The Green-Kubo Relations

Pour cream into coffee and it spreads; push a current through a wire and it heats; shear a fluid and it resists. Each irreversible flow is governed by a transport coefficient — a diffusion constant, an electrical conductivity, a shear viscosity, a thermal conductivity — that you would think you must measure by driving the system out of equilibrium. The Green-Kubo relations say otherwise. They express every one of these coefficients as the time integral of an equilibrium autocorrelation function of the corresponding microscopic flux. The system never has to be pushed: the spontaneous thermal fluctuations of the flux, running quietly at equilibrium, already encode exactly how the system will dissipate when you do push it. The physical bridge is Onsager's regression hypothesis — a spontaneous fluctuation decays, on average, by the same law that governs the relaxation of a deliberately imposed one — combined with linear response theory. Switch on a weak field conjugate to a variable, track the induced flux to first order, take the zero-frequency limit of the response, and the transport coefficient emerges as an equilibrium flux-flux correlation integrated over time. Dissipation is thereby fixed entirely by how fast fluctuations forget themselves. This is the transport-coefficient generalization of the fluctuation-dissipation theorem, and it contains the Einstein relation $D=\mu k_BT$ as its diffusive special case: integrating the velocity autocorrelation gives the same $D$ as the mobility route or the mean-square-displacement slope. In practice, Green-Kubo is the foundation of equilibrium molecular dynamics: simulate an undriven system, record the autocorrelation of the stress tensor or heat current, integrate, and read off the coefficient — provided you respect the hydrodynamic long-time tails that make the integrals delicate.

Key takeaways

A transport coefficient equals the time integral of an equilibrium autocorrelation function of its conjugate microscopic flux — no external driving is needed to measure dissipation.
The generic template is $L=\frac{1}{k_BTV}\int_0^\infty\langle J(0)J(t)\rangle_\mathrm{eq}\,\mathrm{d}t$ , specialized to $D$ , $\eta$ , $\kappa$ , $\sigma$ by choosing the appropriate flux $J$ and prefactor.
It follows from linear response: the transport coefficient is the zero-frequency limit of the response function, which the Kubo identity turns into an equilibrium flux correlation.
Green-Kubo is the dynamical face of the fluctuation-dissipation theorem and reduces to the Einstein relation $D=\mu k_BT$ for diffusion; the velocity-autocorrelation integral and the mean-square-displacement slope encode the same $D$ .
It is the engine of equilibrium molecular dynamics for transport, with the caveat of algebraic hydrodynamic long-time tails ( $C_v\sim t^{-d/2}$ ) that can make the integrals converge slowly or, in $d=2$ , diverge.
Key idea: dissipation is memory. A transport coefficient is the total area under the equilibrium autocorrelation of its flux, $L=\frac{1}{k_BTV}\int_0^\infty\langle J(0)J(t)\rangle\,\mathrm{d}t$ ; how fast the flux forgets itself sets how much the system resists a gradient.

The core insight

A transport coefficient is the time integral of the equilibrium autocorrelation function of its conjugate microscopic flux, $L=\frac{1}{k_BTV}\int_0^\infty\langle J(0)J(t)\rangle_\mathrm{eq}\,\mathrm{d}t$ . Dissipation is encoded entirely in how fast spontaneous flux fluctuations decay — the system reveals its transport coefficients while sitting at rest.

Where this lives today

Green-Kubo is the everyday workhorse for computing transport coefficients from atomistic simulation. To get the thermal conductivity of a thermoelectric material or a glass, a researcher runs an equilibrium molecular-dynamics trajectory, records the microscopic heat current at every step, computes its autocorrelation, and integrates — no temperature gradient, no driven boundaries. The method shines precisely where the Boltzmann/relaxation-time picture fails: strongly anharmonic crystals and amorphous solids, where phonon quasiparticles are ill-defined but the heat-current correlation is still perfectly well-posed. The same relations power modern machine-learning interatomic potentials: once a neural-network potential reproduces forces, its Green-Kubo viscosity and conductivity become predictions testable against experiment, and the long-time-tail subtleties re-enter as questions about finite simulation-box size and trajectory length. The 2D logarithmic divergence, once a curiosity, is now a practical concern for transport in graphene and other atomically thin materials, where "the shear viscosity" must be quoted with a system size.

