Kinetic Langevin Diffusion

• Kinetic Langevin Diffusion is a family of stochastic processes that integrates momentum to enhance sampling efficiency compared to overdamped methods.
• Discrete-time implementations like the GJF–2GJ thermostat yield robust configurational and kinetic statistics across a range of physical systems.
• The method’s leap-frog updates and companion velocity ensure accurate thermodynamic averages and time-step independent diffusion metrics.

Kinetic Langevin Diffusion is a family of stochastic dynamical systems that augment classical overdamped (first-order) Langevin dynamics with velocity (momentum-like) variables, yielding a second-order or underdamped diffusion process. This framework is widely used in computational statistical mechanics, MCMC sampling, and molecular dynamics, and is the basis for a number of high-performance algorithms and analytical tools. The kinetic Langevin approach enables improved mixing rates, dimensionally favorable convergence properties, time-step robust discretizations, and more accurate configurational and kinetic statistics compared to classical overdamped schemes. Discrete-time kinetic Langevin methods, such as the Grønbech-Jensen–Farago (GJF) thermostat, yield robust samplers for both equilibrium and transport statistics over a wide class of physical and probabilistic systems (Jensen et al., 2019).

1. Mathematical Formulation of Kinetic Langevin Diffusion

Kinetic Langevin diffusion describes the evolution of the phase-space vector $(r, v)$ (position and velocity) as a solution to the coupled stochastic differential equations: $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$ where $m$ is mass, $\alpha$ the friction coefficient, $f(r_t)$ a deterministic force field (often $-\nabla V(r_t)$), and $W_t$ standard Brownian motion. The system admits the canonical Boltzmann–Gibbs invariant measure: $$\mu(dr\,dv) = Z^{-1}\exp\left(-\beta[V(r)+(1/2)m|v|^2]\right) dr\,dv,$$ where $\beta = (k_B T)^{-1}$. Assuming sufficient regularity, this SDE enjoys exponential ergodicity and hypocoercivity properties under mild convexity and Poincaré conditions (Stoltz et al., 2016).

2. Discrete-Time Kinetic Langevin: GJF Thermostat and Companion Velocity

Discrete-time schemes are essential for computational implementation. The GJF thermostat (Jensen et al., 2019) is a discrete integrator for Langevin dynamics that maintains robust configurational and kinetic sampling:

Update scheme (with time step $dt$):

$$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$0

where $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$1, $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$2, and $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$3 is Gaussian noise.

Companion half-step velocity:

$$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$4

This companion velocity yields time-step-independent kinetic statistics, exactly reproducing the Maxwell–Boltzmann average kinetic energy even for large $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$5. The method supports a leap-frog formulation efficiently updating both position and velocity (Jensen et al., 2019).

3. Exact Kinetic and Configurational Statistics, Green–Kubo Diffusion

A hallmark of the GJF–2GJ (GJF with companion velocity) scheme is its statistical accuracy in both configurational and kinetic observables:

• $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$6 is exact for all $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$7 in harmonic systems.
• The scheme yields exact time-step-independent variance of kinetic energy and correct discrete-time autocorrelations.
• The discrete-time Green–Kubo formula evaluated via the autocorrelation function of the companion velocity yields:

$$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$8

independently of $$\begin{cases} dr_t = v_t\,dt, \ m\,dv_t = f(r_t)\,dt - \alpha\,v_t\,dt + \sqrt{2\alpha k_B T}\,dW_t, \end{cases}$$9, provided stability constraints hold (Jensen et al., 2019).

4. Extension to Nonlinear and Many-Body Systems

Empirical studies confirm that GJF–2GJ maintains statistical accuracy beyond the harmonic case:

• For nonlinear oscillators (non-quadratic potentials), the method produces kinetic and configurational temperatures and fluctuations accurate to within 1% up to the stability threshold.
• In three-dimensional molecular dynamics with hundreds of particles (Lennard-Jones solids and liquids), the method preserves correct phase-space and thermodynamic observables as $m$0 approaches the maximum stable step, outperforming classical integrators such as BBK and PBS which exhibit systematic drift under similar conditions (Jensen et al., 2019).

5. Algorithmic Summary: Leap-Frog and Stability Constraints

The scheme is efficiently implemented using a leap-frog update:

Step	Update Equation	Notes
Half-step velocity update	$m$1
Position update	$m$2
Force evaluation	$m$3

The stability domain is governed by $m$4 and $m$5, where $m$6 is the highest oscillator frequency. For maximal efficiency, $m$7 should approach but not reach the stability boundary.

6. Comparison with Continuous-time Langevin and Other Discretizations

Continuous-time Langevin yields exact equipartition at any moment. However, naive discretizations introduce $m$8 errors in either configuration or kinetic averages. The GJF–2GJ method, by contrast, ensures exact statistics (energies, fluctuations, diffusion) up to the method's stability limit and aligns with continuous dynamics in the zero time-step limit.

In practice:

• BBK, PBS, and conventional splitting schemes show systematic time-step-dependent deviations in kinetic or configurational averages.
• GJF–2GJ exhibits weak second-order accuracy (error $m$9) in generic systems, and error-free sampling in the harmonic limit (Finkelstein et al., 2021).

7. Practical Recommendations and Broader Context

For effective simulation:

• Choose $\alpha$0 to be 5–10% below the strict analytic stability bound.
• Friction $\alpha$1 ($\alpha$2) should be commensurate with the slowest relevant timescale; moderate values avoid overdamping while ensuring good sampling.
• For precise diffusion estimates, use Riemann-summed velocity autocorrelations.
• The method's validity and robustness extend to complex, nonlinear, and many-body systems without additional computational cost.

Kinetic Langevin discretizations, particularly in the GJF–2GJ class, provide a practical, thermodynamically sound standard for phase-space sampling, molecular dynamics, and stochastic optimization where accurate kinetic, configurational, and transport properties are required (Jensen et al., 2019).