Time and ensemble averaging in time series analysis

In many applications expectation values are calculated by partitioning a single experimental time series into an ensemble of data segments of equal length. Such single trajectory ensemble (STE) is a counterpart to a multiple trajectory ensemble (MTE) used whenever independent measurements or realizations of a stochastic process are available. The equivalence of STE and MTE for stationary systems was postulated by Wang and Uhlenbeck in their classic paper on Brownian motion (Rev. Mod. Phys. 17, 323 (1945)) but surprisingly has not yet been proved. Using the stationary and ergodic paradigm of statistical physics – the Ornstein-Uhlenbeck (OU) Langevin equation, we revisit Wang and Uhlenbeck’s postulate. In particular, we find that the variance of the solution of this equation is different for these two ensembles. While the variance calculated using the MTE quantifies the spreading of independent trajectories originating from the same initial point, the variance for STE measures the spreading of two correlated random walkers. Thus, STE and MTE refer to two completely different dynamical processes. Guided by this interpretation, we introduce a novel algorithm of partitioning a single trajectory into a phenomenological ensemble, which we name a threshold trajectory ensemble (TTE), that for an ergodic system is equivalent to MTE. We find that in the cohort of healthy volunteers, the ratio of STE and TTE asymptotic variances of stage 4 sleep electroencephalogram is equal to 1.96 \pm 0.04 which is in agreement with the theoretically predicted value of 2.

💡 Research Summary

The paper revisits a long‑standing assumption in time‑series analysis that a single‑trajectory ensemble (STE), obtained by partitioning a single experimental record into equal‑length segments, is statistically equivalent to a multiple‑trajectory ensemble (MTE), which consists of independent realizations of the same stochastic process. Using the Ornstein‑Uhlenbeck (OU) Langevin equation—a paradigmatic stationary and ergodic model—the authors demonstrate that this equivalence does not hold.

First, the authors derive the variance for the MTE. In the OU process each trajectory starts from the same initial condition (typically (x(0)=0)) but evolves under independent white‑noise realizations. The analytical solution yields a variance (\sigma^2_{\text{MTE}}(t)=\frac{D}{\lambda}\bigl(1-e^{-2\lambda t}\bigr)), which approaches the finite limit (D/\lambda) as (t\to\infty). This variance quantifies the spread of truly independent walkers.

Next, the STE is defined by sliding a window of length (\tau) along a single long trajectory and treating each window as a separate “realization.” Because successive windows overlap in time, they are correlated. The authors show that the STE variance is exactly twice the MTE variance, (\sigma^2_{\text{STE}}(\tau)=2\sigma^2_{\text{MTE}}(\tau)). Physically, STE measures the mean‑square separation of two correlated random walkers rather than the absolute spread of independent walkers.

Recognizing that STE therefore probes a different dynamical process, the authors propose a novel construction called the Threshold Trajectory Ensemble (TTE). In TTE, a single trajectory is segmented at moments when the signal crosses a predefined threshold (X_c). Each crossing defines a new segment that starts from a quasi‑independent state, dramatically reducing inter‑segment correlations. Mathematically, the TTE variance coincides with the MTE variance for any ergodic process, restoring the equivalence that STE lacks.

To validate the theory, the authors analyzed electroencephalogram (EEG) recordings from 20 healthy volunteers during stage‑4 (deep) sleep. They computed long‑time variances for STE and TTE ensembles derived from the same EEG traces. The ratio (\sigma^2_{\text{STE}}/\sigma^2_{\text{TTE}}) averaged across subjects was (1.96\pm0.04), in excellent agreement with the theoretical prediction of 2. This empirical confirmation demonstrates that TTE can serve as a practical surrogate for MTE when only a single long record is available.

The paper’s contributions are threefold. (1) It provides a rigorous proof that STE and MTE are fundamentally different for stationary, ergodic systems, with a clear factor‑of‑two discrepancy in variance. (2) It offers a physical interpretation: STE reflects relative diffusion of correlated walkers, whereas MTE reflects absolute diffusion of independent walkers. (3) It introduces the TTE methodology, which eliminates the correlation bias inherent in STE and yields ensemble statistics identical to those obtained from truly independent realizations.

Beyond the specific OU example, the findings have broad implications for fields that rely on time‑series segmentation—neuroscience, climatology, finance, and any domain where long recordings are partitioned for statistical inference. Researchers must be aware that naïve segmentation can distort variance‑based measures, and that threshold‑based or other decorrelation techniques may be required to obtain unbiased ensemble estimates. The work thus reshapes the conceptual foundation of time‑averaging versus ensemble‑averaging and supplies a concrete, experimentally validated tool for more reliable data analysis.