Valid Ground Truth Time in Chaotic Dynamics

• VGTT is a metric that specifies the time span during which different ODE solvers yield mutually consistent trajectories in chaotic systems like the Lorenz model.
• It is operationally determined by comparing pairwise discrepancies between solver outputs under strict error tolerances to ensure reliable numerical ground truth.
• VGTT establishes an upper limit for valid prediction time, highlighting that forecasts beyond this interval reflect solver-dependent artifacts rather than true system dynamics.

Searching arXiv for the cited paper and related benchmark/context papers. Valid Ground Truth Time (VGTT) is a numerical-analysis concept introduced to delimit how long a trajectory of a chaotic dynamical system can be treated as solver-independent “ground truth.” In the Lorenz setting studied in "Reservoir computing with large valid prediction time for the Lorenz system" (Hurley et al., 8 Aug 2025), VGTT denotes the time interval over which several common ODE solvers, started from the same initial condition, continue to agree on the trajectory within a practical tolerance. The concept is motivated by the fact that in chaotic systems, exponentially amplified truncation and roundoff errors cause different numerical integrations to separate after sufficiently long times. VGTT therefore functions as an interpretive ceiling on Valid Prediction Time (VPT): once solver disagreement becomes appreciable, a reported trajectory forecast can no longer be regarded as prediction of a uniquely defined reference orbit.

1. Definition and conceptual role

The paper introduces VGTT in its discussion of solver dependence. Its key statement is that the ABM54 solver agrees with a selection of other ODE solvers reasonably suited to the task, such as fourth-order Runge–Kutta, up to around 30–40 Lyapunov times, and that this interval is termed the Valid Ground Truth Time (Hurley et al., 8 Aug 2025). The same discussion states that VGTT offers a computational upper limit for VPT, because comparing VPT values higher than the VGTT is unreasonable if a different solver would yield a different solution.

A formalization consistent with that description is to consider numerical solutions $\{u_i(t)\}$ generated by different solvers from the same initial condition, and define a pairwise discrepancy

$\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$

For a tolerance $\epsilon$, the VGTT may then be written as

$$T_{\text{VGTT}} = \sup\left\{T>0 \,\middle|\, \Delta_{ij}(t)<\epsilon \text{ for all solvers } i,j \text{ and all } 0\le t\le T\right\}.$$

The paper does not present this formula explicitly, but this is the content of its verbal definition: VGTT is the interval over which several common solvers still “agree” on the trajectory.

The motivation is specific to chaotic dynamics. Infinitesimal perturbations in the Lorenz system grow exponentially at a rate set by the largest Lyapunov exponent. Numerical integration inevitably introduces small perturbations through truncation error, roundoff, and local error control. For times shorter than $T_{\text{VGTT}}$, these perturbations remain practically negligible, so the computed trajectory is usable as ground truth. For times beyond $T_{\text{VGTT}}$, the same initial-value problem can yield materially different trajectories under equally reasonable solver choices. Beyond that point, deterministic trajectory-wise “truth” is no longer solver-independent.

2. Numerical origin in the Lorenz experiments

The setting in which VGTT is introduced is the standard chaotic Lorenz system,

$$\frac{dx}{dt} = \sigma (y-x), \qquad \frac{dy}{dt} = x(\rho-z)-y, \qquad \frac{dz}{dt} = xy-\beta z,$$

with $\sigma=10$, $\rho=28$, and $\beta=8/3$. The training and testing data are generated in Julia using DifferentialEquations.jl. The reference solver is ABM54, described as a predictor–corrector Adams–Bashforth–Moulton scheme with a fifth-order Adams–Bashforth explicit predictor, a four-step Adams–Moulton implicit corrector, and starting values from fourth-order Runge–Kutta. The reported settings are absolute tolerance $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$0, relative tolerance $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$1, and output sampling interval $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$2 (Hurley et al., 8 Aug 2025).

VGTT arises by comparing this high-accuracy ABM54 trajectory with trajectories generated by other solvers reasonably suited to the same smooth, non-stiff ODE. The paper explicitly mentions fourth-order Runge–Kutta as an example. Although the supplementary figures are not reproduced in the supplied text, the described procedure is clear: integrate from the same initial condition with ABM54 and alternative solvers, track solver–solver discrepancy over time, and identify the interval for which the solutions remain practically indistinguishable.

Under these conditions, the paper reports that ABM54 agrees with the comparison solvers up to around 30–40 Lyapunov times. That numerical interval is the VGTT for the solver configuration under study. The same discussion notes that using smaller absolute and relative tolerances and a smaller $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$3 can increase the VGTT, at the cost of longer computation times. VGTT is therefore not an invariant of the Lorenz equations alone; it depends on the numerical scheme, tolerances, discretization, and the standard of agreement imposed across solvers.

3. Operational determination

The paper’s description implies a concrete workflow for estimating VGTT. First, one selects a high-quality reference integration, here ABM54 with $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$4 and $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$5. Second, one generates a reference trajectory over many Lyapunov times. Third, one integrates the same initial condition with alternative solvers, such as RK4 and other common methods appropriate to the task. Fourth, one computes a discrepancy between the resulting trajectories at matched sample times. The most natural choice, consistent with the paper’s VPT metric, is a norm on the state difference,

$\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$6

where $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$7 is the Lorenz state from solver $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$8.

The divergence time for a given comparison solver is the first time at which the discrepancy exceeds a chosen tolerance. VGTT is then the minimum such divergence time across the family of solvers under consideration. In words, it is the largest time horizon over which all considered solver outputs remain mutually consistent.

