The Small-Noise Limit of the Most Likely Element is the Most Likely Element in the Small-Noise Limit

Zachary Selk, Harsha Honnappa

TL;DR

This paper resolves a long-standing "conundrum" in stochastic analysis by proving that two different mathematical frameworks for finding the "most likely" path of a system—the Onsager-Machlup function and the Freidlin-Wentzell rate function—are fundamentally consistent. The authors show that as the noise in a system vanishes, the path that maximizes the probability density (Onsager-Machlup) converges to the path that minimizes the cost of a rare deviation (Freidlin-Wentzell).

What Problem Does This Paper Address?

In the world of stochastic processes (systems influenced by random noise), researchers often want to identify the "most likely" behavior or path. However, "most likely" can be defined in two distinct ways:

  1. The Onsager-Machlup (OM) approach: This treats the problem like finding the "mode" of a distribution. Since infinite-dimensional spaces (like the space of all possible paths) don't have a standard uniform measure to define a density against, the OM function acts as a surrogate density.
  2. The Freidlin-Wentzell (FW) approach: This comes from Large Deviation Theory, which studies the exponential decay of probabilities of rare events. The FW rate function acts as a "Lagrangian" or a cost function; the path that minimizes this cost is considered the most likely path in the "small-noise limit."

While both theories claim to find the most likely element, they arise from different mathematical foundations. Previous research had shown they align for specific types of Stochastic Differential Equations (SDEs), but a general proof for broader classes of measures was missing. This paper asks: Is the small-noise limit of the OM-mode actually the same as the FW-mode for any measure equivalent to a Gaussian?

Methods

The authors employ a rigorous analytical approach centered on $\Gamma$-convergence (Gamma-convergence). Unlike standard pointwise convergence, $\Gamma$-convergence is specifically designed for optimization problems; it ensures that if a sequence of functions converges, their minimizers also converge to the minimizer of the limit function.

Key technical components include:

  • Generalized Tilting Lemmas: The authors develop new "tilting" lemmas that allow them to transfer Large Deviation Principles and Onsager-Machlup functions from a simple reference Gaussian measure to more complex, "equivalent" measures.
  • Girsanov’s Theorem: This is used to handle measures that are "equivalent" (meaning they agree on which sets have zero probability), such as the laws of solutions to SDEs compared to pure Brownian motion.
  • Infinite-Dimensional Analysis: The proofs are situated in separable Banach spaces, allowing the results to apply to a wide range of mathematical objects beyond simple paths.

Datasets

As a theoretical mathematics paper, no empirical datasets were used. The authors instead validate their results using mathematical "examples" or models:

  • Path-dependent SDEs: Systems where the future state depends on the entire past history.
  • Systems of Random Algebraic Equations: A non-differential example demonstrating the breadth of the theory.
  • Stochastic Partial Differential Equations (SPDEs): Mentioned as a primary motivation and application area.

Key Insights and Ideas

  • Reconciling the Conundrum: The paper proves that the OM function (scaled by the noise factor $\epsilon^2$) $\Gamma$-converges to the FW rate function. This provides a rigorous bridge between "density-based" likelihood and "large deviation-based" likelihood.
  • The Importance of $\Gamma$-Convergence: The authors highlight that pointwise convergence isn't enough to claim the "most likely" path is preserved. By proving $\Gamma$-convergence, they guarantee that the actual modes (the paths we care about) behave predictably as noise disappears.
  • The "Gaussian Test": In Remark 4.4, the authors note that the difference between the scaled OM function and the FW function can actually serve as a test for whether a measure is Gaussian. If the remainder term is constant, the system behaves quadratically (like a Gaussian); if not, it reveals non-Gaussian characteristics.

Key Results

  • Theorem 1.1: The flagship result. It establishes that for measures equivalent to a Gaussian, the scaled Onsager-Machlup function converges (pointwise and via $\Gamma$-convergence) to the Freidlin-Wentzell rate function as noise $\epsilon \to 0$.
  • Mode Convergence: They prove that any cluster point of the "most likely elements" (minimizers of OM) is a minimizer of the FW rate function.
  • Equicoercivity: The paper shows that if the functions are "equicoercive" (a technical condition ensuring minimizers don't escape to infinity), the limit of the OM-modes is exactly the FW-mode.

Limitations and Future Work

  • Equivalence to Gaussian Measures: The current framework relies on the measure being equivalent to a centered Gaussian measure. While this covers many SDEs and SPDEs, it does not cover systems with more singular perturbations or those that deviate fundamentally from Gaussian noise.
  • Regularity Conditions: The "tilting" lemmas require certain regularity (smoothness) conditions on the density function ($\Phi$). Exploring the limits of these conditions—specifically for "rougher" systems—remains an open area.
  • Computational Application: While the paper provides the theory, the authors suggest that these insights could improve numerical methods for finding transition paths in complex systems (like chemical reactions or climate models), though they do not implement these algorithms here.

Summary updated

Abstract

In this paper, we study the Onsager-Machlup function and its relationship to the Freidlin-Wentzell function for measures equivalent to arbitrary infinite dimensional Gaussian measures.The Onsager-Machlup function can serve as a density on infinite dimensional spaces, where a uniform measure does not exist, and has been seen as the Lagrangian for the "most likely element".The Freidlin-Wentzell rate function is the large deviations rate function for small-noise limits and has also been identified as a Lagrangian for the "most likely element".This leads to a conundrum -what is the relationship between these two functions?We show both pointwise and Γ-convergence (which is essentially the convergence of minimizers) of the Onsager-Machlup function under the small-noise limit to the Freidlin-Wentzell function -and give an expression for both.That is, we show that the small-noise limit of the most likely element is the most likely element in the small noise limit for infinite dimensional measures that are equivalent to a Gaussian.Examples of measures include the law of solutions to path-dependent stochastic differential equations and the law of an infinite system of random algebraic equations.
BibTeX
@inproceedings{selk2024small,
  author = {Zachary Selk and Harsha Honnappa},
  title = {{The Small-Noise Limit of the Most Likely Element is the Most Likely Element in the Small-Noise Limit}},
  booktitle = {Latin},
  year = {2024},
  doi = {10.30757/alea.v21-35},
}
Figure