Underdamped Langevin Dynamics

• Underdamped Langevin Dynamics is a framework described by SDEs that capture both momentum and position evolution under deterministic forces and thermal noise, accurately modeling inertial effects.
• It employs phase-space formulations and hypocoercivity analysis to prove exponential convergence to equilibrium and establish sampling efficiency across diverse regimes.
• Algorithmic implementations using randomized and geometric integrators optimize convergence rates and enhance applications in molecular simulation, Bayesian sampling, and thermodynamic studies.

Underdamped Langevin dynamics refers to a class of stochastic differential equations (SDEs) modeling the evolution of position and momentum variables $(x, p)\in\mathbb{R}^d\times\mathbb{R}^d$ under the influence of both deterministic forces (derived from a smooth potential $U(x)$) and thermal noise, with explicit retention of inertial effects. Unlike overdamped Langevin dynamics, which neglects momentum and models only the slow, friction-dominated regime, the underdamped variant retains a second-order (momentum) term, providing a more accurate description of systems where stochastic evolution occurs on time scales comparable to inertial relaxation. This framework is indispensable in molecular dynamics, Bayesian sampling, statistical physics, stochastic optimization, and non-equilibrium thermodynamics, offering sharper convergence, sampling complexity, and entropy production results across a wide spectrum of confining regimes, algorithmic protocols, and information-theoretic limits.

1. Formulation and Invariant Measure

Standard underdamped Langevin dynamics in $d$ dimensions employs phase-space coordinates $(x,p)$ for position and momentum. The SDE reads: $$\begin{cases} dx_t = p_t\,dt,\ dp_t = -\nabla U(x_t)\,dt - \gamma\,p_t\,dt + \sqrt{2\gamma}\,dW_t, \end{cases}$$ where $U(x)\in C^2(\mathbb{R}^d)$ is the potential, $\gamma>0$ is the friction coefficient, and $W_t$ is standard $d$-dimensional Brownian motion. The generator decomposes as $L = L_H + \gamma L_{\mathrm{FD}}$, with $$L_H = p \cdot \nabla_x - \nabla U(x) \cdot \nabla_p$$ and $L_{\mathrm{FD}} = \Delta_p - p \cdot \nabla_p$.

The resulting process is ergodic under mild growth and regularity assumptions on $U$. Its invariant measure is the Gibbs–Gaussian product: $\mu(dx\,dp) = Z^{-1} e^{-U(x) - |p|^2/2} dx\,dp.$ Marginalizing over $p$ gives the configuration-space equilibrium $\mu_x(dx) = Z_x^{-1} e^{-U(x)} dx$ (Cao et al., 2019).

2. Convergence Analysis and Hypocoercivity

Exponential convergence to equilibrium in $L^2(\mu)$ is governed by Poincaré-type inequalities. If the spatial marginal $\mu_x$ satisfies a Poincaré inequality, i.e.,

$$\int \bigl(g(x) - \mu_x(g)\bigr)^2 d\mu_x \leq C_P \int |\nabla_x g(x)|^2 d\mu_x,$$

then, for sufficiently regular $U$, the semigroup generated by $L$ contracts exponentially: $$\|P_t f - \mu(f)\|_{L^2(\mu)} \leq C e^{-\nu t} \|f - \mu(f)\|_{L^2(\mu)}.$$ The sharp rate is (Cao et al., 2019): $\nu = \frac{m \gamma}{c (\sqrt{m} + R + \gamma)^2}$ with $m$ the Poincaré constant, $R$ determined by convexity ($R=0$ for $U$ convex, $R = \sqrt{K}$ for $\nabla^2 U(x) \geq -K I$, and explicit dependence on the dimension and growth for non-convex cases).

Optimization over $\gamma$ yields maximal rate $\nu_{\max} \asymp \frac{\sqrt{m}}{4c}$ for $m \ll 1$ (strongly confined, small noise), which is $1/\sqrt{m}$ times faster than the overdamped case (rate $O(m)$). This significant speedup depends on preserving inertial effects and tuning friction in correspondence with spectral gap and convexity structure.

3. Weak Confinement: Weighted and Weak Poincaré Regimes

For potentials with fat-tails or sub-exponential growth ($\phi(x)$, $\psi(v)$), classical Poincaré inequalities may fail. Instead, one appeals to weighted Poincaré or weak Poincaré–Lions inequalities, e.g., (Brigati et al., 2024):

• Weighted Poincaré: For a positive weight $W(x)$,

$$\int (f - \mu_W(f))^2\,d\mu_W \leq P_W \int |\nabla_x f|^2\,d\mu$$

where $\mu_W(dx) = Z_W^{-1} W^{-2}(x) d\mu(x)$.

