The Topological Stability Index: A Variance-Based Measure for Persistence Barcodes

This paper introduces the Topological Stability Index (TSI), a variance-based scalar measure for persistence barcodes that quantifies absolute lifetime dispersion and complements entropy-based summaries by capturing structural variability in stochastic fluctuations while remaining insensitive to deterministic trends.

The Gist

Imagine you are a detective trying to understand the shape of a mysterious object by looking at its "fingerprint." In the world of data science, this fingerprint is called a persistence barcode. It's a list of lines (or "bars") where the length of each line tells you how long a specific feature (like a hole or a loop) lasts as you zoom in and out of your data.

For a long time, scientists had a tool called Persistent Entropy to summarize these barcodes. Think of Persistent Entropy like a chef tasting a soup and only caring about the ratio of ingredients. If you have a soup with 1 part salt and 99 parts water, or a soup with 10 parts salt and 990 parts water, the ratio is the same. The chef says, "This tastes the same."

But what if the size of the soup matters? What if one pot is a tiny cup and the other is a giant bathtub? The ratio is the same, but the experience is totally different. The old tools couldn't tell the difference between a tiny, uniform soup and a massive, chaotic one.

This paper introduces a new tool called the Topological Stability Index (TSI) to fix that.

The New Tools: TSI and TSigI

The authors propose a two-part system to describe a barcode, like describing a crowd of people by their average height and their variety of heights.

1. The Topological Signal Index (TSigI): The "Average Height"

   • What it is: This measures the typical size of the bars.
   • The Analogy: Imagine a group of people. TSigI tells you the average height of the group. If everyone is 6 feet tall, the average is 6. If you have one giant and many tiny people, the average might still be 6, but it doesn't tell the whole story. It captures the "signal strength" or the general scale of the features.

2. The Topological Stability Index (TSI): The "Height Variance"

   • What it is: This measures how spread out the bar lengths are. It calculates the variance (the statistical spread).
   • The Analogy: Back to the crowd.
     • Scenario A: Everyone is exactly 6 feet tall. The "spread" is zero. The TSI is low.
     • Scenario B: You have one person who is 7 feet tall and another who is 5 feet tall. The average is still 6, but the group is "messy" or "heterogeneous." The TSI is high.
   • Why it matters: The TSI is sensitive to the absolute differences. It can tell you if a barcode has a few huge, dominant features and many tiny ones (high TSI), versus a barcode where all features are roughly the same size (low TSI).

The Secret Connection: The "Normalized" Version

The authors also created a "normalized" version called cvTSI.

• The Analogy: Imagine you want to compare the "messiness" of a small puddle to a massive ocean. You can't just measure the raw spread of waves because the ocean is naturally bigger. You have to normalize it.
• The Magic Link: The paper proves that this normalized messiness (cvTSI) is mathematically linked to a concept from information theory called Rényi Entropy.
  • Think of it like two different languages describing the same story. One language (Entropy) uses logarithms to compress the story, while the other (cvTSI) uses a straight line (variance). They tell you the same thing about the distribution of the bars, but they emphasize different details. The paper shows you can translate perfectly between them.

What the Experiments Showed

The authors tested these tools on synthetic data (like computer-generated shapes and random time series) to see how they behave compared to the old tools.

1. Deterministic vs. Random:

   • When they added a steady, predictable trend (like a straight line going up) to their data, the old tools (Entropy) and the new tools (TSI) didn't change much. They are good at ignoring boring, predictable patterns.
   • However, when they added random noise (like static on a radio or shaking a camera), the TSI jumped up. It is very good at detecting "chaos" or random fluctuations. It tells you, "Hey, the features are all over the place!"

2. The "Short Bar" Problem:

   • The paper admits a quirk: If you add a tiny, almost invisible bar to your list, the TSI changes. It's like adding one very short person to a room of giants; the "variance" of the room changes instantly.
   • The old Entropy tool is smoother and doesn't care as much about adding a tiny bar.
   • The Takeaway: TSI is great for seeing big structural changes and random noise, but it's a bit "jumpy" if your data has a lot of tiny, noisy features.

Summary in Plain English

• Old Way (Entropy): "How evenly are the features distributed?" (Ignores the actual size).
• New Way (TSI + TSigI): "How big are the features on average?" (TSigI) AND "How much do they vary in size?" (TSI).
• The Result: The new tools give you a better picture of structural variability. They can tell the difference between a system that is uniformly chaotic and one that has a few dominant features mixed with noise. They are particularly good at spotting random fluctuations in data, which the old tools sometimes miss.

In short, the paper gives data scientists a new ruler (TSI) to measure the "messiness" of their data's shape, complementing the old ruler that only measured the "balance" of the shape.

Technical Summary

Technical Summary: The Topological Stability Index

Problem Statement

Topological Data Analysis (TDA) utilizes persistence diagrams and barcodes to represent the evolution of topological features across scales. While these representations are rich and stable, integrating them with standard statistical tools remains challenging due to the lack of a simple linear or convex structure in the space of persistence diagrams.

Existing scalar summaries, such as persistent entropy, address this by mapping barcodes to single values. However, persistent entropy relies on the normalized distribution of persistence lifetimes (relative weights). Consequently, it is scale-invariant and fails to capture absolute dispersion or differences in the magnitude of persistence lifetimes. In many applications, absolute differences in scale and variability are meaningful indicators of structural heterogeneity, yet they are lost in entropy-based summaries. There is a need for a scalar measure that quantifies the absolute dispersion of persistence lifetimes while remaining sensitive to structural heterogeneity.

