The Generalized Langevin Equation and Memory Effects

The ordinary Langevin equation models a heavy particle in a fluid as subject to instantaneous viscous drag $-\gamma_0 v$ plus delta-correlated white noise — the Markovian picture, valid only when the bath relaxes infinitely fast compared with the particle. When the surrounding medium has its own dynamical timescales — a viscoelastic fluid, a polymer network, a dense colloid, a protein interior — the friction acquires memory: the force the bath exerts now reflects how the particle moved over a finite past interval. The generalized Langevin equation (GLE) captures this by replacing the instantaneous $\gamma_0 v(t)$ with a convolution $\int_0^t\gamma(t-t')\,v(t')\,\mathrm dt'$ and the white noise with a stationary colored noise $\xi(t)$ . Crucially, the two modifications are not independent. Equilibrium statistical mechanics demands the second fluctuation-dissipation theorem, which ties the noise autocorrelation directly to the friction memory kernel: $\langle\xi(t)\xi(t')\rangle = mk_BT\,\gamma(|t-t'|)$ . Friction with memory and colored noise are two faces of one object — the memory kernel $\gamma(t)$ . Without this constraint the particle would not equilibrate to the Boltzmann distribution: the stationary state would violate equipartition, energy would be spuriously injected or drained, and the second law would fail for a passive bath. The GLE is thus the unique non-Markovian generalization consistent with thermal equilibrium. The GLE is not merely a phenomenological postulate. The Mori–Zwanzig projection-operator formalism derives it exactly from a Hamiltonian system: starting from a particle bilinearly coupled to a continuum of harmonic oscillators — the Caldeira–Leggett / Zwanzig model — one integrates out the bath and obtains a GLE in which the memory kernel is the bath's spectral density and the noise is the freely-evolving bath coordinates. The memory is therefore a faithful record of the eliminated dynamics, not an approximation. Solving the GLE relies on the convolution theorem: in Laplace space the equation becomes algebraic, so the velocity autocorrelation function, mobility, and mean-square displacement are all determined by $\hat\gamma(s)$ , the Laplace transform of the kernel. This makes the GLE the natural language for microrheology, anomalous diffusion, and the celebrated $t^{-d/2}$ hydrodynamic long-time tail — and it connects to the Fokker–Planck equation only in the Markovian limit, because $v(t)$ obeying a memory kernel is genuinely non-Markovian.

Key takeaways

The generalized Langevin equation (GLE) replaces the instantaneous friction of the ordinary Langevin equation with a memory kernel $\gamma(t)$ , so the drag at time $t$ depends on the entire history of the velocity.
The kernel and the colored noise are locked together by the second fluctuation-dissipation theorem, $\langle\xi(t)\xi(t')\rangle=mk_BT\,\gamma(|t-t'|)$ , which guarantees relaxation to thermal equilibrium.
The GLE is derived exactly by projecting out a bath of harmonic oscillators (Mori–Zwanzig / Caldeira–Leggett); memory encodes the finite response time of the eliminated degrees of freedom.
A short-correlation (white-noise) kernel $\gamma(t)=2\gamma_0\delta(t)$ recovers the Markovian Langevin equation; long-tailed kernels produce anomalous (sub- or super-) diffusion.
Solutions follow from Laplace transforms: the velocity autocorrelation $\hat C_v(s)=(k_BT/m)/(s+\hat\gamma(s)/m)$ and the mean-square displacement are fixed entirely by $\hat\gamma(s)$ .
$v(t)$ alone is not Markov, but it can be embedded in a higher-dimensional Markov system by adding auxiliary memory variables (one per exponential in the kernel).
Key idea: friction with memory and colored noise are two faces of one object — the kernel $\gamma(t)$ sets both the dissipation and, through the second FDT $\langle\xi(t)\xi(0)\rangle=mk_BT\,\gamma(t)$ , the fluctuations, so the GLE is the unique non-Markovian generalization consistent with thermal equilibrium.

The core insight

Friction with memory and colored noise are two faces of one object: the kernel $\gamma(t)$ sets both the dissipation and — through the second fluctuation-dissipation theorem $\langle\xi(t)\xi(0)\rangle=mk_BT\,\gamma(t)$ — the fluctuations, so the GLE is the unique non-Markovian generalization consistent with thermal equilibrium.

