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 given by the two halves of the torus, whose partial sums 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 eigenfunctions, that the symbol is not an 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 and times the Bernoulli value, and for a benchmark pair, with , the excess is produced by the order of the bounded inter-event intervals, not by their distribution. The first-passage exponent is with a non-Bernoulli amplitude, and worst-case Birkhoff averages decay as for a uniquely ergodic system. For the Green-Kubo sum cancels, , yet keeps growing slowly, sign changes are rare and the correlations of Fourier modes decay as . We identify the zero-sum words, which become pure twists at , as resonances that trap orbits with a computable law, set the crossover time at small , 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 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 [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
where for and for . In a previous study [10] we found evidence of ergodic and mixing behavior in a generic, non-resonant case, where the symbolic sums
scale as , and much slower growth, slow correlation decay and strong arithmetic sensitivity for . 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 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 (, Green–Kubo sums, the spectrum of ); 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 , 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 . The generic triangle map is normally diffusive, with Gaussian fluctuations and first-passage exponent , but its diffusion coefficient is not that of a coin: for the benchmark pair , three times the Bernoulli value, and varies between and away from the edges of the parameter square, with deep minima at rational . For the benchmark, is set by the order of the inter-event intervals, which multiplies the value implied by their distribution by (– for other pairs), and the persistence amplitude is not Bernoulli. For the Green–Kubo sum cancels and , while keeps growing slowly, almost independently of , sign changes of are rare and Fourier modes decorrelate as . A small restores diffusion after a time of order .
Three analytical observations frame and partly explain the numerics: the Koopman operator has no non-constant eigenfunctions (Proposition 1), so the symbol is not an coboundary; runs of equal symbols are uniformly bounded (Lemma 1); and at every zero-sum word is a pure twist on its cylinder, carrying segments of periodic points that trap nearby orbits for times with a law and an exactly computable coefficient. These resonances fix the crossover time , 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 , 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 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 . One iteration is the composition of the discontinuous kick , the translation and the parabolic shear . Away from the tangent map is
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 , and the shear converts this momentum difference into a growing separation in 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 . (c) The shear converts the momentum difference into separation in , and further cuts generate new symbolic splittings.
The symbolic sum is a transport observable. On the lift, iteration gives
with , so that . Equation (4) also gives a non-accumulating integrator: and are integers, and is recomputed at every step in double-double arithmetic, with an error of order 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 is stationary with . With , and , stationarity gives
so is the Cesàro mean of the Green–Kubo partial sums and differs from at finite . If , both converge to the diffusion coefficient
Note the normalization , without the factor 2 common in physics; the momentum diffusion coefficient in the convention is . The two-sided spectral density satisfies
so when is continuous at zero. For the Bernoulli benchmark for , and .
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 eigenfunctions and no 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 alone were exhibited for rational .
Proposition 1. If is irrational and is arbitrary, every eigenfunction of the Koopman operator , for any eigenvalue including , is constant.
Proof. Write , and suppose . On the two continuity domains the derivative of is , and the chain rule gives, almost everywhere, and . Thus and . Since is unitary on , for all , so and . The eigenvalue equation on the two half-tori gives , so is invariant under rotation by and hence constant.
Since , its global weak derivatives coincide with the branchwise ones, and the jumps of do not contribute. The proposition does not prove weak mixing, whose spectral criterion concerns all 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 a coboundary if almost everywhere for a measurable , so that .
Corollary 1. (i) If is a coboundary, then is an eigenvalue of , with eigenfunction ; in particular is not ergodic. (ii) For irrational and any , is not a coboundary with .
Proof. Since , satisfies . If , then , contradicting Proposition 1.
By a classical criterion for ergodic [18 Theorem 11.8], is a measurable coboundary if and only if the laws of are tight; for transfer functions, if and only if is bounded [14]. Hence, if is ergodic, cannot stay bounded in probability.
Since , , so and
The time-alternating symbol is therefore a function of the path of the walk alone: a step from level contributes times its direction. This identity is used in § 5.2.
Bounded runs and the word map
On the half-torus , , the map is the skew shift
Lemma 1. If is irrational, there is such that every orbit of contains no run of more than consecutive symbols equal to .
Proof. For irrational the skew shift is minimal [19]: , and for the sequence is a polynomial in with irrational leading coefficient, hence equidistributed by Weyl’s criterion. Let be the interior of . The open sets , , cover , and by compactness finitely many, , already do. A run of symbols has for , so .
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 () shows. It excludes broad (infinite-variance) waiting-time distributions, and it makes the interval statistics of § 4.2 exactly computable.
The bound is with a single function , the longest run of in a half-torus, so at but not in general (for the generic benchmark , ). is small for most and diverges, as , only at the rationals for which has a periodic orbit inside one half-torus, such as , , , $1/2 (Figure S1 in the supplementary material).
Figure S1. The run bound of Lemma 1 (Experiment 21). (a) Longest run of the skew shift in a half-torus, computed exactly for 20 000 values of ; dotted lines mark rationals, dashed the law near . Dots: of the generic benchmark and for with . (b) with in the plane; the dot is the generic benchmark , the squares the two cases.
