Linear Response Theory

Poke a system in thermal equilibrium — apply a small voltage, a weak magnetic field, a gentle push — and it stirs. Linear response theory answers, exactly and universally, how much it stirs to first order in the poke. The astonishing content of the answer is that you never have to solve for the complicated driven, out-of-equilibrium state at all: the entire first-order response is encoded in the spontaneous equilibrium fluctuations of the undriven system. The drum's ring after a single tap is governed by the very same modes that rattle when thermal noise alone shakes it. Compute one equilibrium correlation function and you have predicted every weak-field experiment. The setup is spare and general. A Hamiltonian $H_0$ sits in equilibrium at temperature $T=1/k_B\beta$ ; a weak field $F(t)$ couples to an observable $A$ through $H'(t)=-A\,F(t)$ ; we ask for the induced change $\delta\langle B(t)\rangle$ of some observable $B$ . Linearising the density matrix in $F$ produces a convolution $\delta\langle B(t)\rangle=\int^t\chi_{BA}(t-t')F(t')\,\mathrm{d}t'$ , and the response kernel $\chi_{BA}$ turns out to be the Kubo formula — a commutator average (quantum) or a time-derivative correlation (classical) evaluated purely in the equilibrium ensemble. Three structural pillars give the theory its reach. Causality — no response before the kick, $\chi(t<0)=0$ — makes $\chi(\omega)$ analytic in the upper half $\omega$ -plane and forces the Kramers–Kronig relations. Fluctuation–dissipation ties the absorptive part $\chi''(\omega)$ to the equilibrium noise spectrum. Onsager regression declares that a fluctuation, whether born by chance or imposed by a field, decays back by the same average law. From this single framework flow every transport coefficient — conductivity, viscosity, diffusion, mobility — as Green–Kubo integrals over equilibrium currents, and the Langevin and Fokker–Planck descriptions of Brownian motion inherit their noise–friction balance from it.

Key takeaways

Linear response theory gives the first-order change of an observable under a weak time-dependent perturbation $-A\,F(t)$ entirely in terms of an equilibrium correlation function — the driven nonequilibrium state never has to be solved.
The central result is the Kubo response function $\chi_{BA}(t)=\tfrac{i}{\hbar}\langle[B(t),A]\rangle_0\,\Theta(t)$ (quantum) or $-\beta\langle\dot A(0)B(t)\rangle_0\,\Theta(t)$ (classical), derived by first-order perturbation of the density matrix.
Causality — response only after the kick, $\chi(t)=0$ for $t<0$ — makes $\chi(\omega)$ analytic in the upper half-plane and forces the Kramers–Kronig relations between its reactive ( $\chi'$ ) and dissipative ( $\chi''$ ) parts.
The static limit is a pure thermodynamic fluctuation: $\chi(0)=\beta\,\mathrm{Var}(A)$ , the isothermal susceptibility as an equilibrium variance.
It unifies transport coefficients (conductivity, viscosity, diffusion, mobility) as Green–Kubo integrals of current–current correlations and underlies the fluctuation–dissipation theorem and the Einstein relation.
Key idea: to first order in a weak field a system forgets it is being driven — its response equals an equilibrium correlation function of the coupling operator with the observed operator, computed once and for all in the unperturbed ensemble.

The core insight

To first order in a weak field a system forgets it is being driven: its response is the equilibrium correlation function of the coupling operator with the observed operator, computed once and for all in the unperturbed ensemble.

Where this lives today

Linear response theory is the quiet engine beneath most of what condensed-matter and chemical physics actually measures. Every spectroscopy — dielectric, optical, magnetic resonance, inelastic neutron and X-ray scattering — reports a dissipative susceptibility $\chi''(\mathbf{q},\omega)$ or a correlation spectrum, and the Kramers–Kronig relations are used daily to reconstruct refractive indices and optical constants from absorption data. In simulation, the Green–Kubo route computes ionic conductivities, viscosities, and thermal conductivities of liquids and electrolytes directly from equilibrium molecular-dynamics trajectories, without ever imposing a field. The frontier is where the framework breaks. In glasses, sheared suspensions, biological active matter, and driven quantum systems, the equilibrium fluctuation–dissipation relation is violated, and the size of the violation defines a frequency-dependent effective temperature that has become a standard diagnostic of how far a system sits from equilibrium. Extending Kubo-type identities to nonequilibrium steady states — with explicit entropy-production corrections — is an active branch of stochastic thermodynamics, and it descends directly from the causal, equilibrium theory laid out here.

