Tortuosity Entropy: Theory and Applications

• Tortuosity entropy is an information-theoretic measure that quantifies the spatial disorder of curves by applying Shannon entropy to symbolized local directional changes.
• It utilizes methodologies like chain coding and quantized local differences to construct empirical probability distributions that capture both local unpredictability and long-range structural patterns.
• Practical applications span animal movement analysis to medical imaging, demonstrating its capacity to distinguish between deterministic and stochastic behaviors through robust complexity metrics.

Tortuosity entropy is an information-theoretic quantification of the spatial complexity and unpredictability in curves, trajectories, or movement paths. Unlike purely geometric metrics such as curvature or chord-arc ratio, tortuosity entropy measures the disorder or new information in local directional or morphological changes, grounding the assessment in Shannon entropy and related statistical complexity metrics. Multiple frameworks exist, varying in symbolic encoding, pattern extraction, and underlying probability models, but all share the defining feature of mapping geometric or locomotor complexity onto entropy-based magnitudes.

1. Formal Definitions and Core Concepts

Tortuosity entropy generalizes the notion of curve tortuosity—intuitively, the extent of “wiggle,” “twist,” or spatial disorder—by applying entropy-based analysis to representations derived from trajectories or curves. Given a parameterized curve or a discrete time series $x_1, x_2,\dots, x_N$, the process involves:

• Transforming raw trajectory data into a symbolic or real-valued sequence reflecting local changes (e.g., direction, deviation, or speed).
• Constructing empirical probability distributions over these local changes or their patterns.
• Computing Shannon entropy or related information-theoretic quantities of the resulting distribution.

Major theoretical variants include:

• Symbolic sequence entropy over chain-coded or discretized series (Peña-Mendieta et al., 2024, Liu et al., 2013).
• Entropy of pattern distributions for pattern vectors extracted from symbolized time series (Liu et al., 2013).
• Information entropy of deviation distributions between a curve and a biologically meaningful reference (Wang et al., 24 Jul 2025).

2. Symbolic Encoding and Patternization

Symbolization is central in tortuosity entropy computation. Approaches include:

• Chain Coding: For continuous curves in $\mathbb{R}^n$, a regular grid of spacing $l$ is imposed. Each crossing is snapped to a grid vertex and the direction of motion is coded using a finite alphabet. Freeman’s 8-direction code $\Sigma = \{0,1,\dots,7\}$ is used for 2D, while 26-neighbor coding applies in 3D. The output is a finite-alphabet chain code $f = f_1, f_2,\dots,f_N$ (Peña-Mendieta et al., 2024).
• Quantized Local Differences: In animal movement analysis, the difference $x_{i+1} - x_i$ is quantized into a finite set (e.g., $a=2$ for up/down or $a=3$ for up/no-change/down with threshold) to form a symbol sequence $S = \{s_1, \dots, s_{N-1}\}$ (Liu et al., 2013).

Subsequent embeddings construct overlapping pattern vectors $V(i) = [s_i, s_{i+1}, \dots, s_{i+m-1}]$, whose observed frequencies yield empirical probabilities for entropy evaluation.

3. Entropy Measures and Complexity Quantification

Shannon entropy is the foundational measure, but other derived quantities emphasize distinct aspects of tortuosity:

• Pattern Shannon Entropy: $\mathbb{R}^n$0, where $\mathbb{R}^n$1 is the empirical probability of pattern $\mathbb{R}^n$2 in the m-pattern vector representation. Optional normalization by $\mathbb{R}^n$3 yields a scale-comparable metric $\mathbb{R}^n$4 (Liu et al., 2013).
• Entropy Density ($\mathbb{R}^n$5): Using chain code sequences, the block entropy rate (Shannon entropy rate) is computed as $\mathbb{R}^n$6. In practice, Lempel–Ziv factorization approximates $\mathbb{R}^n$7 for long codes, where $\mathbb{R}^n$8 and $\mathbb{R}^n$9 is the number of LZ factors in $l$0 (Peña-Mendieta et al., 2024).
• Excess Entropy ($l$1): Quantifies the mutual information or pattern content, $l$2, where $l$3 is the block entropy and $l$4 is the entropy density. Random-shuffle comparisons provide nonparametric estimates (Peña-Mendieta et al., 2024).
• Deviation Distribution Entropy: For medical curve analysis, the empirical distribution of pointwise deviations $l$5 between a target and a reference curve is modeled; entropy is evaluated as $l$6. The Information Entropy-Based Framework (IEBF) defines a tortuosity index $l$7 involving scaled cumulative probabilities of $l$8 (Wang et al., 24 Jul 2025).
• Kullback–Leibler Divergence: $l$9 compares the target deviation distribution to a reference, providing a generalized non-symmetric measure of difference (Wang et al., 24 Jul 2025).

4. Interpretation and Theoretical Properties

Tortuosity entropy measures address distinct dimensions of spatial and directional complexity:

• Entropy density ($\Sigma = \{0,1,\dots,7\}$0) or TorEn captures unpredictability or fine-scale disorder in local behaviors (e.g., stepwise changes, heading, velocity). A perfectly straight trajectory has $\Sigma = \{0,1,\dots,7\}$1; highly stochastic motion approaches the uniform-symbol entropy $\Sigma = \{0,1,\dots,7\}$2 (Peña-Mendieta et al., 2024, Liu et al., 2013).
• Excess entropy ($\Sigma = \{0,1,\dots,7\}$3) reflects long-range correlations, motif repetition, or global structure—reflecting pattern richness beyond stepwise unpredictability (Peña-Mendieta et al., 2024).
• IEBF tortuosity quantifies the overall "disorder" in deviations from a meaningful reference, integrating local and global irregularities into a single consistent score (Wang et al., 24 Jul 2025).