The argument extends to words. For with partial sums , the lift gives on the cylinder
with . If is irrational, repetitions of are uniformly bounded. At , however, every zero-sum word () has : on the block map is a pure twist, and the lines carry sets of period- points. These are the horizontal segments of periodic points of Ref. [10]; for they lie at and , with lengths and . A point at momentum distance from a segment of length is translated along it by per block, i.e. by per step, and repeats until it leaves the segment, for a time 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 periodically (about copies of for a stretch of steps), is therefore at least steps long if , i.e. in a strip of area along the segment, and likewise along its images under , which are translates of it. Since Lebesgue measure is invariant, the probability that a given time belongs to a repetition of of length at least (the fraction of time averaged over Lebesgue-distributed initial conditions) is the area of these strips,
Figure S2. The period-two twist at , . (a) The cylinder (grey), the period-two segments (solid) and their images under one step (dotted). (b) at even for orbits at momentum distance from the segment (shaded): they drift by per block and repeat until they leave it; the stay length is compared with the remaining length of the segment divided by . (c) : the momentum drift bends the motion into a parabola; the longest stay fits it into the wider segment.
the sum running over the segments of periodic points of itself, the maximal intervals of on the resonant lines , . The images of these segments are the segments of the cyclic rotations of , so is the same for all rotations, and totals over words count each cyclic class once. The law holds for ; at finite the ends of the segments, where the strip meets the boundary of , give corrections (§ 5.2). For , . For small the drift makes grow linearly in time, so that after steps the position along the segment is ; the longest repetition is the one whose parabola just fits in the longer segment, of length , which gives the maximal length . The lift also yields exact word probabilities; for instance, at , for every (supplementary material, Sec. S1).
Generic parameters and a correction
We call generic if , , are linearly independent over . Then every is irrational, every word has uniformly bounded repetitions, and the map has no periodic orbits, since a period- orbit with word requires . All generic pairs used below satisfy this condition, by the transcendence of and the linear independence over of square roots of distinct square-free integers, except : the rational independence of , , would follow from the algebraic independence of and , which is open.
Correction to Ref. [10]. There the generic case was defined by irrational with , and Lemma 1 states that the map then has no periodic orbits. The necessary condition for a period- orbit excludes only for ; in general it gives , so the hypothesis must be strengthened to rational independence. For example, , satisfy the original condition, yet has , and the line carries a segment of length 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 -extensions of rotations by step cocycles [21]; both arguments need a factor. Here the lifted momentum is a rotation only modulo the dense subgroup : 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 (supplementary material, Sec. S1), and the renormalization schemes for kicked oscillators [23], of which the triangle map is the parabolic member , 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 , of our earlier work.
Diffusion coefficient
For 4096 orbits followed to , the local log-log slopes of and of the mean running maximum are 0.996 and 0.502, and at the largest checkpoint; the amplitude matches the Brownian value . Correlations from 64 trajectories of length agree with the values computed without sampling from the polygonal cylinders (App. § 11) for : the mean squared deviation in units of the bootstrap error is 0.88 per lag, as expected for purely statistical scatter. One has exactly for all parameters (for fixed , the uniform makes uniform, and is the image of under a two-to-one map that preserves Lebesgue measure), but the next correlations are large, , (supplementary material, Figure S3), and reaches 2.04 at and 2.66 at (Figure 3, middle left). The rest of the excess is built by a band of weaker, sign-oscillating correlations at , after which is flat: , . By Eq. (5), with , which explains the slow approach of the finite-time variance. A non-oscillating tail , as reported for the smooth observable [12], is not seen ( over ). The averaged periodogram is flat below at (bottom left). All estimators agree (Figure 3, left):
Figure 3. Two transport regimes. Left: generic , . Right: , . Top: (2048 orbits). Middle: Green–Kubo partial sums (64 trajectories of length , generic; 1024 of length , ). Bottom: log-binned spectral density (32 trajectories of length ). Bands are one standard error; dashed lines mark , 0 and ; dotted lines the Bernoulli value 1. All data from the exact-phase integrator.
Figure S3. Symbol correlations, generic regime (, ; 64 exact-phase trajectories of length ). (a) for : measured (bars) and computed without sampling from the polygonal cylinders (circles); . (b) Deviations in units of the standard error. (c) RMS of in logarithmic bins of with the noise level subtracted (solid) and the noise level (dashed); grey: the band .
The value is not universal: away from the edges of the parameter square ranges from 0.2 to 14 (median 2.3), with dark lenses at rational and bright bands near low-order resonances (supplementary material, Figure S4); four other generic pairs give – (supplementary material, § 14), and the pair A of § 7 gives , 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). at , 1024 orbits per point, on a grid; computed for and reflected through (dotted line), by the exact symmetry . (a) The fundamental square , with the benchmark pair and the pairs A, B of Sec. 7 folded into it. (b) Zoom near the skew-shift limit ; labels mark rational .
The fluctuations are Gaussian on the explored scales: the Kolmogorov distance from the standard Gaussian of the normalized sum is 0.013–0.017 for , at the level of sampling noise, the skewness decays as , and the correlation between and is below 0.02 for . 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 be the intervals between consecutive occurrences of , observed from an occurrence at time 0, and let be the time of the -th return. By Kac’s lemma the mean interval is 2, and since contains symbols and symbols , exactly. Hence with , , whereas an independent renewal process with the same interval law has [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 | |||
| generic (benchmark pair) | 17/7 | 8 | 1.7085 | 0.854 |
| , | 11/11 | 12 | 2.1493 | 1.075 |
| , | 12/12 | 13 | 2.1254 | 1.063 |
| Bernoulli | 2 | 1 |
Table 1.