The physical picture

Tap a drum once and the membrane rings for all later times in a way fixed by its intrinsic modes — the same modes that hum when thermal noise alone rattles it. Linear response makes this identity precise: the driven response and the spontaneous ringing are governed by one object, the equilibrium correlation function. The field only decides how hard and when you tap; the shape of the reply is a property of equilibrium. Causality is the hidden structural constraint that does most of the work. Because nothing responds before it is kicked, the response function is one-sided in time, $\chi(t)=0$ for $t<0$ . A one-sided function has a Fourier transform analytic in half the complex-frequency plane, and that analyticity is exactly the Kramers–Kronig relations linking refraction (the reactive $\chi'$ ) to absorption (the dissipative $\chi''$ ). Dispersion and dissipation are the two faces of one causal object — a theme carried forward in Kramers-Kronig relations and generalized susceptibility and spectral functions. The dissipative part is where energy goes to die. Driving at frequency $\omega$ deposits power $\propto\omega\,\chi''(\omega)|F|^2$ , and the fluctuation-dissipation theorem then declares that this same $\chi''$ measures how loudly the system fluctuates on its own at that frequency: friction and noise are inseparable, tied by the temperature.

Figure

A single impulse at τ = 0 and the causal response: nothing before the kick, a relaxing tail after — the sharp onset is what forces analyticity

Onsager's regression hypothesis gives the picture a thermodynamic soul. A macroscopic disturbance — however it arose, by an experimenter's field or by a rare spontaneous fluctuation — decays back to equilibrium following the same average law. You cannot tell, from the decay alone, whether the bump was designed or accidental. This is why transport coefficients are integrals over the autocorrelation of spontaneous currents, and why response and relaxation functions are one object seen from different angles.

Figure

Driven decay vs spontaneous decay: Onsager regression says a field-imposed displacement and a chance fluctuation relax by the identical average law

The core equations

The perturbation is a weak classical field switched on in time, coupling to a chosen observable of the equilibrium system. This single line fixes the entire structure of what follows; every result is a functional of the pair $(A,B)$ and the temperature.

H(t)=H_0-A\,F(t) \qquad (1)

where: $H_0$ — unperturbed Hamiltonian, in thermal equilibrium at $T=1/k_B\beta$ ; $A$ — operator (or phase-space function) that the field couples to; $F(t)$ — weak classical (c-number) external field, switched on adiabatically from $t\to-\infty$ . Reading it physically: $A$ is the "handle" the field grabs (dipole for an electric field, magnetization for a magnetic field), and $B$ is whatever you choose to measure. To first order in $F$ , the induced change of the observed quantity is a causal convolution of the field with a response kernel. The kernel depends only on the equilibrium system, not on $F$ — that is the whole point.

\delta\langle B(t)\rangle=\int_{-\infty}^{t}\chi_{BA}(t-t')\,F(t')\,\mathrm{d}t' \qquad (2)

where: $\delta\langle B(t)\rangle$ — first-order change in the expectation of $B$ ; $\chi_{BA}(\tau)$ — response (after-effect) function, causal: $\chi_{BA}(\tau)=0$ for $\tau<0$ . The upper limit $t$ (not $+\infty$ ) is causality made manifest: only the past of the field can matter. The quantum Kubo formula identifies the kernel as an equilibrium average of a commutator of Heisenberg operators. Its structure — a commutator, not a plain product — is the fingerprint of the perturbation entering through $[H',\rho_0]$ .

\chi_{BA}(\tau)=\frac{i}{\hbar}\,\big\langle[B(\tau),A(0)]\big\rangle_0\,\Theta(\tau) \qquad (3)

where: $[\,,\,]$ — commutator of Heisenberg operators evolved under $H_0$ ; $\langle\cdot\rangle_0$ — average in the unperturbed canonical ensemble $\rho_0=e^{-\beta H_0}/Z$ ; $\Theta(\tau)$ — Heaviside step enforcing causality. The commutator measures how $B$ at time $\tau$ fails to commute with $A$ at time $0$ — precisely the quantum content of "how the kick propagates." In the classical (high-temperature) limit the commutator becomes a Poisson bracket and the response collapses to the time-derivative of a correlation function — the form that powers all classical transport.

