Fluctuation Theorems

A central question of non-equilibrium statistical mechanics is whether anything exact survives once a system is pushed far from equilibrium, where the linear-response machinery — the fluctuation-dissipation theorem, the Green–Kubo formulas — no longer applies. Fluctuation theorems answer yes. They are exact identities, not approximations, that constrain the entire probability distribution of a fluctuating thermodynamic quantity such as the entropy production or the work done by an external protocol. Discovered in the 1990s by Evans, Cohen, Morriss, Gallavotti, Jarzynski, and Crooks, they reorganized the whole subject by revealing that the second law of thermodynamics is merely the leading inequality buried inside a deeper equality. The architecture is always the same. One defines a trajectory-level fluctuating quantity — the dimensionless entropy production $\Sigma$ accumulated along a stochastic path, or the work $W$ . Microscopic reversibility — the time-reversal symmetry of the underlying Hamiltonian, or local detailed balance of the stochastic dynamics — relates the probability of a forward trajectory to that of its time-reversed twin. Summing over all trajectories with a fixed $\Sigma$ yields a symmetry of the distribution $P(\Sigma)$ , the detailed fluctuation theorem $P(\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}$ . Integrating over all $\Sigma$ collapses it to the integral fluctuation theorem $\langle e^{-\Sigma/k_B}\rangle=1$ , whose convexity immediately delivers $\langle\Sigma\rangle\ge0$ — the second law, now a theorem. The results split into transient theorems (system prepared in equilibrium, then driven — Evans–Searles, Jarzynski, Crooks) and steady-state theorems (system held in a non-equilibrium steady state — Gallavotti–Cohen). They hold for deterministic thermostatted dynamics, for Langevin and Fokker–Planck processes, for discrete master equations, and, with care, for quantum systems. The price of this generality is that the relevant averages are dominated by rare, exponentially unlikely trajectories — the same trajectories that momentarily "violate" the second law. Beyond their elegance, fluctuation theorems are a working experimental tool: dragging colloids through fluid, stretching RNA hairpins with optical tweezers, and counting electrons through quantum dots all furnish measured distributions whose symmetry has been confirmed and used to recover free energies. They sit at the meeting point of stochastic thermodynamics, single-molecule biophysics, and the foundations of irreversibility.

Key takeaways

Fluctuation theorems are exact symmetry relations for the probability distribution of entropy production, work, or heat exchanged by a driven system, valid arbitrarily far from equilibrium where linear response fails.
The detailed fluctuation theorem $P(+\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}$ says entropy-producing trajectories are exponentially more likely than entropy-consuming ones — a quantitative arrow of time.
The integral fluctuation theorem $\langle e^{-\Sigma/k_B}\rangle=1$ follows by integrating it, and by Jensen's inequality forces $\langle\Sigma\rangle\ge0$ — the second law becomes a theorem, not an axiom.
A single trajectory can have $\Sigma<0$ (a transient "second-law violation"); its probability is exponentially small, $\sim e^{-|\Sigma|/k_B}$ , and observable only for small systems over short times.
Transient (Evans–Searles) versus steady-state (Gallavotti–Cohen) theorems differ by their initial ensemble; the work relations Jarzynski and Crooks are the corollaries owned by the jarzynski-crooks-relations article.
Key idea: time-reversal symmetry of the microscopic dynamics, plus a definite initial ensemble, forces the probability ratio of a path to its reverse to equal $e^{\Sigma/k_B}$ — turning the second-law inequality into the exact equality $\langle e^{-\Sigma/k_B}\rangle=1$ that holds arbitrarily far from equilibrium.

The core insight

Entropy production is the log-likelihood ratio between a trajectory and its time-reverse. Because that ratio has a universal exponential-tilt symmetry $P(\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}$ , its exponential average is exactly one — and a Kullback–Leibler divergence being non-negative is the second law.

