Paper: arXiv 2512.16411

Authors: Matthieu Garcin, Louis Perot

Abstract

Relative entropy, as a divergence metric between two distributions, can be used for offline change-point detection and extends classical methods that mainly rely on moment-based discrepancies. To build a statistical test suitable for this context, we study the distribution of empirical relative entropy and derive several types of approximations: concentration inequalities for finite samples, asymptotic distributions, and Berry-Esseen bounds in a pre-asymptotic regime. For the latter, we introduce a new approach to obtain Berry-Esseen inequalities for nonlinear functions of sum statistics under some convexity assumptions. Our theoretical contributions cover both one- and two-sample empirical relative entropies. We then detail a change-point detection procedure built on relative entropy and compare it, through extensive simulations, with classical methods based on moments or on information criteria. Finally, we illustrate its practical relevance on two real datasets involving temperature series and volatility of stock indices.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper is mathematically dense, featuring theoretical contributions like Berry-Esseen bounds and concentration inequalities for nonlinear functions of sum statistics, indicating high math complexity. It also demonstrates empirical rigor through extensive simulations comparing the method to baselines and applications to real-world datasets (temperature and volatility series), making it backtest-ready.

Research Flowchart

  flowchart TD
  G["Research Goal:<br>Distribution of Empirical Relative Entropy for Change-Point Detection"] --> M["Methodology:<br>Derive Concentration Inequalities, Asymptotic Distributions, and Berry-Esseen Bounds"]
  M --> C["Computational Process:<br>Simulation of Offline Change-Point Detection Procedure"]
  C --> D["Data/Inputs:<br>Temperature Series &<br>Stock Market Volatility"]
  C --> E["Comparisons:<br>Moment-based &<br>Information Criterion Methods"]
  D & E --> F["Key Findings/Outcomes:<br>Theoretical Bounds & Practical Efficacy in Detecting Changes"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an offline change‑point detection test based on the empirical relative entropy between two sub‑samples.

Source sentence
The method we put forward here is based on the empirical relative entropy DKL (bpn∥bqn), with the two distributions bpn and bqn estimated on the two sub‑samples.

Assumes 6

  • Under the null hypothesis the whole sequence is i.i.d.; under the alternative each half follows a different discrete distribution. statistical
    Source sentence
    Under the null hypothesis, X1, ..., X2n are identically distributed. Under the alternative hypothesis, there exists t⋆∈J2, 2n −1K such that X1, ..., Xt⋆are identically distributed and follow a discrete probability q, whereas Xt⋆+1, ..., X2n are also identically distributed but follow another discrete probability p.
  • The test is applied in an offline setting with the change‑point at the centre (t⋆ = n). statistical
    Source sentence
    In the offline case, one generally focuses on t⋆= n [47], so that people consider a divergence statistic between the probabilities of the two sub‑samples.
  • The two sub‑samples have equal size n. data
    Source sentence
    Our test thus consists in X1, ..., Xn ∼q and Xn+1, ..., X2n ∼p, with H0 corresponding to p = q and H1 to p̸ = q.
  • The number of categories k is limited to k ≤ 8 in the two‑sample case to avoid empty categories. model
    Source sentence
    In the two‑sample case, we also constrain k to be less than or equal to 8; otherwise, simulations sometimes lead to empty categories, and consequently to zero estimated probabilities bqn, when n = 100.
Show 2 more
  • Agrawal’s condition x > (k−1)/n is always satisfied in the reported results. statistical
    Source sentence
    Agrawal’s condition x > (k−1)/n is always verified in our results and we have to discard the case k = 2 for Mardia’s bound.
  • The analysis focuses on relatively small sample sizes (n < 50) where concentration inequalities are tighter than Berry‑Esseen bounds. other
    Source sentence
    In addition, as we will see in Section 4.1, concentration inequalities give tighter bounds than the Berry‑Esseen approach when n is small.

Data

  • Embrun temperature series · Météo‑France macro public
    1999‑2024
    Source sentence
    We use a public dataset of temperatures at Embrun (Hautes‑Alpes, France), between January 1999 and December 2024, obtained from the website of
  • Realized volatility series of major stock indices · Oxford‑Man Institute of Quantitative Finance Realized Librar other public
    2000‑2021 · AEX, CAC 40, Nikkei 225, Oslo Exchange All‑share, S&P 500, VIX
    Source sentence
    time series of realized volatility computed with a five‑minute discretization of prices of stock indices, imported from the formerly available Oxford‑Man Institute of Quantitative Finance Realized Library

Recognised sources: Bloomberg

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The relative‑entropy based test maintains the nominal confidence level while moments‑based tests (t‑test, F‑test) lose power in some scenarios. statistical
    Source sentence
    The t‑test is now unable to detect change‑points when they exist. The ∆AIC method is liberal again, whereas all the other methods have an appropriate actual confidence level.
  • Concentration inequalities provide tighter finite‑sample bounds than Berry‑Esseen bounds for small n. theoretical
    Source sentence
    In addition, as we will see in Section 4.1, concentration inequalities give tighter bounds than the Berry‑Esseen approach when n is small.
  • The asymptotic quantile slightly underestimates the true quantile for n < 50, with larger underestimation in the tail. statistical
    Source sentence
    We observe that the quantile of the asymptotic distribution is very accurate: it only slightly underestimates the true quantile when n is small (less than 50) and this underestimation becomes more pronounced the further we move into the tail of the distribution.
  • Sanov’s bounds largely overestimate the quantile, while Agrawal’s bounds are close to the true quantile in both one‑ and two‑sample cases. theoretical
    Source sentence
    Sanov’s bounds largely overestimate the quantile. Mardia’s bound is only slightly better than the best Sanov’s bound, whereas the two Agrawal’s bounds, which are very close to each other both in the one‑ and the two‑sample cases…
  • Applications to climate and finance data show the method reveals climate change and changes in volatility dependence linked to macro‑economic events. performance
    Source sentence
    Two applications to real data, namely climate and finance datasets, emphasize the practical relevance of this method. It makes the climate change visible and unveils modifications in the serial dependence of volatility that we can relate to macroeconomic events.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 39 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.