Jarzynski Equality

We take from the jarzynski-crooks-relations article (SM10) the joint derivation of the two nonequilibrium work relations — the Jarzynski equality $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ and its parent, the Crooks fluctuation theorem $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ — from microscopic reversibility. That combined article owns the derivation and the statement that Jarzynski is the integral of Crooks. Here we take the Jarzynski equality as our subject and ask the questions a practitioner asks: what does the exponential average form buy you, and why is it so treacherous to estimate? The Jarzynski equality is astonishing on its face. It says the equilibrium free-energy difference $\Delta F$ between two states — a purely equilibrium quantity, fixed by two partition functions — can be extracted from measurements of work $W$ done along a process driven arbitrarily far from equilibrium, at any speed. Drive a system fast and you waste work on dissipation, so $\langle W\rangle>\Delta F$ ; drive it slowly and $W\to\Delta F$ . But the specific exponential average $\langle e^{-\beta W}\rangle$ recovers $e^{-\beta\Delta F}$ regardless. It is the exponential weighting, and only the exponential weighting, that cancels the protocol-dependent dissipation. That same exponential is the source of the equality's practical curse. The average $\langle e^{-\beta W}\rangle$ is dominated by rare trajectories in which $W$ dips below $\Delta F$ — transient "second-law violations" — and those events become exponentially rare as the typical dissipation grows. A naive arithmetic estimate from a finite sample systematically over-estimates $\Delta F$ (it is biased) and converges exquisitely slowly. Understanding this rare-event structure — and the bidirectional (Bennett acceptance ratio) remedy that uses the Crooks partner — is the SM11 register of the subject, and the reason the equality is powerful in principle yet delicate in the laboratory.

Key takeaways

The Jarzynski equality $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ recovers an equilibrium free-energy difference from work measured along a process driven arbitrarily far from equilibrium, at any speed.
We recall the derivation from the SM10 jarzynski-crooks-relations owner; here we own the exponential-average form and its estimator.
The second law $\langle W\rangle\ge\Delta F$ follows by Jensen's inequality; the gap $W_{\mathrm{diss}}=\langle W\rangle-\Delta F\ge0$ is the mean dissipated work.
For a Gaussian work distribution, $\Delta F=\langle W\rangle-\beta\sigma_W^2/2$ exactly — the equality's near-equilibrium content.
Jarzynski is the integral of Crooks: integrating $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ over $W$ gives the exponential average (see the crooks-fluctuation-theorem article).
The average is dominated by rare $W<\Delta F$ trajectories; the naive estimator is biased high and converges exponentially slowly — the practical curse.
Key idea: the exponential weighting $e^{-\beta W}$ is a precision instrument — it cancels all protocol-dependent dissipation to leave the equilibrium invariant $\Delta F$ exactly, but it does so by up-weighting exponentially rare low-work trajectories, so the very structure that makes the equality exact makes it hard to sample.

The core insight

The exponential weighting $e^{-\beta W}$ is a precision instrument: it cancels all protocol-dependent dissipation to leave the equilibrium invariant $\Delta F$ exactly, for any driving speed. But it does so by up-weighting exponentially rare low-work trajectories — the transient second-law "violations" — so the very structure that makes the equality exact is what makes a finite-sample estimate biased and slow to converge.

Where this lives today

The Jarzynski equality is a standard tool wherever free energies must be computed from driven processes. In molecular dynamics, "fast-switching" or "steered MD" simulations pull a ligand out of a binding pocket many times and average the work to estimate binding free energies, trading long equilibration for many short trajectories. In single-molecule biophysics, optical-tweezer and AFM pulling of nucleic acids and proteins uses the equality (usually via the Crooks/Bennett route) to read equilibrium unfolding free energies from irreversible pulls. The rare-event bias analysed here drives active research into optimal driving protocols and variance-reduction schemes, and the equality is the entry point to the broader stochastic-thermodynamics framework of trajectory work, heat, and entropy.