If the intervals were independent with the same law, the walk would have : 0.85 in the generic regime and 1.07 at , 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 [Figure 2(c)]. For the benchmark pair they are positive (long intervals tend to follow long ones) and grows from 0.85 to at , a factor of about 3.7, consistent with the more precise obtained from the symbol sum; 80% of the increase occurs within twenty events. At they alternate in sign, starting with (a long interval is followed by a short one that compensates it), and falls from 1.07 to 0.05–0.15 within about ten events; the remainder, at , 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 , Eq. (13): generic (solid) and (dashed) with Wilson 68% bands, and for . (b) Effective persistence exponent over factor-four windows; the horizontal line is the Sparre Andersen value . (c) Event-time Green–Kubo sums of the intervals between events; dotted lines mark the renewal values computed exactly from the interval law.
Persistence
Let be the probability that the walk has not entered the positive (negative) half-line up to time , i.e. that for all . For a Bernoulli walk (independent symmetric steps , hence the subscript) the two coincide and the Sparre Andersen theory [15, 25] gives
the lattice counterpart of the universal law for continuous symmetric steps. The involution preserves Lebesgue measure, reverses and conjugates to , so exactly. For the generic parameters ( orbits to ) the effective exponent is compatible with for , but the amplitudes are not Bernoulli and differ between the sides, and [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 [25]; fed with the measured (and for ), it reproduces simulated Bernoulli walks within 1%, but for the map the measured and exceed its prediction by factors 1.38 and 1.77 at : the excess persistence comes from temporal correlations, not from the distribution of .
At persistence is anomalously strong: with orbits to , reaches 28, and decreases from 0.29 at to for , identically for and ; 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 double-precision orbits are not adequate: their symbol sequences differ from the exact ones from , and for an with large partial quotients the double-precision variance is three times too large at (supplementary material, § 15). We followed 2048 orbits to for and for two values with large partial quotients, and (continued-fraction expansion), with replicate ensembles and one run to . The variance curves are nearly independent of : , 33–38, 65–73 and 95–103 at , , and (Figure 3, right), with no feature at the convergent denominators. The local exponent decreases from 0.3 to 0.1 without a plateau, and increases, so the growth is slower than any fixed accessible power and faster than . Beyond this the law is not determined: , , and fit equally well, with model-dependent parameters (supplementary material, § 15).
The Green–Kubo sums explain the suppression. From 1024 trajectories of length , falls from 1 to 0.45 at and 0.09 at , then oscillates around zero with decreasing amplitude (, 0.022, at , , , errors 0.002–0.004): the positive instantaneous variance is almost exactly cancelled by integrated anticorrelations, . The spectral density vanishes at low frequency with local exponent close to one and arithmetic oscillations; scale by scale matches the local variance exponent, as Eq. (7) requires, and a log-corrected linear spectrum , which gives , is consistent with both drifts without being singled out. Exact Markov approximations of order , built without sampling on the return map to the starts of runs (supplementary material, Sec. S4), show the same structure on short scales (): at their diffusion coefficient falls as (), against in the generic case, and their transfer function, the solution of the Poisson equation, grows only logarithmically in instead of as , so that at every finite order is close to a coboundary, as in Conjecture 2.
Thus, on the explored scales (, , ), , and , while grows without saturation. If the growth persists, is not an coboundary (§ 3: an transfer function would keep bounded), although its Green–Kubo sum vanishes as for a coboundary. The bulk of the distribution widens (interquartile range 4 12 from to ), and has a stable, -independent, non-Gaussian law with kurtosis .
Parity and zero-sum-word resonances
The walk moves on the integer lattice, and each step crosses one edge of it ( a level, not a symbol). By the parity identity (8), the alternating sum adds for every crossing of the edge , in either direction: for edges with an even lower end and for edges with an odd one. A walk that oscillates over few edges, as in a repetition of , therefore has an almost ballistic , whereas a walk spreading over many levels crosses both kinds equally often. Since is an ergodic average for the unitary operator , the mean ergodic theorem gives in , the projection of on the eigenspace ; equivalently, , the mass at the frequency , that of the period-two oscillation , of the spectral measure of , defined by . A positive limit would give an eigenfunction with , hence a non-constant -invariant function: would not be ergodic, the same obstruction that a coboundary would create (Corollary 1). Conjecture 2 asserts that neither occurs. In the generic case is constant. At , decreases slowly, from 0.051 to 0.0091 () between and , as , and has a peak at of height against a median of 0.8, with weaker peaks at , , , . At the momentum stays on the lattice , and by Eq. (8) the sign of each burst is fixed by the level on which it occurs, so successive bursts add coherently at .
These resonances are seen directly in the orbits. Let be the fraction of time spent inside repetitions, at least steps long, of zero-sum words of length ; Eq. (11) predicts , independent of . At each orbit keeps its momentum lattice and samples a single distance from each segment, so we use , which lets the momentum drift while cutting repetitions only beyond . Over , for and , we find –, –, –, – against the exact , , , ; decreases with towards , the excess at being a finite- correction [the same law for the deformed map is shown in Figure 6(b)]. For the longest repetition agrees with the cutoff within 1–5% [supplementary material, Figure S2]. The do not decrease with : resonant twists exist at all lengths.
Figure 6. Asymmetric deformation, . (a) : for generic , for (zero mean drift), and for (two seeds of orbits; interquartile scale from ). (b) , : for stays of period two and four, with the exact and (dashed). (c) , : split into the parts accumulated outside and inside the stays; dashed, the slope of Eq. (16) anchored at .
