -2(N_n-n\varepsilon), with NnN_n the number of visits to [0,ε)[0,\varepsilon), and the Fano factor of these counts, F=Var⁡(Nn)/E Nn=Var⁡(Sn)/(4εn)F = \operatorname{Var}(N_n)/\mathbb{E}\,N_n = \operatorname{Var}(S_n)/(4\varepsilon n) (Poisson: F=1F = 1; Bernoulli: F=1−εF = 1-\varepsilon).

Generic β\beta. For small ε\varepsilon the kicks are the returns of the quadratic sequence q0+np0+c+n(n+1)/2q_0 + np_0 + c+n(n+1)/2 to a short interval, whose statistics are Poissonian for generic coefficients [26, 27]. We consider two generic pairs, A=(2−1,3−1)A = (\sqrt{2}-1,\sqrt{3}-1) and B=((5−1)/2,3−1)B = ((\sqrt{5}-1)/2,\sqrt{3}-1). Numerically FF stays within 0.96–1.12 (A) and 0.96–1.22 (B) for ε<0.04\varepsilon< 0.04 [Fig. 5(a),(b)], and the pp-modes follow the Poisson prediction exp⁡[−nε(1−cos⁡4πlα)]\exp[-n\varepsilon(1-\cos4\pi l\alpha)] for ε≤0.03\varepsilon\leq0.03 (not shown). At larger ε\varepsilon the Fano factor is strongly non-monotone and reproducible between independent ensembles (Figure 5): for pair A it peaks at F=3.18F = 3.18 at ε=0.29\varepsilon= 0.29, and the diffusion coefficient D=4εFD = 4\varepsilon F peaks at 4.09 at ε=0.374\varepsilon= 0.374 [panel (c)]. The triangle map with the same parameters (ε=1/2\varepsilon= 1/2) is thus far from the most diffusive member of the family: for pair A its D=1.19D = 1.19 (against 3.09 for the benchmark pair of § 4) is close to the Bernoulli value 1 and 3.4 times smaller than the maximum. At ε=1/4\varepsilon= 1/4 the variance saturates, Var⁡(Xn)/n=2.62±0.01\operatorname{Var}(X_n)/n = 2.62 \pm0.01 from n≃213n \simeq2^{13} to 2242^{24}.

Figure 5. Fano factor of the deformed map at generic β\beta, 250 values of ε\varepsilon, n=216n = 2^{16}, two independent ensembles of 32 768 orbits per pair. (a), (b) F=Var⁡(Sn)/(4εn)F = \operatorname{Var}(S_n)/(4\varepsilon n) for pairs A and B: mean of the two ensembles (band: their spread; dotted: FF at n/16n/16); dashed, the Poisson value 1; dash-dotted, the Bernoulli value 1−ε1 - \varepsilon. (c) D=4εFD = 4\varepsilon F.

β=0\beta= 0: ballistic stays and a logarithmic term. At β=0\beta= 0 and ε≠1/2\varepsilon\ne1/2 the variance per step grows. At ε=1/4\varepsilon= 1/4 (10510^5 orbits to 2242^{24}, two seeds) the raw variance is dominated by a few orbits (one orbit carries up to 18% of ∑Xn2\sum X_n^2, Hill index of ∣Xn∣\lvert X_n\rvert about 2), while the interquartile scale is reproducible and grows linearly in ln⁡n\ln n [Fig. 6(a)]. Each extreme orbit sits, for up to ≳2.5×106\gtrsim2.5 \times10^6 steps, on a periodic stretch of period two (+−+-) or four (++−−++--), i.e. on a zero-sum-word resonance. During such a stay SnS_n does not move while the centring term does, so XnX_n changes by −(1−2ε)-(1-2\varepsilon) per step: a ballistic excursion. At ε=1/2\varepsilon= 1/2 the same stays do not displace the walk.

The stays obey the law (11), with KwK_w computed with the threshold ε\varepsilon and no free parameter [Figure 6(b)]: in 20 000 orbits of length 2242^{24} (α=2−1\alpha= \sqrt{2}-1),

ε=0.1\varepsilon= 0.1ε=0.2\varepsilon= 0.2ε=0.25\varepsilon= 0.25
K2K_2 (theory / measured)0.0800 / 0.0800.3200 / 0.318–0.3200.4216 / 0.424
K4K_4 (theory / measured)0.0800 / 0.078–0.0790.3200 / 0.3220.5000 / 0.499–0.501

Table 2.

The number of stays of length at least ℓ\ell per step is then K/(2ℓ2)K/(2\ell^2), and a stay of length ℓ\ell adds (1−2ε)2ℓ2(1-2\varepsilon)^2\ell^2 to X2X^2: the second moment is marginal (hence the Hill index ≃2\simeq2 above), and

Var⁡(Xn)n≃D0+c(ε)ln⁡n,c(ε)=(1−2ε)2Ktot,Ktot=∑wKw.(16)\frac{\operatorname{Var}(X_n)}{n} \simeq D_0+c(\varepsilon)\ln n,\qquad c(\varepsilon)=(1-2\varepsilon)^2K_{\mathrm{tot}},\qquad K_{\mathrm{tot}}=\sum_w K_w. \tag*{(16)}

For α=2−1\alpha=\sqrt{2}-1 and ε≤0.2\varepsilon\leq0.2 the enumeration finds only m=2,4m=2,4, with K2=K4=8ε2K_2=K_4=8\varepsilon^2, so c(ε)=16ε2(1−2ε)2c(\varepsilon)=16\varepsilon^2(1-2\varepsilon)^2. Above ε≃0.2\varepsilon\simeq0.2 longer words appear, and an exact enumeration of the primitive zero-sum words up to length 60 gives the converged values Ktot=0.976K_{\mathrm{tot}}=0.976 (c=0.244c=0.244) at ε=1/4\varepsilon=1/4 and 1.6891.689 (c=0.270c=0.270) at ε=0.3\varepsilon=0.3; for ε≥0.325\varepsilon\geq0.325 the sum is not converged at length 60 (supplementary material, Sec. S6).

To test Eq. (16) without the heavy-tailed variance, we split Xn=Xnrest+XnstayX_n=X_n^{\mathrm{rest}}+X_n^{\mathrm{stay}} along each orbit, XstayX^{\mathrm{stay}} being the displacement accumulated inside the stays. The centered Var⁡(Xnrest)/n\operatorname{Var}(X_n^{\mathrm{rest}})/n is flat over sixteen octaves (slopes in ln⁡n\ln n of 0.001±0.0010.001\pm0.001 and 0.004±0.0020.004\pm0.002 at ε=0.1,0.2\varepsilon=0.1,0.2; errors are bootstrap over orbits here and below), so the whole logarithmic growth is in the stays [Fig. 6(c)], whose slope for n≤218n\leq2^{18} is 0.064±0.0270.064\pm0.027 and 0.215±0.0790.215\pm0.079 against 0.1020.102 and 0.2300.230 from Eq. (16). At larger nn the stays that matter have length comparable to nn and are too rare to sample. With stays of all even periods up to 14, XrestX^{\mathrm{rest}} is flat also at ε=0.25\varepsilon=0.25 and 0.30.3, but not at 0.350.35, where longer words matter; the stay slopes lie systematically below Eq. (16) (0.15±0.050.15\pm0.05 against 0.270.27 at ε=0.3\varepsilon=0.3), so Eq. (16) is tested only to ∼40%\sim40\% beyond ε=0.25\varepsilon=0.25.

Finally, the deformation moves the zero mean drift to β=−α(1−2ε)\beta=-\alpha(1-2\varepsilon). A stay keeps SS on the slope sm/ms_m/m of its word; if this equals the mean slope $1-2\varepsilon$, the stay does not displace XX, otherwise it does. At ε=1/4\varepsilon=1/4 on the zero-drift line the diffusion is suppressed [Var⁡(Xn)/n[\operatorname{Var}(X_n)/n falls from 0.460.46 at n=28n=2^8 to 0.2388±0.00110.2388\pm0.0011 at 2242^{24}, Fig. 6(a)]$, whereas for irrational ε=(3−1)/3\varepsilon=(\sqrt{3}-1)/3, where no word has the right slope, it does not seem to be [a single run gave Var⁡(Xn)/n=1.47\operatorname{Var}(X_n)/n=1.47 on the line against 1.461.46 at a distance 10−310^{-3} from it]. The triangle map at β=0\beta=0 is the case in which the two features coincide.