\chi_{BA}(\tau)=-\beta\,\big\langle\dot A(0)\,B(\tau)\big\rangle_0\,\Theta(\tau)=\beta\,\frac{\mathrm{d}}{\mathrm{d}\tau}\langle A(0)\,B(\tau)\rangle_0\,\Theta(\tau) \qquad (4)

where: $\dot A=\{A,H_0\}$ — time derivative of $A$ along unperturbed trajectories; $\beta=1/k_BT$ — inverse temperature. The temperature enters as a single factor $\beta$ : warmer systems fluctuate more, so they respond more, and the response is literally $\beta$ times the correlation's slope. The frequency-domain susceptibility is the Fourier transform of the causal kernel; because $\chi$ vanishes for $\tau<0$ the transform runs from $0$ , splitting the answer into an even reactive part and an odd dissipative part.

\chi(\omega)=\int_{0}^{\infty}\chi(\tau)\,e^{i\omega\tau}\,\mathrm{d}\tau=\chi'(\omega)+i\chi''(\omega) \qquad (5)

where: $\chi'(\omega)$ — reactive (in-phase, dispersive) part, even in $\omega$ ; $\chi''(\omega)$ — dissipative (out-of-phase, absorptive) part, odd in $\omega$ . The lower limit $0$ is causality; it is why $\chi'$ and $\chi''$ cannot be chosen independently. Causality's payoff is a dispersion relation: the reactive part is a principal-value transform of the dissipative part. This is the Kramers–Kronig relation, developed in full in the sibling article, quoted here because it is forced by the analytic structure above.

\chi'(\omega)=\frac{1}{\pi}\,\mathcal{P}\!\!\int_{-\infty}^{\infty}\frac{\chi''(\omega')}{\omega'-\omega}\,\mathrm{d}\omega' \qquad (6)

where: $\mathcal{P}$ — Cauchy principal value. Physically: measure absorption ( $\chi''$ ) at all frequencies and you have determined dispersion ( $\chi'$ ) everywhere, with no extra input. At zero frequency the whole apparatus collapses to a thermodynamic identity — the static susceptibility is an equilibrium variance. This is the bridge between dynamics and ordinary thermodynamics.

\chi(0)=\beta\big(\langle A^2\rangle_0-\langle A\rangle_0^2\big)=\beta\,\mathrm{Var}(A) \qquad (7)

where: $\mathrm{Var}(A)$ — equilibrium variance of the coupling observable. Special cases: density fluctuations give the isothermal compressibility, magnetization fluctuations give the magnetic susceptibility — response coefficients are fluctuations.

Derivation AThe Kubo response function from first-order perturbation of the density matrix

The goal is to derive the response kernel $\chi_{BA}$ from scratch, showing why the perturbed, driven ensemble drops out and leaves an equilibrium commutator average. We work in the interaction picture, solve the Liouville–von Neumann equation to first order in the field, and use the cyclic property of the trace to move the perturbation off the density matrix and onto the observables. The strategy is to expand $\rho=\rho_0+\delta\rho$ , keep only the term linear in $F$ , and recognize the result as a trace over the unperturbed $\rho_0$ — which is exactly why one never solves the hard driven problem.

Step 1. Write the quantum Liouville equation for the full density matrix, $i\hbar\,\partial_t\rho=[H_0+H'(t),\rho]$ , with $H'(t)=-A\,F(t)$ , and split $\rho=\rho_0+\delta\rho$ where $[H_0,\rho_0]=0$ (equilibrium is stationary). Why: the equilibrium ensemble commutes with $H_0$ , so all the time dependence lives in the correction $\delta\rho$ , which is what we solve for.

Step 2. Keep terms linear in $F$ : subtract the zeroth-order equation to get $i\hbar\,\partial_t\delta\rho=[H_0,\delta\rho]+[H'(t),\rho_0]$ , with inhomogeneous source $[H'(t),\rho_0]=-F(t)[A,\rho_0]$ . Why: the product $[H',\delta\rho]$ is second order in the field and is dropped — this linearisation is the "linear" in linear response.