Crossover at small
For and and seven values (1024 orbits to each), the variance follows the curve up to a crossover time and then grows linearly. The ratio is approximately a function of (scatter , not monotone in , larger than the statistical errors), with crossover time –. The relevant time is not , at which the drift is of order one, but the time at which the quadratic phase of Eq. (4) is. It is the lifetime of the resonances: for the block map of a zero-sum word acquires the drift , and the longest repetition is . At the smallest ,
with ratio of the two sides 1.00–1.07 () and 0.8–1.25 () for : the variance accumulated by the dynamics up to the crossover is then spread diffusively.
Correlations of Fourier modes and mixing
Mixing is equivalent to the decay of for all zero-mean , and it suffices to check the Fourier modes . Their correlations were measured in double precision in [12, 10], with decay close to (generic) and (). We repeated the measurement with the exact lift, recomputing and at every step, for the modes , , , , , with – steps per parameter pair (the mode coincides with because ).
In the generic case the correlations decay quickly (Figure 4): the rms is for and for , with effective exponents 1.4–2, and the signal is below the noise floor beyond ; this is consistent with over about one decade. At all modes decay slowly but steadily up to , for and , with effective exponent 0.19–0.21 for , , , and no plateau. The correlations are real: the symmetry conjugates the map with to the map with , so at every is real. (For any the modes are real as well, because the map is invariant with the -preserving reversor ; the other modes are complex for They carry the parity of the walk: the -modes nearly vanish at odd lags, and the -modes alternate in sign [Figure 4(c)].
Figure 4. Correlations of Fourier modes (exact lift). (a), (b) Root-mean-square over logarithmic lag bins, after subtraction of the statistical noise, for the modes and , shown where the signal exceeds three standard errors (generic noise floor ). Grey lines: (dashed) and (dash-dotted). (c) Parity structure at , : real part of at even lags and its modulus at odd lags, and at even lags compared with at odd lags.
The -modes have a direct meaning for the cocycle: at ,
is the characteristic function of the increments of . If were a coboundary and mixing, it would tend to , generically non-zero. The steady decay of , from 0.39 at to 0.094 at , is evidence against tightness. Finally, worst-case Birkhoff averages of the Fourier modes over random, near-cut and adversarially optimized initial conditions decay as for two generic pairs, like the uniquely ergodic skew shift, while the same search recovers the periodic segments at and the period-four orbits of the counterexample above (supplementary material, Sec. S5).
An asymmetric deformation: resonances made visible
At 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 by
in the same map interpolates between the triangle map (, up to and ) and the skew shift (); the lift (4) and the exact integrator are unchanged. Now , and we consider the centered walk , with the number of visits to , and the Fano factor of these counts, (Poisson: ; Bernoulli: ).
Generic . For small the kicks are the returns of the quadratic sequence to a short interval, whose statistics are Poissonian for generic coefficients [26, 27]. We consider two generic pairs, and . Numerically stays within 0.96–1.12 (A) and 0.96–1.22 (B) for [Fig. 5(a),(b)], and the -modes follow the Poisson prediction for (not shown). At larger the Fano factor is strongly non-monotone and reproducible between independent ensembles (Figure 5): for pair A it peaks at at , and the diffusion coefficient peaks at 4.09 at [panel (c)]. The triangle map with the same parameters () is thus far from the most diffusive member of the family: for pair A its (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 the variance saturates, from to .
Figure 5. Fano factor of the deformed map at generic , 250 values of , , two independent ensembles of 32 768 orbits per pair. (a), (b) for pairs A and B: mean of the two ensembles (band: their spread; dotted: at ); dashed, the Poisson value 1; dash-dotted, the Bernoulli value . (c) .
: ballistic stays and a logarithmic term. At and the variance per step grows. At ( orbits to , two seeds) the raw variance is dominated by a few orbits (one orbit carries up to 18% of , Hill index of about 2), while the interquartile scale is reproducible and grows linearly in [Fig. 6(a)]. Each extreme orbit sits, for up to steps, on a periodic stretch of period two () or four (), i.e. on a zero-sum-word resonance. During such a stay does not move while the centring term does, so changes by per step: a ballistic excursion. At the same stays do not displace the walk.
The stays obey the law (11), with computed with the threshold and no free parameter [Figure 6(b)]: in 20 000 orbits of length (),
| (theory / measured) | 0.0800 / 0.080 | 0.3200 / 0.318–0.320 | 0.4216 / 0.424 |
| (theory / measured) | 0.0800 / 0.078–0.079 | 0.3200 / 0.322 | 0.5000 / 0.499–0.501 |
Table 2.
The number of stays of length at least per step is then , and a stay of length adds to : the second moment is marginal (hence the Hill index above), and
For and the enumeration finds only , with , so . Above longer words appear, and an exact enumeration of the primitive zero-sum words up to length 60 gives the converged values () at and () at ; for the sum is not converged at length 60 (supplementary material, Sec. S6).
To test Eq. (16) without the heavy-tailed variance, we split along each orbit, being the displacement accumulated inside the stays. The centered is flat over sixteen octaves (slopes in of and at ; errors are bootstrap over orbits here and below), so the whole logarithmic growth is in the stays [Fig. 6(c)], whose slope for is and against and from Eq. (16). At larger the stays that matter have length comparable to and are too rare to sample. With stays of all even periods up to 14, is flat also at and , but not at , where longer words matter; the stay slopes lie systematically below Eq. (16) ( against at ), so Eq. (16) is tested only to beyond .