Discussion and conjectures

Random-walk scaling does not identify a Bernoulli process, and decorrelation does not identify diffusion. In the generic triangle map the exponent of the variance, the Gaussian fluctuations and the first-passage exponent are those of a coin, but the diffusion coefficient is three times larger and is set by the order of bounded waiting times, and the persistence amplitude is not determined by the marginals. At β=0\beta=0 the same type of correlations cancels the Green–Kubo sum, while the walk keeps spreading slowly. The zero-sum-word resonances, pure twists at β=0\beta=0, control the quantities we can compute: the trapping law, the crossover time β−1/2\beta^{-1/2} and, in the deformed map, the logarithmic term (16). We collect the picture in three conjectures, in decreasing order of evidence.

Conjecture 1 (generic map). If 11, α\alpha, β\beta are rationally independent, Tα,βT_{\alpha,\beta} is uniquely ergodic and mixing, the Green–Kubo series converges to D=G∞>0D=G_\infty>0, and, for Lebesgue-distributed initial conditions, Sn/nS_n/\sqrt{n} converges in law to the centered Gaussian of variance DD.

The evidence is the absence of periodic orbits and the bound on all word repetitions under rational independence (Sec. 3), the N−1/2N^{-1/2} decay of worst-case Birkhoff averages, the decay of the Fourier correlations, and the statistics of Sec. 4: Kolmogorov distance from the Gaussian at the sampling-noise level for 103≤n≤3.4×10710^3\leq n\leq3.4\times10^7, running maximum with the Brownian amplitude 2D/π\sqrt{2D/\pi} and persistence exponent 1/21/2, which also support the functional version (convergence of S⌊nt⌋/DnS_{\lfloor nt\rfloor}/\sqrt{Dn} to Brownian motion). The counterexample of Sec. 3 shows that rational independence cannot be weakened to the condition of [10].

Conjecture 2 (intermediate cohomology at β=0\beta= 0). For irrational α\alpha and β=0\beta= 0, T2T^2 is ergodic.

Consequently, by Corollary 1, XX is not a measurable coboundary, the laws of SnS_n are not tight, and the spectral measure of XX has no atom at π\pi, although Var⁡(Sn)=o(n)\operatorname{Var}(S_n) = o(n).

The symbol XX would then sit strictly between the two familiar classes of observables. A coboundary X=g−g∘TX = g - g \circ T has Sn=g−g∘TnS_n = g - g \circ T^n, whose laws are tight, and a spectral density vanishing at ω=0\omega= 0; an observable with D>0D > 0 has Var⁡(Sn)∼Dn\operatorname{Var}(S_n) \sim Dn. At β=0\beta= 0, XX would share D=0D = 0 with the first class and unbounded fluctuations with the second: SnS_n spreads without limit, but more slowly than n\sqrt{n}. The data are consistent: the bulk of the law of SnS_n widens, Var⁡(Sn)\operatorname{Var}(S_n) shows no plateau up to 10810^8, the characteristic function (15) decays, and Var⁡(An)/n2\operatorname{Var}(A_n)/n^2 decreases, although slowly (∝n−0.2\propto n^{-0.2}, the same effective exponent as the Fourier correlations), and the transfer functions of the exact Markov approximations stay small, growing only as ln⁡J\ln J (§ 5). Whether Sn/σnS_n/\sigma_n has a non-Gaussian limit, as its stable shape suggests, is a further question.

Conjecture 3 (resonance renormalization). At β=0\beta= 0 the growth law of Var⁡(Sn)\operatorname{Var}(S_n) is determined by the hierarchy of zero-sum-word twists, through a renormalization relating the resonances of words of length mm to those of longer words.

Every zero-sum word is a twist at β=0\beta= 0, resonances exist at all lengths, a small β\beta destroys them on the time scale at which the variance crosses over to diffusion, and the resonant momenta (j−αW)/m(j-\alpha W)/m exist for every α\alpha, which would explain the insensitivity to the arithmetic of α\alpha. In the deformed map the hierarchy provably sets the logarithmic coefficient through KtotK_{\mathrm{tot}}; at ϵ=1/2\epsilon= 1/2 the stays carry no displacement, and the growth must come from the walk between them, which is the open case. A derivation along these lines would replace the choice among ln⁡2n\ln^2 n, (ln⁡n)p(\ln n)^p and nγn^\gamma, which the data cannot make. Renormalization schemes of this kind are established for interval exchange maps [28, 29], for piecewise rotations and polygon exchanges [23, 30, 31] and, for the quadratic phase of the lift, for theta sums and Heisenberg nilflows [32, 33]; none applies directly here, because the cylinders of the triangle map are polygons cut by the discontinuity lines and the block maps are twists rather than translations.

Other questions remain open: the dependence of DD on (α,β)(\alpha,\beta) and the origin of the maxima of F(ϵ)F(\epsilon); the convergence of KtotK_{\mathrm{tot}} for ϵ>0.3\epsilon> 0.3; and a theory of persistence at β=0\beta= 0, which the resonances do not explain, since long-surviving orbits are almost never repetitions of a periodic word. A further direction is to use observables adapted to the finite-time continuity domains and their successive cuts, rather than global Fourier modes, for which the recurrence analysis of the cuts in Ref. [17] is a natural starting point.

Supplementary Material

The supplementary material contains a table of what each diagnostic can and cannot establish, the word statistics and the twist geometry at β=0\beta= 0, further data for the generic regime (correlations, other parameter pairs and the map of D(α,β)D(\alpha,\beta)), the β=0\beta= 0 variance fits and Markov approximations, the worst-case Birkhoff averages, details of the asymmetric deformation, and the phase-space coverage tests.

Author Declarations

Conflict of Interest

The author has no conflicts to disclose.

Author Contributions

Mirko Degli Esposti: all roles (Conceptualization, Methodology, Software, Formal analysis, Investigation, Validation, Visualization, Writing – original draft, Writing – review & editing).

Data availability

The data that support the findings of this study and the code used to generate them are openly available in Zenodo at https://doi.org/10.5281/zenodo.23623347 [34].

Numerical methods

Integrator. The integers SnS_n and WnW_n of Eq. (4) are updated exactly; qnq_n is recomputed at every step from q0q_0, p0p_0 (double precision) and α\alpha, β\beta (double-double, ∼10−32\sim10^{-32}), with an absolute error of order 10−1610^{-16} that does not accumulate, so XnX_n can be wrong only if qnq_n falls within this distance of a cut. The construction requires ∣Wn∣|W_n|, n(n+1)/2<253n(n+1)/2 < 2^{53}, i.e. n≲108n \lesssim10^8. Symbol sequences coincide with 60-digit mpmath orbits over 2×1042 \times10^4 steps, at β=0\beta= 0 and for generic parameters, and the cost is 1.3 times that of double-precision iteration.

Polygonal cylinders. For a fixed word, qkq_k is affine in (q0,p0)(q_0,p_0), so each cylinder is a finite union of convex polygons. The partition into cylinders of length nn is built by clipping with the lines q0+kp0=constq_0 + kp_0 = \mathrm{const}, k<nk < n (pieces below 10−1610^{-16} in area dropped; lost mass <10−13< 10^{-13}). Then Var⁡(Sn)=∑w∣[w]∣sn(w)2\operatorname{Var}(S_n) = \sum_w |[w]|s_n(w)^2, with ∣[w]∣|[w]| the area of [w][w], is a finite sum of polygon areas, and stationarity gives C(k)=[Var⁡(Sk+1)−2Var⁡(Sk)+Var⁡(Sk−1)]/2C(k) = [\operatorname{Var}(S_{k+1}) - 2\operatorname{Var}(S_k) + \operatorname{Var}(S_{k-1})]/2. The number of pieces grows as ≃0.7n3\simeq0.7n^3, of the order of the cubic growth reported in Ref. [10], which limits exact computations to n≲100n \lesssim100 (to n=256n = 256 for the survival probability Q+(n)Q_+(n), for which only the cylinders with Sk>0S_k > 0 for all kk are refined). Word probabilities, Var⁡(Sn)\operatorname{Var}(S_n), C(k)C(k) and Q+(n)Q_+(n) obtained in this way agree with Monte Carlo estimates. The trap coefficients KwK_w are computed exactly for all primitive zero-sum words of length m≤60m \le60 by a depth-first search over intervals of qq on each resonant line p=(j−αW)/mp = (j - \alpha W)/m.