Step 3. Solve the linear ODE with adiabatic switch-on: $\delta\rho(t)=-\tfrac{1}{i\hbar}\int_{-\infty}^{t}e^{-iH_0(t-t')/\hbar}[A,\rho_0]e^{+iH_0(t-t')/\hbar}\,F(t')\,\mathrm{d}t'$ . Why: the homogeneous evolution $e^{-iH_0\tau/\hbar}(\cdots)e^{+iH_0\tau/\hbar}$ propagates the source forward, and the lower limit $-\infty$ implements the field being off in the distant past.

Step 4. Compute the observable, $\delta\langle B(t)\rangle=\mathrm{Tr}[B\,\delta\rho(t)]$ , and use cyclicity $\mathrm{Tr}(B[A,\rho_0])=\mathrm{Tr}(\rho_0[B,A])$ to shift the perturbation onto the operators. Why: this is the crucial move — the trace is now against the equilibrium $\rho_0$ , so the answer is an equilibrium expectation of a commutator, no driven state required.

Step 5. Identify the kernel: comparing with $\delta\langle B(t)\rangle=\int^t\chi_{BA}(t-t')F(t')\mathrm{d}t'$ gives $\chi_{BA}(\tau)=\tfrac{i}{\hbar}\langle[B(\tau),A(0)]\rangle_0\,\Theta(\tau)$ . In the classical limit $\tfrac{i}{\hbar}[\,,\,]\to\{\,,\,\}$ and an integration by parts against $\rho_0\propto e^{-\beta H_0}$ turns the Poisson bracket into $-\beta\langle\dot A(0)B(\tau)\rangle_0$ . Why: the equilibrium weight $e^{-\beta H_0}$ supplies the factor $\beta$ and the time derivative, delivering the classical Kubo form (4). *Verified: verify_aq-sm-linear-response-theory.py confirms the classical kernel $\chi=-\beta\Theta\,\mathrm{d}C/\mathrm{d}\tau$ is the derivative of the correlation, that $\chi(0)=\beta\,\mathrm{Var}(A)$ follows by integrating it, and the interaction-picture sign conventions.*

Go deeper · Derivation B · 1 minCausality forces analyticity and the Kramers-Kronig dispersion relation

Derivation BCausality forces analyticity and the Kramers-Kronig dispersion relation

The goal is to show that the single fact $\chi(\tau)=0$ for $\tau<0$ makes the susceptibility $\chi(\omega)$ analytic in the upper half of the complex-frequency plane, and that Cauchy's theorem then couples its real and imaginary parts into a principal-value dispersion integral. We treat $\omega$ as complex, use the one-sided time integral to prove analyticity, close a contour in the upper half-plane, and apply the Sokhotski–Plemelj identity to extract the physical (real-axis) relation. The upshot is that dispersion and absorption are not independent — measuring one determines the other.

Step 1. Note that $\chi(\omega)=\int_0^\infty\chi(\tau)e^{i\omega\tau}\mathrm{d}\tau$ converges and is analytic whenever $\mathrm{Im}\,\omega>0$ , because $e^{i\omega\tau}=e^{i\mathrm{Re}(\omega)\tau}e^{-\mathrm{Im}(\omega)\tau}$ is exponentially damped for $\tau>0$ . Why: the one-sidedness in time (causality) is precisely what kills the upper-half-plane divergence, so $\chi(\omega)$ has no poles there.

Step 2. Apply Cauchy's theorem to $\chi(\omega')/(\omega'-\omega)$ on a contour running along the real axis and closing in the upper half-plane; with $\chi\to0$ at large $|\omega'|$ the arc drops out, leaving $\chi(\omega)=\tfrac{1}{2\pi i}\oint\tfrac{\chi(\omega')}{\omega'-\omega}\mathrm{d}\omega'$ . Why: analyticity inside the contour means the only contribution is the pole pushed just below the real axis, which is what carries the physics.