Finally, the deformation moves the zero mean drift to . A stay keeps on the slope of its word; if this equals the mean slope $1-2\varepsilon$, the stay does not displace , otherwise it does. At on the zero-drift line the diffusion is suppressed falls from at to at , Fig. 6(a)]$, whereas for irrational , where no word has the right slope, it does not seem to be [a single run gave on the line against at a distance from it]. The triangle map at 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 the same type of correlations cancels the Green–Kubo sum, while the walk keeps spreading slowly. The zero-sum-word resonances, pure twists at , control the quantities we can compute: the trapping law, the crossover time 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 , , are rationally independent, is uniquely ergodic and mixing, the Green–Kubo series converges to , and, for Lebesgue-distributed initial conditions, converges in law to the centered Gaussian of variance .
The evidence is the absence of periodic orbits and the bound on all word repetitions under rational independence (Sec. 3), the 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 , running maximum with the Brownian amplitude and persistence exponent , which also support the functional version (convergence of 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 ). For irrational and , is ergodic.
Consequently, by Corollary 1, is not a measurable coboundary, the laws of are not tight, and the spectral measure of has no atom at , although .
The symbol would then sit strictly between the two familiar classes of observables. A coboundary has , whose laws are tight, and a spectral density vanishing at ; an observable with has . At , would share with the first class and unbounded fluctuations with the second: spreads without limit, but more slowly than . The data are consistent: the bulk of the law of widens, shows no plateau up to , the characteristic function (15) decays, and decreases, although slowly (, the same effective exponent as the Fourier correlations), and the transfer functions of the exact Markov approximations stay small, growing only as (§ 5). Whether has a non-Gaussian limit, as its stable shape suggests, is a further question.
Conjecture 3 (resonance renormalization). At the growth law of is determined by the hierarchy of zero-sum-word twists, through a renormalization relating the resonances of words of length to those of longer words.
Every zero-sum word is a twist at , resonances exist at all lengths, a small destroys them on the time scale at which the variance crosses over to diffusion, and the resonant momenta exist for every , which would explain the insensitivity to the arithmetic of . In the deformed map the hierarchy provably sets the logarithmic coefficient through ; at 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 , and , 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 on and the origin of the maxima of ; the convergence of for ; and a theory of persistence at , 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 , further data for the generic regime (correlations, other parameter pairs and the map of ), the 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 and of Eq. (4) are updated exactly; is recomputed at every step from , (double precision) and , (double-double, ), with an absolute error of order that does not accumulate, so can be wrong only if falls within this distance of a cut. The construction requires , , i.e. . Symbol sequences coincide with 60-digit mpmath orbits over steps, at and for generic parameters, and the cost is 1.3 times that of double-precision iteration.
Polygonal cylinders. For a fixed word, is affine in , so each cylinder is a finite union of convex polygons. The partition into cylinders of length is built by clipping with the lines , (pieces below in area dropped; lost mass ). Then , with the area of , is a finite sum of polygon areas, and stationarity gives . The number of pieces grows as , of the order of the cubic growth reported in Ref. [10], which limits exact computations to (to for the survival probability , for which only the cylinders with for all are refined). Word probabilities, , and obtained in this way agree with Monte Carlo estimates. The trap coefficients are computed exactly for all primitive zero-sum words of length by a depth-first search over intervals of on each resonant line .
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 to with 50% overlap, binned logarithmically and combined across segment lengths; the lowest frequency is .
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 and 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.
| Diagnostic | What it measures | What it cannot establish |
| , , , | second-order structure; and finite-size corrections on the explored scales | asymptotic limits; any non-Gaussian or higher-order property |
| CLT diagnostics | one-time distribution of , two-time increments | a functional limit theorem |
| First passages | persistence, a multi-time property | a limiting exponent beyond the explored window |
| Inter-event intervals | distribution versus order of the waiting times | the asymptotics of the order effect |
| Exact cylinders, Markov hierarchy | word probabilities without sampling; finite-order models | the asymptotic behavior of the map |
| Fourier modes, worst-case averages | decay of correlations; uniformity of Birkhoff averages | mixing 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 , , where is the longest run of the skew shift inside a half-torus; it is the same for both halves, since commutes with and exchanges them, and by the involution . Hence , and at , but not in general. Figure S1 shows , computed exactly from the polygonal run cylinders for 20 000 values of . It is piecewise constant and small (6–7 on most of ) and diverges only at the rationals for which has a periodic orbit contained in one half-torus: among those of small denominator, , , and , but not , , or . Near them (local exponents 0.41–0.49 on ), which for reflects . In the plane the resonances form the two families of diagonals . For the generic benchmark, is close to and , whereas gives . Irrationality of with does not make both irrational: gives .
The period-two twist and its traps. For at , a point with and (mod 1) is mapped by to : on the cylinder the block map is a pure twist, the identity on the lines . Imposing the symbols gives the two segments of Figure S2, , and , , together with their images under one step, which carry the rotation . At maximum distance the block map is the translation , so an orbit slides along the segment and repeats for the time it needs to cross it, steps [Figure S2: 1415, 567, 944 steps against 1414, 566, 943]. For the block map adds to the momentum, so grows linearly in time and the motion along the segment becomes a parabola; the longest repetition is the one whose parabola, of depth after steps, just fits in the longer segment, of length , i.e. [Figure S2: 483 steps against 484 at ].
at . With and (independent and uniform),
where is an arc of length . Its Fourier coefficients in are proportional to
for every integer , so
for every . The identity was first found numerically, to , in the exact cylinder computations of Sec. S4; it fails for .