Correlations and spectra. Correlations are estimated per trajectory with the exact mean 0, by FFT on zero-padded segments, and averaged over independent trajectories. Spectra are averaged periodograms of Hann-windowed segments of lengths 2152^{15} to 2242^{24} with 50% overlap, binned logarithmically and combined across segment lengths; the lowest frequency is ω≃3×10−6\omega\simeq3 \times10^{-6}.

Statistics. The statistical unit is always an independent initial condition drawn from Lebesgue measure. Errors on variances are bootstrap or batch standard deviations over independent orbits; survival probabilities carry Wilson 68% intervals; effective exponents are computed over windows of fixed ratio. Replicate ensembles with independent seeds agree within errors, except for the heavy-tailed raw variances of § 7, for which we report robust scales. The same pipeline applied to Bernoulli sequences reproduces D=1D = 1 and SX(0)=1S_X(0) = 1 within 4% (supplementary material, § 14).

Use of AI tools. The simulation and analysis code was written and refactored with the assistance of Claude Opus 5.5 and Claude Sonnet 5.5 (Anthropic), used through the Claude Code agent, to speed up code development and to obtain independent critical readings of the drafts. The same tools proposed numerical tests, drafted and revised passages of the manuscript, produced simulated referee reports on earlier drafts, assisted with the literature search, and checked quantitative statements in the text against the project’s own logs. The author reviewed and edited all content, checked the mathematical arguments, and takes full responsibility for the article. Every bibliographic entry was checked against the publisher’s or a bibliographic database record, and every numerical result reported here is produced by the deposited code [34].

Supplementary Material

Experiment numbers refer to the scripts experiments/NN_*.py of the code archive (see Data availability); they follow the chronological research log, so not every number appears here. Table S1 summarizes what each diagnostic of the main text can and cannot establish.

DiagnosticWhat it measuresWhat it cannot establish
Var⁡(Sn)\operatorname{Var}(S_n), C(k)C(k), GKG_K, SX(ω)S_X(\omega)second-order structure; DD and finite-size corrections on the explored scalesasymptotic limits; any non-Gaussian or higher-order property
CLT diagnosticsone-time distribution of SnS_n, two-time incrementsa functional limit theorem
First passages Q±(n)Q_{\pm}(n)persistence, a multi-time propertya limiting exponent beyond the explored window
Inter-event intervalsdistribution versus order of the waiting timesthe asymptotics of the order effect
Exact cylinders, Markov hierarchyword probabilities without sampling; finite-order modelsthe asymptotic behavior of the map
Fourier modes, worst-case averagesdecay of correlations; uniformity of Birkhoff averagesmixing or unique ergodicity

Table S1. Diagnostics used in this work and their scope. Variance, correlations and spectrum are three estimators of the same second-order structure; their agreement is a consistency check rather than independent evidence.

Words, runs and exact probabilities

Run lengths near rational drifts (Experiment 21). The bound of Lemma 1 is Ls=L(cs)L_s = L(c_s), cs=β+sαc_s = \beta+ s\alpha, where L(c)L(c) is the longest run of the skew shift FcF_c inside a half-torus; it is the same for both halves, since q↦q+1/2q \mapsto q + 1/2 commutes with FcF_c and exchanges them, and L(c)=L(−c)=L(1−c)L(c) = L(-c) = L(1-c) by the involution (q,p)↦(−q,−p)(q,p) \mapsto(-q,-p). Hence L+(α,β)=L−(α,−β)L_+(\alpha,\beta) = L_-(\alpha,-\beta), and L+=L−L_+ = L_- at β=0\beta= 0, but not in general. Figure S1 shows L(c)L(c), computed exactly from the polygonal run cylinders for 20 000 values of cc. It is piecewise constant and small (6–7 on most of [0.1,0.3][0.1,0.3]) and diverges only at the rationals a/ba/b for which Fa/bF_{a/b} has a periodic orbit contained in one half-torus: among those of small denominator, c=0c = 0, 1/31/3, 2/52/5 and 1/21/2, but not 1/51/5, 1/41/4, 2/72/7 or 3/73/7. Near them L∝∣c−a/b∣−1/2L \propto\lvert c-a/b\rvert^{-1/2} (local exponents 0.41–0.49 on 10−4≤∣c−a/b∣≤10−210^{-4} \leq\lvert c-a/b\rvert\leq10^{-2}), which for c→0c \to0 reflects qk−q0≃kp0+ck2/2q_k-q_0 \simeq kp_0+ck^2/2. In the (α,β)(\alpha,\beta) plane the resonances form the two families of diagonals β±α=a/b\beta\pm\alpha= a/b. For the generic benchmark, c+=β+α≃0.986c_+ = \beta+\alpha\simeq0.986 is close to c=0c = 0 and L+=17L_+ = 17, whereas c−≃0.250c_- \simeq0.250 gives L−=7L_- = 7. Irrationality of α,β\alpha,\beta with β∉Z+αZ\beta\notin\mathbb{Z}+\alpha\mathbb{Z} does not make both c±c_{\pm} irrational: β=12−α\beta= \frac{1}{2}-\alpha gives c+=12c_+ = \frac{1}{2}.

The period-two twist and its traps. For w=+w=+ at β=0\beta=0, a point with q∈[0,1/2)q \in[0,1/2) and q+p+α∈[1/2,1)q+p+\alpha\in[1/2,1) (mod 1) is mapped by T2T^2 to (q+2p+α,p)(q+2p+\alpha,p): on the cylinder [+−][+-] the block map is a pure twist, the identity on the lines 2p+α∈Z2p+\alpha\in\mathbb{Z}. Imposing the symbols gives the two segments of Figure S2, p=1−α/2p=1-\alpha/2, q∈[12−α2,12)q \in[\frac{1}{2}-\frac{\alpha}{2},\frac{1}{2}) and p=12−α2p=\frac{1}{2}-\frac{\alpha}{2}, q∈[0,12−α2)q \in[0,\frac{1}{2}-\frac{\alpha}{2}), together with their images under one step, which carry the rotation −+-+. At maximum distance δ\delta the block map is the translation q↦q+2δq \mapsto q+2\delta, so an orbit slides along the segment and repeats +−+- for the time it needs to cross it, L/∣δ∣L/\lvert\delta\rvert steps [Figure S2: 1415, 567, 944 steps against 1414, 566, 943]. For β>0\beta>0 the block map adds 2β2\beta to the momentum, so δ\delta grows linearly in time and the motion along the segment becomes a parabola; the longest repetition is the one whose parabola, of depth βn2/8\beta n^2/8 after nn steps, just fits in the longer segment, of length Lmax⁡=max⁡(α,1−α)/2L_{\max}=\max(\alpha,1-\alpha)/2, i.e. n=22Lmax⁡/βn=2\sqrt{2L_{\max}/\beta} [Figure S2: 483 steps against 484 at β=10−5\beta=10^{-5}].

P(−+−+)=1/12P(-+-+)=1/12 at β=0\beta=0. With r=q0r=q_0 and x=p0−αx=p_0-\alpha (independent and uniform),

P(−+−+)=∫01dx∫01dr 1Ax(r)1Ax(r+2x+α),P(-+-+)=\int_0^1 dx\int_0^1 dr\,\mathbf{1}_{A_x}(r)\mathbf{1}_{A_x}(r+2x+\alpha),

where AxA_x is an arc of length ∥x∥\lVert x\rVert. Its Fourier coefficients in α\alpha are proportional to

∫01sin⁡2(πk∥x∥)e4πikx dx=0\int_0^1 \sin^2(\pi k\lVert x\rVert)e^{4\pi i kx}\,dx=0

for every integer k≠0k\ne0, so

P(−+−+)=∫01∥x∥2 dx=1/12P(-+-+)=\int_0^1 \lVert x\rVert^2\,dx=1/12

for every α\alpha. The identity was first found numerically, to 10−1010^{-10}, in the exact cylinder computations of Sec. S4; it fails for β≠0\beta\ne0.