The physical picture

Take a small patch of fluid at equilibrium, with no gradient imposed. Molecules constantly cross any imagined plane, so the local momentum current — the shear stress $\sigma_{xy}$ — is never exactly zero; it fluctuates. The crucial and slightly counter-intuitive point is that the faster those stress fluctuations decorrelate, the less viscous the fluid is. Viscosity is memory: a fluid that remembers its stress for a long time resists shear strongly. The Green-Kubo integral literally sums the lifetime of stress correlations, weighted by their strength, to produce $\eta$ . Onsager's regression hypothesis is the conceptual key that licenses this. A fluctuation you deliberately create with a weak field decays, on average, exactly as a spontaneous equilibrium fluctuation of the same variable would. So the response to driving — which is what a transport coefficient measures — is computable entirely from never-driven equilibrium correlations. The system's own thermal noise is a complete diagnostic of its dissipative behaviour. Diffusion makes the idea concrete. Write $D=\frac{1}{d}\int_0^\infty\langle\mathbf v(0)\cdot\mathbf v(t)\rangle\,\mathrm dt$ . A particle that quickly randomizes its velocity has a small velocity-memory area and diffuses slowly; a near-ballistic particle carries velocity memory far and diffuses fast. The integral is precisely that memory area. This connects to the whole stochastic circle — Brownian motion, the Langevin equation, the Ornstein-Uhlenbeck process — because the velocity autocorrelation is exactly the object those theories compute.

Figure

Punchline: the velocity autocorrelation and its area — the diffusion constant, suppressed by a back-scattering dip

Figure

Green-Kubo as the ω→0 corner of linear response: the FDT circle

The core equations

Green-Kubo is a corollary of linear response theory. When a weak field $F(t)$ couples to a variable $A$ , first-order time-dependent perturbation theory of the phase-space distribution gives the induced average of any observable $B$ as a convolution of the field with an equilibrium correlation function. This is the master formula from which every transport coefficient descends.

\langle B(t)\rangle - \langle B\rangle_0 = \beta\int_{-\infty}^{t}\mathrm{d}t'\,F(t')\,\langle \dot A(t')\,\Delta B(t)\rangle_0 \qquad (1)

where: $H(t)=H_0-F(t)A$ is the perturbed Hamiltonian; $\langle\cdot\rangle_0$ is the equilibrium average over the unperturbed dynamics; $\beta=1/k_BT$ ; $\Delta B=B-\langle B\rangle_0$ . Read it as: the response of $B$ now is the field in the past, weighted by how strongly the flux $\dot A$ then correlates with $B$ now — causality restricts the integral to $t'<t$ . Stationarity of equilibrium correlations — they depend only on time differences — lets the time derivative be moved between the two factors, producing the after-effect (response) function in either of two equivalent forms.

\phi_{BA}(t)=\beta\langle\dot A(0)\,B(t)\rangle_0=-\beta\langle A(0)\,\dot B(t)\rangle_0 \qquad (2)

where: $\phi_{BA}(t)$ is nonzero only for $t>0$ (causality); $\dot A$ is the flux conjugate to the perturbation. This is the object whose zero-frequency limit is a transport coefficient. The linear-response framework it lives in is established separately; here it is the launching point. Taking the static ( $\omega\to0$ ) limit of the response and identifying the driving gradient with the conjugate flux gives the generic Green-Kubo relation — the single template behind all four transport coefficients.

L=\lim_{\omega\to0}\mathrm{Re}\,\tilde\chi(\omega)=\frac{1}{k_BTV}\int_0^\infty\langle J(0)\,J(t)\rangle_0\,\mathrm{d}t \qquad (3)

where: $L$ is the transport coefficient; $J$ the microscopic flux conjugate to the gradient; $V$ the volume; $\tilde\chi(\omega)$ the frequency-dependent susceptibility. The one-sided ( $t>0$ ) integral is where irreversibility enters an otherwise time-symmetric correlation: dissipation is the retarded half of a symmetric memory. For diffusion the flux is the single-particle velocity, and the same $D$ emerges two ways — as the velocity-autocorrelation integral (Green-Kubo) or as the long-time slope of the mean-square displacement (Einstein).