Where this lives today

The generalized Langevin equation is the working language of passive microrheology: by tracking a tracer bead and measuring its mean-square displacement, the generalized Einstein relation $\mu(s)=1/[ms+\hat\gamma(s)]$ is inverted (via $\hat\gamma(s)=6\pi a\,\eta^*(s)$ , a frequency-dependent Stokes drag) to extract the complex shear modulus $G^*(\omega)$ of the surrounding material across decades of frequency — turning a Brownian particle into a rheometer for cytoplasm, gels, and polymer solutions. In protein dynamics and reaction-rate theory, the friction on a reaction coordinate has picosecond-scale memory, and Grote–Hynes theory generalizes Kramers' escape rate by evaluating $\hat\gamma$ at the reactive barrier frequency rather than the static friction — memory makes barrier crossing faster than the high-friction Kramers prediction. Anomalous subdiffusion in crowded cells — tracers in cytoplasm, lipid membranes, and chromatin — is captured by a power-law memory kernel (the fractional Langevin equation), which reproduces the observed $\langle x^2\rangle\sim t^\alpha$ with correct thermal equilibration, distinguishing memory-driven subdiffusion from continuous-time random walks by its persistent anti-correlated velocity statistics. And the very same harmonic-bath derivation underlies the quantum theory of dissipation: the spectral density $J(\omega)$ classifies baths as Ohmic, sub-Ohmic, or super-Ohmic and governs macroscopic quantum tunneling, decoherence, and the spin–boson model — so the classical GLE is the shadow of the Caldeira–Leggett quantum machinery.

The physical picture

Picture pushing a spoon through honey versus through a cornstarch-and-water suspension. In honey the drag responds instantly to your speed — Markovian friction. In the viscoelastic cornstarch the medium "remembers" how you just moved and pushes back based on recent history. That history-dependence is exactly the convolution $\int\gamma(t-t')\,v(t')\,\mathrm dt'$ : the drag at this instant is a weighted sum over the velocity's past, with the memory kernel $\gamma$ as the weight. The same bath that dissipates energy is the bath that kicks the particle around. A bath with a slow internal timescale dissipates with memory and kicks with temporal correlations. The fluctuation-dissipation theorem says these are quantitatively the same function: you cannot have memory friction without colored noise of matched spectrum, on pain of violating the second law. Memory is simply what remains after you delete degrees of freedom — the bath oscillators evolve, store the particle's past influence, and feed it back later. "Markovian" just means the bath forgets faster than you can measure; "non-Markovian" means the bath's clock is comparable to yours.

Figure

Memory friction versus instantaneous drag: colored noise and the memory kernel are one object

Mathematically, non-Markovianity costs you the Markov property of $v(t)$ alone, but you can always buy it back by adding auxiliary variables — the memory modes. The price of a closed description is extra dimensions. An exponential kernel needs just one auxiliary variable (an Ornstein–Uhlenbeck memory force); a power-law kernel needs a continuum, giving Mittag–Leffler relaxation rather than simple exponentials.

Figure

Velocity autocorrelation: the memory kernel's back-scattering dip is absent in the Markovian curve

The core equations

The force-free GLE replaces instantaneous drag by a causal convolution of the velocity with the memory kernel, driven by colored noise. Only the past enters — the integral runs from $0$ to $t$ .

m\,\dot v(t)=-\int_0^t\gamma(t-t')\,v(t')\,\mathrm dt'+\xi(t) \qquad (1)

where: $m$ — mass of the tagged particle; $v(t)$ — velocity; $\gamma(t)$ — memory (friction) kernel; $\xi(t)$ — stationary, zero-mean colored random force; the causal integral means the drag depends on past velocities only. Setting $\gamma(t)=2\gamma_0\delta(t)$ collapses this to the ordinary Langevin equation, which establishes the instantaneous $-\gamma_0 v$ drag and white noise. The second fluctuation-dissipation theorem locks the noise autocorrelation to the kernel — this is the article's central owned result and the constraint that makes equilibration possible.