Where this lives today

Fluctuation theorems have become a standard experimental instrument. Optical tweezers unfold RNA and DNA hairpins at finite speed, and the crossing point of the forward and reverse work histograms reads off the folding free energy $\Delta F$ even when every individual pull is strongly irreversible — the technique that Collin, Ritort, Jarzynski, Smith, Tinoco and Bustamante validated in 2005. Counting single electrons tunneling through a quantum dot yields the full distribution of transferred charge and confirms the steady-state (Gallavotti–Cohen) theorem for mesoscopic transport. In molecular simulation, Jarzynski's equality lets one compute free energies from many fast steered-molecular-dynamics pulls instead of slow reversible integration. And the theorems have spawned a whole family of descendants — most importantly the thermodynamic uncertainty relations, which bound a molecular machine's precision by its dissipation, and which are reshaping how we think about the efficiency limits of biological and artificial nanomachines.

The physical picture

Film a colloidal bead being dragged through water, then play the film backward. The second law says the reverse usually looks wrong — heat appears to flow uphill, the drag appears to push the bead forward. The fluctuation theorem makes this quantitative: the forward film is more probable than its reverse by exactly $e^{\Sigma/k_B}$ , where $\Sigma$ is the entropy produced. For macroscopic $\Sigma$ the reverse is astronomically unlikely, and we recover the ordinary, absolute arrow of time. But for a micron-sized bead over a few milliseconds, $\Sigma\sim k_B$ , and transient time-reversed events — trajectories that momentarily consume entropy — are actually observable in the lab.

:::svg[Entropy-production distribution P(Σ): the small negative-Σ tail is the outlier; ln[P(Σ)∕P(−Σ)] is a straight line of slope 1∕k_B]

::: The Jarzynski equality is a "free lunch" that isn't. The trajectories that dominate $\langle e^{-W/k_BT}\rangle$ are precisely the rare ones where the system extracted work or absorbed little — the single-shot second-law-violating events. The equality holds because these rare events are up-weighted exponentially, exactly cancelling the typical dissipation. Think of entropy production as a log-likelihood ratio between forward and time-reversed dynamics: this is why $\langle\Sigma\rangle\ge0$ on average, because a relative entropy (Kullback–Leibler divergence) between two path measures cannot be negative. The fluctuation theorem is then nothing but the statement that this log-likelihood ratio has a distribution obeying a universal exponential-tilt symmetry.

Figure

Near-equilibrium consistency: a Gaussian Σ obeying the FT must have ⟨Σ⟩ = Var(Σ)∕2k_B — the fluctuation-dissipation relation

The core equations

The foundational statement is the detailed fluctuation theorem: it is an exact symmetry of the probability distribution of the total (system + reservoir) entropy production over a process, and it holds arbitrarily far from equilibrium.

\frac{P(\Sigma)}{P(-\Sigma)} = e^{\Sigma/k_B} \qquad (1)

where: $P(\Sigma)$ — probability that the total entropy produced over the process equals $\Sigma$ ; $\Sigma$ — total entropy production (system + medium), carrying dimensions of $k_B$ ; $k_B$ — Boltzmann constant. Physically: a trajectory producing entropy $\Sigma$ is exponentially more likely than the one producing $-\Sigma$ . Integrating the detailed theorem against $e^{-\Sigma/k_B}$ and using normalization collapses it to a single number — the integral fluctuation theorem, the most compact exact statement of the second law.

\big\langle e^{-\Sigma/k_B}\big\rangle = 1 \qquad (2)

where: $\langle\cdot\rangle$ — average over the forward-process ensemble. Physically: the exponential average of the (negative) entropy production is pinned to unity, so no matter how the system is driven, its dissipation obeys this exact constraint. Because the exponential is convex, Jensen's inequality applied to Eq. (2) immediately yields the second law as an inequality — not assumed, but derived.

\langle\Sigma\rangle \ge 0, \qquad \langle\Sigma\rangle = k_B\,D_{\mathrm{KL}}\!\big(\mathcal{P}_F\,\|\,\mathcal{P}_R\big) \qquad (3)

where: $D_{\mathrm{KL}}$ — Kullback–Leibler divergence between the forward and time-reversed path measures. Physically: mean entropy production is the statistical distinguishability of a process from its time-reverse; it vanishes only at reversibility. For a driven system prepared in equilibrium, the fluctuating quantity is the work $W$ , and the same reversibility argument gives the two celebrated work relations — stated here, derived and owned in the jarzynski-crooks-relations article as corollaries of this theorem.

