Xingyu Wang, Chang-Han Rhee
5 min
Abstract
In this paper, we study the first-exit time of stochastic difference equation $X^η_{j+1}(x) = X^η_{j}(x) + ηa\big( X^η_{j}(x)\big) + ησ\big( X^η_{j}(x)\big)Z_{j+1}$ and its truncated variant $X^{η|b}_{j+1}(x) = X^{η|b}_{j}( x) + φ_b\big(ηa\big( X^{η|b}_{j}( x)\big) + ησ\big( X^{η|b}_{j}( x)\big) Z_{j+1}\big)$, where $φ_b(x) = (x/|x|)\min\{|x|, b\}$ and the law of the noise $Z_t$ is multivariate regularly varying. The truncation operator $φ_b(\cdot)$ is often introduced as a modulation mechanism in heavy-tailed systems, such as stochastic gradient descent algorithms in deep learning. By developing a framework that connects large deviations with metastability, we leverage the locally uniform sample-path large deviations for both processes in Wang and Rhee (2024) to obtain precise characterizations of the joint distributions of the first exit times and exit locations. The resulting limit theorem unveils a discrete hierarchy of phase transitions (i.e., exit times) as the truncation threshold $b$ varies, and manifests the catastrophe principle, whereby key events or metastable behaviors in heavy-tailed systems are driven by catastrophic behavior in a few components while the rest of the system behaves nominally. These developments lead to a comprehensive heavy-tailed counterpart of the classical Freidlin-Wentzell theory.
Alex: And where does the process land after escape?
Sam: The exit spot follows a specific measure that integrates over J-fold heavy-tail distributions times path lengths. It maps the drift path plus those J truncated jumps to points just outside the basin. They develop a general framework for Markov chains using asymptotic atoms—reliable reset zones the process hits quickly, like respawning at checkpoints in a game level. These atoms cover the basin densely, and exits behave uniformly from them.
Alex: So the atoms let them prove the joint law sharply, by breaking escapes into regeneration cycles. For the truncated dynamics, they verify atoms exist with the right scale.
Sam: Yes. To make it rigorous, they shrink the basin slightly—creating a smaller safe zone where every point is a tiny distance ε away from the edge. They prove lemmas showing paths with exactly J minus one truncated jumps land deep inside, requiring another jump to escape.
Alex: Those lemmas glue the multi-jump paths to the general theory, ensuring scaled exit time goes exponential and location to the measure.
Sam: In machine learning, this ties to SGD where gradients get clipped: if the update exceeds norm b, project it back to length b—like capping a vector pull to safe strength. The numerics trace those predicted phase jumps as clip tightens.
Alex: What's the main takeaway for someone following machine learning training?
Sam: The paper delivers a heavy-tailed version of classical theory tailored to truncated dynamics like clipped SGD. It shows escapes from basins happen via a fixed number of jumps J, creating discrete shifts in time scales that match simulations. This links rare-event math to why models dwell longer in wide flat regions.
Alex: Instead of smooth escapes, it's these stepped phases. Any caveats in the setup?
Sam: The analysis assumes smooth drift and diffusion functions, and multivariate regular variation for the noise tails. Numerics are only in one dimension so far, with global behavior covered in a companion paper.
Alex: That's a meaningful synthesis from multi-jump math to why training behaves this way in clipped setups. Thanks, Sam—this has been a clear dive into the mechanics.
Sam: My pleasure, Alex. The work stands as a notable advance in rare-event analysis for optimization.
Alex: And that's our look at first-exit times in truncated heavy-tailed systems. Thanks for joining ResearchPod.