Word complexity (Experiment 20). From the exact cylinder partition we count, for n≤110n\le110, the distinct admissible words p(n)p(n) (a cylinder made of several polygons is counted once), the number of polygons, and the admissible zero-sum words z(n)z(n), i.e. those with Sn=0S_n=0 (Table S2). All words are admissible up to n=5n=5 (generic) and n=6n=6 (β=0\beta=0). The first difference satisfies s(n)/(3n2)=0.661, 0.704, 0.708s(n)/(3n^2)=0.661,\ 0.704,\ 0.708 at n=110n=110, in agreement with p/n3p/n^3: the growth is cubic, with coefficient ≃0.66\simeq0.66 (generic) and ≃0.70\simeq0.70–0.710.71 (β=0\beta=0, still rising slowly). The second difference b(n)b(n) gives b/(6n)b/(6n) fluctuating between 0.500.50 and $0.93, and cubic fits on n≥30, 50, 70n\ge30,\ 50,\ 70 give 0.570.57–0.740.74, so neither the cubic coefficient nor the lower-order terms are determined. The number of polygons exceeds p(n)p(n) by less than 0.03%0.03\%: nearly every cylinder is connected. The topological entropy vanishes. At β=0\beta=0 the zero-sum words are a constant fraction ≃23%\simeq23\% of all words, z(n)∝n3.0z(n)\propto n^{3.0}, against z(n)∝n2.1z(n)\propto n^{2.1} and a falling fraction (4% at n=110n=110) in the generic case, where the spread n\sqrt{n} of SnS_n alone would give n2.5n^{2.5}. This is consistent with the zero-sum twists of § 3.2, but we have no derivation.

casepppolygonsp/n3p/n^3zzz/pz/p
generic (α=1/e\alpha=1/e, β=(5−1)/2\beta=(\sqrt{5}-1)/2)881 378881 6200.66234 3560.039
β=0\beta=0, α=2−1\alpha=\sqrt{2}-1938 742938 9420.705215 1380.229
β=0\beta=0, α=1/e\alpha=1/e943 784943 8360.709217 2630.230

Table S2. Word complexity at n=110n=110 (Experiment 20). pp: distinct words; zz: distinct words with Sn=0S_n=0.

Generic regime: further data

Long runs. With 4096 initial conditions to n=108n=10^8 the late-time exponents are γvar=0.9957\gamma_{\mathrm{var}}=0.9957 and γmax⁡=0.5024\gamma_{\max}=0.5024, with Var⁡(Sn)/n=3.07854\operatorname{Var}(S_n)/n=3.07854 and ⟨Smax⁡(n)⟩/n=1.4077\langle S_{\max}(n)\rangle/\sqrt{n}=1.4077 at the largest checkpoint; the run to n=107n=10^7 with 2048 orbits gave 1.0086, 0.5062 and 3.10155. The ensemble variance of a few thousand orbits has a relative error of about 3% at every nn, whereas long-trajectory correlations determine DnD_n for n≲104n\lesssim10^4 ten to forty times more precisely. An earlier truncated Green–Kubo value 3.425 at K=2×105K=2\times10^5, from eight double-precision trajectories, is not confirmed (G1.3×105=3.08±0.04G_{1.3\times10^5}=3.08\pm0.04). Figure S3 shows the individual correlations: the measured C(k)C(k) coincide with the exact cylinder values for k<80k<80, with deviations distributed as statistical noise, and the RMS correlation falls by two decades across the band 102≲k≲2×10310^2\lesssim k\lesssim2\times10^3 before reaching the noise level ≃2×10−5\simeq2\times10^{-5}.

Finite-size corrections. The first moment M=∑kkC(k)M=\sum_k kC(k) converges on the accessible scales (M≃95M\simeq95–100 for 2×103≤K≤1042\times10^3\le K\le10^4), so Dn=D−2M/n+o(1/n)D_n=D-2M/n+o(1/n) reaches DD to 1% only for n≳2×104n\gtrsim2\times10^4. A non-oscillating tail C(k)≃ak−3/2C(k)\simeq ak^{-3/2} would give Dn−D∝n−1/2D_n-D\propto n^{-1/2}, GK−D∝K−1/2G_K-D\propto K^{-1/2} and a cusp SX(ω)−D∝∣ω∣S_X(\omega)-D\propto\sqrt{|\omega|}. A fit restricted to k<103k<10^3 suggests a≃0.6a\simeq0.6–1, but over 2×103≤k≤1042\times10^3\le k\le10^4 the data give a=0.08±0.16a=0.08\pm0.16. The k−3/2k^{-3/2} decay reported for the smooth observable cos⁡πy\cos\pi y [12] therefore does not carry over to θ\theta on these scales. The periodogram decreases roughly linearly in ω\sqrt{\omega} for 10−3≲ω≲10−110^{-3}\lesssim\omega\lesssim10^{-1}, the image of the correlation band at k∼102k\sim10^2–10310^3, and is flat below.

Other generic pairs (Experiment 12). For (α,β)=(2−1,g)(\alpha,\beta)=(\sqrt{2}-1,g), (1/e,2−1)(1/e,\sqrt{2}-1), (π−3,1/e)(\pi-3,1/e) and (g,3−1)(g,\sqrt{3}-1), g=(5−1)/2g=(\sqrt{5}-1)/2, Green–Kubo sums from 32 trajectories of length 2242^{24} give D=2.19D=2.19, 2.94, 2.15 and 2.33 (errors ≃0.03\simeq0.03; the ensemble variance at n=220n=2^{20} agrees). The ratio of DD to the renewal value v/2v/2 is 2.4–3.7.

Map of D(α,β)D(\alpha,\beta) (Experiment 24). Besides α→α+1\alpha\to\alpha+1, α→−α\alpha\to-\alpha (q→q+12q\to q+\frac{1}{2}) and β→−β\beta\to-\beta (σ\sigma), the identity θ/2≡12(mod1)\theta/2\equiv\frac{1}{2}\pmod{1} gives Tα+1/2,β=Tα,β+1/2T_{\alpha+1/2,\beta}=T_{\alpha,\beta+1/2}, hence D(α,β)=D(12−α,12−β)D(\alpha,\beta)=D(\frac{1}{2}-\alpha,\frac{1}{2}-\beta): the fundamental domain is the triangle α,β≥0\alpha,\beta\geq0, α+β≤12\alpha+\beta\leq\frac{1}{2}. Figure S4 shows Dn=Var⁡(Sn)/nD_n=\operatorname{Var}(S_n)/n at n=214n=2^{14} on a 200×\times200 grid with irrational offsets (1024 orbits per point, statistical error ≃4%\simeq4\%); DnD_n changes by ≃5%\simeq5\% between n=210n=2^{10} and 2142^{14}, except at β≃0\beta\simeq0 and near α=0,12\alpha=0,\frac{1}{2} (0.4% of the points change by more than 50%). Away from the edges (α,β≥0.02\alpha, \beta\ge0.02, α≤0.48\alpha\le0.48) DD ranges from 0.20 to 14 (1st–99th percentiles 0.7–6.8, median 2.3) and is below the Bernoulli value on 6% of the points; towards the skew-shift limit α→0\alpha\to0 and near β=0\beta= 0, DnD_n is still growing at n=214n = 2^{14} (up to 53 at α≃β≃10−3\alpha\simeq\beta\simeq10^{-3}) and is not a converged value. Dark lenses, with DD well below 1, sit at rational β=j/m\beta= j/m with small mm and open from the skew-shift limit α=0\alpha= 0 [panel (b)]: there zero-sum words whose length is a multiple of mm have cw∈Zc_w \in\mathbb{Z}, the mechanism that gives D=0D = 0 at β=0\beta= 0. Bright bands follow the low-order resonances mβ+sα∈Zm\beta+ s\alpha\in\mathbb{Z} with s≠0s \ne0 (e.g. β=α\beta= \alpha and α+β=12\alpha+ \beta= \frac{1}{2}) near the corners of the square but leave them in the interior.

