Crooks Fluctuation Theorem

We take from jarzynski-crooks-relations (the SM10 combined owner) the joint discovery that the full statistics of nonequilibrium work encode an equilibrium free-energy difference exactly, and its two headline identities — the Jarzynski equality ⟨e^{−βW}⟩=e^{−βΔF} and the Crooks relation P_F(W)/P_R(−W)=e^{β(W−ΔF)}. This article owns the distribution-level half of that pair: the Crooks fluctuation theorem as a statement about the shapes of forward and reverse work histograms, its practical payoff — the histogram-crossing free-energy readout — and the maximum-likelihood Bennett acceptance ratio estimator built on it. The integral (exponential-average) half is developed in the sibling article jarzynski-equality; Jarzynski is the $W$ -integral of Crooks, and we make that reduction explicit. Consider a system in contact with a heat bath at inverse temperature $\beta=1/k_BT$ , with a control parameter $\lambda(t)$ (an optical-trap position, a field) driven from $\lambda_A$ to $\lambda_B$ over a time $\tau$ . The forward experiment starts in canonical equilibrium of $H(x,\lambda_A)$ ; the reverse runs the mirror protocol $\lambda(\tau-t)$ from equilibrium of $H(x,\lambda_B)$ . The work $W$ is a functional of the stochastic trajectory, so it fluctuates from pull to pull. Crooks proved that the ratio of forward to time-reflected reverse work densities is the exponential of the dissipated work $W-\Delta F$ . The theorem follows from microscopic reversibility — a forward trajectory and its conjugate reverse trajectory have probabilities related by the Boltzmann factor of the heat exchanged. This holds for any Markov dynamics satisfying instantaneous detailed balance (Metropolis Monte Carlo, Langevin, master-equation dynamics). Beyond clarifying the statistical meaning of the second law, Crooks is a tool: because $P_F$ and $P_R(-W)$ cross exactly at $W=\Delta F$ regardless of how hard you drive, two histograms of an irreversible process read off the equilibrium free-energy difference — turning waste into measurement.

Key takeaways

The Crooks fluctuation theorem (CFT) states that the forward and reverse work distributions of a driven system obey $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ .
It quantifies the arrow of time as a probability statement: 'transient second-law violations' (dissipated work $<0$ ) are exponentially rare but never forbidden.
The two distributions cross exactly at $W=\Delta F$ — a robust, bias-free free-energy readout from nonequilibrium single-molecule pulling.
Integrating CFT over the forward distribution gives the Jarzynski equality $\langle e^{-\beta W}\rangle_F=e^{-\beta\Delta F}$ (its sibling article), and Jensen then gives $\langle W\rangle\ge\Delta F$ .
CFT rests on microscopic reversibility (detailed balance of the driven dynamics) and canonical-equilibrium starts for each direction; it is exact for arbitrarily fast driving.
The Bennett acceptance ratio (BAR) is the maximum-likelihood free-energy estimator built directly on Crooks, optimally combining forward and reverse work samples.
Key idea: the forward and reverse work distributions satisfy $P_F(+W)=P_R(-W)\,e^{\beta(W-\Delta F)}$ ; they cross where $W=\Delta F$ , and integrating recovers Jarzynski's equality and hence the second law.

The core insight

For a system driven by a protocol $\lambda(t)$ between two equilibrium endpoints, the forward and reverse work distributions satisfy $P_F(+W)=P_R(-W)\,e^{\beta(W-\Delta F)}$ ; they cross where $W=\Delta F$ , and integrating over the work axis recovers Jarzynski's equality and hence the second law. Microscopic reversibility is the engine — it makes the ratio of a trajectory's forward and reverse probabilities exactly $e^{\beta(W-\Delta F)}$ .

Where this lives today

The Crooks relation is the theoretical backbone of modern free-energy computation. In drug discovery, alchemical binding-affinity calculations run forward and reverse transformations of a ligand and feed the work values to BAR (or its multistate generalization MBAR) — the workhorse of relative binding free-energy pipelines used across the pharmaceutical industry. In single-molecule biophysics, the histogram-crossing readout extracts folding and binding free energies of nucleic acids and proteins from irreversible pulling, sidestepping the impossibly slow reversible pulls that would otherwise be required. And in stochastic thermodynamics, Crooks is the prototype of the detailed fluctuation theorems that now quantify the efficiency fluctuations of molecular motors and the thermodynamic cost of computation and sensing.