D=\frac{1}{d}\int_0^\infty\langle\mathbf v(0)\cdot\mathbf v(t)\rangle\,\mathrm{d}t=\lim_{t\to\infty}\frac{\langle|\mathbf r(t)-\mathbf r(0)|^2\rangle}{2d\,t} \qquad (4)

where: $\mathbf v(t)$ is the particle velocity, $\mathbf r(t)$ its position, $d$ the spatial dimension. The two forms are one identity read forwards and backwards; the derivation below shows an integration by parts converts one into the other. Two microscopic fluxes make the template concrete for real materials. The off-diagonal stress tensor drives viscosity; it has a kinetic (momentum-carrying) part and a virial (force-carrying) part.

\sigma_{\alpha\beta}=\frac{1}{V}\left[\sum_i m_i v_{i\alpha}v_{i\beta}+\sum_{i<j}r_{ij\alpha}\,f_{ij\beta}\right] \qquad (5)

where: the first sum is the ideal-gas momentum flux, the second the interaction contribution ( $r_{ij}$ = pair separation, $f_{ij}$ = pair force). In a dilute gas the kinetic part dominates; in a dense liquid the virial part does. Finally, the velocity autocorrelation does not decay purely exponentially: hydrodynamic backflow gives an algebraic long-time tail that makes the Green-Kubo integrals subtle and, in two dimensions, ill-defined.

C_v(t)\sim\frac{k_BT}{m}\frac{1}{[4\pi(D+\nu)t]^{d/2}}\quad(t\to\infty) \qquad (6)

where: $\nu=\eta/\rho$ is the kinematic viscosity; the tail exponent is $-d/2$ . In $d=2$ this is $t^{-1}$ , so $\int^\infty C_v\,\mathrm dt$ diverges logarithmically — a genuine physical statement that 2D transport coefficients are not strictly finite.

Derivation AGreen-Kubo formula for electrical conductivity

The goal is to derive the Kubo expression for the DC electrical conductivity as the time integral of the equilibrium current-current correlation, starting from the linear-response master formula and taking the static-field limit. The strategy is to identify the perturbation (a uniform field coupled to the polarization) and the flux it conjugates (the electric current), insert them into the master formula, then hold the field constant and integrate — the one subtlety is keeping track of extensive versus intensive current so the volume factor lands correctly. Watch that the conjugate variable is the polarization $A=\sum_i q_i x_i$ , whose rate is the current, not the current itself.

Step 1. Identify the perturbation and its conjugate flux. A weak uniform field along $x$ perturbs by $H'(t)=-A\,E(t)$ with $A=\sum_i q_i x_i$ , so the conjugate rate is $\dot A=\sum_i q_i v_{ix}=V J^e_x$ , the extensive current.

Step 2. Insert into the master formula, measuring $B=\dot A=VJ^e_x$ and using $\phi_{BA}=\beta\langle\dot A(0)\dot A(t)\rangle_0$ :

\langle VJ^e_x(t)\rangle=\beta\int_{-\infty}^t\mathrm dt'\,E(t')\,\langle\dot A(t')\,\dot A(t)\rangle_0.

Step 3. Take the static limit. Set $E(t')=E$ constant and substitute $\tau=t-t'$ :

\langle VJ^e_x\rangle=\beta E\int_0^\infty\langle\dot A(0)\,\dot A(\tau)\rangle_0\,\mathrm d\tau.

Step 4. Convert to current density and divide by the field. Using $\dot A=VJ^e_x$ and $\sigma=\langle J^e_x\rangle/E$ (Ohm's law):

\sigma=\frac{\beta}{V}\int_0^\infty\langle\dot A(0)\dot A(\tau)\rangle\,\mathrm d\tau=\frac{V}{k_BT}\int_0^\infty\langle J^e_x(0)J^e_x(\tau)\rangle\,\mathrm d\tau.