Distribution of SnS_n. The skewness of SnS_n is 0.65–0.77 for n≤30n \le30, 0.41 at n=102n = 10^2 and 0.19 at n=103n = 10^3, decaying roughly as n−1/2n^{-1/2} and compatible with zero from n∼104n \sim10^4. The involution (q,p)↦(−q,−p)(q, p) \mapsto(-q, -p) maps SnS_n for (α,β)(\alpha, \beta) to −Sn-S_n for (α,−β)(\alpha, -\beta), so the sign of the skewness is reversed by β↦−β\beta\mapsto-\beta; for the benchmark SnS_n is slightly more often negative, consistent with Q+>Q−Q_+ > Q_-. The correlation between SnS_n and S2n−SnS_{2n} - S_n is 0.07±0.020.07 \pm0.02 at n=102n = 10^2.

Bernoulli control. The same finite-sample pipeline applied to independent unbiased Bernoulli sequences, on the same scales, gives Var⁡(Sn)/n=0.9676\operatorname{Var}(S_n)/n = 0.9676, G200000=0.9856G_{200000} = 0.9856, and a low-frequency spectrum SX(0)≃1.015S_X(0) \simeq1.015 with local exponent η=0.0069\eta= 0.0069, consistent with D=1D = 1 and SX≡1S_X \equiv1. With few trajectories the Green–Kubo partial sum is more sensitive to accumulated noise than the spectral estimate.

The β=0\beta= 0 regime: further data

Precision. The earlier study of this case already used extended precision [10]. On the same initial conditions, double-precision symbol sequences first differ from the exact ones at n∼106n \sim10^6–7×1067 \times10^6 at β=0\beta= 0 (at n∼8×104n \sim8 \times10^4 in the generic regime). For most α\alpha the variances still agree within the sampling noise, and a double-precision run with 4096 orbits gives Var⁡(S107)=88.0\operatorname{Var}(S_{10^7}) = 88.0 and Var⁡(S108)=112.6\operatorname{Var}(S_{10^8}) = 112.6, consistent with the exact data; for α=[0,2,30,2,300,2,3000,1,1,…]\alpha= [0, 2, 30, 2, 300, 2, 3000, 1, 1, \ldots], however, the double-precision variance is three times too large for n≳5×106n \gtrsim5 \times10^6.

Growth law. The six values of α\alpha and the replicate ensembles give the same curves within errors (about 5% at the largest scale); α≃0.49194\alpha\simeq0.49194 shows a short transient at n∼20n \sim20–6060, reflecting α≃1/2\alpha\simeq1/2, and no feature is visible at the convergent denominators 33102 and 74646 of π−3\pi- 3. The models a+bln⁡2na + b\ln^2 n, a+cln⁡n+bln⁡2na + c\ln n + b\ln^2 n, a+b(ln⁡n)pa + b(\ln n)^p and a+cnγa + cn^\gamma fit all ensembles equally well for n≥102n \ge10^2, with b≃0.2b \simeq0.2–0.350.35, p≃1.5p \simeq1.5–2.12.1 and γ≃0.05\gamma\simeq0.05–0.10.1 depending on the model and on α\alpha. They are numerically degenerate over seven to eight decades (nγ=1+γln⁡n+12γ2ln⁡2n+⋯n^\gamma= 1 + \gamma\ln n + \frac{1}{2}\gamma^2\ln^2 n + \cdots), the coefficient of ln⁡2n\ln^2 n is not a robust quantity, and the offset a≃6a \simeq6–8.58.5 differs between independent ensembles. The logarithmic growth of the extremes Smax⁡(n)S_{\max}(n) of single orbits reported for quadratic irrational α\alpha in Ref. [10] concerns a different quantity and is not in conflict with these results. On the explored scales we cannot exclude a very slow crossover to diffusion with a tiny DD, or a small power γ\gamma.

Spectrum. Binned values of 1−η(ω)1 - \eta(\omega) scatter by about ±0.3\pm0.3, with decade averages mostly between 0 and 0.2; the drift η=1.004→0.861\eta= 1.004 \to0.861 seen in double-precision runs of different sizes is within this scatter. The variance reconstructed from the measured spectrum through Eq. (7) reproduces the shape of the ensemble variance, 10–15% below it. The residual oscillations of GKG_K are of arithmetic origin rather than noise; A(1.3×105)=∑k≤KC(k)=−0.494±0.001A(1.3 \times10^5) = \sum_{k \le K} C(k) = -0.494 \pm0.001.

Parity and the π\pi peak. For α=1/e\alpha= 1/e, Var⁡(An)/n2\operatorname{Var}(A_n)/n^2 decreases from 0.034 to 0.0061 between n=210n = 2^{10} and 2222^{22}, and the peak of SXS_X at π\pi has height ≃680\simeq680. There are no peaks at 2π/52\pi/5 or at irrational frequencies. Per orbit, An/nA_n/n is nearly constant over decades, with abrupt jumps. At rational frequencies other than π\pi the phases of successive bursts are not locked, which is why the π\pi peak dominates although the trap coefficients of the length-four words are larger. The trap law (11) for α=1/e\alpha= 1/e agrees with the exact coefficients within 5–20%; at β=0\beta= 0 itself the estimates scatter more than at β=10−7\beta= 10^{-7} because each orbit’s momentum lattice is frozen.

Persistence by momentum class (Experiment 10). At β=0\beta= 0, pn−αSn=p0(mod1)p_n - \alpha S_n = p_0 \pmod{1}, so p0p_0 labels families of orbits that evolve independently on T×Z\mathbb{T} \times\mathbb{Z}. Sampling p0p_0 on 1024 bins (128 orbits each, α=2−1\alpha= \sqrt{2} - 1): the growth of Var⁡(Sn)\operatorname{Var}(S_n) is shared by all classes (at n=106n = 10^6 the 10% of classes with the largest conditional variance carry 15% of the total), whereas long survival is carried by thin, q0q_0-independent bands of p0p_0: at n=106n = 10^6, 77% of the classes have no survivor, against 28% expected for a p0p_0-independent survival. The bands are selected early, P(τ+>256∣p0)P(\tau_+ > 256 \mid p_0) ranging from 0.003 outside to 0.2–0.3 inside, and this profile is reproduced exactly by the polygonal cylinders pruned when SnS_n becomes positive. Some band edges coincide with the resonances ∥2p0+(2k−1)α∥=0\lVert2p_0 + (2k - 1)\alpha\rVert= 0 of orbits bouncing between two adjacent levels. After the selection the dynamics is common to all classes: conditional on survival to 4096, the probability of surviving to 2202^{20} is 0.25–0.35 in almost every active class, corresponding to θ≃0.22\theta\simeq0.22.

Persistence: null models. For fractional Brownian motion the persistence exponent is 1−H1-H [35]; the local variance exponents 2H≃0.12H \simeq0.1–0.2 would give θ≃0.9\theta\simeq0.9, far from the measured 0.2. The Sparre Andersen identity fed with the measured marginals fails at β=0\beta= 0 in the exponent, not only in the amplitude. For the generic case its extensions to correlated steps exist only under renewal-type assumptions [36]; we use it only as a null model, which reproduces simulated Bernoulli walks within 1% up to n=4096n = 4096.