\langle\xi(t)\,\xi(t')\rangle=m\,k_BT\,\gamma(|t-t'|) \qquad (2)

where: $k_B$ — Boltzmann constant; $T$ — bath temperature; $\langle\cdot\rangle$ — average over noise realizations. Physically: the amount of kicking is exactly the amount of damping, matched frequency by frequency, so the stationary state is Boltzmann with $\langle v^2\rangle=k_BT/m$ . In the frequency domain, the noise spectrum $S_\xi(\omega)=2mk_BT\,\mathrm{Re}\,\tilde\gamma(\omega)$ is set by the dissipative (real) part of the complex friction. The convolution becomes algebra under Laplace transform: the velocity autocorrelation is fixed entirely by the transform of the kernel.

\hat C_v(s)=\frac{k_BT/m}{s+\hat\gamma(s)/m},\qquad C_v(t)=\langle v(t)v(0)\rangle \qquad (3)

where: $\hat C_v(s)$ — Laplace transform of the velocity autocorrelation; $\hat\gamma(s)=\int_0^\infty\gamma(t)e^{-st}\,\mathrm dt$ ; $C_v(0)=\langle v^2\rangle=k_BT/m$ by equipartition. The whole dynamics is compressed into one algebraic expression; the noise term drops because $\langle\xi(t)v(0)\rangle=0$ for $t>0$ by causality. The long-time diffusion constant depends on the kernel only through its zero-frequency value — the DC friction — giving a generalized Einstein relation.

D=\lim_{s\to0}k_BT\,\mu(s)=\frac{k_BT}{\hat\gamma(0)},\qquad \hat\gamma(0)=\int_0^\infty\gamma(t)\,\mathrm dt \qquad (4)

where: $\mu(s)=1/[ms+\hat\gamma(s)]$ — frequency-dependent mobility; $D$ — long-time diffusion constant. If the kernel's time-integral converges, normal diffusion results; if $\hat\gamma(0)$ diverges (a $t^{-\alpha}$ tail), $D\to0$ and one gets subdiffusion. The mean-square displacement is a double time-integral of the velocity autocorrelation, reducing to $2Dt$ at long times (the Green–Kubo statement).

\langle\Delta x^2(t)\rangle=2\int_0^t(t-t')\,C_v(t')\,\mathrm dt' \qquad (5)

where: $\Delta x(t)=x(t)-x(0)$ . As $t\to\infty$ this yields $2Dt$ with $D=\int_0^\infty C_v\,\mathrm dt$ , tying the diffusion constant to the integrated VACF. The kernel itself is not free — the Caldeira–Leggett harmonic-bath model derives it from the bath spectral density, with the second FDT following as a theorem.

\gamma(t)=\sum_i\frac{c_i^2}{m_i\omega_i^2}\cos(\omega_i t)=\frac{2}{\pi}\int_0^\infty\frac{J(\omega)}{\omega}\cos(\omega t)\,\mathrm d\omega \qquad (6)

where: $J(\omega)=\frac{\pi}{2}\sum_i\frac{c_i^2}{m_i\omega_i}\delta(\omega-\omega_i)$ — bath spectral density; $c_i,\omega_i,m_i$ — bath coupling, frequency, mass; the free oscillation $\cos(\omega_i t)$ of each bath mode is the microscopic source of memory. The Mori–Zwanzig formalism produces exactly this structure by projecting out the bath.

Derivation AThe GLE from the Caldeira-Leggett harmonic bath (exact elimination)

The goal is to derive the generalized Langevin equation — kernel and noise together — exactly, by integrating out a bath of harmonic oscillators bilinearly coupled to the particle, and to show that the second fluctuation-dissipation theorem then follows as a theorem, not an assumption. We write the system-bath Hamiltonian with a counter-term, solve each driven bath oscillator exactly, substitute back into the particle equation, integrate the driven part by parts to expose the memory convolution, and finally compute the noise autocorrelation from canonical bath averages.

Step 1. Write the equations of motion: $m\ddot x=-V'(x)+\sum_i c_i(q_i-\frac{c_i}{m_i\omega_i^2}x)$ and $m_i\ddot q_i=-m_i\omega_i^2 q_i+c_i x$ , so each bath mode is a driven linear oscillator. Why: the coupling is bilinear, so the bath responds linearly to the particle's trajectory and can be solved in closed form.