Why the volume factor flips: $\dot A$ is extensive ( $\propto V$ ), so $\dot A^2\propto V^2$ ; dividing by the single $V$ from $A$ 's definition leaves one factor $V$ multiplying the intensive current correlation. *Verified: for a Drude memory $\langle J^e(0)J^e(t)\rangle=(nq^2k_BT/mV)e^{-t/\tau}$ the integral gives $\sigma=nq^2\tau/m$ to 1 part in $10^9$ — verify_green-kubo-relations.py, matching Kubo (1957).*

Go deeper · Derivation B · 1 minEquivalence of the Green-Kubo and Einstein forms of D

Derivation BEquivalence of the Green-Kubo and Einstein forms of D

The goal is to prove that the velocity-autocorrelation integral (Green-Kubo) and the mean-square-displacement slope (Einstein) give the same diffusion constant, and that both equal $k_BT/m\gamma$ , tying the whole result to the Einstein relation $D=\mu k_BT$ . The strategy is to write the displacement as a time integral of velocity, expand the mean square into a double time integral of the velocity autocorrelation, reduce it with stationarity, then differentiate to expose the long-time slope. Watch the change of variables that turns the double integral into a single one with a $(t-\tau)$ weight — that weight is what makes the slope, not the value, the diffusion constant.

Step 1. Write the mean-square displacement as a double integral of the velocity autocorrelation:

\langle|\Delta\mathbf r(t)|^2\rangle=\int_0^t\!\!\int_0^t\langle\mathbf v(t')\cdot\mathbf v(t'')\rangle\,\mathrm dt'\,\mathrm dt''.

Step 2. Use stationarity, $\langle\mathbf v(t')\cdot\mathbf v(t'')\rangle=C_v(t''-t')$ , and fold the double integral into a single one:

\langle|\Delta\mathbf r(t)|^2\rangle=2\int_0^t(t-\tau)\,C_v(\tau)\,\mathrm d\tau.

Step 3. Differentiate in $t$ ; the $(t-\tau)$ weight differentiates to $1$ and the long-time limit picks out the full integral:

\frac{\mathrm d}{\mathrm dt}\langle|\Delta\mathbf r|^2\rangle=2\int_0^tC_v(\tau)\,\mathrm d\tau\xrightarrow{t\to\infty}2d\,D.

Step 4. Evaluate on the Ornstein-Uhlenbeck velocity correlation $C_v(t)=\tfrac{d\,k_BT}{m}e^{-\gamma t}$ (which the Langevin equation and the Ornstein-Uhlenbeck process establish): $D=\tfrac1d\int_0^\infty C_v=\tfrac{k_BT}{m\gamma}=\mu k_BT$ with mobility $\mu=1/m\gamma$ . Why all three routes agree: the MSD slope, the velocity-integral, and the mobility are three linear functionals of the same exponential memory, and integrating $e^{-\gamma t}$ once gives the $1/\gamma$ that all share. *Verified: numerical integration of the OU correlation reproduces $D=k_BT/m\gamma$ to 1 part in $10^9$ , and the symbolic MSD slope equals $2\int_0^\infty C_v$ — verify_green-kubo-relations.py.*

Limiting cases

The single template $L=\frac{1}{k_BTV}\int_0^\infty\langle J(0)J(t)\rangle\,\mathrm dt$ specializes to each coefficient by choosing the flux $J$ and prefactor:

Case	Result	Meaning
Self-diffusion	$D=\frac{1}{d}\int_0^\infty\langle\mathbf v_i(0)\cdot\mathbf v_i(t)\rangle\,\mathrm dt$	single-particle velocity flux; no $1/V$ (one-particle property); $d$ = dimension
Shear viscosity	$\eta=\frac{V}{k_BT}\int_0^\infty\langle\sigma_{xy}(0)\sigma_{xy}(t)\rangle\,\mathrm dt$	off-diagonal stress (kinetic + virial); averaging 3 independent components cuts variance ( $\sigma_{xy}$ here is extensive)
Thermal conductivity	$\kappa=\frac{1}{k_BT^2V}\int_0^\infty\langle J^q_x(0)J^q_x(t)\rangle\,\mathrm dt$	heat-current flux; note $T^2$ — heat couples to $\nabla(1/T)=-\nabla T/T^2$
Electrical conductivity	$\sigma=\frac{1}{k_BTV}\int_0^\infty\langle J^e_x(0)J^e_x(t)\rangle\,\mathrm dt$	charge current $J^e_x=\sum_iq_iv_{ix}$ ; the $\omega\to0$ Kubo conductivity