Step 3. Resolve the pole with Sokhotski–Plemelj, $\tfrac{1}{\omega'-\omega-i0}=\mathcal{P}\tfrac{1}{\omega'-\omega}+i\pi\delta(\omega'-\omega)$ , and separate real and imaginary parts. Why: the principal value carries the dispersive coupling while the delta reproduces $\chi$ itself, and matching real and imaginary parts yields the two coupled Kramers–Kronig relations, e.g. $\chi'(\omega)=\tfrac{1}{\pi}\mathcal{P}\int\tfrac{\chi''(\omega')}{\omega'-\omega}\mathrm{d}\omega'$ .

Step 4. Confirm consistency on a solvable model: a single-relaxation (Debye) response $\chi(\omega)=\chi_0/(1-i\omega\tau_R)$ has $\chi'=\chi_0/(1+\omega^2\tau_R^2)$ and $\chi''=\chi_0\,\omega\tau_R/(1+\omega^2\tau_R^2)$ , which satisfy the dispersion integral exactly. Why: this closed form is the workhorse example — its $\chi''$ peaks at $\omega\tau_R=1$ and its $\chi'$ is the KK transform of that peak, checked numerically. *Verified: verify_aq-sm-linear-response-theory.py confirms upper-half-plane analyticity of the Debye form, that $\chi''$ peaks at $\omega\tau_R=1$ , and that the KK principal-value integral reproduces $\chi'$ to better than 1%.*

Limiting cases

Linear response is a leading-order theory; its regime of validity and its special cases are sharply defined and each recovers a familiar law.

Case	Result	Meaning
Weak-field (linear) regime	$\lvert A\,F\rvert\ll k_BT$	The perturbation energy must be small versus thermal energy so first order dominates; else nonlinear susceptibilities $\chi^{(2)},\chi^{(3)},\dots$ enter.
Static (DC) limit	$\chi(0)=\beta\big(\langle AB\rangle-\langle A\rangle\langle B\rangle\big)$	The zero-frequency susceptibility is the equilibrium covariance times $\beta$ — a thermodynamic fluctuation formula (compressibility, magnetic susceptibility).
Classical (high- $T$ ) limit	$\chi_{BA}(\tau)=-\beta\langle\dot A(0)B(\tau)\rangle_0\Theta(\tau)$	The commutator $\tfrac{i}{\hbar}\langle[B,A]\rangle$ becomes a Poisson-bracket/time-derivative correlation when $\hbar\omega\ll k_BT$ .
Markovian / single-relaxation	$\chi(\omega)=\chi_0/(1-i\omega\tau_R)$	One relaxation time gives Debye/Lorentzian response; $\chi''$ peaks at $\omega\tau_R=1$ . The basis of the Einstein relation and the Ornstein–Uhlenbeck process.

Worked examples

Example 1 — The Debye response of a relaxing dipole, and where it dissipates. This computes a real susceptibility curve and locates the frequency of maximum energy absorption — the practical heart of dielectric spectroscopy.

Step 1. Take a single-relaxation-time response $\chi(\tau)=\dfrac{\chi_0}{\tau_R}e^{-\tau/\tau_R}\Theta(\tau)$ with $\chi_0=1$ and $\tau_R=1\times10^{-9}\,\mathrm{s}$ (a nanosecond dielectric relaxation).

Step 2. Fourier transform: $\chi(\omega)=\int_0^\infty\tfrac{\chi_0}{\tau_R}e^{-\tau/\tau_R}e^{i\omega\tau}\mathrm{d}\tau=\dfrac{\chi_0}{1-i\omega\tau_R}$ , so $\chi'(\omega)=\dfrac{\chi_0}{1+\omega^2\tau_R^2}$ and $\chi''(\omega)=\dfrac{\chi_0\,\omega\tau_R}{1+\omega^2\tau_R^2}$ .

Step 3. Find the loss peak by setting $\mathrm{d}\chi''/\mathrm{d}\omega=0$ : the maximum sits at $\omega\tau_R=1$ , i.e. $\omega_{\max}=1/\tau_R=1\times10^{9}\,\mathrm{rad/s}$ ( $f_{\max}=\omega_{\max}/2\pi\approx1.59\times10^{8}\,\mathrm{Hz}\approx159\,\mathrm{MHz}$ ).

Step 4. At the peak $\chi''(\omega_{\max})=\chi_0/2=0.5$ while $\chi'(\omega_{\max})=\chi_0/2=0.5$ : half the static response has rotated from reactive into dissipative. The material absorbs most strongly at $f\approx159\,\mathrm{MHz}$ , where $\chi''=0.5\chi_0$ — the fingerprint used to read off relaxation times from a loss spectrum.