Step 2. Solve each oscillator: $q_i(t)=q_i(0)\cos\omega_i t+\frac{p_i(0)}{m_i\omega_i}\sin\omega_i t+\frac{c_i}{m_i\omega_i}\int_0^t\sin[\omega_i(t-t')]\,x(t')\,\mathrm dt'$ — homogeneous (initial conditions) plus particular (driven by $x$ ). Why: this is the standard Green's-function solution of a forced harmonic oscillator.

Step 3. Substitute back and integrate the driven $\sin$ term by parts; it produces a $\cos$ kernel convolved with $\dot x$ plus a boundary term that exactly cancels the counter-term $\frac{c_i^2}{m_i\omega_i^2}x$ . Why: the counter-term is included precisely to prevent a spurious harmonic renormalization of the particle's potential; without it $V(x)$ would be shifted.

Step 4. Identify the memory kernel $\gamma(t)=\sum_i\frac{c_i^2}{m_i\omega_i^2}\cos(\omega_i t)$ and the noise $\xi(t)=\sum_i c_i[(q_i(0)-\frac{c_i}{m_i\omega_i^2}x(0))\cos\omega_i t+\frac{p_i(0)}{m_i\omega_i}\sin\omega_i t]$ — the freely-evolving bath modes. Why: the history-dependent piece collects into the convolution, and what remains is driven only by the bath's initial conditions.

Step 5. Draw $\{q_i(0),p_i(0)\}$ from the canonical distribution and compute $\langle\xi(t)\xi(t')\rangle=k_BT\sum_i\frac{c_i^2}{m_i\omega_i^2}\cos[\omega_i(t-t')]=k_BT\,\gamma(t-t')$ , using equipartition $\langle q_i^2\rangle=k_BT/(m_i\omega_i^2)$ . Why: the noise and the kernel share the identical coupling combination $c_i^2/(m_i\omega_i^2)$ , so the second FDT is forced by the thermal bath — a theorem, not a postulate. *Verified: the harmonic-bath kernel and the FDT $\langle\xi(t)\xi(0)\rangle=k_BT\,\gamma(t)$ per mode are checked symbolically from the canonical average $\langle q_i^2\rangle=k_BT/(m_i\omega_i^2)$ ; verify_aq-sm-generalized-langevin-equation.py.*

Go deeper · Derivation B · 1 minLaplace-space solution and the exponential-memory Markovian embedding

Derivation BLaplace-space solution and the exponential-memory Markovian embedding

The goal is to solve the linear GLE in closed form via the convolution theorem, extract the velocity autocorrelation and the generalized Einstein relation, and then show that a single-exponential kernel is equivalent to a two-dimensional Markov process — the route to simulating memory. We Laplace-transform the correlation equation, solve algebraically, read off $D$ from the $s\to0$ limit, and then introduce one auxiliary memory variable that obeys its own Ornstein–Uhlenbeck equation.

Step 1. Multiply the GLE by $v(0)$ and average; the noise term drops ( $\langle\xi(t)v(0)\rangle=0$ , causality), leaving $m\dot C_v(t)=-\int_0^t\gamma(t-t')C_v(t')\,\mathrm dt'$ . Why: the initial velocity is uncorrelated with the future noise, so only the deterministic memory-friction survives in the correlation equation.

Step 2. Laplace-transform using the convolution theorem: $m[s\hat C_v-C_v(0)]=-\hat\gamma(s)\hat C_v$ , giving $\hat C_v(s)=(k_BT/m)/(s+\hat\gamma(s)/m)$ with $C_v(0)=k_BT/m$ . Why: convolution becomes multiplication under Laplace transform, turning the integro-differential equation into algebra.

Step 3. Read the diffusion constant from the $s\to0$ limit: $D=\hat C_v(0)=k_BT/\hat\gamma(0)$ , the generalized Einstein relation. Why: the Laplace transform of $C_v$ at $s=0$ is its time-integral, which is exactly the Green–Kubo diffusion constant.

Step 4. For the exponential kernel $\gamma(t)=(\gamma_0/\tau_c)e^{-t/\tau_c}$ , define the memory variable $z(t)$ so that $m\dot v=z$ and $\dot z=-z/\tau_c-(\gamma_0/\tau_c)v+(1/\tau_c)\eta(t)$ with white $\eta$ . Why: differentiating the exponential convolution generates $-z/\tau_c$ plus the instantaneous $-\gamma_0 v/\tau_c$ , closing the system in $(v,z)$ .