Common trapThe Green-Kubo integral does not always converge — long-time tails are physical, not numerical

The velocity autocorrelation has a hydrodynamic long-time tail $C_v(t)\sim t^{-d/2}$ (Alder-Wainwright, 1970): backflow vortices feed momentum back to a tagged particle. In $d=2$ this is $t^{-1}$ and the diffusion/viscosity integrals diverge logarithmically, so two-dimensional transport coefficients are not strictly finite — a genuine physical statement, not a simulation artifact. Two further traps: the equilibrium average must be over the unperturbed dynamics (never drive the system to measure its Green-Kubo coefficient), and the microscopic heat current must include the full interaction-energy transport, not just kinetic pieces, or $\kappa$ comes out wrong for dense or many-body systems.

Worked examples

Example 1 — Diffusion constant from a velocity autocorrelation (the real number). A tagged argon-like atom ( $m=6.63\times10^{-26}\,$ kg) at $T=300\,$ K has an Ornstein-Uhlenbeck velocity autocorrelation $C_v(t)=\frac{k_BT}{m}e^{-\gamma t}$ with friction rate $\gamma=1.0\times10^{12}\,$ s $^{-1}$ . Find the diffusion constant.

Step 1. The mean-square velocity is $\langle v^2\rangle_{1\mathrm D}=k_BT/m=(1.381\times10^{-23}\times300)/(6.63\times10^{-26})=6.25\times10^{4}\,$ m $^2$ /s $^2$ (equipartition per component).

Step 2. The Green-Kubo integral of a single exponential is $\int_0^\infty e^{-\gamma t}\,\mathrm dt=1/\gamma$ , so

D=\int_0^\infty\frac{k_BT}{m}e^{-\gamma t}\,\mathrm dt=\frac{k_BT}{m\gamma}=\frac{6.25\times10^{4}}{1.0\times10^{12}}\ \mathrm{m^2/s}.

Step 3. Evaluate: $D=\mathbf{6.25\times10^{-8}\ m^2/s}$ . This equals $\mu k_BT$ with mobility $\mu=1/m\gamma=1.51\times10^{13}\,$ s/kg, the Einstein relation, and gives an MSD slope $\mathrm d\langle\Delta x^2\rangle/\mathrm dt=2D=1.25\times10^{-7}\,$ m $^2$ /s.

Example 2 — DC conductivity from a Drude current correlation. A metal has $n=1.0\times10^{28}\,$ m $^{-3}$ carriers of charge $q=1.602\times10^{-19}\,$ C, mass $m=6.63\times10^{-26}\,$ kg, and scattering time $\tau=1.0\times10^{-14}\,$ s, so the current autocorrelation is $\langle J^e_x(0)J^e_x(t)\rangle=(nq^2k_BT/mV)e^{-t/\tau}$ (unit volume). Find $\sigma$ .

Step 1. Apply the electrical Green-Kubo relation $\sigma=\frac{V}{k_BT}\int_0^\infty\langle J^e_x(0)J^e_x(t)\rangle\,\mathrm dt$ . The $V/k_BT$ prefactor times the $k_BT/V$ inside cancels except for the exponential integral.

Step 2. The integral $\int_0^\infty e^{-t/\tau}\,\mathrm dt=\tau$ , leaving the Drude result $\sigma=nq^2\tau/m$ .