The physical picture

Drag a colloidal bead through water by moving an optical trap from A to B. Do it slowly and you barely disturb equilibrium; the work you spend equals the free-energy difference $\Delta F$ . Do it fast and viscous drag wastes extra work, so on average $\langle W\rangle>\Delta F$ . Every individual pull gives a scattered work value. The Jarzynski equality says: don't average $W$ — average $e^{-\beta W}$ , and the answer is exactly $e^{-\beta\Delta F}$ , no matter how violently you pulled. Why the exponential and not the mean? Because $e^{-\beta W}$ is a Boltzmann factor for the work. Equilibrium statistical mechanics weights microstates by $e^{-\beta H}$ ; driving reshuffles the weights, but microscopic reversibility conserves the total Boltzmann measure, and $e^{-\beta W}$ is the bookkeeping device that tracks it from the initial ensemble to the final one. The mean work obeys only an inequality; the exponential average obeys an equality, because it is the physically conserved quantity.

Figure

The work distribution and the rare-event tail that carries the exponential average

The forward work distribution P(W). Its mean ⟨W⟩ lies to the RIGHT of the free-energy difference ΔF; the gap is the mean dissipated work W_diss. The exponential average ⟨e^(−βW)⟩ that yields e^(−βΔF) is dominated not by the bulk near ⟨W⟩ but by the shaded red far-left tail of trajectories with W below ΔF — the rare transient "second-law violations". These unsampled events are the outlier that carries the whole average, which is exactly why a finite sample is biased and slow to converge.

The equality is an exact sum rule on the work distribution: $\int P(W)e^{-\beta W}\,\mathrm{d}W=e^{-\beta\Delta F}$ for every protocol and every speed. This is why a single equilibrium number can be pulled from messy nonequilibrium data — and why, when the far-left tail is under-sampled, the pull goes wrong.

Figure

Two routes to ΔF: the exponential average (this article) and the Crooks histogram crossing (sibling)

Two estimators for the same ΔF. This article owns the left route — the Jarzynski exponential average of forward work — which is exact but dominated by the rare low-work tail and therefore biased for finite samples. The sibling crooks-fluctuation-theorem article owns the right route — the crossing of the forward and reverse work histograms at W = ΔF — which reads ΔF robustly without sampling the tail. Jarzynski is the integral of Crooks.

The core equations

The equality relates a nonequilibrium average to an equilibrium ratio of partition functions. We recall the setup from the jarzynski-crooks-relations owner: a system in a bath at inverse temperature $\beta$ , driven by a control parameter $\lambda(t)$ from $\lambda_A$ to $\lambda_B$ , started in equilibrium at $\lambda_A$ .

\big\langle e^{-\beta W}\big\rangle=e^{-\beta\,\Delta F} \qquad (1)

where: $W$ — the inclusive thermodynamic work along a trajectory (a fluctuating quantity); $\beta=1/k_BT$ — inverse temperature of the bath; $\Delta F=F(\lambda_B)-F(\lambda_A)$ — equilibrium Helmholtz free-energy difference; $\langle\cdot\rangle$ — average over trajectories started in equilibrium at $\lambda_A$ and driven by the fixed protocol. Reading: the exponentially weighted work average equals a pure equilibrium quantity, for any driving speed. The second law is a corollary, not an assumption: apply Jensen's inequality to the convex map $x\mapsto e^{-\beta x}$ .

\langle W\rangle-\Delta F\;=\;W_{\mathrm{diss}}\;\ge\;0 \qquad (2)

where: $W_{\mathrm{diss}}$ — mean dissipated (irreversible) work; the inequality is the Clausius statement, saturated only in the reversible (non-fluctuating) limit. Reading: on average you always spend at least $\Delta F$ ; the surplus is dissipation. Near equilibrium, or whenever the work distribution is Gaussian, the equality collapses to an exact mean–variance relation for $\Delta F$ .