Step 5. The pair $(v,z)$ is Markovian and admits a 2D Fokker–Planck equation; as $\tau_c\to0$ , $(\gamma_0/\tau_c)e^{-t/\tau_c}\to2\gamma_0\delta(t)$ and $z$ is adiabatically eliminated, recovering the ordinary Langevin equation with $\langle\xi\xi\rangle=2\gamma_0 k_BT\delta(t)$ . Why: one auxiliary mode per exponential restores the Markov property in an enlarged space; the delta limit collapses it back to instantaneous friction. *Verified: $\hat C_v(s)=(k_BT/m)/(s+\hat\gamma(s)/m)$ , $\hat\gamma(0)=\gamma_0$ , $D=k_BT/\gamma_0$ , the Markovian VACF $C_v(t)=(k_BT/m)e^{-\gamma_0 t/m}$ , and the MSD long-time slope $2D$ are checked symbolically; verify_aq-sm-generalized-langevin-equation.py.*

Limiting cases

The kernel $\gamma(t)$ interpolates between distinct physical regimes:

Case	Result	Meaning
White-noise (Markovian) limit	$\gamma(t)=2\gamma_0\,\delta(t)$	Recovers the ordinary Langevin equation; $v(t)$ becomes Markov (Ornstein–Uhlenbeck), $\langle\xi\xi\rangle=2\gamma_0 k_BT\delta$ .
Exponential memory	$\gamma(t)=\dfrac{\gamma_0}{\tau_c}e^{-t/\tau_c}$	Single bath timescale; equivalent to a two-variable Markov system (velocity + one auxiliary mode).
Power-law (long-tailed) memory	$\gamma(t)\sim t^{-\alpha},\ 0<\alpha<1$	Scale-free bath; subdiffusion $\langle x^2\rangle\sim t^\alpha$ (fractional Langevin dynamics, $\hat\gamma(s)\sim s^{\alpha-1}$ ).
Hydrodynamic Basset memory	$\gamma(t)\sim t^{-1/2}\ (t\to\infty)$	Momentum diffusion in the solvent; yields the long-time tail $C_v(t)\sim t^{-3/2}$ in 3D, absent in the Markovian model.

Worked examples

Example 1 — Exponential memory and the Markovian corner. Take $\gamma(t)=(\gamma_0/\tau_c)e^{-t/\tau_c}$ ; find the Laplace-space friction, the diffusion constant, and the $\tau_c\to0$ recovery.

Step 1. Laplace-transform the kernel: $\hat\gamma(s)=\int_0^\infty(\gamma_0/\tau_c)e^{-t/\tau_c}e^{-st}\,\mathrm dt=\gamma_0/(1+s\tau_c)$ .

Step 2. The DC friction is $\hat\gamma(0)=\gamma_0$ — independent of $\tau_c$ — so the diffusion constant is $D=k_BT/\hat\gamma(0)=k_BT/\gamma_0$ , the same as the Markovian result. The memory only reshapes the approach to diffusion, not its asymptotic rate.

Step 3. As $\tau_c\to0$ , $\hat\gamma(s)\to\gamma_0$ (frequency-independent), so $\hat C_v(s)\to(k_BT/m)/(s+\gamma_0/m)$ , which inverts to the Lorentzian VACF $C_v(t)=(k_BT/m)e^{-\gamma_0 t/m}$ . The finite bath correlation time $\tau_c$ sets a corner frequency $\sim1/\tau_c$ beyond which memory matters; below it, the ordinary Langevin picture is recovered exactly.

Example 2 — Reading the long-time tail. A hydrodynamic bath gives $\hat\gamma(s)=\gamma_0(1+\alpha\sqrt{s\tau_\nu}+\dots)$ near $s=0$ ; deduce the velocity-autocorrelation tail in three dimensions.

Step 1. The non-analytic $\sqrt{s}$ term controls the small- $s$ (long-time) behavior of $\hat C_v(s)=(k_BT/m)/(s+\hat\gamma(s)/m)$ .

Step 2. By the Tauberian theorem, $\hat f(s)\sim s^{\beta}$ as $s\to0$ corresponds to $f(t)\sim t^{-\beta-1}$ ; with $\beta=1/2$ this gives $C_v(t)\sim t^{-3/2}$ in 3D (generally $t^{-d/2}$ in $d$ dimensions).

