Initial-State Precision as a Predictive Resource: From Tori to Strange Nonchaotic Attractors and Chaos

Prediction is often characterized by forward error growth: nearby states separate, finite-size perturbations amplify, and Lyapunov exponents or related quantities summarize that instability. A physical measurement problem is naturally posed in the opposite direction. Given a target, an allowed error, and a finite prediction horizon, how accurately must the initial condition be specified?

We address that inverse problem with a deliberately restricted resource architecture. The resource is the local physical resolution of the initial state or of a declared initial coordinate. It is not an unrestricted source code: an encoder that already knows the exact state and may transmit the future target directly would bypass the dynamical prediction problem. The question is instead how the required initial measurement precision scales with the horizon.

This viewpoint is related to finite-size Lyapunov analysis and $\varepsilon$-dependent entropy [1, 2], predictive rate-distortion theory [3], metric mean dimension [4], and finite-rate state estimation [5], but the resource considered here is different: all of it is spent at the initial time. A complementary formulation optimizes the shape of a fixed uncertainty budget among predictive directions [6]; here we ask how the amount of required resolution grows with prediction horizon.

Strange nonchaotic attractors (SNAs) provide a particularly useful test. They are geometrically strange while the largest nontrivial Lyapunov exponent can remain nonpositive [7, 8]. At the same time, their response to the phase of quasiperiodic forcing can become unbounded, as quantified by the phase-sensitivity construction of Pikovsky and Feudel [9]. Thus a smooth torus, an SNA, and ordinary chaos provide three qualitatively different forms of sensitivity against which an initial-precision resource can be calibrated.

The paper is organized around that comparison. We first define a horizon-wide one-shot precision resource and record two elementary consequences needed for its interpretation. An exactly solvable expanding circle/Chebyshev system then provides a closed-form chaotic calibration. The main numerical result is obtained for the quasiperiodically forced logistic map: representative torus, SNA, and chaotic regimes show bounded, logarithmic-like, and linear-like precision growth, respectively. We then explain why algebraic phase sensitivity produces logarithmic precision demand, test the numerical worst-case approximation with phase grids up to thousands of samples, and finally study the very long activation horizons near the torus–SNA onset. Additional accuracy crossovers and broader parameter scans are reported in the Appendices.

Let $(X,d_{X})$ be a metric state space, $F:X\to X$ a deterministic one-step map, and $g:X\to\mathbb{R}$ the scalar prediction target. Starting from $x_{0}\in X$,

For a horizon $N$, define the target trajectory

with

A sensing architecture is specified by a nested family of local uncertainty cells $\mathcal{U}_{\delta}(x)\subset X$, with $x\in\mathcal{U}_{\delta}(x)$ and larger $\delta$ representing coarser initial information. A metric ball $\mathcal{U}_{\delta}(x)=\{x^{\prime}:d_{X}(x,x^{\prime})\leq\delta\}$ is the standard isotropic example, but the formulation also allows a declared coordinate-selective sensor, such as phase-only uncertainty in the forced system below. The resource is therefore always conditional on the stated sensing architecture.

Three associated error quantities are useful. The worst-case trajectory diameter is

The minimax prediction radius is

where $\widehat{\bm{y}}$ is the trajectory estimate chosen for the uncertainty cell. If instead the trajectory from the central state is used as predictor, the nominal-trajectory error is

They obey

For scalar trajectory sets under the sup norm, $E_{N}=D_{N}/2$.

Fix a reference half-width $\delta_{\rm ref}>0$. For a minimax tolerance $\varepsilon>0$, define

and the one-shot initial precision resource

For a one-dimensional uniform sensor, $B_{N}$ is the additional number of binary resolution bits required relative to $\delta_{\rm ref}$, up to integer rounding. The full horizon dependence is retained because a linear rate $\limsup B_{N}/N$ cannot distinguish bounded from logarithmic or other sublinear growth.

Consider the expanding circle map

with the circle metric

and observable

For this calibration the declared uncertainty cell is the phase ball $\mathcal{U}_{\delta}(\theta)=\{\theta^{\prime}:d_{\mathbb{T}}(\theta,\theta^{\prime})\leq\delta\}$. Thus two admissible phases may differ by as much as $2\delta$.

If $m^{N}\delta\leq 1/4$, the exact horizon-wide output diameter is

Once $m^{N}\delta\geq 1/4$, the uncertainty interval contains a half-turn separation at or before horizon $N$, and the cosine reaches its full diameter $2a$. For the minimax architecture, therefore, every fixed fine tolerance $0<\varepsilon<a$ has

and, for fixed $0<\delta_{\rm ref}\leq 1/4$,

Thus all fixed fine accuracies eventually have the same linear resource class and differ by a constant number of bits.