The physical picture

Imagine stretching a folded RNA hairpin with optical tweezers. When you pull, you usually do MORE work than $\Delta F$ (you waste energy as heat); when you release along the reverse protocol you usually recover LESS. Crooks says the rare forward realizations where you got lucky and did little work occur with exactly the probability that the corresponding reverse trajectory occurs — weighted by the Boltzmann factor of the heat. The asymmetry between forward and backward IS the dissipation. Where the two histograms cross, that asymmetry vanishes: the exponential factor is unity, and there $W=\Delta F$ .

Figure

Forward P_F(W) and reflected reverse P_R(−W) work histograms crossing exactly at W = ΔF — the free-energy readout that survives any driving speed

The two Gaussian-like histograms of equal width cross at exactly $W=\Delta F$ (red). The forward mean $\langle W\rangle_F$ sits to the right of $\Delta F$ by the dissipated work $\langle W_d\rangle\ge0$ — the second law — yet the crossing location does not move as driving speeds up; only the widths change.

The second law is recovered as the statement that the forward distribution sits, on average, to the right of $\Delta F$ . But the tail extending to $W<\Delta F$ is never empty: these 'transient second-law violations' occur with exponentially small probability $\sim e^{-\beta|W-\Delta F|}$ . In a macroscopic system $W$ scales with particle number, so these excursions become unobservable and the inequality looks absolute.

Figure

Microscopic reversibility: a trajectory and its time-reverse differ in probability only by the Boltzmann factor of heat dumped to the bath — the bookkeeping Crooks sums along every path

Microscopic reversibility is the engine: a path $\gamma$ and its mirror $\tilde\gamma$ differ in probability only by $e^{\beta Q}$ , the Boltzmann factor of the heat delivered to the bath. Crooks is just this accounting, summed over all paths that share a given work value.

The core equations

The central statement compares two distributions. Fix the protocol $\lambda(t)$ and its mirror; run each many times from the appropriate equilibrium start. Crooks relates the forward work density $P_F$ to the time-reflected reverse density $P_R(-W)$ point-by-point.

\frac{P_F(+W)}{P_R(-W)}=e^{\beta(W-\Delta F)} \qquad (1)

where: $P_F(W)$ — density of work $W$ in the forward process (start in equilibrium at $\lambda_A$ ); $P_R(W)$ — work density in the reverse process (start in equilibrium at $\lambda_B$ , run $\lambda(\tau-t)$ ); $\beta=1/k_BT$ — inverse bath temperature; $\Delta F=F(\lambda_B)-F(\lambda_A)$ — equilibrium free-energy difference; $W$ — work done ON the system in one realization. Reading: the log-ratio is linear in $W$ with slope $\beta$ and zero-crossing at $W=\Delta F$ . The work along a single trajectory is the energy injected by changing $\lambda$ (not the heat, which is bath exchange at fixed $\lambda$ ).

W=\int_0^\tau \dot\lambda(t)\,\frac{\partial H(x(t),\lambda)}{\partial\lambda}\bigg|_{\lambda(t)}\,\mathrm{d}t \qquad (2)

where: $H(x,\lambda)$ — system Hamiltonian depending on phase-space point $x$ and control parameter $\lambda$ ; $x(t)$ — stochastic trajectory; $\dot\lambda$ — protocol rate. The first law per realization reads $\Delta H=W+Q$ : work is the explicit- $\lambda$ part of $\mathrm{d}H$ , heat the motion-of- $x$ part (Sekimoto's stochastic energetics — recalled from stochastic-thermodynamics, not re-derived here). Setting the exponential in (1) to unity gives the estimator that makes the theorem practical.

