Bandt–Pompe Symbolization Methodology

• Bandt–Pompe symbolization is a rank-based method that converts time series data into discrete ordinal patterns, facilitating robust complexity quantification.
• It constructs overlapping delay vectors and maps them to permutations, making the approach resilient to measurement noise and monotonic transformations.
• The method is widely applied in fields like biomedical signal processing and nonlinear dynamics, supporting advanced analyses of complex systems.

The Bandt–Pompe symbolization methodology is a rank-based approach for transforming real-valued time series into symbolic sequences of ordinal patterns, facilitating the quantitative characterization of complex dynamics using information-theoretic measures such as permutation entropy and statistical complexity. Ordinal analysis, as introduced by Bandt and Pompe in 2002, has become a foundational tool in the study of time series from diverse domains, leveraging the temporal ordering relations within embedded vectors to yield robust, easily interpretable statistics that are insensitive to monotonic transformations and measurement noise.

1. Formal Definition and Symbolization Procedure

The methodology begins with the construction of overlapping delay vectors from a given real-valued time series $X = \{x_t\}_{t=1}^N$, using two key parameters: embedding dimension $D \geq 2$ and temporal delay $\tau \geq 1$ (Aquino et al., 2015, Rosso et al., 2016, Leyva et al., 2022, Gonzalez et al., 2024, Avhale et al., 27 Aug 2025). For $t$ ranging from $1$ to $N-(D-1)\tau$,

$$\mathbf{X}_t = [x_t,\, x_{t+\tau},\, \ldots,\, x_{t+(D-1)\tau}]$$

Each vector is mapped to an ordinal pattern (permutation) $\pi = (r_0, r_1, ..., r_{D-1})$ such that

$$x_{t+r_0\tau} \leq x_{t+r_1\tau} \leq \cdots \leq x_{t+r_{D-1}\tau}$$

Ties are resolved by prescribing that earlier time indices are ranked lower (i.e., $r_i < r_{i+1}$ if $D \geq 2$0) (Rosso et al., 2016, Leyva et al., 2022, Gonzalez et al., 2024).

The mapping of all such vectors yields a sequence of symbols (ordinal patterns) taking as values the $D \geq 2$1 possible permutations of $D \geq 2$2 indices.

2. Permutation Probability Distribution and Information-Theoretic Quantifiers

The empirical distribution of observed patterns $D \geq 2$3 is defined as

$D \geq 2$4

Permutation entropy $D \geq 2$5 quantifies the diversity of observed patterns via the Shannon entropy (Aquino et al., 2015, Rosso et al., 2016, Gonzalez et al., 2024, Avhale et al., 27 Aug 2025):

$D \geq 2$6

$D \geq 2$7

To capture both randomness and structure, statistical complexity $D \geq 2$8 is constructed by combining normalized entropy with a disequilibrium measure such as the normalized Jensen–Shannon divergence $D \geq 2$9 relative to the uniform distribution $\tau \geq 1$0:

$\tau \geq 1$1

$\tau \geq 1$2

$\tau \geq 1$3

This construction ensures $\tau \geq 1$4 and is maximized for distributions that are neither completely random nor perfectly ordered (Aquino et al., 2015, Rosso et al., 2016, Avhale et al., 27 Aug 2025). Alternative complexity quantifiers, such as LMC complexity and, for discrete distributions, Fisher information, are also widely employed (Rosso et al., 2016, Montani et al., 2013).

3. Practical Considerations: Parameter Selection and Tie Handling

Selection of the embedding dimension $\tau \geq 1$5 and delay $\tau \geq 1$6 is informed by trade-offs between pattern diversity and statistical reliability. Bandt and Pompe recommend $\tau \geq 1$7, with $\tau \geq 1$8 (e.g., $\tau \geq 1$9) to ensure all patterns are well-sampled (Aquino et al., 2015, Rosso et al., 2016, Leyva et al., 2022, Gonzalez et al., 2024, Avhale et al., 27 Aug 2025). The time delay $t$0 is often set to $t$1 for consecutive samples but can be increased to probe slower dynamics or specific time scales indicated by autocorrelation or mutual information analyses (Aquino et al., 2015, Gonzalez et al., 2024).

Ties in floating-point data are rare, but real-world measurement constraints (e.g., quantization or plateaus) can produce equal values. The canonical approach resolves ties by temporal order—earlier indices are ranked lower (Leyva et al., 2022, Gonzalez et al., 2024, Aquino et al., 2015). In high-frequency tied data such as HRV time series, alternative tie-breaking or imputation strategies are required. Bayesian missing data imputation, which infers a tie-case's permutation based on observed pattern frequencies from complete cases, yields minimally biased entropy estimates (Traversaro et al., 2017).