These quantities complement, rather than replace, classical geometric indices (fractal dimension, curvature integrals). For example, the fractal dimension $\Sigma = \{0,1,\dots,7\}$4 measures space-filling, but is blind to the sequential unpredictability quantified by $\Sigma = \{0,1,\dots,7\}$5. A trajectory may have high $\Sigma = \{0,1,\dots,7\}$6 (dense filling) but low $\Sigma = \{0,1,\dots,7\}$7 (regularity), or vice versa.

Key comparative points are summarized in the following table:

Measure	Input	What it Captures
$\Sigma = \{0,1,\dots,7\}$8 (entropy density)	Symbolized code	Step-to-step unpredictability
$\Sigma = \{0,1,\dots,7\}$9 (excess entropy)	Symbolized code	Long-range structural pattern
$f = f_1, f_2,\dots,f_N$0 (fractal dim.)	Chain code	Geometric space-filling
$f = f_1, f_2,\dots,f_N$1	Deviations $f = f_1, f_2,\dots,f_N$2	Disorder vs. reference curve

5. Computational Methodologies

Representative computational pipelines include:

• TorEn (Animal Movement):

1. Symbolize local differences in trajectory (quantization, selection of $f = f_1, f_2,\dots,f_N$3, $f = f_1, f_2,\dots,f_N$4).
2. Construct pattern vectors $f = f_1, f_2,\dots,f_N$5, compute pattern counts and empirical probabilities.
3. Compute entropy $f = f_1, f_2,\dots,f_N$6 and normalized $f = f_1, f_2,\dots,f_N$7.
4. Optionally, apply sliding windows to study local changes along the trajectory (Liu et al., 2013).

• Chain Code/LZ76 (General Curves):

1. Discretize trajectory using grid-based chain coding (e.g., Freeman code).
2. Compute code length at varying grid resolutions for fractal dimension.
3. Estimate $f = f_1, f_2,\dots,f_N$8 via Lempel–Ziv factorization, estimate $f = f_1, f_2,\dots,f_N$9 using original vs. shuffled sequences (Peña-Mendieta et al., 2024).

• IEBF (Medical Curves):

1. Extract target and reference curves, compute pointwise differences $x_{i+1} - x_i$0.
2. Perform optional frequency filtering to separate spatial scales.
3. Estimate density $x_{i+1} - x_i$1 via kernel methods.
4. Map probabilities and compute $x_{i+1} - x_i$2 using Eq. (17) (Wang et al., 24 Jul 2025).

Efficient pattern statistics and entropy estimation are essential for performance, especially with large data or high embedding dimensions.

6. Case Studies and Experimental Validation

Tortuosity entropy has been validated on diverse spatiotemporal domains:

• Animal Movement: Tortuosity entropy (TorEn) detects behavioral transitions (e.g., straight migration vs. foraging), discriminates between deterministic and stochastic trajectories, and is robust under variable sampling. Real GPS tracks of Galapagos albatross show TorEn accurately distinguishing directed travel from tortuous search phases (Liu et al., 2013).
• Physiological Gait/Posture: On PhysioNet foot-force and center-of-pressure trajectories, entropy density ($x_{i+1} - x_i$3) and excess entropy ($x_{i+1} - x_i$4) distinguish between Parkinson’s disease, healthy controls, and fallers, with PD patients showing higher $x_{i+1} - x_i$5 (irregularity) and lower $x_{i+1} - x_i$6 (less pattern) (Peña-Mendieta et al., 2024).
• Chaotic Dynamics: In Hénon–Heiles orbital simulations, entropy density and excess entropy track transitions from regular (low $x_{i+1} - x_i$7, high structure) to chaotic (higher $x_{i+1} - x_i$8, lower $x_{i+1} - x_i$9) dynamical regimes (Peña-Mendieta et al., 2024).
• Medical Imaging: IEBF successfully quantifies tortuosity in meibomian gland atrophy, outperforming curvature- and arc-chord-based metrics in ROC analysis for discriminating pathological cases, and shows sensitivity to both local and global curve irregularities. Performance is robust to noise and parameter choices (Wang et al., 24 Jul 2025).

7. Practical Recommendations, Limitations, and Extensions

Recommendations for effective use include:

• Selection of symbolization/coarse-graining parameters (e.g., $a=2$0 or $a=2$1, threshold $a=2$2 for angles) to balance sensitivity and noise robustness (Liu et al., 2013).
• Embedding dimension $a=2$3 to trade off pattern richness and data sufficiency (Liu et al., 2013).
• Chain code grid refinement for targeted spatial resolution, and use of Lempel–Ziv factorization for reliable entropy rate estimation (Peña-Mendieta et al., 2024).
• Reference-curve selection in medical or anatomical contexts to ensure meaningful baseline irregularity discrimination (Wang et al., 24 Jul 2025).
• Frequency-domain decomposition for multi-scale tortuosity analysis (Wang et al., 24 Jul 2025).

Tortuosity entropy is inherently model-free and does not presuppose stationarity or deterministic vs. stochastic dynamics, making it broadly applicable to real-world, noisy, and heterogeneous data. The methodology is generalizable to various domains—retinal vessels, nerve fiber bundles, and vascular morphologies—provided suitable representation and, in comparative contexts, a reference “standard” curve for meaningful entropy evaluation (Wang et al., 24 Jul 2025).

A plausible implication is that as data scales grow and multi-scale structure becomes the norm, the integration of geometric, dynamical, and information-theoretic tortuosity metrics will become increasingly central in the quantitative analysis of biological and physical curves.