Word complexity (Experiment 20). From the exact cylinder partition we count, for , the distinct admissible words (a cylinder made of several polygons is counted once), the number of polygons, and the admissible zero-sum words , i.e. those with (Table S2). All words are admissible up to (generic) and (). The first difference satisfies at , in agreement with : the growth is cubic, with coefficient (generic) and – (, still rising slowly). The second difference gives fluctuating between and $0.93, and cubic fits on give –, so neither the cubic coefficient nor the lower-order terms are determined. The number of polygons exceeds by less than : nearly every cylinder is connected. The topological entropy vanishes. At the zero-sum words are a constant fraction of all words, , against and a falling fraction (4% at ) in the generic case, where the spread of alone would give . This is consistent with the zero-sum twists of § 3.2, but we have no derivation.
| case | polygons | ||||
| generic (, ) | 881 378 | 881 620 | 0.662 | 34 356 | 0.039 |
| , | 938 742 | 938 942 | 0.705 | 215 138 | 0.229 |
| , | 943 784 | 943 836 | 0.709 | 217 263 | 0.230 |
Table S2. Word complexity at (Experiment 20). : distinct words; : distinct words with .
Generic regime: further data
Long runs. With 4096 initial conditions to the late-time exponents are and , with and at the largest checkpoint; the run to 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 , whereas long-trajectory correlations determine for ten to forty times more precisely. An earlier truncated Green–Kubo value 3.425 at , from eight double-precision trajectories, is not confirmed (). Figure S3 shows the individual correlations: the measured coincide with the exact cylinder values for , with deviations distributed as statistical noise, and the RMS correlation falls by two decades across the band before reaching the noise level .
Finite-size corrections. The first moment converges on the accessible scales (–100 for ), so reaches to 1% only for . A non-oscillating tail would give , and a cusp . A fit restricted to suggests –1, but over the data give . The decay reported for the smooth observable [12] therefore does not carry over to on these scales. The periodogram decreases roughly linearly in for , the image of the correlation band at –, and is flat below.
Other generic pairs (Experiment 12). For , , and , , Green–Kubo sums from 32 trajectories of length give , 2.94, 2.15 and 2.33 (errors ; the ensemble variance at agrees). The ratio of to the renewal value is 2.4–3.7.
Map of (Experiment 24). Besides , () and (), the identity gives , hence : the fundamental domain is the triangle , . Figure S4 shows at on a 200200 grid with irrational offsets (1024 orbits per point, statistical error ); changes by between and , except at and near (0.4% of the points change by more than 50%). Away from the edges (, ) 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 and near , is still growing at (up to 53 at ) and is not a converged value. Dark lenses, with well below 1, sit at rational with small and open from the skew-shift limit [panel (b)]: there zero-sum words whose length is a multiple of have , the mechanism that gives at . Bright bands follow the low-order resonances with (e.g. and ) near the corners of the square but leave them in the interior.
Distribution of . The skewness of is 0.65–0.77 for , 0.41 at and 0.19 at , decaying roughly as and compatible with zero from . The involution maps for to for , so the sign of the skewness is reversed by ; for the benchmark is slightly more often negative, consistent with . The correlation between and is at .
Bernoulli control. The same finite-sample pipeline applied to independent unbiased Bernoulli sequences, on the same scales, gives , , and a low-frequency spectrum with local exponent , consistent with and . With few trajectories the Green–Kubo partial sum is more sensitive to accumulated noise than the spectral estimate.
The 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 – at (at in the generic regime). For most the variances still agree within the sampling noise, and a double-precision run with 4096 orbits gives and , consistent with the exact data; for , however, the double-precision variance is three times too large for .
Growth law. The six values of and the replicate ensembles give the same curves within errors (about 5% at the largest scale); shows a short transient at –, reflecting , and no feature is visible at the convergent denominators 33102 and 74646 of . The models , , and fit all ensembles equally well for , with –, – and – depending on the model and on . They are numerically degenerate over seven to eight decades (), the coefficient of is not a robust quantity, and the offset – differs between independent ensembles. The logarithmic growth of the extremes of single orbits reported for quadratic irrational 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 , or a small power .
Spectrum. Binned values of scatter by about , with decade averages mostly between 0 and 0.2; the drift 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 are of arithmetic origin rather than noise; .
Parity and the peak. For , decreases from 0.034 to 0.0061 between and , and the peak of at has height . There are no peaks at or at irrational frequencies. Per orbit, is nearly constant over decades, with abrupt jumps. At rational frequencies other than the phases of successive bursts are not locked, which is why the peak dominates although the trap coefficients of the length-four words are larger. The trap law (11) for agrees with the exact coefficients within 5–20%; at itself the estimates scatter more than at because each orbit’s momentum lattice is frozen.
Persistence by momentum class (Experiment 10). At , , so labels families of orbits that evolve independently on . Sampling on 1024 bins (128 orbits each, ): the growth of is shared by all classes (at the 10% of classes with the largest conditional variance carry 15% of the total), whereas long survival is carried by thin, -independent bands of : at , 77% of the classes have no survivor, against 28% expected for a -independent survival. The bands are selected early, ranging from 0.003 outside to 0.2–0.3 inside, and this profile is reproduced exactly by the polygonal cylinders pruned when becomes positive. Some band edges coincide with the resonances 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 is 0.25–0.35 in almost every active class, corresponding to .