\big\langle e^{-W/k_BT}\big\rangle = e^{-\Delta F/k_BT}, \qquad \frac{P_F(W)}{P_R(-W)} = e^{(W-\Delta F)/k_BT} \qquad (4)

where: $W$ — work done by the protocol $\lambda(t)$ ; $\Delta F=F_B-F_A$ — equilibrium free-energy difference between endpoints; $P_F,P_R$ — forward and reverse work distributions. Physically: the full work statistics encode the equilibrium free energy exactly, however violent the driving. Finally, the trajectory-level entropy balance that makes $\Sigma$ concrete: it splits into a system (Shannon) part and a medium (heat) part, each defined per stochastic trajectory.

\Sigma = \Delta S_{\text{sys}} + \frac{Q}{T}, \qquad \Delta S_{\text{sys}} = -k_B\ln\frac{p(x_\tau,\tau)}{p(x_0,0)} \qquad (5)

where: $Q$ — heat delivered to the medium along the path; $\dot S_{\text{med}}=Q/T$ — the bath's entropy change; $\Delta S_{\text{sys}}$ — trajectory change of stochastic (Shannon) system entropy. Physically: the total that obeys the fluctuation theorem is system plus medium, not either alone.

Derivation AThe detailed and integral fluctuation theorems from path time-reversal

The goal is to derive the two master identities — the detailed theorem $P(\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}$ and the integral theorem $\langle e^{-\Sigma/k_B}\rangle=1$ — from a single structural input: entropy production defined as the log-ratio of a forward path weight to the weight of its time-reverse. Everything follows from that definition plus normalization; no near-equilibrium assumption enters, which is exactly why the results hold arbitrarily far from equilibrium. We define $\Sigma$ , bin trajectories by it, substitute the definition, and integrate.

Step 1. Define the entropy production of a forward path $\gamma$ as $\Sigma[\gamma]=k_B\ln\big(\mathcal{P}_F[\gamma]/\mathcal{P}_R[\tilde\gamma]\big)$ , where $\tilde\gamma$ is the time-reverse and $\mathcal{P}_F,\mathcal{P}_R$ are the forward and reverse path measures. Why: this identifies entropy production with the statistical irreversibility of the path — the microscopic-reversibility content of the dynamics.

Step 2. Write the distribution as a sum over paths, $P(\Sigma)=\int\!\mathcal{D}\gamma\,\mathcal{P}_F[\gamma]\,\delta(\Sigma-\Sigma[\gamma])$ , then substitute $\mathcal{P}_F=e^{\Sigma/k_B}\mathcal{P}_R$ from the definition. Why: replacing the forward measure by the reverse measure times $e^{\Sigma/k_B}$ is the one algebraic move the whole theorem rests on.

Step 3. Change the integration variable $\gamma\to\tilde\gamma$ ; since $\Sigma[\tilde\gamma]=-\Sigma[\gamma]$ , the delta-function flips to $\delta(\Sigma+\Sigma[\tilde\gamma])$ , giving $P(\Sigma)=e^{\Sigma/k_B}P(-\Sigma)$ — the detailed FT, Eq. (1). Why: the reverse path produces the negative entropy, so binning the reverse ensemble at $-\Sigma$ reconstructs $P(-\Sigma)$ .

Step 4. Integrate the detailed FT against $e^{-\Sigma/k_B}$ : $\langle e^{-\Sigma/k_B}\rangle=\int\!P(\Sigma)e^{-\Sigma/k_B}\mathrm{d}\Sigma=\int\!P(-\Sigma)\mathrm{d}\Sigma=1$ , using $P(\Sigma)e^{-\Sigma/k_B}=P(-\Sigma)$ then normalization. Why: the exponential exactly cancels the tilt, leaving the reverse distribution, which integrates to one — the integral FT, Eq. (2).