Step 3. In $d=2$ the tail is $t^{-1}$ , whose time-integral $\int_0^\infty C_v\,\mathrm dt$ diverges logarithmically, so the diffusion constant of a two-dimensional fluid is ill-defined — a striking failure of the naive Markovian Einstein relation that only the GLE reveals. The moving particle sets up a vortex ring in the fluid that diffuses back and re-pushes it, sustaining velocity correlations far longer than the exponential Markovian decay (Alder & Wainwright, 1970).

Origins · 1 minmemory kernel and how this was discovered

memory kernel

non-Markovian friction, 1965–1973 · The memory kernel

\gamma(t)

is the function that replaces the instantaneous friction coefficient of the ordinary Langevin equation, making the drag depend on the entire history of the velocity. It is uniquely tied to the colored-noise autocorrelation by the second fluctuation-dissipation theorem,

\langle\xi(t)\xi(0)\rangle=mk_BT\,\gamma(t)

. The kernel emerges from the Mori–Zwanzig projection formalism (Mori 1965) and from the exact elimination of a harmonic bath (Zwanzig 1973 / Caldeira–Leggett); its Laplace transform controls the velocity autocorrelation, mobility, and diffusion. A delta kernel is Markovian; a power-law kernel produces anomalous diffusion.

History

1908

— Langevin writes the stochastic equation of motion for a Brownian particle with instantaneous friction

1965

— Mori derives the generalized Langevin equation and continued-fraction expansion by linear projection (Prog. Theor. Phys. 33, 423)

1970

— Alder and Wainwright discover the

t^{-d/2}

long-time tail of the velocity autocorrelation in molecular-dynamics simulations

1973

— Zwanzig gives the exact harmonic-bath elimination yielding the GLE and second FDT (J. Stat. Phys. 9, 215)

1983

— Caldeira and Leggett extend the system-bath model to quantum dissipation and macroscopic tunneling

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 15 (Brownian motion, fluctuation-dissipation) · Kardar, Statistical Physics of Particles, Ch. 9 · Reif, Fundamentals of Statistical and Thermal Physics, Ch. 15 · Sethna, Statistical Mechanics: Entropy, Order Parameters and Complexity, Ch. 10 · Zwanzig, Nonequilibrium Statistical Mechanics (2001), Ch. 1, 8 · Kubo, Toda, Hashitsume, Statistical Physics II, Ch. 1–2 · Mori, Prog. Theor. Phys. 33, 423 (1965); Zwanzig, J. Stat. Phys. 9, 215 (1973).

Provenance. All equations are standard and cross-checked against Zwanzig Nonequilibrium Statistical Mechanics (Ch. 1, 8), Kubo–Toda–Hashitsume Statistical Physics II (Ch. 1–2), Pathria & Beale Ch. 15, and Kardar Ch. 9. A verifier should: (1) reproduce the harmonic-bath elimination giving $\gamma(t)=\sum_i(c_i^2/m_i\omega_i^2)\cos\omega_i t$ and the matching $\langle\xi\xi\rangle=k_BT\gamma$ from canonical bath averages $\langle q_i^2\rangle=k_BT/(m_i\omega_i^2)$ ; (2) confirm the Laplace-space VACF $\hat C_v(s)=(k_BT/m)/(s+\hat\gamma(s)/m)$ ; (3) verify $D=k_BT/\hat\gamma(0)$ as the $s\to0$ limit; (4) check the Tauberian inversion of $\sqrt s$ to a $t^{-3/2}$ tail in 3D. This article owns the second-kind fluctuation-dissipation theorem $\langle\xi\xi\rangle=mk_BT\gamma$ ; it recalls, by name, the langevin-equation (its Markovian limit, which establishes the instantaneous $-\gamma_0 v$ drag), the fluctuation-dissipation-theorem (the first-kind FDT this generalizes), and the mori-zwanzig-formalism (which derives the GLE exactly). No numerical constants beyond $k_B$ , $T$ , equipartition $k_BT/m$ , and the dimensional exponents $t^{-d/2}$ are asserted.

Prerequisites: the Langevin equation · the fluctuation-dissipation theorem · the Wiener process and white noise · the Ornstein–Uhlenbeck process. Leads to: the Fokker–Planck equation · Kramers' escape rate theory · the Einstein relation · Brownian motion.