The familiar Chebyshev form follows by defining

Using $T_{k}(\cos u)=\cos(ku)$,

The orbit is therefore available in closed form. Here the Chebyshev representation serves as an exactly solvable calibration of the initial-precision resource in ordinary chaos: exponential instability gives $B_{N}=O(N)$, while changing a fixed fine accuracy changes only the additive offset.

We study

where $x_{n}\in[0,1]$ is the fiber coordinate, $\phi_{n}\in\mathbb{T}$ is the phase of the quasiperiodic drive, $\alpha$ is the logistic control parameter, $\epsilon_{\rm d}$ is the drive amplitude, and

Following the standard parametrization of the forced-logistic SNA literature, we use

Along this line a fractalization route from a smooth torus to an SNA occurs near $\alpha_{F}\simeq 2.6526$ [10, 8].

The fiber Lyapunov exponent is

Representative numerical values are negative for the torus ($\alpha=2.50$, $\Lambda_{x}\simeq-0.216$), weakly negative for the SNA ($\alpha=2.70$, $\Lambda_{x}\simeq-0.025$), and positive for chaos ($\alpha=2.75$, $\Lambda_{x}\simeq+0.017$); only the sign and regime assignment are used below.

For phase sensitivity define

Differentiating Eq. (23),

with

For one initial state,

The conventional phase-sensitivity function takes an attractor-wide minimum over sampled initial states,

where $\mathcal{S}$ denotes the sampled post-transient initial-state set. Its unbounded growth is an established SNA diagnostic [9, 8].

At the common numerical tolerance $\eta=0.1$, Figure 1 compares the three representative regimes on the axes that expose their observed horizon dependence. The torus saturates, the SNA is approximately linear when plotted against $\log_{2}N$, and chaos is approximately linear in $N$ over the resolved finite-perturbation window. The figure therefore displays the central numerical hierarchy directly: bounded, logarithmic-like, and linear-like precision demand.

Thus the numerical hierarchy is

The class labels in Eq. (38) describe the resolved horizon windows of the forced-logistic calculation and are not asymptotic classifications. The circle/Chebyshev benchmark in Eq. (20) is the analytic result; the three forced-logistic scaling labels are finite-window numerical observations.

The SNA result is the central nontrivial case. A negative fiber Lyapunov exponent does not imply bounded initial phase precision. Direct finite-perturbation fits give logarithmic slopes of order a few bits per doubling of horizon, with the coefficient depending on tolerance, fitting window, and worst-case aggregation. The growth form is more robust than any one fitted coefficient.

Because a finite maximum can underestimate a true supremum, especially on an SNA, we test the sampled worst case by nested phase-grid refinement. For an $M$-point grid, let $R_{N,M}^{\max}(\delta)$ denote the maximum, over its base phases, of the horizon-wide deviation obtained from the four offsets in Eq. (35). For each $\alpha$, $M_{\max}=4096$ equally spaced initial phases were propagated for $3\times 10^{4}$ transient iterations from $x=0.2$ to generate base states. Nested dyadic subsets with

were then evaluated using the four offsets in Eq. (35). For the most sensitive $\delta=10^{-8}$ check the grid was extended to $M=8192$.

For the SNA at $\alpha=2.70$ and $N=32768$, the sampled maximum at $\delta=10^{-8}$ changes from $0.4630$ at $M=32$ to $0.46959$ at $M=256$, $0.46979$ at $M=4096$, and $0.46982$ at $M=8192$. The $4096\to 8192$ change is about $7\times 10^{-5}$ relative. At $\delta=10^{-10}$, the larger change associated with rarer extreme events is already only $2.5\times 10^{-4}$ relative between $M=2048$ and $4096$. For chaos at $\alpha=2.75$ and $N=4096$, the corresponding maxima stabilize near $0.6163$, with changes below $4\times 10^{-4}$ relative from $M=2048$ to $4096$ for all three tested perturbation scales. Figure 3 visualizes the SNA refinement directly: the sampled maximum rises at coarse phase resolution and then forms a clear plateau, showing that the reported long-horizon envelope is numerically stable to further phase-grid refinement.

This convergence study does not prove the mathematical supremum over phase. It does, however, show that the long-horizon envelope near $0.47$ is not an artifact of a 16–256 point phase sample. The complete convergence table, including the chaotic case, is provided in Appendix C.

The second main numerical result concerns the time needed for SNA sensitivity to reach a finite operational scale. Immediately above the reported fractalization point, derivative-based phase sensitivity can already grow while a finite perturbation remains microscopically small for extremely long horizons.

