On revolutions

The paper addresses the longstanding challenge of defining and detecting "revolutions"—swift, profound, multivariate changes—across politics, culture, and science. While historians and scientists often invoke revolutions or paradigm shifts, empirical identification is contentious, with debates about continuity versus discontinuity. The authors propose giving the detection of revolutions a statistical foundation: define revolutions as periods with significantly elevated multivariate rates of change relative to background. They argue that revolutions typically involve many variables changing in concert and at speed, and that multivariate time series offer the most direct way to estimate rates and detect such accelerations. Building on Foote novelty, they introduce a non-parametric, permutation-based method to locate revolutionary and conservative periods in multivariate time series and apply it to diverse historical, political, and cultural datasets.

The paper situates its contribution within broad scholarly discourse on revolutions across domains: political revolutions (e.g., French Revolution), scientific revolutions (Kuhn), and debates critiquing or minimizing such claims (Schama; Shapin; skepticism of a Darwinian Revolution). It highlights analogous debates in historical natural sciences (e.g., the "Human Revolution" in archaeology; punctuated equilibrium in paleontology and its contentious empirical support). The notion of punctuated change has diffused into linguistics, technology, and socio-political evolution, with models explaining punctuated dynamics. Existing methods detect regime/phase shifts in univariate or parametric multivariate contexts; evolutionary biology often relies on phylogenetic methods to infer rate variation. However, revolutions commonly involve many variables changing concurrently, motivating a multivariate, time-series-based non-parametric approach that can capture coordinated, rapid change without assuming time-invariant parametric relationships.

Overview: Revolutions are defined as statistically significant local increases in multivariate rate of change. The method computes a temporal distance matrix over all pairs of time periods using any suitable multivariate distance metric. Revolutionary periods manifest as checkerboard patterns in the distance matrix: low within-block variability (before and after) and high cross-block variability (between periods). Foote novelty is used to quantify these patterns.

Foote novelty computation: For each target time t, a checkerboard-like kernel with half-width k (size 2k+1) is centered on the main diagonal of the distance matrix D. The kernel has negative weights on local (diagonal) blocks and positive weights on cross (off-diagonal) blocks, with two modifications: (1) a radially symmetric Gaussian taper (SD = 2 × 0.4k) to reduce edge effects and weight nearer distances more heavily; (2) a central cross of zeros, aligning the kernel center with time t. The Foote novelty score F_t is the Hadamard product sum of the kernel with the corresponding submatrix of D. Sliding across t yields F_t over time for each k. Different k capture different temporal scales; small k resolve short-term boundaries; large k capture longer spans.

Statistical inference: To distinguish extraordinary from background variation, a non-parametric test compares observed F_t with a null distribution generated by permuting each diagonal (excluding the zero main diagonal) of D independently, preserving matrix symmetry and more of D's structure than axis permutations. This corresponds to block-wise time permutations of varying lengths ≥1. For each k and valid t, p-values are obtained from the permutation distribution; significance is assessed (authors use α = 0.05/2 with Bonferroni-style consideration across tests). Conservative periods (unusually low change) can also be detected.

Rate index: To visualize scale-agnostic relative rate fluctuations, the rate index R_t is computed by standardizing F_t^k by its mean over time for each k and averaging across a set K of k values valid at t. R_t > 1 indicates above-average local rate; R_t < 1 indicates below-average. R_t is descriptive; inference relies on F_t significance.

Variable contributions: To attribute revolutions to variables, re-run analysis after removing individual variables (or combinations) and assess changes in significant F_t within the revolution window.

Sensitivity/specificity via simulations: The authors simulate multivariate time series varying persistence p (AR(1) from 0 to 1), revolution strength s, revolution length l, number of variables n, and kernel half-width k. They evaluate Type I error (no revolution) and Type II error (with revolution), and recommend differencing when persistence exceeds ~0.25 to control false positives. They note reduced power for short/weak revolutions with few variables and for very small k. Practical guidance: estimate persistence p; if p ≤ 0.25, analyze levels; otherwise, analyze first differences. Examine multiple k and identify revolutions consistent across scales; use smallest k that detects the revolution to estimate boundaries most precisely.

Simulations and test performance:

• Type I error: On level series without revolutions (s=0), false positives align with α when p=0 but rise with persistence, reaching ~16% for random walks (p≈1). After first differencing, false positives are at or below α across p. Threshold where risk rises: p≈0.25.
• Type II error: On level series with revolutions (s>0), mean false negative rate ≈22% across p. Differencing reduces power in stationary series but only slightly in highly persistent series. The method often fails for short (l≤6 time units) and weak (s≤0.5) revolutions with few variables (n≤10), especially with very small k (k=1).