What is the difference between the first and second fluctuation-dissipation theorems?

The first FDT relates a system's linear response function to the equilibrium autocorrelation of an observable — how the system responds to an external perturbation is fixed by its spontaneous fluctuations. The second FDT is internal to the Langevin description: it ties the friction memory kernel $\gamma(t)$ to the autocorrelation of the random force, $\langle\xi(t)\xi(0)\rangle=mk_BT\,\gamma(t)$ . The first concerns observable response; the second guarantees the modeled noise has exactly the strength needed to keep the particle in thermal equilibrium. The GLE owns the second FDT; the first is a separate, complementary statement.

Why does an exponential memory kernel make the dynamics solvable while a power-law kernel is hard?

An exponential $e^{-t/\tau_c}$ has a single rational Laplace transform $\gamma_0/(1+s\tau_c)$ , so the GLE maps onto a finite-dimensional linear Markov system — velocity plus one auxiliary mode, an Ornstein–Uhlenbeck pair with closed-form correlations. A power-law kernel $t^{-\alpha}$ has a non-rational, branch-cut Laplace transform $\sim s^{\alpha-1}$ , which requires an infinite continuum of auxiliary modes and yields Mittag–Leffler relaxation rather than simple exponentials. This is why exponential memory is the workhorse of numerical GLE integration: each exponential in a multi-exponential fit costs just one extra variable.

What is the role of the counter-term in the Caldeira-Leggett Hamiltonian?

The counter-term $\frac{c_i}{m_i\omega_i^2}x$ inside each bath potential cancels a static, $x$ -dependent boundary term that appears when the driven bath response is substituted back into the particle equation. Without it, that term would add a spurious harmonic potential $-\frac12(\sum_i c_i^2/m_i\omega_i^2)x^2$ that renormalizes — and could destabilize — the physical potential $V(x)$ . Including the counter-term guarantees the particle feels only its bare potential plus genuine memory friction and noise, so the model has a well-defined equilibrium.

How does the memory kernel produce anomalous diffusion?

Through its zero-frequency behavior. Normal diffusion needs a finite DC friction $\hat\gamma(0)=\int_0^\infty\gamma\,\mathrm dt$ , giving $D=k_BT/\hat\gamma(0)$ . If the kernel has a slow power-law tail $\gamma(t)\sim t^{-\alpha}$ with $0<\alpha<1$ , then $\hat\gamma(0)$ diverges, $D\to0$ , and the mean-square displacement grows sublinearly as $\langle x^2\rangle\sim t^\alpha$ — subdiffusion. A kernel that decays too fast (or has $\hat\gamma(0)\to0$ ) gives superdiffusion. The detailed shape of $\gamma$ at finite frequency only affects the crossover; the asymptotic exponent is set by the small- $s$ singularity of $\hat\gamma(s)$ .

Is the colored noise in the GLE Gaussian, and does it have zero mean?

Yes on both counts, when the bath is harmonic and canonically distributed. In the Caldeira–Leggett derivation the noise $\xi(t)$ is a linear combination of the bath's initial coordinates and momenta, drawn from a Gaussian (canonical) distribution; a linear combination of Gaussians is Gaussian, and since the equilibrium averages of $q_i(0),p_i(0)$ vanish, $\langle\xi(t)\rangle=0$ . What makes it colored rather than white is the temporal correlation $\langle\xi(t)\xi(0)\rangle=mk_BT\,\gamma(t)$ , which is nonzero over the bath's memory time $\tau_c$ rather than a delta spike.

How is the GLE related to the Mori-Zwanzig formalism?

The Mori–Zwanzig projection-operator formalism derives the GLE exactly from Hamiltonian dynamics: projecting the full motion onto the chosen slow variable (e.g. velocity) produces precisely $\dot A=i\Omega A-\int_0^t K(t-s)A(s)\,\mathrm ds+F(t)$ , with the memory kernel $K$ and random force $F$ defined by the projection. The Caldeira–Leggett harmonic bath is a concrete realization where the projection can be carried out in closed form. So the GLE is not a phenomenological guess — it is the exact reduced equation of motion, and the Mori–Zwanzig formalism is the general proof that such a reduction always exists.