Define the sampled activation horizon

with $N_{*}=\infty$ if the threshold is not crossed during the simulated interval. For $\delta=10^{-8}$ and $\eta=0.05$, several parameter values in $2.6527\lesssim\alpha\lesssim 2.658$ did not cross the operational scale by

At $\alpha=2.660$ the sampled crossing occurred near $N\simeq 3.4\times 10^{4}$, while at $\alpha=2.6625$ it occurred near $N\simeq 2.6\times 10^{3}$. The dependence on $\alpha$ is strongly nonmonotonic, so no critical power law is inferred from these data. Figure 4 makes both features visible: a broad near-onset interval remains unactivated up to the maximum simulated horizon, while nearby parameter values can activate orders of magnitude earlier.

The relevant classification is therefore not exhausted by a dynamical label. For a finite-precision observer one must distinguish

A system can already be strange in an asymptotic diagnostic while remaining operationally torus-like at a fixed finite tolerance for a very long time.

The central result is the resource hierarchy in Eq. (38). The exact circle/Chebyshev map shows that exponential instability produces a linear initial-precision cost. In the forced-logistic calculation, the representative smooth torus shows bounded phase-precision demand over the resolved horizons, while the representative SNA exhibits the intermediate logarithmic-like burden associated with algebraic phase sensitivity. This is an operational classification: it does not redefine the underlying attractors, and it does not identify the fitted resource coefficient with a universal fractal dimension or Lyapunov quantity.

Accuracy provides a second operational control, distinct from dynamical class. Proposition 1 shows that fixed accuracies inside one common scale-separable regime differ only by a constant bit offset. The fine-accuracy curves indeed show this approximate offset equivalence. At much coarser tolerances, however, finite-scale structure can become operationally irrelevant. At fixed SNA dynamics ($\alpha=2.70$) the sampled crossover is near $\eta\simeq 0.47$, while at fixed chaotic dynamics ($\alpha=2.75$) it lies near $\eta\simeq 0.615$–$0.620$ in the present nominal-trajectory convention. These are finite-horizon masking scales, not claimed asymptotic invariants; detailed scans are reported in Appendix A.

The inverse horizon interpretation is useful experimentally. If

with logarithmic slope $\mu_{B}>0$ and tolerance-dependent offset $C(\eta)$, then in an SNA-like regime a fixed precision budget $B_{0}$ gives

One extra initial bit therefore multiplies the accessible horizon. In a chaotic regime with linear slope $\kappa_{B}>0$,

one instead obtains

so additional bits extend the horizon additively.

The framework connects directly to several established ideas. In particular, the phase-sensitivity diagnostic of SNAs [9] is inverted here into a one-shot initial measurement requirement. Finite-size Lyapunov exponents already describe scale-dependent error growth [2, 1]; here the classified quantity is the initial resolution required by a fixed horizon-wide task. Predictive rate-distortion and metric mean dimension allow much more general encodings [3, 4], whereas the present architecture intentionally keeps a local physical sensor. Restoration entropy concerns a continuing communication rate [5]; $B_{N}$ is paid once, at the initial time.

Several limitations delimit the scope of these results. The phase grid remains finite, so the true supremum is not proved. The order of $\delta\to 0$ and $N\to\infty$ is delicate near the SNA onset. The resource depends on the declared sensing coordinate and metric; if the forcing phase is supplied by an exact external clock rather than measured, phase precision is not the relevant resource. Finally, the present requirement is horizon-wide; terminal-time prediction can behave differently in folded bounded maps.

Initial precision can be treated as an operational resource for prediction. In a common scale-separable regime, changing a fixed tolerance changes only a constant number of bits. The exactly solvable circle/Chebyshev map provides the calibration $B_{N}=N\log_{2}m+O(1)$ for ordinary exponential chaos.

The forced logistic map adds the nontrivial intermediate case. Under a declared phase-only sensing architecture, a smooth torus, an SNA, and chaos exhibit bounded, logarithmic-like, and linear-like precision demand over the resolved horizons. For the SNA, the mechanism is the conversion of algebraic phase sensitivity into a logarithmic cost in inverse initial scale. Nested phase grids up to thousands of samples show that the extreme-value envelopes used in this conclusion are numerically stable, while the near-onset calculations reveal activation horizons that can exceed one million iterations.

Finite-precision prediction therefore depends on more than the asymptotic dynamical label. The measurement architecture, required accuracy, and horizon determine which dynamical structure must actually be resolved.

This work was supported by the research grant SP016 from SOBIN Institute LLC.