An exact Markov hierarchy

Let Σ={θ(q0)=+1,θ(q0−p0)=−1}\Sigma= \{\theta(q_0) = +1, \theta(q_0 - p_0) = -1\} be the set of points where a run of +1+1 begins; μ(Σ)=1/4\mu(\Sigma) = 1/4 exactly and the mean return time is 4. By Lemma 1 the first-return map RR to Σ\Sigma is piecewise affine with finitely many branches, one for each cycle (+1)a(−1)b(+1)^a(-1)^b (67 generic, 58 for β=0\beta= 0, α=1/e\alpha= 1/e, 63 for β=0\beta= 0, α=2−1\alpha= \sqrt{2} - 1); on each branch R(q,p)=(q+np+βn(n+1)/2+αW,p+nβ+α(a−b))R(q,p) = (q + np + \beta n(n+1)/2 + \alpha W, p + n\beta+ \alpha(a-b)), n=a+bn = a + b. The increment of SS over a cycle is φ=a−b\varphi= a - b, and D=σ2(φ;R)/4D = \sigma^2(\varphi; R)/4. Cylinders of RR of length J+1J + 1 are obtained by intersecting affine images of branches with the branch partition, and define the order-JJ Markov chain on cylinders of length JJ, in the spirit of the Markov approximations of Ref. [10] but without sampling. Its asymptotic variance is σJ2=2⟨φ,x⟩π−⟨φ,φ⟩π\sigma_J^2 = 2\langle\varphi,x\rangle_\pi- \langle\varphi,\varphi\rangle_\pi, where xx solves the Poisson equation (I−P)x=φ(I-P)x = \varphi (sparse direct solver), and DJ=σJ2/4D_J = \sigma_J^2/4. For J≤28J \le28 (≃2.7×105\simeq2.7 \times10^5 states) DJ≃2.9D_J \simeq2.9 in the generic regime, while for β=0\beta= 0 DJ≃0.78/JD_J \simeq0.78/J (14≤J≤2814 \le J \le28) for both values of α\alpha. Since the chain retains a memory of about 4J4J steps, 4JDJ≃3.14J D_J \simeq3.1 should be compared with dVar⁡(Sn)/dln⁡n≃3.2d\operatorname{Var}(S_n)/d\ln n \simeq3.2–3.7 at n≃102n \simeq10^2. The transfer function χ\chi has root-mean-square size growing like J\sqrt{J} in the generic regime but only logarithmically at β=0\beta= 0 (xrms≃3x_{\mathrm{rms}} \simeq3 at J=28J = 28), suggestive of a near-coboundary structure. A finite-order chain does not determine the asymptotic behavior of the map.

Worst-case Birkhoff averages

Unique ergodicity is equivalent to sup⁡x∣N−1∑n<Nf(Tnx)∣→0\sup_x \lvert N^{-1}\sum_{n<N} f(T^n x)\rvert\to0 for a dense set of continuous ff (here the Fourier modes, with the usual care for the discontinuity of TT), which typical-orbit statistics do not probe. We estimate the supremum by the maximum over 32 768 Lebesgue initial conditions, over 8192 initial conditions within 10−1410^{-14}–10−310^{-3} of the cuts, and by a greedy adversarial search at N=210,214,216N = 2^{10}, 2^{14}, 2^{16} (64 parents ×\times 63 children, radii 10−210^{-2}–10−710^{-7}), whose worst points are then followed to N=220N = 2^{20} (Experiment 15, exact lift).

For both generic pairs the worst case decays as N−1/2N^{-1/2} (exponent 0.49 over three decades), like the uniquely ergodic skew shift, with a ratio max/median ≃5\simeq5–6 that is stable in NN (it would be ≃4\simeq4 for Gaussian fluctuations): the worst case is the extreme of a fixed, somewhat heavy-tailed distribution of ∼3×104\sim3 \times10^4 diffusive fluctuations. Adversarially selected points are transients and become typical by N=220N=2^{20}; points next to the cuts behave like random ones. The method finds the known obstructions: at β=0\beta=0 it recovers both period-two segments, p=1−α/2p=1-\alpha/2 with average cos⁡(πα/2)=0.7957\cos(\pi\alpha/2)=0.7957 and p=12−α2p=\frac{1}{2}-\frac{\alpha}{2} with cos⁡(π(1−α)/2)=0.6057\cos(\pi(1-\alpha)/2)=0.6057, constant to N=106N=10^{6}; for the counterexample pair it finds the period-four orbits, with exact average 0.42487. Typical β=0\beta=0 orbits converge slowly (median ∝N−0.08\propto N^{-0.08}), consistent with the n−0.2n^{-0.2} correlations. This is numerical evidence for unique ergodicity under rational independence, not a proof: a finite search cannot exclude very small exceptional sets.

N=210N = 2^{10}2142^{14}2182^{18}2202^{20}
generic (1/e,g)(1/e,g), random max / median0.179 / 0.0350.049 / 0.0090.012 / 0.00230.0062 / 0.0011
generic (1/e,g)(1/e,g), near cuts0.2050.0460.0110.0051
generic (1/e,g)(1/e,g), adversarial at 2102^{10}0.2130.0190.00730.0026
generic (2−1,g)(\sqrt{2}-1,g), random max0.2740.0680.0110.0050
generic (2−1,g)(\sqrt{2}-1,g), adversarial at 2102^{10}0.2930.0420.00800.0035
skew shift (0,g)(0,g), random max0.0600.0150.00350.0016
β=0\beta= 0 (2−1)(\sqrt{2}-1), random max / median0.795 / 0.0970.606 / 0.0750.587 / 0.0610.451 / 0.055
β=0\beta= 0, adversarial at 2142^{14}0.7960.7960.7960.796
counterexample (1/e,(1−α)/2)(1/e,(1-\alpha)/2), adversarial0.4250.4250.4250.425

Table S3. Largest ∣N−1∑n<Ne2πiqn∣|N^{-1}\sum_{n<N}e^{2\pi iq_n}| found (g=(5−1)/2g=(\sqrt{5}-1)/2).

The asymmetric deformation: further data

Generic β\beta. The correlation at lag one vanishes for every ε\varepsilon (since p0p_0 is uniform given q0q_0), whereas at lag two it vanishes only at ε=1/2\varepsilon=1/2. For ε<0.04\varepsilon<0.04, F=0.96F=0.96–1.121.12 (pair A) and 0.960.96–1.221.22 (pair B), with statistical error ≃0.015\simeq0.015; at ε=0.1\varepsilon=0.1–0.20.2 the tail of the pp-modes already deviates from the Poisson prediction, with a momentum relaxation time ∼1/ε\sim1/\varepsilon. The two ensembles of Figure 5 differ by an rms of 0.021 in both pairs, against 0.022 (A) and 0.021 (B) expected, so the structure is reproducible at every grid point; FF at n/16n/16 is smaller by 1.5% on average (1.8% rms), and no extrapolation in nn is attempted. Pair B has maxima of the smoothed FF near ε≃0.10\varepsilon\simeq0.10, 0.20, 0.34, 0.41, DD up to 4.80 at ε=0.41\varepsilon=0.41 and D(1/2)=2.37D(1/2)=2.37 (Experiment 12: 2.33). Two zoomed windows, ε∈[0.28,0.30]\varepsilon\in[0.28,0.30] (400 points) and 1/3±0.0051/3\pm0.005 (401 points), each with two independent ensembles of 32 768 orbits at n=216n=2^{16}, show a smooth profile: the seeds agree within statistics (rms zz-difference 1.00 and 1.02), polynomials of degree 8–12 fit with χ2/dof\chi^2/\mathrm{dof} 1.04–1.17, and F(1/3)=2.50F(1/3)=2.50 shows no feature at the rational. The jaggedness of the wide grid is a sum of bumps of width 0.005–0.01 plus noise; no self-similar structure is seen down to a step 5×10−55\times10^{-5} at noise 0.02.

Tails at β=0\beta=0, ε=1/4\varepsilon=1/4. In two seeds of 10510^5 orbits to 2242^{24}, one orbit carries 3–18% of ∑Xn2\sum X_n^2, the kurtosis is 10210^{2}–4×1034\times10^{3}, the Hill index of ∣Xn∣|X_n| is 2.0–2.3 from n=214n=2^{14} to 2242^{24}, and the raw Var⁡(Xn)/n\operatorname{Var}(X_n)/n at 2242^{24} differs between seeds (3.51 and 5.17). The interquartile scale σIQR2/n\sigma_{\mathrm{IQR}}^2/n, σIQR=(Q75−Q25)/1.349\sigma_{\mathrm{IQR}}=(Q_{75}-Q_{25})/1.349, equals 1.34, 1.89, 2.26, 2.66 at n=210,216,220,224n=2^{10},2^{16},2^{20},2^{24} in both seeds, with slope 0.133±0.0010.133\pm0.001 in ln⁡n\ln n; it is not the variance and is not compared with c(ε)c(\varepsilon). The extreme orbits have uniform (q0,p0)(q_0,p_0), are ordinary at earlier times (Spearman correlation of ∣X∣|X| between 2162^{16} and 2242^{24}: 0.002), and are all negative. The same run at ε=1/2\varepsilon=1/2 has light tails (kurtosis 4.7–5.3, Hill index 5–9) and Var⁡(S224)=94.8\operatorname{Var}(S_{2^{24}})=94.8, against 95.0 from the fit a+bln⁡2na+b\ln^2 n. In Experiment 17 the stays of lengths 128 to 2162^{16} contribute 1.41 to EX2/n\mathbb{E}X^2/n at ε=1/4\varepsilon=1/4, against 1.44 predicted, and the law (11) holds within 1–2% up to ℓ≃212\ell\simeq2^{12}.