Step 5. Apply Jensen's inequality to the convex function $e^{-x}$ : $1=\langle e^{-\Sigma/k_B}\rangle\ge e^{-\langle\Sigma\rangle/k_B}$ , so $\langle\Sigma\rangle\ge0$ — the second law. Why: convexity turns the exact equality into the familiar inequality, deriving (not postulating) the second law. Verified in verify_aq-sm-fluctuation-theorems.py: for a Gaussian $\Sigma$ with $\langle\Sigma\rangle=\sigma^2/2k_B$ the DFT holds identically, $\langle e^{-\Sigma/k_B}\rangle=1$ numerically, and $\langle\Sigma\rangle\ge0$ .

Go deeper · Derivation B · 1 minEntropy production as relative entropy: a clean proof of the second law

Derivation BEntropy production as relative entropy: a clean proof of the second law

Beyond the inequality, the mean entropy production has an exact information-theoretic identity: it equals $k_B$ times the Kullback–Leibler divergence between the forward and reverse path measures. This upgrades the second law from "a number is non-negative" to "the distinguishability of a process from its time-reverse is non-negative," which is manifestly true and vanishes only at reversibility. We average the log-ratio definition and recognize the relative entropy.

Step 1. Average the definition from the previous card: $\langle\Sigma\rangle=k_B\int\!\mathcal{D}\gamma\,\mathcal{P}_F[\gamma]\ln\big(\mathcal{P}_F[\gamma]/\mathcal{P}_R[\tilde\gamma]\big)$ . Why: taking the forward-ensemble expectation of the per-path log-ratio is precisely the definition of a KL divergence.

Step 2. Recognize the integral as $k_B\,D_{\mathrm{KL}}(\mathcal{P}_F\|\mathcal{P}_R)$ , the relative entropy of the forward path measure with respect to the reverse. Why: $\int p\ln(p/q)$ with $p=\mathcal{P}_F$ , $q=\mathcal{P}_R$ is the textbook KL divergence.

Step 3. Use $D_{\mathrm{KL}}\ge0$ (Gibbs' inequality), with equality iff $\mathcal{P}_F=\mathcal{P}_R$ , to conclude $\langle\Sigma\rangle\ge0$ , vanishing only when forward and reverse ensembles coincide — i.e. at reversibility. Why: a relative entropy is non-negative by the log-sum (Jensen) inequality, giving the second law its cleanest proof. Verified in verify_aq-sm-fluctuation-theorems.py: $D_{\mathrm{KL}}\ge0$ on a two-outcome test measure, vanishing exactly when the two measures are equal.

Limiting cases

Fluctuation theorems contain the familiar results as limits:

Case	Result	Meaning
Second law	$\langle e^{-\Sigma/k_B}\rangle=1\Rightarrow\langle\Sigma\rangle\ge0$	Jensen's inequality on the convex exponential yields non-negative mean entropy production — the second law as a theorem.
Near equilibrium (Gaussian)	$\langle\Sigma\rangle=\tfrac{1}{2k_B}\,\mathrm{Var}(\Sigma)$	A Gaussian $\Sigma$ consistent with the FT must have variance $=2k_B$ times its mean — the fluctuation-dissipation relation recovered.
Quasi-static driving	$P_F(W)\to\delta(W-\Delta F)$	Reversible limit: work has no spread and equals $\Delta F$ ; Jarzynski and Crooks become trivial identities.
Long time / steady state	$\lim_{t\to\infty}\tfrac{1}{t}\ln\tfrac{P(\sigma t)}{P(-\sigma t)}=\sigma/k_B$	Gallavotti–Cohen steady-state form for the entropy-production rate $\sigma$ , governing its large-deviation rate function.

Worked examples

Example 1 — how often does a colloid "violate" the second law? A micron bead is dragged through water so that the mean entropy production over the measurement window is $\langle\Sigma\rangle=2\,k_B$ , with a Gaussian distribution consistent with the DFT. What fraction of trajectories have $\Sigma<0$ (transient second-law "violations")?

Step 1. The DFT fixes the Gaussian: consistency requires $\mathrm{Var}(\Sigma)=2k_B\langle\Sigma\rangle$ , so in units of $k_B$ , $\mu=2$ and $\sigma^2=2\cdot2=4$ , i.e. $\sigma=2$ .