\Delta F=\langle W\rangle-\frac{\beta}{2}\,\sigma_W^2 \qquad (3)

where: $\sigma_W^2$ — variance of the work; the second cumulant supplies the dissipation $W_{\mathrm{diss}}=\beta\sigma_W^2/2$ . Reading: the free-energy difference is the mean work minus a fluctuation correction; the wider the work distribution, the more you dissipated. This is the operational form used to correct free-energy estimates in simulation and single-molecule experiments. (Sign: $\Delta F$ is below $\langle W\rangle$ , consistent with $W_{\mathrm{diss}}\ge0$ .) Finally, the equality is the integral of its Crooks parent — the relation that ties this article to its sibling.

\frac{P_F(W)}{P_R(-W)}=e^{\beta(W-\Delta F)}\;\;\xrightarrow{\int\,\mathrm{d}W}\;\;\langle e^{-\beta W}\rangle_F=e^{-\beta\Delta F} \qquad (4)

where: $P_F(W)$ — forward work distribution; $P_R(-W)$ — reverse-protocol work distribution at $-W$ ; the crooks-fluctuation-theorem article owns the distribution form and the histogram-crossing readout. Reading: Jarzynski is what you get by integrating Crooks over all work values.

Derivation AThe Gaussian work distribution and the exact dissipation correction

The goal of this derivation is to extract the operationally most useful consequence of the Jarzynski equality: when the work distribution is Gaussian — which happens near equilibrium, or by the central limit theorem when the work accumulates from many small independent contributions — the equality gives an exact, closed-form relation between the free-energy difference, the mean work, and the work variance. This is the formula practitioners actually use to correct raw work measurements, and it exhibits both the second-law inequality and the dissipation in a single line. We assume $P(W)$ is Gaussian and evaluate the exponential average directly; nothing about the microscopic derivation (owned by SM10) is repeated.

Step 1. Take $W\sim\mathcal{N}(\langle W\rangle,\sigma_W^2)$ and compute the exponential average as a Gaussian moment-generating integral: $\langle e^{-\beta W}\rangle=\int \frac{e^{-(W-\langle W\rangle)^2/2\sigma_W^2}}{\sqrt{2\pi\sigma_W^2}}\,e^{-\beta W}\,\mathrm{d}W$ . Why a Gaussian: the MGF of a normal variable is exactly $\exp(-\beta\langle W\rangle+\tfrac12\beta^2\sigma_W^2)$ , so the integral is done in closed form.

Step 2. Evaluate the integral: $\langle e^{-\beta W}\rangle=\exp\!\big(-\beta\langle W\rangle+\tfrac12\beta^2\sigma_W^2\big)$ . Why the $+\tfrac12\beta^2\sigma_W^2$ term appears: completing the square in the exponent shifts the mean and leaves the variance contribution — the standard Gaussian MGF.

Step 3. Impose the Jarzynski equality $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ and take logs: $-\beta\Delta F=-\beta\langle W\rangle+\tfrac12\beta^2\sigma_W^2$ . Divide by $-\beta$ : $\Delta F=\langle W\rangle-\tfrac{\beta}{2}\sigma_W^2$ . Why this is the dissipation correction: the fluctuation term subtracts exactly the mean dissipated work $W_{\mathrm{diss}}=\tfrac{\beta}{2}\sigma_W^2\ge0$ .

Step 4. Read off the second law and its saturation. Since $\sigma_W^2\ge0$ , we have $\Delta F\le\langle W\rangle$ , i.e. $\langle W\rangle\ge\Delta F$ , with equality iff $\sigma_W=0$ (non-fluctuating work = reversible limit). Why this is Jensen: the same inequality follows generally from convexity of $e^{-\beta W}$ ; the Gaussian case makes the gap explicitly $\tfrac{\beta}{2}\sigma_W^2$ . Verified symbolically and numerically in verify_aq-sm-jarzynski-equality.py: the Gaussian $\Delta F=\langle W\rangle-\beta\sigma_W^2/2$ (with the correct sign), the worked numbers $\langle W\rangle=5,\sigma_W=2\,k_BT\Rightarrow\Delta F=3,W_{\mathrm{diss}}=2\,k_BT$ , and the Jensen bound.