The total trap coefficient (Experiment 19). Table S4 gives Ktot(ε)K_{\mathrm{tot}}(\varepsilon) for α=2−1\alpha=\sqrt{2}-1 from all primitive zero-sum words of length m≤60m\leq60. Ktot=16ε2K_{\mathrm{tot}}=16\varepsilon^2 holds exactly for ε≤0.2\varepsilon\leq0.2 (only m=2,4m = 2, 4); the first longer word, m=10m = 10, appears at ε≃0.2075\varepsilon\simeq0.2075, then m=8m = 8 at 0.229. K2K_2 saturates at 3α2=9−623\alpha^2 = 9 - 6\sqrt{2} for 0.3≤ε≤0.50.3 \le\varepsilon\le0.5. The identity Ktot=16ε2K_{\mathrm{tot}} = 16\varepsilon^2 holds for ε≤0.05\varepsilon\le0.05 for five values of α\alpha (2−1\sqrt{2} - 1, 1/e1/e, 3−1\sqrt{3} - 1, (5−1)/2(\sqrt{5} - 1)/2, π−3\pi- 3), with an α\alpha-dependent range (π−3\pi- 3 deviates at 0.1, the others at 0.15–0.2). For ε≥0.325\varepsilon\ge0.325 the partial sums still grow at m=60m = 60 and cc is only a lower bound; whether KtotK_{\mathrm{tot}} is finite there, and how it behaves as ε→1/2\varepsilon\to1/2, is open.

ε\varepsilonK(m≤30)K(m \le30)K(m≤60)K(m \le60)16ε216\varepsilon^2ccconverged
0.1000.16000.16000.16000.1024yes
0.2000.64000.64000.64000.2304yes
0.2250.78550.78550.81000.2376yes
0.2500.97600.97601.00000.2440yes
0.2751.25791.25791.21000.2547yes
0.3001.68891.68891.44000.2702yes
0.3252.38362.47551.6900≥0.303\ge0.303no
0.3503.09073.54801.9600≥0.319\ge0.319no
0.4004.34355.53442.5600≥0.221\ge0.221no
0.4506.515411.47583.2400≥0.115\ge0.115no

Table S4. KtotK_{\mathrm{tot}} from zero-sum words of length m≤30m \le30 and m≤60m \le60, and c=(1−2ε)2Ktot(m≤60)c = (1 - 2\varepsilon)^2K_{\mathrm{tot}}(m \le60); α=2−1\alpha= \sqrt{2} - 1.

Stay decomposition (Experiment 18). Stays are maximal stretches of period two (sn=sn−2s_n = s_{n-2}, at least 66 steps) or four (at least 132 steps, without a period-two sub-stretch), 20 000 orbits to 2242^{24} per ε\varepsilon. Centered variances are essential, since outside the stays the mean step is not zero. Var⁡(Xnrest)/n\operatorname{Var}(X_n^{\mathrm{rest}})/n goes 0.58 → 0.61 (ε=0.1\varepsilon= 0.1), 1.16 → 1.28 (0.2), 1.17 → 1.33 (0.25). At larger nn the stay slope is lower than predicted (0.099 ± 0.051 at ε=0.2\varepsilon= 0.2 on [218,224][2^{18}, 2^{24}]), as expected: in 2×1042 \times10^4 orbits fewer than one stay with ℓ≥n/8\ell\ge n/8 is expected at n=224n = 2^{24} (≃ 0.02 at ε=0.2\varepsilon= 0.2), so the sample variance underestimates its mean; the data support Eq. (16) for n≤218n \le2^{18} only, and D0D_0 is not determined. With stays of all even periods up to 14, at ε=0.25\varepsilon= 0.25, 0.3, 0.35, the slopes of XrestX^{\mathrm{rest}} are −0.001 ± 0.002, 0.002 ± 0.002 and 0.150 ± 0.029 (words longer than 14 are missing at 0.35), and the stay slope at ε=0.3\varepsilon= 0.3 is 0.15 ± 0.05 on [212,218][2^{12}, 2^{18}], about 2σ2\sigma below c=0.270c = 0.270. Hence Eq. (16) is not contradicted, but it is tested to about 40% only for ε>0.25\varepsilon> 0.25; at ε=0.25\varepsilon= 0.25 the raw variance is too noisy for a slope.

Zero drift and small β\beta. On the line β=−α(1−2ε)\beta= -\alpha(1 - 2\varepsilon) at ε=1/4\varepsilon= 1/4, the slope of Var⁡(Xn)/n\operatorname{Var}(X_n)/n in ln⁡n\ln n goes from −0.046 (282^8–2122^{12}) to −0.0023 ± 0.0006 (2202^{20}–2242^{24}); whether the limit is finite, ≃ 0.23, is not decided. At ε=(3−1)/3\varepsilon= (\sqrt{3} - 1)/3 the same line gives 1.47 against 1.46 at β\beta shifted by 10−310^{-3}. For β=10−3\beta= 10^{-3} and 10−210^{-2} at ε=1/4\varepsilon= 1/4 (2048 orbits, n≤222n \le2^{22}) the variance is flat in nn, the trap cutoff being ∝β−1/2\propto\beta^{-1/2}.

Phase-space exploration at finite resolution

To probe the exploration of the torus directly, we revisit the cell-coverage test of Casati and Prosen [12]. Partition the torus into N=m2N = m^2 equal square cells and let nN(t;z0)n_N(t; z_0) be the number of distinct cells visited by z0,…,Tt−1z0z_0, \ldots, T^{t-1}z_0; we estimate rN(t)=Ez0[nN(t;z0)/N]r_N(t) = \mathbb{E}_{z_0}[n_N(t; z_0)/N] from 4096 independent uniform initial conditions, counting coverage per orbit before averaging. For independent uniform draws,

rNiid(t)=1−(1−1N)t⟶1−e−s,s=tN,(17)r_N^{\mathrm{iid}}(t) = 1 - \left(1 - \frac{1}{N}\right)^t \longrightarrow1 - e^{-s}, \qquad s = \frac{t}{N}, \tag*{(17)}

and we use the exact finite-NN expression. We use the parameters of the original study, which in our unit-torus convention read αcov=(g−e−1)/4\alpha_{\mathrm{cov}}=(g-e^{-1})/4, βcov=(g+e−1)/4\beta_{\mathrm{cov}}=(g+e^{-1})/4, g=(5−1)/2g=(\sqrt{5}-1)/2 (the change of coordinates q=(x/2) mod 1q=(x/2)\bmod1, p=(y/2) mod 1p=(y/2)\bmod1 from the torus [−1,1)2[-1,1)^2 halves both parameters and preserves the partition for even mm), and compare them with (αcov,0)(\alpha_{\mathrm{cov}},0). All coverage orbits are in double precision; the β=0\beta=0 data, to t=1.6×106t=1.6\times10^6, are therefore indicative only.

Generic coverage. For m=100,200,400,800m=100,200,400,800 and t≤10Nt\leq10N the coverage nearly coincides with the independent reference (Figure S5), with a small resolved deficit dN(1)=rNiid−rN=(3.226,1.495,0.640,0.201)×10−3d_N(1)=r_N^{\mathrm{iid}}-r_N=(3.226,1.495,0.640,0.201)\times10^{-3} at s=1s=1 (standard errors (0.145,0.075,0.038,0.020)×10−3(0.145,0.075,0.038,0.020)\times10^{-3}; a repetition at m=400m=400 with another seed gives 0.677±0.0390.677\pm0.039). N dN(1)\sqrt{N}\,d_N(1) decreases from 0.3230.323 to 0.1610.161, so a simple N−1/2N^{-1/2} correction is not supported, and no asymptotic exponent is assigned.

Figure S5. Generic phase-space coverage, 4096 initial conditions per grid. (a) Mean covered fraction against s=t/Ns=t/N, with the exponential reference. (b) Signed deviation from the exact finite-NN independent reference, ×103\times10^3; shading, two pointwise standard errors.

