The Cost of Learning under Multiple Change Points

We consider an online learning problem in environments with multiple change points. In contrast to the single change point problem that is widely studied using classical “high confidence” detection schemes, the multiple change point environment presents new learning-theoretic and algorithmic challenges. Specifically, we show that classical methods may exhibit catastrophic failure (high regret) due to a phenomenon we refer to as endogenous confounding. To overcome this, we propose a new class of learning algorithms dubbed Anytime Tracking CUSUM (ATC). These are horizon-free online algorithms that implement a selective detection principle, balancing the need to ignore “small” (hard-to-detect) shifts, while reacting “quickly” to significant ones. We prove that the performance of a properly tuned ATC algorithm is nearly minimax-optimal; its regret is guaranteed to closely match a novel information-theoretic lower bound on the achievable performance of any learning algorithm in the multiple change point problem. Experiments on synthetic as well as real-world data validate the aforementioned theoretical findings.

💡 Research Summary

The paper tackles the problem of online learning in environments where the underlying data‑generating distribution undergoes multiple abrupt changes, a setting that has received far less attention than the classic single‑change‑point scenario. The authors first formalize the task as tracking a piecewise‑constant mean µₜ of a sub‑Gaussian sequence {Xₜ} over a horizon T with an unknown number S of change points. Performance is measured by dynamic regret, i.e., the expected cumulative squared error Σₜ(µ̂ₜ‑µₜ)², which directly captures the cost of delayed or inaccurate adaptation.

A central contribution is the identification and rigorous quantification of “endogenous confounding”: when a change point is missed, the learner continues to use outdated samples in its reference statistics, creating a mixture of pre‑ and post‑change distributions. This mixture reduces the signal‑to‑noise ratio of subsequent detection statistics, making later change points progressively harder to detect. Lemma 3.1 formalizes how the inclusion of stale data inflates variance and biases the estimator, thereby increasing future regret.

To address this, the authors propose the Anytime Tracking CUSUM (ATC) algorithm. ATC maintains a CUSUM‑style statistic ˆD_{r,k,t} for every candidate split point k between the most recent restart time r and the current time t. The statistic is a standardized difference of block means, analogous to a generalized likelihood‑ratio test for unknown means. Crucially, ATC employs a time‑varying detection threshold η_t that decays roughly as √{log t}. This adaptive threshold implements a “selective detection” principle: large shifts quickly exceed the threshold and trigger a restart, while small or short‑lived shifts remain below the threshold and are deliberately ignored, incurring only a modest regret penalty.

Theoretical analysis yields two complementary results. Theorem 4.1 proves a non‑asymptotic upper bound on dynamic regret: R_T ≤ C·σ²·(S+1)·log T, where σ² is the sub‑Gaussian variance proxy and C is a problem‑dependent constant. This bound holds without any detectability assumptions on the magnitude or spacing of the changes, demonstrating that ATC remains robust even when some shifts are statistically indistinguishable. Theorem 4.2 establishes an information‑theoretic lower bound of Ω(σ²·(S+1)·log(T/(S+1))) (up to lower‑order log‑log terms), showing that ATC’s regret is within a logarithmic factor of the minimax optimum.

Empirical validation is performed on both synthetic data—where change magnitudes, intervals, and noise levels are varied—and real‑world datasets, including ride‑hailing demand time series and binary source‑coding streams. Across all settings, ATC consistently outperforms traditional high‑confidence change‑detection methods (e.g., δ‑PAC), sliding‑window approaches, and discounting schemes. In particular, ATC mitigates the cascade of errors caused by endogenous confounding, maintaining low cumulative squared error even when change points are frequent or subtle.

Overall, the paper makes a substantial contribution to the literature on non‑stationary online learning. By exposing the pitfalls of naïvely extending single‑change‑point detection to multiple changes and by providing a principled, horizon‑free algorithm with near‑optimal regret guarantees, it opens new avenues for applications such as real‑time demand forecasting, adaptive resource allocation, and online compression where abrupt regime shifts are the norm. Future work may extend the framework to high‑dimensional settings, incorporate contextual information, or explore alternative loss functions, but the core insight—balancing selective detection against the cost of endogenous confounding—remains a powerful guiding principle.