P_F(\Delta F)=P_R(-\Delta F) \qquad (3)

where: the forward and time-reflected reverse histograms intersect exactly at the equilibrium free-energy difference. This is model-free — no fit to the distribution shapes is needed, only the crossing abscissa. Averaging (1) over the forward ensemble collapses the distribution statement to the integral (Jarzynski) statement, developed fully in the sibling jarzynski-equality.

\langle e^{-\beta W}\rangle_F=e^{-\beta\Delta F}\qquad\text{(Jarzynski = integral of Crooks)} \qquad (4)

where: $\langle\cdot\rangle_F$ — average over forward realizations. Jarzynski is literally the $W$ -integral of the rearranged Crooks relation; the two are not independent results.

Derivation ACrooks from master-equation microscopic reversibility

The goal is to obtain the distribution-level Crooks relation $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ from the most elementary dynamics that still captures the physics — a discrete-time Markov chain whose relaxation steps obey instantaneous detailed balance. We discretize the protocol into work steps (updating $\lambda$ at fixed configuration) alternating with heat steps (stochastic relaxation at fixed $\lambda$ ), factorize the trajectory probability via the Markov property, and take the ratio of a forward path weight to that of its conjugate reverse path. Every relaxation transition ratio contributes a Boltzmann factor of its energy change; the product telescopes into the total heat, and the equilibrium start/end factors supply the partition-function ratio that is $\Delta F$ . Coarse-graining over all paths of a given work $W$ then lifts the path statement to the distribution statement.

Step 1. Write the forward path weight as the equilibrium start times a product of relaxation steps: $P_F[\gamma]=p_A^{eq}(x_0)\prod_{i}w_{\lambda_{i+1}}(x_i\to x_{i+1})$ , with $p_A^{eq}(x_0)=e^{-\beta H(x_0,\lambda_A)}/Z_A$ . Why: the Markov property factorizes any trajectory probability into single-step transition probabilities, and the forward run begins in canonical equilibrium at $\lambda_A$ .

Step 2. Impose instantaneous detailed balance at each fixed $\lambda$ : $w_\lambda(x\to x')/w_\lambda(x'\to x)=e^{-\beta[H(x',\lambda)-H(x,\lambda)]}$ . Why: the relaxation steps are the coupling to the bath; requiring them to preserve $e^{-\beta H}$ at frozen $\lambda$ is exactly the statement that each such step exchanges heat $-\Delta H$ with a bath at temperature $T$ .

Step 3. Write the reverse path weight for the mirror protocol started from equilibrium at $\lambda_B$ : $P_R[\tilde\gamma]=p_B^{eq}(x_N)\prod_{i}w_{\lambda_{i+1}}(x_{i+1}\to x_i)$ . Why: the conjugate trajectory retraces the same configurations in reverse order under the reversed protocol, and the reverse experiment starts canonical at $\lambda_B$ .

Step 4. Take the ratio. The product of detailed-balance factors accumulates $e^{-\beta\sum\Delta H_{\rm relax}}=e^{-\beta Q}$ , the total heat to the bath, so $P_F[\gamma]/P_R[\tilde\gamma]=\dfrac{e^{-\beta H(x_0,\lambda_A)}/Z_A}{e^{-\beta H(x_N,\lambda_B)}/Z_B}\,e^{-\beta Q}$ . Why: only the relaxation (heat) steps carry a directional detailed-balance ratio; the work steps change $H$ at fixed $x$ and cancel between numerator and denominator except through the endpoint energies.

Step 5. Insert the first law $W=\Delta H_{\rm total}-Q=[H(x_N,\lambda_B)-H(x_0,\lambda_A)]-Q$ and $\Delta F=-k_BT\ln(Z_B/Z_A)$ . The endpoint energies combine with $Q$ into $W$ ; the partition functions give $\Delta F$ ; the result is $P_F[\gamma]/P_R[\tilde\gamma]=e^{\beta(W-\Delta F)}$ . Why: the algebra is just regrouping $-\beta[H(x_N)-H(x_0)]-\beta Q+\ln(Z_B/Z_A)$ into $\beta(W-\Delta F)$ .