Example 2 — Static magnetic susceptibility as a spin-fluctuation variance. This shows the DC limit is literally a thermodynamic fluctuation, not merely proportional to one.

Step 1. Take $N$ non-interacting spin- $\tfrac12$ moments $\mu$ in a field, so the total magnetization is $M=\mu\sum_i s_i$ with $s_i=\pm1$ ; the coupling observable is $A=M$ .

Step 2. The equilibrium variance at zero field is $\mathrm{Var}(M)=\mu^2\sum_{i,j}\langle s_i s_j\rangle=\mu^2 N$ (independent spins, $\langle s_i s_j\rangle=\delta_{ij}$ , $\langle s_i^2\rangle=1$ ).

Step 3. Apply the static Kubo identity $\chi(0)=\beta\,\mathrm{Var}(M)=\beta\mu^2 N=N\mu^2/k_BT$ .

Step 4. Recognize the Curie law: the static susceptibility $\chi(0)=N\mu^2/k_BT\propto1/T$ emerges purely as an equilibrium fluctuation variance — the $1/T$ Curie divergence is the growth of magnetization fluctuations as thermal disordering weakens.

Origins · 1 minlinear response theory and how this was discovered

linear response theory

framework crystallized in the 1950s · Linear response theory is the framework that turns "how does an equilibrium system react to a weak push?" into a calculable equilibrium correlation function. Its modern form was assembled in the early 1950s from several threads: Callen and Welton's 1951 general fluctuation–dissipation theorem, Green's 1952–54 correlation-function transport expressions, and Ryogo Kubo's 1957 unification, which recast every linear transport coefficient and susceptibility as a retarded equilibrium commutator average — the Kubo formula. The framework rests on three pillars that predate it: Einstein's 1905 diffusion–mobility link, Nyquist's 1928 thermal-noise argument, and Onsager's 1931 regression hypothesis and reciprocal relations. Its power is that one equilibrium computation predicts every weak-field experiment, from optical conductivity to neutron scattering.

History

1905

— Einstein links diffusion to mobility, the first fluctuation–dissipation statement

1928

— Nyquist derives thermal (Johnson) noise from equilibrium arguments

1931

— Onsager states the regression hypothesis and reciprocal relations

1951

— Callen and Welton prove the general fluctuation–dissipation theorem

1954

— Green expresses transport coefficients as time-correlation integrals

1957

— Kubo unifies linear response as retarded equilibrium commutators (the Kubo formula)

1966

— Kubo's Reports on Progress in Physics review consolidates the framework

Going further

References. Kubo, Toda & Hashitsume, Statistical Physics II: Nonequilibrium Statistical Mechanics · Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 15 · Kardar, Statistical Physics of Particles, ch. on response · Reif, Fundamentals of Statistical and Thermal Physics · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity · Landau & Lifshitz, Statistical Physics Part 1, §123–126 · Chandler, Introduction to Modern Statistical Mechanics, Ch. 8

Provenance. All formulas are standard. A verifier recomputes: (1) the interaction-picture derivation of the Kubo kernel $\chi_{BA}=\tfrac{i}{\hbar}\Theta\langle[B(\tau),A]\rangle_0$ and its classical form $-\beta\Theta\,\dot C$ (Kubo–Toda–Hashitsume Ch. 4; Pathria & Beale §15); (2) upper-half-plane analyticity and the Kramers–Kronig relation via Cauchy + Sokhotski–Plemelj (Landau–Lifshitz §123); (3) the Debye response $\chi(\omega)=\chi_0/(1-i\omega\tau_R)$ with $\chi''$ peaking at $\omega\tau_R=1$ ; (4) the static identity $\chi(0)=\beta\,\mathrm{Var}(A)$ and the Curie law $N\mu^2/k_BT$ . Numbers quoted are textbook standard; no nonstandard values are asserted.

Builds on: Hamiltonian dynamics and the density matrix · Equilibrium statistical ensembles · Fourier analysis and complex analyticity · The Langevin equation. Leads to: The Fluctuation-Dissipation Theorem · The Kubo Formula · Kramers-Kronig Relations · The Einstein Relation · The Fokker-Planck Equation.