β=0.\boldsymbol{\beta=0.} Setting β=0\beta=0 at the same α\alpha changes coverage drastically (Figure S6): at t=10Nt=10N, rN=0.11309±0.00054r_N=0.11309\pm0.00054, 0.06046±0.000280.06046\pm0.00028, 0.03312±0.000170.03312\pm0.00017 for m=100,200,400m=100,200,400, against 0.999950.99995 for independent draws; for m=100m=100, coverage rises only from 0.0430.043 to 0.0810.081 and 0.1130.113 for t=103t=10^3, 10410^4, 10510^5. Neither a logarithmic law nor a non-zero asymptotic unvisited fraction is established. Ergodicity implies eventual visits to every cell but no rate, and near-complete coverage does not imply ergodicity or independence.

Figure S6. Slow coverage for β=0\beta= 0 at α=αcov\alpha= \alpha_{\mathrm{cov}}: coverage on a linear scale against logarithmic s=t/Ns = t/N (shading: two pointwise standard errors); the independent reference, not shown, would be ≃99.995%\simeq99.995\% at s=10s = 10.

References

  1. [1]Giulio Casati and Tomaž Prosen. Mixing property of triangular billiards. Physical Review Letters, 83(23):4729–4732, 1999.DOI
  2. [2]G. Casati, C. Tsallis, and F. Baldovin. Linear instability and statistical laws of physics. Europhysics Letters, 72(3):355–361, 2005.DOI
  3. [3]Roberto Artuso, Tulio M. de Oliveira, and Cesar Manchein. Records and occupation time statistics for area-preserving maps. Entropy, 25(2):269, 2023.DOI
  4. [4]Howard Masur and Serge Tabachnikov. Rational billiards and flat structures. In Boris Hasselblatt and Anatole Katok, editors, Handbook of Dynamical Systems, volume 1A, pages 1015–1089. Elsevier, 2002.
  5. [5]S. Kerckhoff, H. Masur, and J. Smillie. Ergodicity of billiard flows and quadratic differentials. Annals of Mathematics, 124(2):293–311, 1986.DOI
  6. [6]Jon Chaika and Giovanni Forni. Weakly mixing polygonal billiards. Annals of Mathematics, 203(3):695–735, 2026.arxiv.org/abs/2003.00890
  7. [7]OpenAI. Ergodicity of triangular billiards with an irrational angle, 2026. OpenAI Math Release preprint, September 25, https://github.com/openai/math/blob/main/preprints/Ergodicity-of-triangular-billiards-with-an-irrational-angle-September-25-2026/Ergodicity-of-triangular-billiards-with-an-irrational-angle-September-25-2026.pdf.hub.valency.io/works/W_sn7apvb4
  8. [8]OpenAI. Weak mixing of triangular billiards with an irrational angle, 2026. OpenAI Math Release preprint, October 5, https://github.com/openai/math/blob/main/preprints/Weak-mixing-of-triangular-billiards-with-an-irrational-angle-October-5-2026/weak-mixing-of-triangular-billiards.pdf.hub.valency.io/works/W_7qhykhwq
  9. [9]Giovanni Forni and Nelson Moll. Cohomological equation for geodesic flows on flat surfaces, 2025. arXiv:2510.18128.arxiv.org/abs/2510.18128
  10. [10]Martin Horvat, Mirko Degli Esposti, Stefano Isola, Tomaž Prosen, and Leonid Bunimovich. On ergodic and mixing properties of the triangle map. Physica D: Nonlinear Phenomena, 238(4):395–415, 2009.DOI
  11. [11]Kensuke Yoshida, Giulio Casati, Shingo Watanabe, and Akira Shudo. Sublinear diffusion in the generalized triangle map. Physical Review E, 106(1):014206, 2022.DOI
  12. [12]Giulio Casati and Tomaž Prosen. Triangle map: A model of quantum chaos. Physical Review Letters, 85(20):4261–4264, 2000.pubmed.ncbi.nlm.nih.gov/11060613
  13. [13]Rainer Klages and Nickolay Korabel. Understanding deterministic diffusion by correlated random walks. Journal of Physics A: Mathematical and General, 35(23):4823–4836, 2002.DOI
  14. [14]V. P. Leonov. On the dispersion of time-dependent means of a stationary stochastic process. Theory of Probability and its Applications, 6(1):87–93, 1961.DOI
  15. [15]E. Sparre Andersen. On the fluctuations of sums of random variables. Mathematica Scandinavica, 1:263–285, 1953.DOI
  16. [16]E. G. Altmann, G. Cristadoro, and M. Degli Esposti. On the origin of long-range correlations in texts. Proceedings of the National Academy of Sciences, 109(29):11582–11587, 2012.DOI
  17. [17]Mirko Degli Esposti and Stefano Galatolo. Recurrence near given sets and the complexity of the Casati–Prosen map. Chaos, Solitons & Fractals, 23(4):1275–1284, 2005.
  18. [18]Klaus Schmidt. Cocycles on Ergodic Transformation Groups, volume 1 of Macmillan Lectures in Mathematics. Macmillan Company of India, Delhi, 1977.
  19. [19]H. Furstenberg. Strict ergodicity and transformation of the torus. American Journal of Mathematics, 83:573–601, 1961.DOI
  20. [20]H. B. Keynes and D. Newton. A “minimal”, non-uniquely ergodic interval exchange transformation. Mathematische Zeitschrift, 148:101–105, 1976.DOI
  21. [21]W. A. Veech. Strict ergodicity in zero dimensional dynamical systems and the Kronecker–Weyl theorem mod 2. Transactions of the American Mathematical Society, 140:1–33, 1969.DOI
  22. [22]Michael Boshernitzan. A condition for unique ergodicity of minimal symbolic flows. Ergodic Theory and Dynamical Systems, 12:425–428, 1992.
  23. [23]K. L. Kouptsov, J. H. Lowenstein, and F. Vivaldi. Quadratic rational rotations of the torus and dual lattice maps. Nonlinearity, 15:1795–1842, 2002.DOI
  24. [24]S. M. Duarte Queirós. On the role of ergodicity and mixing in the central limit theorem for Casati–Prosen triangle map variables, 2008. arXiv:0802.0406 (v4, 2009).
  25. [25]E. Sparre Andersen. On the fluctuations of sums of random variables II. Mathematica Scandinavica, 2, 1954.DOI
  26. [26]Z. Rudnick and P. Sarnak. The pair correlation function of fractional parts of polynomials. Communications in Mathematical Physics, 194:61–70, 1998.DOI
  27. [27]J. Marklof. Pair correlation densities of inhomogeneous quadratic forms. Annals of Mathematics, 158:419–471, 2003.DOI
  28. [28]G. Rauzy. Échanges d’intervalles et transformations induites. Acta Arithmetica, 34:315–328, 1979.DOI
  29. [29]W. A. Veech. Gauss measures for transformations on the space of interval exchange maps. Annals of Mathematics, 115:201–242, 1982.DOI
  30. [30]J. H. Lowenstein, K. L. Kouptsov, and F. Vivaldi. Recursive tiling and geometry of piecewise rotations by π/7. Nonlinearity, 17:371–395, 2004.DOI
  31. [31]W. P. Hooper. Renormalization of polygon exchange maps arising from corner percolation. Inventiones Mathematicae, 191:255–320, 2013.arxiv.org/abs/1105.6137
  32. [32]J. Marklof. Limit theorems for theta sums. Duke Mathematical Journal, 97:127–153, 1999.DOI
  33. [33]L. Flaminio and G. Forni. Equidistribution of nilflows and applications to theta sums. Ergodic Theory and Dynamical Systems, 26:409–433, 2006.DOI
  34. [34]M. Degli Esposti. Code and data for “Random-walk scaling without Bernoulli statistics: transport, persistence and resonances in the triangle map”. Zenodo, version v1.0, 2026. https://doi.org/10.5281/zenodo.23623347.
  35. [35]G. M. Molchan. Maximum of a fractional Brownian motion: probabilities of small values. Communications in Mathematical Physics, 205:97–111, 1999.DOI
  36. [36]R. Artuso, G. Cristadoro, M. Degli Esposti, and G. Knight. Sparre-Andersen theorem with spatiotemporal correlations. Physical Review E, 89:052111, 2014.DOI

Paper details

vlnc:W_ts9tfvu3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 13

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents