Path Integral and Onsager-Machlup Action for Fluctuations

A stochastic differential equation fixes only the instantaneous statistics of the noise, yet the questions that matter in non-equilibrium physics are almost always about whole histories: the probability of a rare excursion, the dominant route a particle takes to escape a trap, the entropy produced along an irreversible path. The Onsager–Machlup formalism answers these by assigning a probability to each entire trajectory. The chance of observing a history $\{x(t)\}$ is $P[x]\,\mathcal{D}x \propto e^{-S[x]}\,\mathcal{D}x$ , where $S$ — the Onsager–Machlup action — is a real, Euclidean-weighted functional built from the drift, the diffusion, and the rate of energy dissipation. It is the classical, dissipative cousin of the Feynman path integral, with $D$ playing the role that $\hbar$ plays in quantum mechanics. Onsager and Machlup introduced this "principle of least dissipation" in 1953 for linear fluctuations near equilibrium, extending Onsager's reciprocal relations from static forces to time-dependent paths. The modern derivation writes the noise-averaged probability as a functional delta-function enforcing the Langevin equation, then integrates over the Gaussian noise. The subtle ingredient is the measure: because Brownian paths are nowhere differentiable, the continuum action is shorthand for a discretized limit, and the Jacobian of the change of variables from noise to path contributes a term proportional to the divergence of the drift — a term whose precise form depends on whether one discretizes à la Itô or Stratonovich, and which must never be dropped. Two payoffs make the action worth the care. In the weak-noise limit the path integral localizes onto the single trajectory that minimizes $S$ — the "instanton" or optimal fluctuation path — connecting Onsager–Machlup to large-deviation theory and to Kramers' escape rate. And the asymmetry of the action under time reversal turns out to equal the entropy produced along the path: $\ln(P[x]/P[\tilde x]) = \Delta S_{\text{medium}}/k_B$ . That single identity is the microscopic seed from which the fluctuation theorems grow, so the Onsager–Machlup action is the natural bridge between the dynamics of a single noisy particle and the exact far-from-equilibrium symmetries of stochastic thermodynamics.

Key takeaways

The Onsager–Machlup (OM) action assigns a weight $\propto e^{-S[x]}$ to each trajectory $x(t)$ of a Langevin system, turning stochastic dynamics into a real-weighted path integral $P[x]\,\mathcal{D}x$ .
For overdamped dynamics $\dot x = a(x) + \sqrt{2D}\,\eta$ the action is $S=\int\!\big[\tfrac{(\dot x-a)^2}{4D} + \tfrac12 a'(x)\big]\mathrm{d}t$ ; the $\tfrac12 a'$ Jacobian term is convention-dependent and mandatory.
The Lagrangian is a "least-dissipation" functional: the noise-free relaxation extremizes it ( $\delta S=0$ gives the deterministic equation of motion), and departures cost action exponentially.
Time-reversing the action gives the entropy produced: $\ln(P[x]/P[\tilde x])=\Delta S_{\text{med}}/k_B$ — the origin of the fluctuation theorems.
In the weak-noise limit $D\to0$ the integral is dominated by the minimum-action instanton, giving large-deviation rates and Kramers' $e^{-\Delta V/k_BT}$ .
Key idea: average over noise the constraint that $x(t)$ solves the Langevin equation; the Gaussian noise integral converts the dynamics into a path probability $e^{-S[x]}$ whose action is a quadratic cost for departing from deterministic relaxation, plus a drift-divergence Jacobian fixed by the stochastic-calculus convention.

The core insight

Every history of the particle is a string in spacetime carrying a weight $e^{-S[x]}$ . The action $S$ charges the path for the mismatch between its actual velocity and the drift the force alone would produce; the least-action path is the deterministic one, and the difference between a path's action and its time-reverse is exactly the entropy dumped to the bath.

Where this lives today

The Onsager–Machlup action is the workhorse of rare-event science. Numerically minimizing it — the geometric minimum-action method and the string method — finds the most probable transition pathways and their rates in high-dimensional systems that could never be sampled by brute force: protein folding and misfolding, crystal nucleation, chemical reaction coordinates, and even tipping points in climate models. Promoting the trajectory $x(t)$ to a field $\phi(\mathbf r,t)$ turns the same action into the Martin–Siggia–Rose–Janssen field theory that underlies the renormalization-group analysis of dynamic critical phenomena, where the dynamic critical exponent $z$ is computed from exactly this response-field formalism. And because the action's time-reversal asymmetry is the trajectory entropy production, it is the technical heart of modern stochastic thermodynamics — the language in which the fluctuation theorems and thermodynamic uncertainty relations are stated and proved.

The physical picture

Picture every possible history of the particle as a string drawn in the plane of position against time. The deterministic, noise-free trajectory — steepest descent on the potential — costs zero excess action. Any wiggle the noise must "pay for" costs action proportional to the squared mismatch $(\dot x - a)^2$ between the real velocity and the velocity the drift alone would produce. The path integral sums $e^{-\text{cost}}$ over all strings, so the most probable history is the one of least dissipation, and improbable fluctuations are exponentially suppressed — exactly as Boltzmann weights suppress high-energy configurations, but now in the space of trajectories rather than states.

Figure

Trajectory ensemble weighted by e^{−S∕2k_B}: the least-action drift path is bold, a rare uphill fluctuation is the faint red outlier

Onsager's original insight was that the integrand is literally a dissipation function: relaxation along the path of least dissipated energy is the most probable, and fluctuations away from it are exponentially rare. The Jacobian term is the price of curved coordinates — changing variables from the flat white-noise measure to the path measure deforms the volume element by the divergence of the drift, $a'(x)$ , which is why it survives in the continuum. In the small-noise limit the sum over paths collapses onto the single minimizing instanton, just as a quantum amplitude localizes onto a classical trajectory as $\hbar\to0$ ; here $D$ plays the role of $\hbar$ , and large-deviation rates are the classical actions of these optimal fluctuation paths.

Figure

Time-reversal asymmetry: the forward action and the reverse action differ by the heat dumped to the bath

The core equations

The starting point is the overdamped Langevin equation — the stochastic equation of motion whose trajectory-level statistics we want to package. In the register of stochastic dynamics, the drift $a(x)$ carries the deterministic force and the white noise $\eta$ carries the thermal kicks; the diffusion constant is tied to the mobility by the Einstein relation established for the langevin-equation, $D=\mu k_BT$ .

\dot x(t) = a(x) + \sqrt{2D}\,\eta(t), \qquad \langle \eta(t)\eta(t')\rangle = \delta(t-t') \qquad (1)

where: $x(t)$ — the stochastic trajectory; $a(x)=-\mu\,\partial_x V$ — the drift in a force field; $D$ — diffusion constant, $D=\mu k_BT$ ; $\eta(t)$ — unit-strength Gaussian white noise; $\mu$ — mobility (inverse friction). Physically: the particle drifts downhill under $a$ while being jostled by $\eta$ . Because the noise is Gaussian, its own probability weight is a simple quadratic functional — the Wiener measure. Every question about the path can be answered by pushing this weight forward through the change of variables $\eta\to x$ .

P[\eta]\,\mathcal{D}\eta \propto \exp\!\Big(-\tfrac12\!\int_0^T \eta^2\,\mathrm{d}t\Big)\,\mathcal{D}\eta \qquad (2)

where: $\eta(t)$ — the white-noise field being integrated out; $T$ — total observation time. Physically: each independent time-mode of the noise is an independent unit Gaussian, so the joint weight exponentiates the summed variance. Carrying that measure through the Langevin constraint gives the central object: a real-weighted probability for the whole path, controlled by the Onsager–Machlup action. In the Stratonovich (midpoint) convention the action carries the mandatory drift-divergence Jacobian.

P[x]\,\mathcal{D}x \propto e^{-S_{\mathrm{OM}}[x]}\,\mathcal{D}x, \qquad S_{\mathrm{OM}}[x]=\int_0^T\!\Big[\frac{(\dot x - a)^2}{4D} + \frac12\,a'(x)\Big]\mathrm{d}t \qquad (3)

where: $S_{\mathrm{OM}}$ — the OM action functional; $(\dot x-a)^2$ — squared deviation of the actual from the deterministic velocity; $a'(x)=\partial_x a$ — the Jacobian (drift-divergence) term; $4D$ — normalization from the noise variance $2D$ . Physically: the first term is the "cost" of deviating from deterministic relaxation; the second corrects the path measure so the propagator matches Fokker–Planck. Specializing to a potential makes the structure transparent and exposes the state-function boundary term that encodes equilibrium.

S_{\mathrm{OM}} = \frac{1}{4D}\int_0^T (\dot x + \mu V')^2\,\mathrm{d}t + \frac{\mu}{2}\int_0^T V''(x)\,\mathrm{d}t \qquad (4)

where: $V(x)$ — potential energy, force $=-V'$ ; $V''$ — curvature $=$ drift divergence; $\mu$ — mobility, $D=\mu k_BT$ . Physically: the cross term inside the square is a total time-derivative $\propto \Delta V$ , so the "bulk" of the action is symmetric and only boundaries break time-reversal. Finally, the quantity that gives the article its reach: the time-reversal asymmetry of the action equals the entropy dumped to the medium along the path. This is the identity the fluctuation-theorems promote to an exact symmetry.

\Sigma_{\text{med}}[x] \equiv k_B\big(S_{\mathrm{OM}}^{\text{rev}} - S_{\mathrm{OM}}^{\text{fwd}}\big) = \frac{1}{T}\int_0^T F\circ \dot x\,\mathrm{d}t \qquad (5)

where: $\Sigma_{\text{med}}$ — heat-related medium entropy production along the path; $F=-V'$ — force; $\circ$ — the Stratonovich product; rev/fwd — the reversed and forward trajectories. Physically: the log-ratio of watching the film forwards versus backwards is the dissipation, in units of $k_B$ .

