📊 statistics

Transcripts and Algebraic Distances in Time Series: Stochastic Properties and Nonparametric Dependence Tests

This paper introduces a novel nonparametric framework for testing serial dependence in time series by analyzing transcripts and algebraic distances between successive ordinal patterns, demonstrating that these new statistics offer superior power compared to existing methods.

Original authors: Christian H. Weiß, José M. Amigó

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christian H. Weiß, José M. Amigó

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a river flow. You want to understand if the water is moving randomly (like a chaotic splash) or if there is a hidden pattern (like a steady current or a repeating ripple).

This paper introduces a new way to "listen" to that river, not by measuring the exact height of the water, but by looking at the order in which the water levels rise and fall.

Here is a breakdown of the paper's ideas using simple analogies:

1. The "Ordinal Pattern" (The Snapshot)

First, the authors take a small window of time (say, three moments in a row) and look at the water levels. They don't care about the exact numbers (like 10.5 meters or 12.2 meters); they only care about the ranking.

Did it go Low → Medium → High?
Did it go High → Low → Medium?

They call these rankings "Ordinal Patterns" (OPs). Think of this like taking a photo of a race and only recording who finished 1st, 2nd, and 3rd, ignoring their exact times. This makes the data robust against noise (like a sudden splash) and easy to compute.

2. The "Transcript" (The Difference)

The paper's big innovation is looking at what happens between these snapshots.
If Snapshot A is "Low-Medium-High" and the very next Snapshot B is "High-Medium-Low," how did the pattern change?

The authors invented a concept called a "Transcript."

The Analogy: Imagine the patterns are words in a language. If you have the word "CAT" and the next word is "ACT," the "transcript" is the specific set of moves needed to turn "CAT" into "ACT."
In math terms, the transcript is the "difference" or the "transformation" between two consecutive patterns. It tells you exactly how the river's behavior shifted from one moment to the next.

3. The "Edit Distances" (Measuring the Shift)

Once they have the "transcript" (the transformation), they measure how "big" that change is using two specific rulers:

Cayley Distance: Counts the minimum number of swaps needed to fix the pattern. (Like swapping two letters in a word to make it right).
Kendall Distance: Counts the minimum number of adjacent swaps needed. (Like sliding letters next to each other to fix the order).

These distances act like a speedometer for change. If the river is truly random, these distances should jump around in a predictable, average way. If the river has a hidden pattern (serial dependence), these distances will behave differently—either staying too small (stuck in a loop) or jumping too wildly.

4. The "Detective Work" (Testing for Patterns)

The authors used these tools to build a new set of statistical tests.

The Goal: To catch if a time series is truly random (Independent and Identically Distributed, or i.i.d.) or if it has a hidden memory (dependence).
The Method: They calculated the average "distance" between patterns and the distribution of all possible "transcripts." They then compared real data against what a purely random river should look like.

5. The Results: Why It's Better

The paper ran thousands of simulations (creating fake rivers with known patterns) and tested their new tools against old tools.

The Finding: The new "Transcript" and "Distance" tools were often more powerful at spotting hidden patterns than the older methods. They were better at catching both subtle trends and sudden shifts.
The Real-World Test: They applied this to temperature data from Hamburg, Germany. After removing the obvious yearly seasons, they tested if the remaining daily changes were random.
- Result: The new tests successfully detected that the temperature changes were not random; they still had a hidden memory (dependence), whereas some older tests missed it or were less certain.

Summary

Think of the old way of analyzing time series as looking at a movie frame-by-frame to see if the actors are moving randomly.
This paper suggests a smarter way: Watch how the actors' positions change from one frame to the next. By measuring the specific "moves" (transcripts) required to get from one scene to the next, you can detect a script (a pattern) much faster and more accurately than by just looking at the scenes themselves.

The authors conclude that these "transcript-based" tools are a powerful new addition to the toolbox for anyone trying to find order in chaotic data.

Technical Summary: Transcripts and Algebraic Distances in Time Series

Problem Statement
The paper addresses the detection of serial dependence in univariate, continuously distributed time series. While the use of Ordinal Patterns (OPs)—symbolic representations based on the rank order of values in a sliding window—has become a standard tool for analyzing dependence structures, existing methods primarily rely on the marginal frequencies of these patterns. The authors argue that the algebraic structure of OPs (viewed as elements of a symmetric group) has not been fully exploited for dependence testing. Specifically, the "difference" or relationship between successive OPs, which captures the dynamics of how the rank order evolves over time, remains underutilized in nonparametric testing frameworks.

Methodology
The proposed approach transforms the original time series into a sequence of transcripts and algebraic distances.