This procedure makes explicit why VGTT is a property of a numerical experiment rather than a purely analytical object. The exact tolerance $\Delta_{ij}(t) = \|u_i(t)-u_j(t)\|.$9 is not stated in the paper, but the intended interpretation is: within the VGTT window, solver differences are negligible relative to the attractor scale and to the forecasting task; outside that window, the notion of a single trajectory-valued ground truth becomes dependent on numerical conventions. A plausible implication is that two studies using the same dynamical system but different integration standards may report different VGTT values even if their forecasting models are identical.

4. Relation to Valid Prediction Time and Lyapunov growth

The paper distinguishes sharply between VGTT and VPT. VPT compares a model forecast with a selected reference trajectory, whereas VGTT compares solver outputs with one another. The VPT error is defined as the normalized squared error

$\epsilon$0

with $\epsilon$1 for Lorenz and the denominator given by the total variance of the true trajectory over time. VPT is then the first time for which $\epsilon$2 exceeds a threshold, taken as $\epsilon$3 following Jaurigue’s benchmark (Hurley et al., 8 Aug 2025).

The paper reports that, on a log-scale plot of the normalized mean squared error, both the reservoir computer and the benchmark have nearly the same slope, specifically $\epsilon$4 and $\epsilon$5, respectively, and that this slope is close to but slightly less than the maximum Lyapunov exponent of the Lorenz system, $\epsilon$6. This is consistent with the standard expectation of exponential error amplification in chaotic systems. The paper’s conclusion is that because the slopes are nearly identical, the VPT is decided by the jump in error that occurs in the first few time steps. This in turn supports the claim that, using knowledge of the Lyapunov exponent, VPT can be predicted from the error in the first few prediction steps, providing a computationally efficient evaluation method.

The conceptual distinction can be summarized as follows.

Quantity	Comparison	Interpretation
VGTT	Solver vs. solver	Time over which the reference trajectory is solver-independent
VPT	Model vs. reference trajectory	Time over which a predictor tracks the chosen trajectory within threshold

The paper’s central interpretive rule is that VPT exceeding VGTT is not meaningful (Hurley et al., 8 Aug 2025). If a predictor attains a nominal VPT beyond the interval over which the reference orbit is solver-stable, then the excess horizon measures agreement with one numerical realization rather than prediction of a uniquely defined continuous-time trajectory. This is not merely a technical caveat; it changes the epistemic status of long-horizon forecast claims in chaotic systems.

5. Consequences for reservoir computing and benchmark interpretation

The study investigates the dependence of reservoir-computing VPT on hyperparameters including the regularization coefficient, reservoir size, and spectral radius, and it reports high VPT values greater than 30 Lyapunov times in a noiseless Lorenz setting. Under carefully chosen conditions, the reservoir computer achieves approximately 70% of a benchmark performance based on the output of a single prediction step used as initial conditions for the Lorenz equations. The paper also identifies two spectral-radius regimes associated with large VPT: a small radius near zero, producing simple but stable operation, and a larger radius operating at the “edge of chaos” (Hurley et al., 8 Aug 2025).

VGTT constrains how these results should be read. Because the Lorenz dataset is noiseless, the paper finds that overfitting can be beneficial: the optimal ridge-regression regularization is extremely small, with $\epsilon$7, essentially a least-squares fit, and nonzero regularization is detrimental to VPT in this setting. The authors explicitly caution that these conditions may not hold for noisy systems, though they may still be useful for real-world applications with limited noise. In the noiseless case, however, the only effective “noise” in the target data is numerical error from the solver. VGTT therefore demarcates the interval over which fitting a particular solver-generated trajectory can still be interpreted as fitting the underlying continuous-time system rather than overfitting solver-specific artifacts.

A common misconception is that very large VPT on a low-dimensional chaotic benchmark automatically implies physically meaningful long-range prediction. The paper directly rejects that interpretation: if the reported VPT reaches or exceeds the interval where alternative solvers already disagree, the result is solver-dependent trajectory reproduction rather than solver-independent deterministic forecasting. The same limitation applies to the benchmark used in the study, because that benchmark also relies on a numerical integration of the Lorenz equations and is therefore subject to the same VGTT ceiling.

6. Generalization, limitations, and reporting standards

Although the study is conducted on the Lorenz system, it states that the conclusions are qualitatively applicable to other noiseless model dynamical systems, while the details for other specific systems and the applicability to noisy, real-world time series are left to future work (Hurley et al., 8 Aug 2025). The generalization is straightforward: for any ODE or PDE without a closed-form solution, one may integrate from the same initial condition using multiple high-quality solvers, compute pairwise discrepancies, and define a solver-agreement horizon analogous to VGTT. For a general state $\epsilon$8, the same formal pattern applies,

$\epsilon$9

The concept is strongest when the evaluation target is trajectory-wise agreement. Beyond the VGTT, only statistical properties such as invariant measures or attractor geometry remain robust in the sense described in the supplied material. This suggests a broader methodological distinction: short- and medium-horizon deterministic forecasting claims should be assessed against solver-stable reference trajectories, whereas longer-horizon claims in chaotic systems may need to shift toward distributional, invariant, or statistical criteria.

The paper also advances a reporting norm. It encourages explicit specification of the numerical methods used to generate chaotic time-series datasets, including solver type, order, tolerances, and sampling interval. The practical reason is immediate: a forecasting metric defined against a numerically generated “ground truth” is only as meaningful as the interval over which that ground truth is itself numerically robust. In this sense, VGTT is both a diagnostic and a disclosure standard. It quantifies the credibility horizon of trajectory-level evaluation and prevents over-interpretation of long VPT values in chaotic forecasting experiments.