Step 6. Coarse-grain: $P_F(W)=\sum_{\gamma:\,W[\gamma]=W}P_F[\gamma]$ , and each such $\gamma$ has reverse $\tilde\gamma$ with work $-W$ . Summing the path identity over all paths of fixed $W$ gives $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ . Why: the exponential factor depends on $\gamma$ only through $W[\gamma]$ , so it pulls out of the sum, leaving the distribution-level relation. Verified by verify_aq-sm-crooks-fluctuation-theorem.py: the ratio equals 1 at $W=\Delta F$ (unique crossing, no width parameter), and a Gaussian pair consistent with (1) crosses exactly at $\Delta F$ .

Go deeper · Derivation B · 1 minThe Gaussian near-equilibrium pair and its crossing

Derivation BThe Gaussian near-equilibrium pair and its crossing

The goal is to show what Crooks forces on the shapes of $P_F$ and $P_R$ in the experimentally common slow-driving (linear-response) regime, where the central-limit theorem makes both work distributions Gaussian. We assume $P_F(W)$ is Gaussian, impose relation (1), and read off three consequences: the two Gaussians must have equal variance, the log-ratio slope must equal $\beta$ (fixing the mean shift in terms of the variance), and the histograms cross at exactly $\Delta F$ . This is the quantitative content behind the crossing-point estimator and links the work variance to the dissipation — a fluctuation–dissipation relation.

Step 1. Take $P_F(W)=\mathcal N(\mu_F,\sigma^2)$ and $P_R(-W)=\mathcal N(-\mu_R,\sigma_R^2)$ (the reverse work distribution reflected). Form $\ln[P_F(W)/P_R(-W)]$ . Why: Crooks (1) says this log-ratio must be exactly linear in $W$ with slope $\beta$ ; matching a ratio of Gaussians to a linear function constrains their parameters.

Step 2. The coefficient of $W^2$ in the log-ratio is $-\tfrac{1}{2\sigma^2}+\tfrac{1}{2\sigma_R^2}$ . Linearity requires it to vanish, so $\sigma_R=\sigma$ . Why: a nonzero quadratic term would make the log-ratio curve, contradicting the linear form $\beta(W-\Delta F)$ ; equal widths is the only escape.

Step 3. With equal widths the slope is $(\mu_F+\mu_R)/\sigma^2$ , which Crooks sets equal to $\beta$ : $\mu_F+\mu_R=\beta\sigma^2$ . Solving the crossing $\ln$ -ratio $=0$ gives $W_{\rm cross}=(\mu_F-\mu_R)/2$ , and using the slope constraint this equals $\Delta F$ . Why: the crossing is where the two equal-width Gaussians meet — the midpoint of their centers — and the slope constraint pins that midpoint to $\Delta F$ .

Step 4. The variance–dissipation link follows: $\sigma_W^2=2k_BT\,(\mu_F-\Delta F)=2k_BT\,\langle W_d\rangle$ . Why: substituting $\Delta F=\mu_F-\tfrac12\beta\sigma^2$ (from $\langle e^{-\beta W}\rangle$ of a Gaussian, Step in the Jarzynski sibling) gives the near-equilibrium fluctuation–dissipation statement that work fluctuations are proportional to mean dissipation. Verified by verify_aq-sm-crooks-fluctuation-theorem.py: equal-variance constraint, slope $=(\mu_F+\mu_R)/\sigma^2$ , crossing at the center-midpoint, and $\sigma_W^2=2k_BT\,W_d$ all check symbolically; the worked-example Gaussian pair crosses at $W=\Delta F=3\,k_BT$ .

Limiting cases

Case	Result	Meaning
Quasi-static / reversible	$P_F(W)\to\delta(W-\Delta F),\ W_d\to0$	Infinitely slow driving: both distributions collapse to a spike at $\Delta F$ ; no dissipation, perfect reversibility.
Near-equilibrium (Gaussian work)	$\sigma_W^2=2k_BT\,\langle W_d\rangle$	Slow but finite driving gives Gaussian $P_F$ ; Crooks forces variance and dissipation to be linked — the fluctuation–dissipation theorem.
Macroscopic / thermodynamic limit	$\langle W\rangle\ge\Delta F$ (sharp)	Work is extensive, relative fluctuations $\sim N^{-1/2}$ ; the sub- $\Delta F$ tail vanishes and the deterministic second-law inequality emerges.
Average over forward (Jarzynski)	$\langle e^{-\beta W}\rangle_F=e^{-\beta\Delta F}$	Integrating Crooks over all $W$ yields the sibling Jarzynski equality, exact for arbitrarily fast driving.