• Weak Poincaré (Röckner-Wang): There exists a decreasing function $\beta_x(s)$ so that, for bounded $f\in H^1(\mu)$,

$$\mathrm{Var}_\mu(f) \leq s \int |\nabla_x f|^2\,d\mu + \beta_x(s) \|f\|^2_{\mathrm{osc}}$$

Algebraic or sub-exponential (stretched-exponential) relaxation rates are then proved constructively. For instance, if $\phi(x) = \langle x \rangle^\alpha$, $\psi(v) = \langle v \rangle^\delta$ with $\alpha,\delta\in(0,1)$, the $L^2$ norm decays as $$C\,\exp(-c\,t^{\frac{\alpha\delta}{2\alpha+2\delta-3\alpha\delta}})$$, while for polynomial tails, algebraic rates are established (Brigati et al., 2024).

Hypocoercivity techniques rely crucially on space-time weighted inequalities and explicit averaging lemmas. These tools allow full quantification of time-averaged energy dissipation even when the measure fails to be strictly log-concave.

4. Algorithmic Implementation and Information-Theoretic Complexity

Algorithmic discretization employs splitting schemes and integrators tailored to both the geometry and noise structure. Notably,

• Randomized midpoint methods attain optimal strong order $3/2$ convergence rates in $L^2$ error, i.e.,

$$\mathrm{err}(N, d) \geq C_{\text{low}} \sqrt{d} N^{-3/2}$$

matching upper bounds for randomized gradient and Brownian-oracle algorithms (Cao et al., 2020).

• Geometric (splitting) integrators such as UBU splitting extend strong order to $p=2$ under global convexity (Ye et al., 2024), with MSE for ergodic averages scaling as $O(h^{2p} + (Nh)^{-1})$ in terms of step size $h$ and iterations $N$.

Information complexity in high dimensions is fundamentally limited by the number of admissible queries to $\nabla U$ and Brownian increments; randomized schemes achieve strictly better rates than deterministic ones when oracle access is restricted.

5. Practical Applications and Extensions

Underdamped Langevin schemes are extensively deployed in molecular simulation (position-dependent friction), MIMO detection (annealed underdamped Langevin with splitting), drug-diffusion modeling (G-JF integrator for non-uniform media), and mean-field optimization (GANs, kernel methods).

• In drug-eluting stent models, the inertial scheme accurately solves Fick’s equation with spatially varying coefficients, robustly propagates across sharp interfaces and semi-permeable membranes, and remains highly stable even under large time steps (Regev et al., 2018).
• In Bayesian sampling and regression, gradient-adjusted underdamped Langevin dynamics (GAUL) demonstrably boosts mixing rates over both classical overdamped and underdamped algorithms, especially for ill-conditioned Gaussian targets—the discrete mixing time scales as $O(\sqrt{\kappa} \log(d/\epsilon))$ in terms of the condition number $\kappa$ (Zuo et al., 2024).

6. Non-Equilibrium Thermodynamics: Thermodynamic Uncertainty Relations

Underdamped Langevin dynamics modifies the classic thermodynamic uncertainty relation by introducing explicit inertial corrections:

• In the steady state, the uncertainty product for a current $J(t)$ obeys

$$\frac{\mathrm{Var}[J(t)]}{\langle J(t)\rangle^2} \geq \frac{2}{\Sigma + \int A(x,v)\,p_{\mathrm{ss}}(x,v) dx dv + \mathrm{Var}_{\mathrm{ss}}[v]}$$

where $\Sigma$ amalgamates the overdamped entropy rate with $\Psi$, a strictly positive correction due to local mean acceleration and velocity fluctuations. These terms vanish in the overdamped limit $m \to 0$, but in the underdamped regime, they permit currents to be more precise than classical bounds would suggest (Dechant, 2022, Vu et al., 2019).

• The entropy production rate $\Sigma$ can be rigorously estimated in terms of cumulant and stochastic currents derived from trajectory data, requiring only the damping-to-mass ratio and diffusion constant—not detailed knowledge of $U$ (Lyu et al., 2024).

7. Statistical Inference and Data-Driven Model Identification

Inference of SDE parameters from trajectory data leverages unbiased finite-difference and orthonormal-basis estimators, correcting for measurement noise and discretization bias.

• Underdamped Langevin inference (ULI) recovers both drift and diffusion coefficients robustly even under strong measurement noise, scales to high-dimensional systems, and provides a principled framework for unbiased estimation via central-difference and stochastic Taylor expansion (Brückner et al., 2020).
• Extensions to colored noise, non-Gaussian processes, sparse bases, and EM-type hidden-variable models are feasible within the same statistical architecture.

---

In summary, underdamped Langevin dynamics integrates foundational stochastic theory, explicit hypocoercivity analysis, complex algorithmic construction, non-equilibrium thermodynamic bounds, and modern statistical inference into a unified mathematical and computational framework. Its versatility and quantitative precision are underpinned by advanced functional inequalities and rigorous probabilistic methods extending far beyond the scope of classical overdamped paradigms.