Step 3. Evaluate: $\sigma=\dfrac{(10^{28})(1.602\times10^{-19})^2(10^{-14})}{6.63\times10^{-26}}=\mathbf{3.87\times10^{1}\ S/m}$ … (more precisely $38.7\,$ S/m for these toy numbers). The full frequency dependence is $\sigma(\omega)=\sigma(0)/(1-\mathrm i\omega\tau)$ ; Green-Kubo is exactly its $\omega\to0$ corner.

Origins · 1 mingreen-kubo relation and how this was discovered

green-kubo relation

named for Melville S. Green (1922–1979) and Ryogo Kubo (1920–1995); relations established 1954 & 1957 · The Green-Kubo relations bear the names of two theorists who, three years apart, expressed transport coefficients as equilibrium time-correlation integrals. Melville S. Green, in 1954, obtained the correlation-function expressions for transport coefficients from a general theory of irreversible processes; Ryogo Kubo, in 1957, derived the same structure as a corollary of his linear-response theory, in which the response of a system to a weak field is a canonical correlation function of the unperturbed system. Kubo's formulation — with its Kubo-transformed correlation functions and the fluctuation-dissipation theorem it contains — became the standard language of linear-response transport theory and is the framework in which the Green-Kubo relations are usually presented today.

History

1866

— Maxwell computes gas viscosity from kinetic theory and predicts it is independent of density

1928

— Nyquist and Johnson relate thermal voltage noise to resistance, the first fluctuation-dissipation statement

1931

— Onsager states the regression hypothesis: spontaneous fluctuations decay by the macroscopic relaxation law

1954

— Melville S. Green expresses transport coefficients as equilibrium time-correlation integrals

1957

— Ryogo Kubo derives the correlation-function form from linear-response theory

1970

— Alder and Wainwright discover the algebraic long-time tail of the velocity autocorrelation in molecular dynamics

Going further

References. Kubo, Toda & Hashitsume, Statistical Physics II: Nonequilibrium Statistical Mechanics, ch. 4 · Pathria & Beale, Statistical Mechanics, 4th ed., ch. on fluctuation-dissipation and transport · Kardar, Statistical Physics of Particles, ch. on kinetic theory and linear response · Reif, Fundamentals of Statistical and Thermal Physics, ch. 15 · Sethna, Statistical Mechanics: Entropy, Order Parameters and Complexity, ch. 10 · Hansen & McDonald, Theory of Simple Liquids, ch. 7-8 · Zwanzig, Nonequilibrium Statistical Mechanics, ch. on linear response.

Provenance. A verifier should: (1) reproduce the linear-response/Kubo master formula from first-order perturbation of the Liouville equation (Kubo-Toda-Hashitsume ch. 4; Zwanzig); (2) confirm the diffusion derivation $D=k_BT/m\gamma$ from the Ornstein-Uhlenbeck velocity correlation and the MSD-slope equivalence (Reif ch. 15; Pathria); (3) check the Drude integral $\sigma=nq^2\tau/m$ ; (4) verify prefactor conventions for $\eta$ ( $V/k_BT$ extensive stress) and $\kappa$ ( $1/k_BT^2V$ ) against Hansen-McDonald ch. 7-8 and Kardar; (5) confirm the long-time-tail exponent $-d/2$ (Alder-Wainwright; Hansen-McDonald). All numbers (Drude conductivity, Einstein relation $D=\mu k_BT$ , tail exponent) are standard textbook values; none fabricated.

Builds on: The Fluctuation-Dissipation Theorem · The Langevin Equation · The Einstein Relation · The Ornstein-Uhlenbeck Process. Leads to: Brownian Motion · The Diffusion Equation · The Fokker-Planck Equation.

Do I have to drive the system out of equilibrium to measure a transport coefficient with Green-Kubo?

No — that is the whole point. You run an undriven equilibrium simulation and record the spontaneous fluctuations of the appropriate flux (velocity for $D$ , stress for $\eta$ , heat current for $\kappa$ , charge current for $\sigma$ ), then integrate their autocorrelation over time. Onsager's regression hypothesis guarantees that these equilibrium fluctuations decay by the same law that governs the relaxation of a deliberately imposed gradient, so the never-driven correlation contains the full transport information. Driving the system would in fact spoil the measurement, because the Green-Kubo average must be taken over the unperturbed dynamics.