4. The Entropy–Complexity Plane and Dynamical Inference

Plotting normalized permutation entropy $t$2 versus statistical complexity $t$3 situates time series within a causality plane whose boundaries are determined by theoretical minima and maxima of $t$4 at each entropy level (Aquino et al., 2015, Leyva et al., 2022, Gonzalez et al., 2024, Lucas et al., 12 Nov 2025). Different regimes cluster in distinct regions:

• Purely stochastic (white noise): $t$5, $t$6
• Deterministic chaos: intermediate $t$7, maximal $t$8
• Regular deterministic: $t$9, $1$0

This framework robustly distinguishes stochastic, chaotic, and regular dynamics, and is resilient to noise contamination. Analyses of forbidden (never occurring in deterministic systems) and missing (unobserved due to finite sampling) patterns further decode the presence of underlying deterministic structure, even under strong noise (Rosso et al., 2011, Aronis et al., 2018).

5. Extensions: Weighted, Global, and Multivariate Variants

Weighted permutation entropy (WPE) enhances classical permutation entropy by incorporating amplitude information, such as local variance within each embedded vector, resulting in a weighted distribution over ordinal patterns (Lucas et al., 12 Nov 2025). WPE improves sensitivity to changes in both ordering and amplitude dynamics.

Global Permutation Entropy (GPE) generalizes the standard consecutive-pattern PE by considering all possible order-$1$1 patterns, including non-consecutive point subsets, leveraging efficient algorithms to scale the enumeration (Avhale et al., 27 Aug 2025). GPE reveals structural information otherwise inaccessible through the classical consecutive embedding approach.

Multivariate generalizations, bivariate and network extensions facilitate ordinal pattern analysis of multidimensional or coupled time series. Permutation mutual information and transfer entropy derived from joint or conditional pattern distributions are increasingly used to probe coupling and causality in networked systems (Lehnertz, 2023).

6. Representative Applications and Impact

Bandt–Pompe analysis has demonstrated value across physics, physiology, neuroscience, geosciences, and engineering. Applications include:

• Biomedical signal analysis: EEG and iEEG-based separation of cognitive or sleep states (Lucas et al., 12 Nov 2025, Carlos et al., 22 Jul 2025, Lucas et al., 2021), HRV analyses for cardiological assessment (Traversaro et al., 2017), detection of seizures (Leyva et al., 2022), or classification in clinical neuroscience (Lehnertz, 2023).
• Dynamical system analysis: Distinguishing stochasticity from determinism in cardiac arrhythmia (Aronis et al., 2018), detecting forbidden patterns and determinism under noise (Rosso et al., 2011), inferring macroscopic behaviors in complex systems (e.g., vehicle traffic, neural activity) (Aquino et al., 2015, Montani et al., 2013).
• Nonlinear signal processing: Feature extraction for classification, clustering, and model identification, including hybrid approaches that combine BP symbolic features with machine learning pipelines (Rosso et al., 2016, Avhale et al., 27 Aug 2025, Leyva et al., 2022, Flores et al., 12 Jul 2025).

A condensed table of core metric definitions is shown below (notation as in text):

Quantifier	Formula	Range
Permutation entropy	$1$2	$1$3
Jensen–Shannon divergence	$1$4	$1$5
Statistical complexity	$1$6	$1$7

7. Open Problems and Limitations

Current limitations include the statistical characterization of ordinal measures under non-stationarity and in finite samples, principled feature reduction for ordinal-pattern-based machine learning, reliable handling of missing or irregularly sampled data, and extensions to higher-dimensional or network time series (Leyva et al., 2022, Lehnertz, 2023). While the methodology is robust to monotonic transformations and moderate noise, it discards amplitude magnitudes, which may limit its discriminative power in settings where both order and value are informative (Flores et al., 12 Jul 2025, Leyva et al., 2022).

Combining Bandt–Pompe features with kernel or deep-learning models is an active area of research, promising further improvements in time series classification and forecasting (Leyva et al., 2022, Rosso et al., 2016). Ongoing studies explore optimal parameterization, efficient computation for high $1$8, and integration with non-parametric or compression-based complexity measures (Zozor et al., 2013, Mateos et al., 2017).

---

References:

(Aquino et al., 2015, Rosso et al., 2016, Leyva et al., 2022, Gonzalez et al., 2024, Avhale et al., 27 Aug 2025, Rosso et al., 2011, Flores et al., 12 Jul 2025, Aronis et al., 2018, Montani et al., 2013, Mateos et al., 2017, Carlos et al., 22 Jul 2025, Zozor et al., 2013, Lehnertz, 2023, Keller et al., 2015, Lucas et al., 2021, Traversaro et al., 2017, Lucas et al., 12 Nov 2025).