Step 2. The fraction with $\Sigma<0$ is $\Phi\!\big((0-\mu)/\sigma\big)=\Phi(-1)=\tfrac12\,\mathrm{erfc}(1/\sqrt2)=\mathbf{0.159}$ , about 16% of trajectories.

Step 3. This is a genuinely observable rate, not a paradox: it is exactly what Wang, Sevick, Mittag, Searles and Evans measured in 2002 for a colloidal bead in an optical trap. As the window lengthens $\langle\Sigma\rangle$ grows, $\sigma/\mu=\sqrt{2/\mu}\to0$ , and the "violations" vanish — recovering the macroscopic arrow of time.

Example 2 — the exponential average is pinned to one. For the same Gaussian ( $\mu=2$ , $\sigma^2=4$ in units of $k_B$ ), verify the integral fluctuation theorem directly.

Step 1. For a Gaussian with mean $\mu$ and variance $\sigma^2$ , $\langle e^{-\Sigma/k_B}\rangle=e^{-\mu+\sigma^2/2}$ (the generating function).

Step 2. Substitute $\mu=2$ , $\sigma^2=4$ : $\;e^{-2+4/2}=e^{-2+2}=e^0=\mathbf{1}$ , exactly as the integral FT demands.

Step 3. The relation $\sigma^2=2\mu$ is not a coincidence — it is forced by the DFT. Any Gaussian entropy-production distribution that obeys the detailed theorem automatically satisfies $\langle e^{-\Sigma/k_B}\rangle=1$ ; this is the near-equilibrium fluctuation-dissipation relation hiding inside the far-from-equilibrium identity.

Origins · 1 minfluctuation theorems and how this was discovered

fluctuation theorems

exact non-equilibrium symmetry relations, 1993–1999 · The fluctuation theorems are a family of exact identities constraining the distribution of entropy production or work in driven systems. The detailed version

P(\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}

was found in deterministic thermostatted simulations by Evans, Cohen and Morriss (1993) and proved for steady states by Gallavotti and Cohen (1995); Jarzynski (1997) and Crooks (1999) supplied the work relations. Together they show the second law is the leading inequality of a deeper equality, valid arbitrarily far from equilibrium — the organizing principle of modern stochastic thermodynamics.

History

1993

— Evans, Cohen and Morriss observe the transient fluctuation theorem in thermostatted molecular-dynamics simulations (Phys. Rev. Lett. 71, 2401)

1994

— Evans and Searles formulate the transient fluctuation theorem for systems starting in equilibrium

1995

— Gallavotti and Cohen prove the steady-state fluctuation theorem for chaotic dynamics (Phys. Rev. Lett. 74, 2694)

1997

— Jarzynski derives the nonequilibrium work equality (Phys. Rev. Lett. 78, 2690)

1999

— Crooks establishes the detailed work fluctuation theorem (Phys. Rev. E 60, 2721)

2002

— Wang, Sevick, Mittag, Searles and Evans measure transient second-law "violations" for a dragged colloid

2005

— Seifert defines trajectory entropy production and its integral fluctuation theorem (Phys. Rev. Lett. 95, 040602)

Going further

Texts. Seifert, "Stochastic thermodynamics, fluctuation theorems and molecular machines," Rep. Prog. Phys. 75, 126001 (2012) · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity · Kardar, Statistical Physics of Particles · Evans & Searles, "The fluctuation theorem," Adv. Phys. 51, 1529 (2002).

Provenance. Equations and derivations follow Evans–Cohen–Morriss PRL 71, 2401 (1993); Gallavotti–Cohen PRL 74, 2694 (1995); Jarzynski PRL 78, 2690 (1997); Crooks PRE 60, 2721 (1999); Seifert Rep. Prog. Phys. 75, 126001 (2012). Verification reproduces (1) the detailed FT $P(\Sigma)=e^{\Sigma/k_B}P(-\Sigma)$ for a Gaussian with $\langle\Sigma\rangle=\sigma^2/2k_B$ ; (2) the integral FT $\langle e^{-\Sigma/k_B}\rangle=1$ ; (3) the Jensen route to $\langle\Sigma\rangle\ge0$ ; (4) the KL-divergence identity and its non-negativity; (5) the worked-example fraction $\Phi(-1)=0.159$ . No numerical constants beyond $k_B$ and standard definitions are asserted.