What is the difference between the Green-Kubo relations and the fluctuation-dissipation theorem?

They are the same physics at different frequencies. The fluctuation-dissipation theorem relates the full frequency-dependent response $\chi(\omega)$ to the spectrum of equilibrium fluctuations. The Green-Kubo relations are the zero-frequency ( $\omega\to0$ ) corner of that statement, integrated over time: a transport coefficient is a $\chi(0)$ -type quantity, so it is the total time integral of the corresponding flux autocorrelation. Green-Kubo is the transport-coefficient specialization; the fluctuation-dissipation theorem is the general frequency statement it descends from.

Why does viscosity get larger when stress correlations last longer? That feels backwards.

It is worth slowing down on. Viscosity measures how strongly a fluid resists shear, and that resistance is momentum transport with memory. If the shear stress fluctuations decorrelate slowly — the fluid "remembers" its stress state for a long time — then a small imposed strain builds up a large accumulated stress response, i.e. high viscosity. The Green-Kubo integral $\int_0^\infty\langle\sigma_{xy}(0)\sigma_{xy}(t)\rangle\,\mathrm dt$ is exactly the accumulated memory. A fluid whose stress fluctuations vanish instantly has zero viscosity (an ideal gas of non-interacting particles in the collisionless limit); one that never forgets is a solid.

Why is the Green-Kubo integral one-sided, from 0 to infinity, if equilibrium correlations are time-symmetric?

The equilibrium correlation function $\langle J(0)J(t)\rangle$ is indeed even in $t$ for time-reversal-even fluxes. The one-sidedness comes from causality in the response function: a field applied now can only affect the flux later, so the response $\phi(t)$ is retarded, nonzero only for $t>0$ . When you take the $\omega\to0$ limit you integrate that retarded response, which is the correlation restricted to $t>0$ . Physically, the arrow of time in a transport coefficient lives entirely in this retarded integration — the underlying correlation is symmetric, but dissipation samples only its future half.

How is the Einstein relation

D=\mu k_BT

a special case of Green-Kubo?

The Einstein relation ties the diffusion constant $D$ to the mobility $\mu$ (drift velocity per unit force). Green-Kubo writes $D$ as the integral of the velocity autocorrelation. For a Brownian particle whose velocity obeys the Ornstein-Uhlenbeck correlation $\langle v(0)v(t)\rangle=(k_BT/m)e^{-\gamma t}$ , that integral gives $k_BT/m\gamma$ ; and since the mobility is $\mu=1/m\gamma$ , this is exactly $\mu k_BT$ . So the Einstein relation is the diffusion instance of the general Green-Kubo template, with the friction rate $\gamma$ setting both the velocity-memory time and the mobility.

What are hydrodynamic long-time tails and why do they matter for Green-Kubo?

In the 1970 molecular-dynamics work of Alder and Wainwright, the velocity autocorrelation was found to decay not exponentially but algebraically, $C_v(t)\sim t^{-d/2}$ , because a moving particle sets up a hydrodynamic vortex that feeds momentum back to it. This tail matters for Green-Kubo because the integral runs to infinity: a $t^{-d/2}$ tail converges only for $d\ge3$ . In two dimensions ( $t^{-1}$ ) the diffusion and viscosity integrals diverge logarithmically, meaning 2D transport coefficients depend weakly on system size and time — they must be quoted with a length scale, not as pure numbers.

Can Green-Kubo handle systems where the Boltzmann relaxation-time picture fails, like glasses?

Yes, and this is a major reason it is used. The Boltzmann transport equation and the relaxation-time approximation assume well-defined quasiparticles (phonons, electrons) with a scattering time. In strongly anharmonic crystals, amorphous solids, and glasses, no such quasiparticles exist, so that framework breaks down. The Green-Kubo heat-current autocorrelation makes no quasiparticle assumption whatsoever — it is defined directly from the microscopic dynamics — so it correctly captures thermal transport in these disordered systems, provided the many-body energy-current operator is defined with its full interaction terms.