Go deeper · Derivation B · 1 minWhy the exponential average is biased: the rare-event curse

Derivation BWhy the exponential average is biased: the rare-event curse

The goal here is to explain the single most important practical fact about the Jarzynski equality: although $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ is exact, any finite-sample estimate of $\Delta F$ built from it is systematically biased, and the bias comes from the same rare low-work trajectories that make the equality work. Understanding this bias — and knowing it always over-estimates $\Delta F$ — is what separates a correct free-energy calculation from a subtly wrong one. We analyse the standard estimator $\widehat{\Delta F}=-\tfrac1\beta\ln\!\big(\tfrac1N\sum_i e^{-\beta W_i}\big)$ and show it is biased upward, then point to the bidirectional remedy.

Step 1. The exponential average $\langle e^{-\beta W}\rangle$ is dominated by trajectories with $W<\Delta F$ , because $e^{-\beta W}$ up-weights small $W$ and these lie in the far-left tail of $P(W)$ . Why they matter so much: even though each such trajectory is rare, its weight $e^{-\beta W}$ is exponentially large, so the average hinges on adequately sampling the tail.

Step 2. With a finite sample the tail is under-represented, so $\tfrac1N\sum_i e^{-\beta W_i}$ under-estimates the true $\langle e^{-\beta W}\rangle$ (it misses the biggest contributions). Why the direction is fixed: the sample mean of a positive, heavy-right-weighted quantity is biased low; missing large summands lowers it.

Step 3. Because $\widehat{\Delta F}=-\tfrac1\beta\ln(\text{under-estimate})$ and $-\ln$ is decreasing, an under-estimated average makes $\widehat{\Delta F}$ larger than the truth: the free-energy estimate is biased high. Why this is dangerous: it looks like the process is more free-energetically costly than it is, and the bias shrinks only as fast as the tail is sampled — exponentially slowly when $W_{\mathrm{diss}}\gg k_BT$ .

Step 4. The remedy is bidirectional. The Crooks partner (sibling article) lets forward and reverse work data be combined via the Bennett acceptance ratio, or via the histogram-crossing at $W=\Delta F$ , neither of which requires sampling the troublesome tail. Why this works: the crossing point is set by the bulk of both distributions, not the tails, so it is statistically robust — this is why single-molecule experiments quote Crooks/Bennett estimates, not raw Jarzynski averages. Verified in verify_aq-sm-jarzynski-equality.py: a Monte-Carlo estimator of $\widehat{\Delta F}$ for Gaussian work at $\langle W\rangle=5,\sigma_W=2\,k_BT$ is biased high (bias $\approx0.20,0.05,0.01\to0$ as $N=50,500,5000,50000$ ), confirming the upward bias and slow convergence.

Limiting cases

Case	Result	Meaning
Quasistatic (reversible)	$\tau\to\infty:\ W\to\Delta F$ , $\sigma_W\to0$	$P(W)\to\delta(W-\Delta F)$ ; the equality becomes trivial and reversible thermodynamics is recovered.
Second-law (Jensen) corollary	$\langle W\rangle\ge\Delta F$	Convexity of $e^{-\beta W}$ gives the Clausius inequality with $W_{\mathrm{diss}}=\langle W\rangle-\Delta F\ge0$ .
Gaussian / near-equilibrium	$\Delta F=\langle W\rangle-\tfrac{\beta}{2}\sigma_W^2$	The variance supplies the exact dissipation correction; the operational estimator.
Instantaneous switch $\tau\to0$	$\langle e^{-\beta\Delta H}\rangle_A=e^{-\beta\Delta F}$	With no relaxation, work is $\Delta H$ on the initial ensemble — the Zwanzig free-energy-perturbation identity.
Strong dissipation	$W_{\mathrm{diss}}\gg k_BT$	The estimator's bias and variance explode; only bidirectional (Crooks/Bennett) methods remain reliable.