Common trapCommon pitfall: ΔF is not ⟨W⟩, and both sides must start in equilibrium

$W$ is the work done ON the system through the protocol; $\Delta F=F_B-F_A$ is the equilibrium free-energy difference, NOT the mean work $\langle W\rangle$ (they differ by the dissipated work). Both experiments must START in true canonical equilibrium of their respective endpoints; the system need not END in equilibrium. Confusing $\Delta F$ with $\langle W\rangle$ , or mixing inclusive vs exclusive work conventions, corrupts the crossing. A subtler trap: the naive histogram-midpoint of the two Gaussian centers equals $\Delta F$ only because Crooks forces equal widths — if you fit two Gaussians of different widths (bad sampling), the crossing shifts. Use BAR, which combines both directions' statistics, rather than eyeballing a single crossing.

Worked examples

Example 1 — Reading ΔF and dissipation from a single-molecule pull (near-equilibrium Gaussian work). An optical-tweezer experiment on an RNA hairpin gives a forward work distribution that is Gaussian with mean $\langle W\rangle_F=5\,k_BT$ and standard deviation $\sigma_W=2\,k_BT$ . Work in units of $k_BT$ , so $\beta=1$ .

Step 1. The Jarzynski relation $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ for a Gaussian gives $\Delta F=\langle W\rangle-\tfrac12\beta\sigma_W^2=5-\tfrac12(1)(2^2)=\mathbf{3\,k_BT}$ .

Step 2. The dissipated work is $\langle W_d\rangle=\langle W\rangle-\Delta F=5-3=\mathbf{2\,k_BT}$ , consistent with $\sigma_W^2=2k_BT\langle W_d\rangle=2(1)(2)=4=(2\,k_BT)^2$ . ✓

Step 3. Crooks fixes the reverse work distribution: its mean is $\langle W\rangle_R=-(\langle W\rangle_F-\beta\sigma_W^2)=-(5-4)=\mathbf{-1\,k_BT}$ , same width $\sigma_W=2\,k_BT$ . The reflected reverse histogram $P_R(-W)$ is then centered at $+1\,k_BT$ .

Step 4. The forward $P_F(W)=\mathcal N(5,2)$ and reflected reverse $P_R(-W)=\mathcal N(1,2)$ — equal widths — cross at the midpoint of their centers, $(5+1)/2=\mathbf{3\,k_BT}=\Delta F$ . The crossing reads the free energy directly, no fit needed. All four numbers reproduced by the verifier.

Example 2 — Bennett acceptance ratio combines both directions. Take the Crooks-consistent pair above: $4\times10^5$ forward works sampled from $\mathcal N(5,2)$ and $4\times10^5$ reverse works from $\mathcal N(-1,2)$ (in $k_BT$ ).

Step 1. BAR solves the implicit equation $\sum_F f\!\big(\beta(W_F-\Delta F)\big)=\sum_R f\!\big(\beta(W_R+\Delta F)\big)$ with Fermi function $f(x)=1/(1+e^{x})$ — the maximum-likelihood estimator on the union of both samples.

Step 2. Root-finding by bisection returns $\Delta F=\mathbf{3.00\,k_BT}$ , matching the analytic value. Reproduced by the verifier, which also confirms the naive single-direction Jarzynski estimate $-\tfrac1\beta\ln\langle e^{-\beta W}\rangle_F$ converges to the same $3\,k_BT$ but far more slowly, since it hangs on the rare left tail of $P_F$ .