Persistence: null models. For fractional Brownian motion the persistence exponent is [35]; the local variance exponents –0.2 would give , far from the measured 0.2. The Sparre Andersen identity fed with the measured marginals fails at 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 .
An exact Markov hierarchy
Let be the set of points where a run of begins; exactly and the mean return time is 4. By Lemma 1 the first-return map to is piecewise affine with finitely many branches, one for each cycle (67 generic, 58 for , , 63 for , ); on each branch , . The increment of over a cycle is , and . Cylinders of of length are obtained by intersecting affine images of branches with the branch partition, and define the order- Markov chain on cylinders of length , in the spirit of the Markov approximations of Ref. [10] but without sampling. Its asymptotic variance is , where solves the Poisson equation (sparse direct solver), and . For ( states) in the generic regime, while for () for both values of . Since the chain retains a memory of about steps, should be compared with –3.7 at . The transfer function has root-mean-square size growing like in the generic regime but only logarithmically at ( at ), 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 for a dense set of continuous (here the Fourier modes, with the usual care for the discontinuity of ), 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 – of the cuts, and by a greedy adversarial search at (64 parents 63 children, radii –), whose worst points are then followed to (Experiment 15, exact lift).
For both generic pairs the worst case decays as (exponent 0.49 over three decades), like the uniquely ergodic skew shift, with a ratio max/median –6 that is stable in (it would be for Gaussian fluctuations): the worst case is the extreme of a fixed, somewhat heavy-tailed distribution of diffusive fluctuations. Adversarially selected points are transients and become typical by ; points next to the cuts behave like random ones. The method finds the known obstructions: at it recovers both period-two segments, with average and with , constant to ; for the counterexample pair it finds the period-four orbits, with exact average 0.42487. Typical orbits converge slowly (median ), consistent with the correlations. This is numerical evidence for unique ergodicity under rational independence, not a proof: a finite search cannot exclude very small exceptional sets.
| generic , random max / median | 0.179 / 0.035 | 0.049 / 0.009 | 0.012 / 0.0023 | 0.0062 / 0.0011 |
| generic , near cuts | 0.205 | 0.046 | 0.011 | 0.0051 |
| generic , adversarial at | 0.213 | 0.019 | 0.0073 | 0.0026 |
| generic , random max | 0.274 | 0.068 | 0.011 | 0.0050 |
| generic , adversarial at | 0.293 | 0.042 | 0.0080 | 0.0035 |
| skew shift , random max | 0.060 | 0.015 | 0.0035 | 0.0016 |
| , random max / median | 0.795 / 0.097 | 0.606 / 0.075 | 0.587 / 0.061 | 0.451 / 0.055 |
| , adversarial at | 0.796 | 0.796 | 0.796 | 0.796 |
| counterexample , adversarial | 0.425 | 0.425 | 0.425 | 0.425 |
Table S3. Largest found ().
The asymmetric deformation: further data
Generic . The correlation at lag one vanishes for every (since is uniform given ), whereas at lag two it vanishes only at . For , – (pair A) and – (pair B), with statistical error ; at – the tail of the -modes already deviates from the Poisson prediction, with a momentum relaxation time . 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; at is smaller by 1.5% on average (1.8% rms), and no extrapolation in is attempted. Pair B has maxima of the smoothed near , 0.20, 0.34, 0.41, up to 4.80 at and (Experiment 12: 2.33). Two zoomed windows, (400 points) and (401 points), each with two independent ensembles of 32 768 orbits at , show a smooth profile: the seeds agree within statistics (rms -difference 1.00 and 1.02), polynomials of degree 8–12 fit with 1.04–1.17, and 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 at noise 0.02.
Tails at , . In two seeds of orbits to , one orbit carries 3–18% of , the kurtosis is –, the Hill index of is 2.0–2.3 from to , and the raw at differs between seeds (3.51 and 5.17). The interquartile scale , , equals 1.34, 1.89, 2.26, 2.66 at in both seeds, with slope in ; it is not the variance and is not compared with . The extreme orbits have uniform , are ordinary at earlier times (Spearman correlation of between and : 0.002), and are all negative. The same run at has light tails (kurtosis 4.7–5.3, Hill index 5–9) and , against 95.0 from the fit . In Experiment 17 the stays of lengths 128 to contribute 1.41 to at , against 1.44 predicted, and the law (11) holds within 1–2% up to .
The total trap coefficient (Experiment 19). Table S4 gives for from all primitive zero-sum words of length . holds exactly for (only ); the first longer word, , appears at , then at 0.229. saturates at for . The identity holds for for five values of (, , , , ), with an -dependent range ( deviates at 0.1, the others at 0.15–0.2). For the partial sums still grow at and is only a lower bound; whether is finite there, and how it behaves as , is open.
| converged | |||||
| 0.100 | 0.1600 | 0.1600 | 0.1600 | 0.1024 | yes |
| 0.200 | 0.6400 | 0.6400 | 0.6400 | 0.2304 | yes |
| 0.225 | 0.7855 | 0.7855 | 0.8100 | 0.2376 | yes |
| 0.250 | 0.9760 | 0.9760 | 1.0000 | 0.2440 | yes |
| 0.275 | 1.2579 | 1.2579 | 1.2100 | 0.2547 | yes |
| 0.300 | 1.6889 | 1.6889 | 1.4400 | 0.2702 | yes |
| 0.325 | 2.3836 | 2.4755 | 1.6900 | no | |
| 0.350 | 3.0907 | 3.5480 | 1.9600 | no | |
| 0.400 | 4.3435 | 5.5344 | 2.5600 | no | |
| 0.450 | 6.5154 | 11.4758 | 3.2400 | no |
Table S4. from zero-sum words of length and , and ; .