Common trapThe exponential average is the danger, not just the magic

It is tempting to treat $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ as a plug-in formula: measure many works, average $e^{-\beta W}$ , take a log. For anything but weak dissipation this is quietly wrong. The average is carried by rare $W<\Delta F$ events you almost never sample, so the finite- $N$ estimate is biased high in $\Delta F$ and converges exponentially slowly in the dissipation. Two further traps: $W$ must be the inclusive thermodynamic work $\int\partial_\lambda H\,\dot\lambda\,\mathrm{d}t$ (not the heat, not the "exclusive" work), and the system must start in true equilibrium at $\lambda_A$ . When dissipation is large, switch to the Crooks histogram-crossing or Bennett acceptance ratio, which use both forward and reverse data and do not depend on the tail.

Worked examples

Example 1 — Gaussian pull: reading ΔF from mean and variance. A single-molecule pulling experiment on an RNA hairpin is fast enough that the forward work is approximately Gaussian with mean $\langle W\rangle=5\,k_BT$ and standard deviation $\sigma_W=2\,k_BT$ . What is the free-energy difference, and how much work was dissipated?

Step 1. Use the Gaussian form $\Delta F=\langle W\rangle-\tfrac{\beta}{2}\sigma_W^2$ . Working in units of $k_BT$ (so $\beta=1/k_BT$ and $\beta\sigma_W^2=\sigma_W^2/k_BT=4\,k_BT$ ): $\Delta F=5-\tfrac12(4)=5-2=\mathbf{3\,k_BT}$ .

Step 2. The dissipated work is $W_{\mathrm{diss}}=\langle W\rangle-\Delta F=5-3=\mathbf{2\,k_BT}$ — exactly $\tfrac{\beta}{2}\sigma_W^2$ , so half the squared width in $k_BT$ units.

Step 3. Cross-check the equality directly: $\langle e^{-\beta W}\rangle=e^{-\beta\langle W\rangle+\beta^2\sigma_W^2/2}=e^{-5+2}=e^{-3}=\mathbf{0.0498}$ , which is $e^{-\beta\Delta F}=e^{-3}$ — the equality holds to machine precision. Naively exponentiating $-\langle W\rangle$ would have given $e^{-5}=0.0067$ , wrong by a factor of $\sim7$ : the variance correction is essential.

Example 2 — The estimator curse in numbers. Suppose you estimate $\Delta F$ for the same system from $N$ fast pulls using $\widehat{\Delta F}=-\tfrac1\beta\ln\!\big(\tfrac1N\sum_i e^{-\beta W_i}\big)$ . How biased is the answer?

Step 1. The true value is $\Delta F=3\,k_BT$ (Example 1). The average $\langle e^{-\beta W}\rangle$ is dominated by trajectories with $W<3\,k_BT$ , which lie about one standard deviation below the mean and further into the tail.

Step 2. A Monte-Carlo study (in the verifier) gives the mean estimate over many repetitions: at $N=50$ pulls the estimate is biased high by $\approx\mathbf{0.20\,k_BT}$ ; at $N=500$ by $\approx0.05\,k_BT$ ; at $N=5000$ by $\approx0.01\,k_BT$ ; only by $N=50\,000$ is the bias negligible. The bias is always positive (over-estimates $\Delta F$ ) and shrinks slowly.

Step 3. The lesson: even for this benign Gaussian with modest dissipation ( $W_{\mathrm{diss}}=2\,k_BT$ ), hundreds to thousands of pulls are needed for a trustworthy Jarzynski estimate. For $W_{\mathrm{diss}}=10\,k_BT$ the required sample explodes exponentially — which is precisely why the Crooks/Bennett bidirectional method (sibling article) is the tool of choice in practice.

Origins · 1 minjarzynski equality and how this was discovered

jarzynski equality

derived by Christopher Jarzynski, 1997 · Christopher Jarzynski proved the equality in a 1997 Physical Review Letter while at Los Alamos, showing that the equilibrium free-energy difference between two states could be recovered exactly from the exponential average of work done in a finite-time, arbitrarily-fast process. A companion paper the same year gave the master-equation proof for stochastic dynamics. The result overturned the intuition that nonequilibrium measurements could yield only inequalities, and — together with Crooks's 1999 fluctuation theorem — launched the modern field of single-molecule and computational free-energy estimation.