Origins · 1 mincrooks fluctuation theorem and how this was discovered

crooks fluctuation theorem

G. E. Crooks, 1999, American physicist · Gavin Crooks, working with David Chandler at UC Berkeley, generalized Jarzynski's 1997 equality into a full distribution-level symmetry. His 1999 Physical Review E paper cast the second law as an exact relation between the work histograms of a driven process and its time-reverse — the sharpest microscopic statement of irreversibility to date, and the foundation of the Bennett-acceptance-ratio free-energy method now standard in molecular simulation.

History

1976

— Bennett introduces the acceptance-ratio method for optimally estimating free-energy differences from two ensembles (J. Comput. Phys. 22, 245)

1993

— Evans, Cohen and Morriss report the first steady-state fluctuation theorem in shearing simulations (Phys. Rev. Lett. 71, 2401)

1997

— Jarzynski proves the nonequilibrium work equality ⟨e^{−βW}⟩=e^{−βΔF} (Phys. Rev. Lett. 78, 2690)

1998

— Crooks derives the path-space fluctuation theorem for Markovian dynamics (J. Stat. Phys. 90, 1481)

1999

— Crooks states the work-distribution relation P_F(W)/P_R(−W)=e^{β(W−ΔF)} (Phys. Rev. E 60, 2721)

2002

— Liphardt, Bustamante and collaborators verify the Jarzynski equality by RNA pulling (Science 296, 1832)

2005

— Collin, Ritort, Jarzynski, Smith, Tinoco and Bustamante verify Crooks and recover RNA folding free energies from the histogram crossing (Nature 437, 231)

Going further

References. Crooks, G. E., 'Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences', Phys. Rev. E 60, 2721 (1999) · Crooks, G. E., 'Nonequilibrium measurements of free energy differences for microscopically reversible Markovian systems', J. Stat. Phys. 90, 1481 (1998) · Jarzynski, C., 'Nonequilibrium equality for free energy differences', Phys. Rev. Lett. 78, 2690 (1997) · Bennett, C. H., 'Efficient estimation of free energy differences from Monte Carlo data', J. Comput. Phys. 22, 245 (1976) · Sethna, J. P., 'Statistical Mechanics: Entropy, Order Parameters, and Complexity', 2nd ed. · Kardar, M., 'Statistical Physics of Particles' · Sekimoto, K., 'Stochastic Energetics', Springer Lecture Notes in Physics 799 · Collin, D. et al., 'Verification of the Crooks fluctuation theorem and recovery of RNA folding free energies', Nature 437, 231 (2005) · Seifert, U., 'Stochastic thermodynamics, fluctuation theorems and molecular machines', Rep. Prog. Phys. 75, 126001 (2012)

Provenance. Central result: Crooks PRE 60, 2721 (1999), Eq. (10); Jarzynski PRL 78, 2690 (1997). A verifier should confirm: (1) the work/heat split via Sekimoto's stochastic energetics (Lecture Notes Phys. 799); (2) the master-equation derivation using instantaneous detailed balance reproduces $P_F/P_R=e^{\beta(W-\Delta F)}$ (Sethna 2nd ed. Ch. 11; Seifert Rep. Prog. Phys. 75, 126001, Sec. 4); (3) Jarzynski equality as the work-integral of CFT and $\langle W\rangle\ge\Delta F$ via Jensen; (4) the Bennett acceptance ratio as the maximum-likelihood estimator (Bennett 1976; Shirts et al., Phys. Rev. Lett. 91, 140601, 2003). The Gaussian relation $\sigma_W^2=2k_BT\langle W_d\rangle$ and the factor $k_B$ in $\beta=1/k_BT$ are standard. Experimental validation: Collin et al. Nature 437, 231 (2005).

Builds on: Detailed Balance and Microscopic Reversibility · The Master Equation · Markov Processes and the Markov Property · Langevin Equation and Stochastic Differential Equations Leads to: Fluctuation–Dissipation Theorem · Monte Carlo Method · Thermostats and Ensemble Control in Molecular Dynamics · Metropolis–Hastings Algorithm

Is the Crooks relation the same as the Jarzynski equality?