Methodology

The authors introduce the Topological Stability Index (TSI), a variance-based scalar measure defined as the sample variance of the multiset of persistence lifetimes.

1. Definition and Core Properties

Let $B$ be a persistence barcode with $n_B$ bars and lifetimes $\ell_i = d_i - b_i$. The TSI is defined as:
$$\text{TSI}(B) := \text{Var}(L_B) = \frac{1}{n_B - 1} \sum_{i=1}^{n_B} \left( \ell_i - \frac{L_B}{n_B} \right)^2$$
where $L_B = \sum \ell_i$ is the total persistence.

Key mathematical properties established include:

• Scaling: The TSI scales quadratically ($c^2$) under uniform scaling of the filtration values.
• Translation Invariance: The TSI is invariant under uniform translation of death times (shifting all lifetimes by a constant), provided the number of bars remains fixed.
• Extremal Characterization: For a fixed number of bars and fixed total persistence, the TSI is minimized (zero) when all lifetimes are equal and maximized when persistence is concentrated in a single bar.
• Update Formulas: Explicit recursive formulas are derived for the TSI under the insertion or deletion of a bar, showing sensitivity to the deviation of the new bar's length from the existing mean.
• Stability: While the TSI is not continuous under the insertion of arbitrarily short bars (due to changes in sample size normalization), it admits quantitative bounds relative to the empty diagram and the bottleneck distance when the number of bars is fixed.

2. Complementary Signal Index

To capture the typical scale of lifetimes, the authors define the Topological Signal Index (TSigI):
$$\text{TSigI}(B) := \frac{\sum \ell_i^2}{\sum \ell_i}$$
This is interpreted as a persistence-weighted mean lifetime. Together, $(\text{TSigI}(B), \text{TSI}(B))$ form a two-dimensional summary encoding both the magnitude (signal strength) and the dispersion (structural variability) of the barcode.

3. Normalized Version and Entropy Connection

To bridge the gap between variance-based and entropy-based summaries, a normalized version, cvTSI, is introduced:
$$\text{cvTSI}(B) := \frac{\text{TSI}(B)}{(\bar{\ell}_B)^2}$$
where $\bar{\ell}_B$ is the mean bar length.

• Scale Invariance: cvTSI is invariant under uniform scaling.
• Relation to Rényi Entropy: The authors prove an exact algebraic relation between cvTSI and the Rényi entropy of order two ($H_2$). Specifically, cvTSI is an affine function of the collision probability $\sum p_i^2$ (where $p_i$ are normalized lifetimes). Thus, cvTSI is a monotone reparametrization of $H_2$.
• Taylor Expansion: Near the uniform distribution, the persistent entropy $E(B)$ can be approximated as a linear function of cvTSI, showing that cvTSI captures the leading quadratic deviation of entropy from its maximum.

Key Results

The paper validates the theoretical properties and practical utility of TSI through numerical experiments on synthetic geometric data and stochastic time series:

1. Geometric Configurations (Circles):

   • In disjoint and intertwined circle models, TSI converges rapidly to an asymptotic value as sampling density increases, demonstrating robustness to sampling density.
   • Unlike persistent entropy, which depends heavily on the convergence of birth times to zero, TSI remains invariant under uniform translations of the barcode (e.g., varying sample size in disjoint circles).
   • TSI is sensitive to local perturbations (short-lived bars), whereas entropy reflects the overall balance of the normalized distribution.

2. Noise Robustness:

   • Under increasing Gaussian or uniform noise, TSI decreases rapidly toward zero as dominant features are destroyed and lifetimes become uniformly small.
   • In contrast, persistent entropy increases monotonically as the distribution of lifetimes becomes more uniform (many short-lived features).
   • cvTSI exhibits non-monotone behavior, peaking when a mixture of prominent and short-lived features exists, before decreasing as noise dominates.

3. Stochastic Time Series (Geometric Brownian Motion):

   • When analyzing GBM, TSI is largely insensitive to deterministic trends (drift) but responds strongly to stochastic fluctuations (volatility).
   • Increasing volatility leads to higher TSI values, reflecting increased dispersion in persistence lifetimes.
   • This contrasts with entropy, which shows only weak dependence on drift and moderate dependence on volatility.

Significance and Claims

The paper claims that the Topological Stability Index provides a necessary complement to existing entropy-based summaries in TDA. Its primary contributions are:

• Capturing Absolute Dispersion: Unlike persistent entropy, TSI quantifies the absolute variability of persistence lifetimes, making it sensitive to heterogeneous feature scales and structural complexity that entropy misses.
• Unified Perspective: Through the normalized cvTSI, the paper establishes a direct mathematical link between variance-based measures and information-theoretic summaries (Rényi entropy), unifying two distinct approaches to scalar summarization.
• Complementary Sensitivity: The experiments demonstrate that TSI and entropy capture different aspects of data structure. TSI is relatively insensitive to deterministic trends but highly responsive to stochastic fluctuations and variations in persistence magnitude.
• Two-Dimensional Summary: The pair $(\text{TSigI}, \text{TSI})$ offers a simple, interpretable two-dimensional summary that encodes both the typical scale of topological features and their structural variability.

The authors conclude that while TSI has limitations regarding continuity under bar insertion and dependence on the number of bars, it serves as a robust descriptor for structural heterogeneity, particularly in scenarios where absolute scale and dispersion are critical. Future work is suggested in developing functional analogues within the persistence-curve framework and studying asymptotic behaviors for statistical inference.