An Eyring--Kramers Law for the Hypoelliptic Third-Order Langevin Diffusion

Technical Summary: An Eyring–Kramers Law for the Hypoelliptic Third-Order Langevin Diffusion

Problem Statement
This paper addresses the metastable transition dynamics of the hypoelliptic third-order Langevin diffusion in the low-temperature limit ($\varepsilon \downarrow 0$). The system is defined by the stochastic differential equation (SDE):
$$\begin{cases} d\theta_t = p_t dt, \\ dp_t = -L^{-1}\nabla U(\theta_t) dt + \gamma r_t dt, \\ dr_t = -\gamma p_t dt - \gamma r_t dt + \sqrt{2\gamma\varepsilon/L} dB_t, \end{cases}$$
where $\theta_t, p_t, r_t \in \mathbb{R}^d$. Unlike the classical overdamped Langevin diffusion (noise on $\theta$) or the underdamped (kinetic) Langevin diffusion (noise on $p$), the noise in this third-order model acts solely on the auxiliary variable $r_t$. Randomness propagates to the position $\theta_t$ through a three-level H"ormander chain: $r \to p \to \theta$.

The study focuses on a double-well potential $U$ with two non-degenerate local minima ($m, s$) separated by a unique index-one saddle ($\sigma$). The primary objective is to determine the sharp asymptotics of the mean transition time from a neighborhood of $m$ to a neighborhood of $s$. Specifically, the authors seek to establish the Eyring–Kramers law, which characterizes the mean hitting time as:
$$\mathbb{E}[\tau] \sim \text{Prefactor} \times \exp\left(\frac{U(\sigma) - U(m)}{\varepsilon}\right).$$
While the Arrhenius exponential scale is well-understood via Freidlin–Wentzell theory, deriving the sharp prefactor for this degenerate, hypoelliptic system requires new analytical tools due to the characteristic boundary geometry and the higher-order noise propagation.

Methodology
The proof combines a weak-capacity framework with detailed local asymptotic analysis. The methodology proceeds through the following steps:

1. Weak-Capacity Framework: The authors utilize results from [HLWZ26] which establish a weak capacity–hitting identity for hypoelliptic diffusions. This identity relates the mean hitting time to the ratio of a global integral involving the adjoint committor and the weak capacity between the metastable sets. This approach avoids the need for strong boundary conditions that are difficult to satisfy for degenerate diffusions.
2. Saddle Capacity Asymptotics:
   • Linearization: The deterministic dynamics are linearized at the saddle point $\sigma$. The stability analysis reveals that the unstable rate is not a simple eigenvalue but the unique positive root $\mu_\sigma$ of a cubic characteristic polynomial derived from the third-order chain.
   • Gaussian Current Calculation: A saddle-adapted test function is constructed using the unstable dual coordinate $\Lambda_\sigma$. The capacity is reduced to a constrained Gaussian integral (a "Gaussian saddle current") across a linear section transverse to the unstable direction.
   • Global Patching: The local saddle profile is patched to the constant values in the potential wells using a global test function. This involves controlling "side layers" where the Hamiltonian sublevel sets connect to the saddle box, ensuring that contributions from these regions are exponentially negligible.
3. Well Asymptotics and Flatness:
   • Mass Localization: The numerator of the capacity identity is shown to concentrate near the starting minimum $m$ using Laplace asymptotics and Hamiltonian sublevel localization.
   • Intrawell Flatness: A uniform Harnack estimate and local hypoelliptic regularity (derived from third-order heat-kernel and Malliavin bounds) are used to prove that the mean transition time is essentially constant within the shrinking neighborhood of the starting well. This allows replacing the boundary average in the weak capacity identity with the value at the specific starting point $(m, 0, 0)$.

Key Contributions

• Derivation of the Third-Order Unstable Rate: The paper identifies that the prefactor is governed by $\mu_\sigma$, the unique positive root of the cubic equation:
  $$\mu^3 + \gamma\mu^2 + \left(\gamma^2 - \frac{\lambda_\sigma}{L}\right)\mu - \frac{\gamma\lambda_\sigma}{L} = 0,$$
  where $-\lambda_\sigma$ is the unique negative eigenvalue of the Hessian $\nabla^2 U(\sigma)$. This contrasts with the quadratic equation governing the underdamped case.
• Explicit Prefactor Formula: The authors prove that as $\varepsilon \downarrow 0$:
  $$\mathbb{E}_{(m,0,0)}[\tau_{S_\varepsilon}] = [1 + o_\varepsilon(1)] \frac{2\pi}{\mu_\sigma} \sqrt{\frac{-\det \nabla^2 U(\sigma)}{\det \nabla^2 U(m)}} \exp\left(\frac{U(\sigma) - U(m)}{\varepsilon}\right).$$
• Comparison with Underdamped Dynamics: Under matched kinetic normalizations, the paper demonstrates that the third-order unstable rate $\mu_\sigma$ is strictly larger than the underdamped rate $\nu_\sigma$. Consequently, the metastable prefactor for the third-order diffusion is strictly smaller than that of the underdamped counterpart.
• Technical Novelty: The work extends the two-level underdamped construction to a three-level chain. This necessitates handling anisotropic short-time scales ($t^{1/2}, t^{3/2}, t^{5/2}$), uniform third-order density estimates, and a more complex global capacity construction involving physical side layers to patch the saddle profile to the well components.

Results

• Theoretical: Theorem 2.7 establishes the Eyring–Kramers law for the third-order Langevin diffusion. The result confirms that the Arrhenius exponent remains determined by the potential barrier height, while the prefactor is determined by the specific geometry of the third-order linearization at the saddle.
• Numerical: A numerical experiment on a one-dimensional double-well potential ($U(\theta) = \frac{1}{4}(\theta^2-1)^2$) validates the theoretical predictions. Simulations of 20,000 trajectories per temperature show that the ratio of the mean transition times between the third-order and underdamped models converges to the theoretical ratio $\nu_\sigma / \mu_\sigma \approx 0.819$ as the temperature decreases, confirming the predicted prefactor improvement.

Significance and Claims
The paper claims to provide the first rigorous derivation of the sharp Eyring–Kramers prefactor for a hypoelliptic third-order Langevin diffusion. Its significance lies in:

1. Completing the Hierarchy: It fills a gap in the understanding of high-order Langevin dynamics, which are increasingly used in accelerated sampling algorithms and non-convex optimization.
2. Quantifying Acceleration: By proving that the third-order prefactor is strictly smaller than the underdamped prefactor, the paper provides a theoretical basis for the observation that higher-order Langevin dynamics can accelerate metastable transitions.
3. Methodological Advancement: The combination of weak-capacity techniques with explicit Gaussian current calculations and intrawell flatness estimates offers a robust framework for analyzing metastability in degenerate, non-reversible diffusions where standard elliptic methods fail.

The authors remain modest regarding broader implications, noting that their result applies specifically to the double-well geometry under Assumptions 2.1 and 2.2, and that the comparison of prefactors is strictly valid under matched kinetic normalizations.