History

1997

— Christopher Jarzynski publishes the equality

\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}

in Phys. Rev. Lett. 78, 2690

1997

— Jarzynski gives the master-equation proof for stochastic dynamics in Phys. Rev. E 56, 5018

1999

— Gavin Crooks derives the fluctuation theorem of which the Jarzynski equality is the integral corollary, Phys. Rev. E 60, 2721

2002

— Liphardt, Dumont, Smith, Tinoco and Bustamante test the equality by mechanically unfolding RNA, Science 296, 1832

2003

— Hummer and Szabo formulate the reconstruction of free-energy surfaces from pulling data using the equality

2005

— Collin, Ritort, Jarzynski, Smith, Tinoco and Bustamante verify the Crooks relation and Bennett method on RNA, Nature 437, 231

Going further

References. Jarzynski, C., 'Nonequilibrium Equality for Free Energy Differences', Phys. Rev. Lett. 78, 2690 (1997) · Jarzynski, C., 'Equilibrium free-energy differences from nonequilibrium measurements: A master-equation approach', Phys. Rev. E 56, 5018 (1997) · Crooks, G. E., 'Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences', Phys. Rev. E 60, 2721 (1999) · Sethna, J. P., 'Statistical Mechanics: Entropy, Order Parameters, and Complexity' (2nd ed.), Oxford · Kardar, M., 'Statistical Physics of Particles', Cambridge · Liphardt, J. et al., 'Equilibrium Information from Nonequilibrium Measurements in an Experimental Test of Jarzynski's Equality', Science 296, 1832 (2002) · Seifert, U., 'Stochastic thermodynamics, fluctuation theorems and molecular machines', Rep. Prog. Phys. 75, 126001 (2012)

Provenance. Equality $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ and its derivation follow Jarzynski PRL 78, 2690 (1997) and PRE 56, 5018 (1997); the Crooks parent from PRE 60, 2721 (1999). This article recalls the joint derivation from the SM10 jarzynski-crooks-relations owner and does not repeat it; its owned contribution is the exponential-average form, the Gaussian dissipation correction, and the estimator/rare-event analysis. A verifier should: (1) confirm the Gaussian result $\Delta F=\langle W\rangle-\beta\sigma_W^2/2$ from the Gaussian moment-generating integral; (2) check the worked numbers $\langle W\rangle=5,\sigma_W=2\,k_BT\Rightarrow\Delta F=3,W_{\mathrm{diss}}=2\,k_BT$ ; (3) verify the Jensen second-law corollary $\langle W\rangle\ge\Delta F$ ; (4) confirm Jarzynski as the $W$ -integral of Crooks with equal-variance Gaussians crossing at $\Delta F$ ; (5) reproduce the upward estimator bias by Monte-Carlo. Experimental context references Liphardt et al. (2002) and Collin et al. (2005). No fabricated constants beyond $k_B,T,\beta$ .

Builds on: Detailed Balance and Microscopic Reversibility · The Master Equation · Markov Processes and the Markov Property · Fokker–Planck Equation Leads to: Crooks Fluctuation Theorem · Fluctuation–Dissipation Theorem · Molecular Dynamics Simulation · Thermostats and Ensemble Control in Molecular Dynamics

How can an equilibrium quantity come out of a nonequilibrium measurement?

Because the exponential weighting $e^{-\beta W}$ is not an arbitrary average — it is a Boltzmann factor for the work. Microscopic reversibility guarantees that the total Boltzmann measure of the system is conserved as the control parameter is driven, and $\langle e^{-\beta W}\rangle$ is precisely the bookkeeping quantity that transports that conserved measure from the initial equilibrium ensemble at $\lambda_A$ to the final one at $\lambda_B$ . The protocol-dependent dissipation contributes to individual work values but cancels exactly inside the exponential average, leaving only the ratio of partition functions $Z(\lambda_B)/Z(\lambda_A)=e^{-\beta\Delta F}$ . The full derivation is owned by the jarzynski-crooks-relations article.