Why is the response a commutator rather than a simple product of operators?

It comes from where the perturbation enters. In first-order perturbation theory the field appears through $[H',\rho_0]$ in the Liouville equation, and $H'=-AF$ carries $A$ . When you compute $\delta\langle B\rangle=\mathrm{Tr}(B\,\delta\rho)$ and use the cyclic property of the trace to move the perturbation off $\rho_0$ and onto the observables, the two operators arrive in a commutator $[B(\tau),A(0)]$ . Physically the commutator measures the extent to which $B$ at a later time fails to commute with $A$ at the moment of the kick — that non-commutation is the propagated response. Classically the commutator becomes a Poisson bracket and, weighted by $e^{-\beta H_0}$ , turns into $-\beta\langle\dot A B\rangle$ .

Where did the driven nonequilibrium state go — don't we need it to compute a response?

That is exactly the miracle of the theory: you don't. The response is first order in the field, so the correction to the density matrix, $\delta\rho$ , is evaluated by tracing against the unperturbed equilibrium $\rho_0$ . The perturbed, genuinely driven ensemble would only be needed at second order and beyond (for nonlinear susceptibilities). To linear order the system "forgets it is being driven": the entire answer is an equilibrium correlation function, computed once in the ensemble you already know. This is why a single equilibrium simulation or measurement predicts every weak-field experiment.

What exactly is the difference between the reactive part χ′ and the dissipative part χ″?

Split the periodic response into the piece in phase with the drive and the piece a quarter-cycle out of phase. The in-phase piece is $\chi'(\omega)$ : it stores energy and gives it back each cycle (dispersion, refraction, reactive susceptibility), and it is even in $\omega$ . The out-of-phase piece is $\chi''(\omega)$ : it absorbs net energy at rate $\propto\omega\chi''|F|^2$ (dissipation, absorption, loss), and it is odd in $\omega$ . They are not independent — Kramers–Kronig ties them through causality. The fluctuation–dissipation theorem singles out $\chi''$ because only the dissipative channel is fed by, and feeds, the equilibrium fluctuations.

How does Onsager's regression hypothesis connect to linear response?

Onsager's hypothesis says a macroscopic deviation from equilibrium decays, on average, by the same law whether it was created by an external field or arose as a spontaneous thermal fluctuation. In the language here, "created by a field then switched off" is the response (its decay is the relaxation function), and "arose spontaneously" is the correlation function of equilibrium fluctuations. The classical Kubo formula $\chi(\tau)=-\beta\,\mathrm{d}C/\mathrm{d}\tau\,\Theta(\tau)$ is the precise, quantitative statement of the hypothesis: the response is literally the time-derivative of the equilibrium correlation. This is why transport coefficients are integrals over autocorrelations of spontaneous currents.

When does linear response theory fail, and how badly?

It fails whenever the reference state is not equilibrium — driven steady states, sheared fluids, aging glasses, active matter — or when the field is strong enough that $|AF|\gtrsim k_BT$ and nonlinear susceptibilities matter. In the nonequilibrium case the failure is structural, not a small correction: the clean identity between response and correlation (the FDT) no longer holds, and one must add an effective temperature or an explicit entropy-production term. The size of the discrepancy is itself useful — it is the standard experimental measure of "distance from equilibrium." For strong fields the fix is systematic: linear response is just the first term of an expansion, and $\chi^{(2)},\chi^{(3)},\dots$ capture harmonic generation and rectification.

Does the same theory really give both optical constants and diffusion coefficients?

Yes — that universality is the whole selling point. Choose the coupling operator $A$ and the observed operator $B$ to match the experiment. For light, $A$ is the polarization and $B$ the current, and $\chi(\omega)$ is the optical conductivity whose real and imaginary parts are absorption and refraction. For a tagged Brownian particle, $A$ is position and $B$ the velocity, and the zero-frequency limit is the mobility while the velocity autocorrelation integral is the diffusion constant — the two sides of the Einstein relation. Same kernel $\chi_{BA}=\tfrac{i}{\hbar}\Theta\langle[B,A]\rangle$ , different labels. The Kubo formula develops the transport-coefficient face in detail.