No — they are two levels of the same physics. Crooks is a distribution-level statement relating the full forward and reverse work histograms point-by-point, $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ . The Jarzynski equality (its sibling article) is the integral consequence: multiply Crooks by $e^{-\beta W}$ and integrate over all $W$ , using that the reverse distribution is normalized, and you get $\langle e^{-\beta W}\rangle_F=e^{-\beta\Delta F}$ . So Jarzynski is the first-moment shadow of Crooks; Crooks carries strictly more information (all moments of the work distribution, plus the crossing readout).

What exactly is 'microscopic reversibility' here?

It is the statement that a forward trajectory $\gamma$ and its time-reversed conjugate $\tilde\gamma$ (run under the reversed protocol) have probabilities related only by the Boltzmann factor of the heat exchanged: $\Pi[\gamma]/\tilde\Pi[\tilde\gamma]=e^{\beta Q[\gamma]}$ . For stochastic Markov dynamics this follows from instantaneous detailed balance of the relaxation steps; for deterministic thermostatted molecular dynamics it follows from time-reversal symmetry of the equations plus phase-space contraction bookkeeping. It is the single ingredient that turns the ratio of path weights into $e^{\beta(W-\Delta F)}$ .

Why does the crossing-point method beat the exponential average?

The Jarzynski average $\langle e^{-\beta W}\rangle$ is dominated by the far-left tail of $P_F$ — rare trajectories with $W<\Delta F$ . As dissipation grows those events become exponentially rare, so a finite sample systematically over-estimates $\Delta F$ (a biased estimator that converges only with exponentially many samples). The Crooks crossing, by contrast, is located where both distributions have appreciable weight (near their overlap), so it is estimated from typical events. The Bennett acceptance ratio makes this optimal, using both forward and reverse samples in a maximum-likelihood way.

How does Crooks reduce to the fluctuation–dissipation theorem?

In the slow-driving (linear-response) regime the central-limit theorem makes $P_F(W)$ Gaussian. Imposing Crooks on a Gaussian forces its variance and mean dissipation to satisfy $\sigma_W^2=2k_BT\langle W-\Delta F\rangle$ — work fluctuations proportional to mean dissipation. This is exactly the fluctuation–dissipation theorem at second order in the perturbation: the same relation between spontaneous fluctuations (variance) and the response to driving (dissipation). Crooks thus contains FDT as its near-equilibrium limit while remaining exact arbitrarily far from equilibrium.

Does Crooks apply to deterministic molecular dynamics, or only to stochastic models?

Both, with care. The cleanest derivation uses stochastic Markov dynamics with instantaneous detailed balance (Metropolis, Langevin, master-equation), where the heat is unambiguous. For deterministic thermostatted molecular dynamics (Nosé–Hoover and friends) the theorem still holds provided the reversed trajectory and the phase-space compression factor are defined consistently — the compression plays the role of the entropy exchanged with the thermostat. In practice Langevin and Metropolis codes reproduce $\ln[P_F/P_R]=\beta(W-\Delta F)$ as a straight line of unit slope $\beta$ , which is a stringent correctness check for simulation code.

What is the dissipated work, and why must it be non-negative on average?

The dissipated work is $W_d=W-\Delta F$ ; per realization it can be negative (a lucky low-work pull), but its ensemble mean satisfies $\langle W_d\rangle=\langle W\rangle-\Delta F\ge0$ . This is the second law, and Crooks makes it precise: $\langle W_d\rangle=k_BT\,D(P_F\|P_R)$ , the Kullback–Leibler divergence between the forward and reverse path ensembles, which is non-negative by construction. So the average dissipation is literally the statistical distinguishability of the process from its time-reverse — a quantitative arrow of time.

Can I use Crooks if my forward and reverse histograms barely overlap?

That is the hard regime. If driving is strong the two histograms separate and their crossing sits far in both tails, where sampling is poor — the crossing estimate becomes noisy and the Jarzynski average essentially fails. The practical fixes are to slow the protocol (narrowing each distribution toward $\Delta F$ ), to use bidirectional BAR/MBAR which extracts the most information from whatever overlap exists, or to stage the transformation through intermediate windows so that neighboring histograms overlap well. Insufficient overlap is the single most common reason a free-energy calculation gives a biased answer.