What exactly is the "work" W in the equality?

It is the inclusive thermodynamic work, $W=\int_0^\tau \partial_\lambda H(x(t),\lambda)\,\dot\lambda\,\mathrm{d}t$ — the energy injected by changing the external control parameter along its protocol. This is distinct from the heat $Q$ exchanged with the bath (which comes from the system's motion at fixed $\lambda$ ), and from the "exclusive" work convention that omits boundary terms. The equality holds specifically for this $W$ ; substituting the heat, or the exclusive work, breaks it. Getting the work definition right is the single most common source of error when applying the equality to a real experiment or simulation.

Why does the estimator over-estimate ΔF rather than under-estimate it?

Because the exponential average $\langle e^{-\beta W}\rangle$ is a positive quantity dominated by its largest terms, which come from rare low-work trajectories. A finite sample almost always misses some of those rare large summands, so the sample mean of $e^{-\beta W}$ is biased low. Since $\widehat{\Delta F}=-\tfrac1\beta\ln(\text{that mean})$ and $-\ln$ is a decreasing function, a mean biased low produces a $\Delta F$ biased high. The direction is therefore fixed: naive Jarzynski estimates systematically report the process as more costly than it is. The bias shrinks only as the low-work tail is progressively sampled, which is exponentially slow when dissipation is large.

When is the Gaussian formula ΔF = ⟨W⟩ − βσ_W²/2 valid?

Whenever the work distribution is well-approximated by a Gaussian. This happens near equilibrium (weak, slow driving) and, more generally, whenever the total work accumulates as a sum of many small, roughly independent contributions so the central limit theorem applies (e.g. pulling a long polymer through many independent segments). For strongly driven single-barrier processes the distribution can be markedly non-Gaussian — often with a fat low-work tail — and the second-cumulant formula then under-corrects. In that regime one either includes higher cumulants or, better, switches to the bidirectional Crooks/Bennett estimator, which makes no Gaussian assumption.

Does the equality let a system extract work below ΔF and beat the second law?

No. Individual trajectories can have $W<\Delta F$ — these are the transient "second-law violations" that carry the exponential average — but the ensemble mean obeys $\langle W\rangle\ge\Delta F$ exactly, by Jensen's inequality. You cannot systematically select the favourable trajectories without a measurement-and-feedback apparatus, and the thermodynamic cost of that information (Landauer's principle, information thermodynamics) exactly restores the balance. The equality quantifies how rare the sub- $\Delta F$ trajectories are — a fraction weighted by $e^{-\beta W}$ — which is precisely why they cannot be harvested for free.

What is the Bennett acceptance ratio and why is it better?

The Bennett acceptance ratio is a bidirectional estimator that combines forward and reverse work data optimally to estimate $\Delta F$ . Because the Crooks relation ties $P_F(W)$ and $P_R(-W)$ together, the two data sets constrain $\Delta F$ jointly, and Bennett's method finds the maximum-likelihood value from both — effectively using the overlap region of the two histograms (near $W=\Delta F$ ) rather than the poorly-sampled tail of either. It is provably the minimum-variance estimator among a broad class and is the standard analysis for single-molecule pulling and alchemical free-energy simulations. The sibling crooks-fluctuation-theorem article develops the crossing readout on which it rests.

How does the instantaneous-switch limit connect to older free-energy methods?

If you switch the control parameter infinitely fast ( $\tau\to0$ ), the system has no time to move, so the work is simply the energy change $W=\Delta H=H(x,\lambda_B)-H(x,\lambda_A)$ evaluated on the initial equilibrium ensemble. The Jarzynski equality then reads $\langle e^{-\beta\Delta H}\rangle_A=e^{-\beta\Delta F}$ , which is exactly the Zwanzig free-energy-perturbation formula from 1954. So the venerable free-energy-perturbation method is the sudden-limit special case of the Jarzynski equality, and the finite-time equality generalises it to any driving speed — with the same rare-event sampling difficulty, now recognised as intrinsic rather than a quirk of perturbation theory.