Applications to real datasets:

1. Global democracy (V-Dem indices: freedom of expression, association, clean elections, elected executive, suffrage; yearly means over extant states): Persistence p≥0.25, so first differences used. Conducted 3192 tests over all k; 208 significant (vs. 78 expected at α=0.05/2), indicating real revolutions. Identified significant high-rate periods (revolutions): 1944–1949, 1962, 1975–1985, 1989–1996. Interpretation: aligns with Huntington’s waves, with the third wave split into two sub-waves. 1962 is a “reverse wave” marked by anti-democratic coups in several countries. Variable contributions: late 1940s revolution driven by political structure changes; 1977–1984 and 1989–1996 driven by increases in personal liberty.

2. US pop music (Billboard Hot 100, 1960–2010; 16 harmonic/timbral features; highly persistent, analyzed in first differences): R_t > 1 during several intervals; 552 tests, 58 significant (vs. 14 expected), yielding revolutions at 1968–1969, 1982–1983, 1986–1988. Interpreted as: mid-1960s rise of rock-related chords and aggressive percussion; early 1980s revival of guitar-heavy rock and drum machines; late 1980s rise of hip hop at expense of rock/pop timbres.

3. Crime rates per 100,000, England and Wales, 1900–2000: 2304 tests; 138 significant (vs. 58 expected). Revolutions: 1965–1978 (general increase across categories) and 1989–1995 (general decrease), consistent with the post-1960 crime rise and the 1990s decline.

4. US newborn girls’ names, 1945–2010: Strong evidence of revolutions, notably 1974–1975 and 1988–1991, reflecting coordinated fashion cycles (e.g., rises of Jessica, Ashley, Lauren, Amanda, Amber; later Emma, Isabella, Olivia, Hannah).

5. Additional datasets (US car models 1950–2010; BMJ articles 1960–2008; English/Irish/American novels 1840–1890): Method applicable; car models showed much change but no revolutions over the observed period; BMJ and novels did not present strong revolutionary signals in the reported summary.

General: Method identifies both revolutionary and conservative periods; revolutions are rare, domain-dependent, and differ in causes and variable contributions.

The study provides a statistical framework to adjudicate debates about revolutionary versus gradual change by focusing on significant local increases in multivariate rates. By using Foote novelty with diagonal-permutation nulls, the method detects coordinated, rapid changes across variables while being agnostic to specific data distributions and accommodating diverse distance measures. The findings validate well-known historical claims (e.g., democratization waves, pop music shifts) and uncover nuanced structures (e.g., sub-waves, reverse waves), demonstrating the method’s interpretive value and potential to attribute revolutions to specific variable sets. Simulations clarify conditions for reliable inference—stationarity or differencing to control false positives, adequate variable count and revolution span/strength for power—and guide practical application. The concept of revolution is sharpened: revolutions are local accelerations relative to background variability; in super-linear evolutions, the cutting edge may appear as a shifting, ongoing revolution. The method also offers descriptive tools (R_t) for visualizing rate landscapes, complementing significance testing, and suggests a path to distinguish structural (variance-covariance shifts) from non-structural revolutions in future work.

The paper introduces a general, non-parametric method to detect revolutions in multivariate time series as statistically significant local increases in the rate of change using Foote novelty and diagonal-permutation testing. It demonstrates applicability across political, social, and cultural domains, recovering known revolutions (e.g., democratization waves, pop music shifts) and identifying less emphasized ones (e.g., coordinated cycles in baby names, crime rate reversals). Simulations establish performance characteristics and practical guidance (assess persistence; difference when necessary; examine multiple temporal scales via k). The approach provides both inferential detection and descriptive visualization (R_t) of rate dynamics. Future research directions include distinguishing structural versus non-structural revolutions via changes in variance-covariance structures, refining multiple-testing adjustments across k and t, extending attribution methods for variable contributions, and applying the framework to biological and cultural datasets where variable salience may change over time.

• Stationarity requirement: On levels, Type I error inflates with persistence (notably above p≈0.25); first differencing is recommended for persistent series but can reduce power in truly stationary data.
• Power limitations: The test frequently fails to detect short (l≤6) and weak (s≤0.5) revolutions, especially with few variables (n≤10) and very small kernels (k=1).
• Edge effects and resolution: Power to detect revolutions diminishes at series boundaries; temporal resolution coarsens at larger k; smallest k that detects a revolution provides most precise boundary estimates.
• Multiple testing: Many tests across t and k increase false discovery risk; Bonferroni-like considerations are applied but may be conservative.
• R_t interpretability: R_t is descriptive and can be elevated without statistical significance, particularly near series ends.
• Dependence on data quality: Requires long, well-sampled multivariate series with low sampling variance relative to overall variability; distance metric choice may affect sensitivity.
• Structural identification: Distinguishing structural (variance-covariance) from non-structural revolutions is not resolved and left for future work.
• Changing variable salience: Emergence/disappearance of variables (e.g., technological shifts) can be handled but complicate interpretation; formal criteria for such cases are not developed.