Related. Prerequisites: the langevin-equation · fokker-planck-equation · master-equation · ito-stratonovich-calculus · fluctuation-dissipation-theorem. Leads to: stochastic-thermodynamics · single-molecule free-energy estimation · thermodynamic-uncertainty-relations. The work-relation corollaries — Jarzynski and Crooks — are derived and owned in jarzynski-crooks-relations; the trajectory heat/work bookkeeping is owned in stochastic-thermodynamics.

How can the second law be "violated" by a single trajectory without contradiction?

The second law is a statement about averages, and the fluctuation theorem makes the average non-negative exactly. A single short trajectory of a small system can have $\Sigma<0$ — it momentarily consumes entropy — but the detailed theorem fixes its probability at $e^{-|\Sigma|/k_B}$ times that of the entropy-producing twin. For a macroscopic $\Sigma$ this factor is astronomically small, so "violations" are never seen; for $\Sigma\sim k_B$ they occur a few percent of the time and have been measured. The ensemble average $\langle\Sigma\rangle$ remains $\ge0$ throughout, so the law itself is never broken.

What is the difference between the detailed and integral fluctuation theorems?

The detailed theorem $P(\Sigma)/P(-\Sigma)=e^{\Sigma/k_B}$ is a statement about the full shape of the distribution — it relates the probability of every value of $\Sigma$ to that of $-\Sigma$ . The integral theorem $\langle e^{-\Sigma/k_B}\rangle=1$ is a single scalar consequence obtained by integrating the detailed one against $e^{-\Sigma/k_B}$ . The detailed theorem is strictly stronger: it implies the integral theorem, but not vice versa. Experimentally the detailed version is verified by measuring the whole histogram and checking the straight-line log-ratio; the integral version is a one-number check.

Why are transient and steady-state theorems distinguished?

They differ in the required initial ensemble. Transient theorems (Evans–Searles, Jarzynski, Crooks) require the system to start in a canonical equilibrium distribution at the initial control value, then be driven; the symmetry holds for the finite-time entropy production or work. Steady-state theorems (Gallavotti–Cohen) require the system to already sit in a non-equilibrium steady state maintained by sustained driving; the symmetry then holds only asymptotically, for the long-time entropy-production rate. Confusing the two — e.g. applying Jarzynski from a non-equilibrium start — breaks the identity.

Why is estimating a free energy from Jarzynski's equality so hard in practice?

Because $\langle e^{-\beta W}\rangle$ is dominated by the rare trajectories with $W$ small or negative — exactly the exponential tail of the forward work distribution. With a finite sample you almost never draw those dominant events, so the estimate of $\Delta F$ is systematically biased (too high) and converges slowly. The cure is to use both directions: the Crooks relation lets forward and reverse data be combined optimally through the Bennett acceptance ratio, and the histogram-crossing point gives $\Delta F$ without trusting either tail. These estimators are owned by the jarzynski-crooks-relations article.

Do fluctuation theorems apply to quantum systems?

Yes, with care. The quantum work is defined through a two-point measurement scheme — measure the energy at the start and end of the protocol — and the Tasaki–Kurchan quantum Jarzynski equality $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ then holds with the same form as the classical one. The subtleties are operator ordering (the work is not a simple observable because $H(0)$ and $H(\tau)$ need not commute) and the disturbance introduced by the initial measurement. Steady-state quantum versions connect to full counting statistics, as confirmed in electron-transport experiments through quantum dots.

How do these exact relations reduce to the familiar fluctuation-dissipation theorem?

Near equilibrium the entropy-production (or work) distribution becomes Gaussian. The detailed theorem then forces a relation between its mean and its variance: $\langle\Sigma\rangle=\mathrm{Var}(\Sigma)/2k_B$ . Expanded in the driving strength, the mean (dissipation) is linear in the perturbation while the variance measures spontaneous equilibrium fluctuations — so this single constraint is exactly the content of the fluctuation-dissipation theorem and, through it, the Green–Kubo formulas. The far-from-equilibrium identity thus contains linear response as its Gaussian limit rather than replacing it.