Conceptual Framework:
- Ordinal Patterns (OPs): A time series is discretized into a sequence of permutations $\pi_t \in S_m$ (the symmetric group of degree $m$ ) based on the relative ordering of $m$ consecutive observations.
- Transcripts: Defined as the group-theoretic "difference" between successive OPs: $\tau_t = \pi_{t+1} \circ \pi_t^{-1}$ . This concept generalizes the subtraction of elements in additive groups to the non-abelian symmetric group.
- Edit Distances: The authors utilize two specific metrics derived from transcripts:
  - Cayley Distance ( $d_C$ ): The minimum number of transpositions required to transform one permutation into another.
  - Kendall Distance ( $d_K$ ): The minimum number of adjacent transpositions required.
- These distances are shown to be norms of the transcript, linking the algebraic structure directly to measurable dissimilarity.
Stochastic Properties:
- The paper derives the marginal distributions of transcripts and the resulting distances for three specific process types under the null hypothesis of serial independence (i.i.d.):
  - Standard i.i.d. processes.
  - Symmetric Random Walks (SRW).
  - Generalized Coin-Tossing (GCT) processes.
- It is demonstrated that the marginal distributions of transcripts and distances differ substantially between i.i.d. processes and serially dependent processes (e.g., AR(1), SRW, GCT). For instance, positive dependence tends to increase the frequency of zero-distance transcripts (repeated OPs), while negative dependence decreases it.
- Bivariate Distributions: The authors derive the joint distributions of transcripts at lag $h=1, 2, 3$ for i.i.d. processes, establishing the dependence structure (specifically, that the transcript series is 3-dependent for i.i.d. inputs).
Asymptotic Theory and Test Statistics:
- Using a Central Limit Theorem for $\alpha$ -mixing (specifically $q$ -dependent) processes, the authors establish the asymptotic normality of the vector of relative frequencies of transcripts and distances.
- Based on these asymptotics, six new test statistics are proposed for testing the null hypothesis of serial independence:
  - Mean Distance Tests: $d_C$ and $d_K$ (sample means of Cayley and Kendall distances).
  - Entropy/Deviance Tests: $H_\tau, H_C, H_K$ (Kullback–Leibler divergence between observed and theoretical frequencies).
  - Goodness-of-Fit Tests: $\Delta_\tau, \Delta_C, \Delta_K$ (Pearson statistics).
- The asymptotic distributions of the entropy and Pearson statistics are derived as quadratic forms of independent chi-squared variables, with eigenvalues explicitly calculated for the i.i.d. case.

Key Results

Theoretical Derivations: The paper provides closed-form expressions for the marginal and bivariate distributions of transcripts and distances for i.i.d., SRW, and GCT processes. It establishes that the mean Cayley and Kendall distances are sensitive to the type of serial dependence (increasing or decreasing depending on the process).
Simulation Study: A comprehensive simulation study compares the new transcript-based tests against:
- The parametric ACF test.
- Existing nonparametric OP-based tests (entropy, Pearson, $\beta$ , and $\tau$ tests from Weiß, 2022).
- Findings:
  - The new tests generally exhibit superior power compared to previous OP-based tests, particularly for AR(1) processes and processes contaminated with additive outliers.
  - The Kendall distance-based tests ( $d_K$ , $H_K$ , $\Delta_K$ ) and transcript-based tests ( $H_\tau$ , $\Delta_\tau$ ) are consistently powerful across linear and nonlinear alternatives.
  - The Cayley distance-based tests ( $d_C$ , $H_C$ , $\Delta_C$ ) show mixed performance, excelling notably for ARCH(1) processes but performing less well for others.
  - All proposed tests demonstrate robustness to outliers, outperforming the ACF test in contaminated scenarios.
Real-World Application: The methods were applied to monthly air temperature data from Hamburg (1936–2024). After seasonal differencing, the transcript-based tests (specifically $H_\tau$ , $H_K$ , $\Delta_K$ ) successfully rejected the null hypothesis of serial independence, detecting dependence structures that some competing OP-based tests missed or found ambiguous.

Significance and Claims
The authors claim that this work extends the utility of Ordinal Patterns by leveraging their algebraic group structure. By treating OPs not just as symbols but as group elements, the introduction of transcripts and edit distances provides a richer representation of the time series' dependence structure.

The paper asserts that:

Transcripts offer a new dimension of analysis: They capture the "difference" between successive states, which is distinct from the marginal frequency of the states themselves.
Improved Performance: The proposed nonparametric tests often outperform former OP-based tests and the standard ACF test, particularly in terms of power against various linear and nonlinear alternatives and robustness to outliers.
Practical Utility: The approach is computationally feasible and provides a valuable complement to existing tools for detecting serial dependence in real-world data.

The authors conclude that while the current framework focuses on univariate self-transcripts, the algebraic approach suggests promising future directions, including cross-transcripts for multivariate/coupled systems and spatial ordinal patterns. They also note that deriving closed-form distributions for more complex processes remains an open area for future research.