Stay decomposition (Experiment 18). Stays are maximal stretches of period two (, at least 66 steps) or four (at least 132 steps, without a period-two sub-stretch), 20 000 orbits to per . Centered variances are essential, since outside the stays the mean step is not zero. goes 0.58 → 0.61 (), 1.16 → 1.28 (0.2), 1.17 → 1.33 (0.25). At larger the stay slope is lower than predicted (0.099 ± 0.051 at on ), as expected: in orbits fewer than one stay with is expected at (≃ 0.02 at ), so the sample variance underestimates its mean; the data support Eq. (16) for only, and is not determined. With stays of all even periods up to 14, at , 0.3, 0.35, the slopes of 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 is 0.15 ± 0.05 on , about below . Hence Eq. (16) is not contradicted, but it is tested to about 40% only for ; at the raw variance is too noisy for a slope.
Zero drift and small . On the line at , the slope of in goes from −0.046 (–) to −0.0023 ± 0.0006 (–); whether the limit is finite, ≃ 0.23, is not decided. At the same line gives 1.47 against 1.46 at shifted by . For and at (2048 orbits, ) the variance is flat in , the trap cutoff being .
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 equal square cells and let be the number of distinct cells visited by ; we estimate from 4096 independent uniform initial conditions, counting coverage per orbit before averaging. For independent uniform draws,
and we use the exact finite- expression. We use the parameters of the original study, which in our unit-torus convention read , , (the change of coordinates , from the torus halves both parameters and preserves the partition for even ), and compare them with . All coverage orbits are in double precision; the data, to , are therefore indicative only.
Generic coverage. For and the coverage nearly coincides with the independent reference (Figure S5), with a small resolved deficit at (standard errors ; a repetition at with another seed gives ). decreases from to , so a simple 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 , with the exponential reference. (b) Signed deviation from the exact finite- independent reference, ; shading, two pointwise standard errors.
Setting at the same changes coverage drastically (Figure S6): at , , , for , against for independent draws; for , coverage rises only from to and for , , . 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 at : coverage on a linear scale against logarithmic (shading: two pointwise standard errors); the independent reference, not shown, would be at .
References
- [1]Giulio Casati and Tomaž Prosen. Mixing property of triangular billiards. Physical Review Letters, 83(23):4729–4732, 1999.DOI
- [2]G. Casati, C. Tsallis, and F. Baldovin. Linear instability and statistical laws of physics. Europhysics Letters, 72(3):355–361, 2005.DOI
- [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]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]S. Kerckhoff, H. Masur, and J. Smillie. Ergodicity of billiard flows and quadratic differentials. Annals of Mathematics, 124(2):293–311, 1986.DOI
- [6]Jon Chaika and Giovanni Forni. Weakly mixing polygonal billiards. Annals of Mathematics, 203(3):695–735, 2026.arxiv.org/abs/2003.00890
- [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]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]Giovanni Forni and Nelson Moll. Cohomological equation for geodesic flows on flat surfaces, 2025. arXiv:2510.18128.arxiv.org/abs/2510.18128
- [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]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]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]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]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]E. Sparre Andersen. On the fluctuations of sums of random variables. Mathematica Scandinavica, 1:263–285, 1953.DOI
- [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]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]Klaus Schmidt. Cocycles on Ergodic Transformation Groups, volume 1 of Macmillan Lectures in Mathematics. Macmillan Company of India, Delhi, 1977.
- [19]H. Furstenberg. Strict ergodicity and transformation of the torus. American Journal of Mathematics, 83:573–601, 1961.DOI
- [20]H. B. Keynes and D. Newton. A “minimal”, non-uniquely ergodic interval exchange transformation. Mathematische Zeitschrift, 148:101–105, 1976.DOI
- [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]Michael Boshernitzan. A condition for unique ergodicity of minimal symbolic flows. Ergodic Theory and Dynamical Systems, 12:425–428, 1992.
- [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]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]E. Sparre Andersen. On the fluctuations of sums of random variables II. Mathematica Scandinavica, 2, 1954.DOI
- [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]J. Marklof. Pair correlation densities of inhomogeneous quadratic forms. Annals of Mathematics, 158:419–471, 2003.DOI
- [28]G. Rauzy. Échanges d’intervalles et transformations induites. Acta Arithmetica, 34:315–328, 1979.DOI
- [29]W. A. Veech. Gauss measures for transformations on the space of interval exchange maps. Annals of Mathematics, 115:201–242, 1982.DOI
- [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]W. P. Hooper. Renormalization of polygon exchange maps arising from corner percolation. Inventiones Mathematicae, 191:255–320, 2013.arxiv.org/abs/1105.6137
- [32]J. Marklof. Limit theorems for theta sums. Duke Mathematical Journal, 97:127–153, 1999.DOI
- [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]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]G. M. Molchan. Maximum of a fractional Brownian motion: probabilities of small values. Communications in Mathematical Physics, 205:97–111, 1999.DOI
- [36]R. Artuso, G. Cristadoro, M. Degli Esposti, and G. Knight. Sparre-Andersen theorem with spatiotemporal correlations. Physical Review E, 89:052111, 2014.DOI