Derivation ADeriving the OM action by integrating out the noise

The goal is to turn the Langevin equation — a statement about instantaneous noise — into a probability for the entire history, and to expose exactly where the drift-divergence Jacobian comes from. We work on a time grid, insert the equation of motion as a delta-functional, change variables from the independent noise increments to the path increments, keep the Jacobian of that change honestly, and take the continuum limit. The single delicate choice is the discretization point at which the drift is evaluated; the midpoint (Stratonovich) rule produces the factor $\tfrac12$ that lands in the final action.

Step 1. Discretize time as $t_n=n\Delta t$ . Each noise increment $\eta_n$ is an independent Gaussian of variance $1/\Delta t$ , so the joint weight is $\prod_n e^{-\tfrac12 \eta_n^2 \Delta t}$ . Why: white noise is delta-correlated, so distinct time-slices are statistically independent and the functional weight factorizes over the grid.

Step 2. Write the Langevin step with the drift evaluated at the midpoint $\bar x_n=(x_n+x_{n+1})/2$ : $\;x_{n+1}-x_n = a(\bar x_n)\Delta t + \sqrt{2D}\,\eta_n\Delta t$ , and solve for $\eta_n = \tfrac{1}{\sqrt{2D}}\big(\tfrac{\Delta x_n}{\Delta t} - a(\bar x_n)\big)$ . Why: the midpoint rule is the Stratonovich convention; it keeps ordinary calculus valid and is what makes the Jacobian nontrivial.

Step 3. The change of variables $\eta_n\to x_{n+1}$ has Jacobian $\big|\partial\eta_n/\partial x_{n+1}\big| = \tfrac{1}{\sqrt{2D}\,\Delta t}\big(1-\tfrac12 a'(\bar x_n)\Delta t\big)$ ; the product over $n$ contributes $\exp\!\big(-\tfrac12\sum_n a'\,\Delta t\big)$ to the weight. Why: differentiating $\eta_n$ with respect to the midpoint-containing $x_{n+1}$ picks up $-\tfrac12 a'$ ; this is the origin of the divergence term.

Step 4. Assemble the exponent by combining the Gaussian noise weight (with $\eta_n$ substituted) and the log-Jacobian: $\sum_n\big[\tfrac{(\Delta x_n/\Delta t - a)^2}{4D} + \tfrac12 a'(\bar x_n)\big]\Delta t$ . Why: $\eta_n^2/2 = (\Delta x_n/\Delta t - a)^2/(4D)$ after substitution, and the Jacobian adds the $\tfrac12 a'$ .

Step 5. Take $\Delta t\to0$ : the sum becomes $\int_0^T\!\big[\tfrac{(\dot x - a)^2}{4D} + \tfrac12 a'\big]\mathrm{d}t = S_{\mathrm{OM}}$ , the boxed action of Eq. (3). Why: the continuum limit of the Riemann sum reproduces the action; the $\tfrac12 a'$ is the Stratonovich Jacobian that survives. Verified in verify_aq-sm-onsager-machlup-action.py: the single-step kernel with the $\tfrac12 a'$ prefactor solves the Fokker–Planck equation to $O(\Delta t)$ with $\langle\Delta x\rangle=a\Delta t$ , $\langle\Delta x^2\rangle=2D\Delta t$ ; dropping the term gives the wrong drift.

Go deeper · Derivation B · 1 minThe free-diffusion action reproduces the diffusion propagator

Derivation BThe free-diffusion action reproduces the diffusion propagator

The cleanest, fully numerical check of the whole construction is the driftless case $a=0$ : there the action is purely kinetic, the least-action path between two endpoints is a straight line, and its weight must reproduce the exact Gaussian diffusion propagator. This both validates the normalization and delivers the article's first worked number. We evaluate the action on the straight-line path, exponentiate, and compare with the known heat kernel.

Step 1. For $a=0$ the action is $S=\tfrac{1}{4D}\int_0^T \dot x^2\,\mathrm{d}t$ (the Jacobian term $a'=0$ vanishes identically). Why: with no drift there is no divergence to correct; only the kinetic cost remains.

Step 2. The action-minimizing path between $x(0)=0$ and $x(T)=L$ has constant velocity $\dot x = L/T$ (Euler–Lagrange for $\dot x^2$ ), so $S_{\min}=\tfrac{1}{4D}\,(L/T)^2\,T = \dfrac{L^2}{4DT}$ . Why: the straight line extremizes $\int \dot x^2$ subject to fixed endpoints; any wiggle only adds positive cost.

Step 3. The path weight in the semiclassical form $P\propto e^{-S_{\min}}$ then gives $P\propto e^{-L^2/4DT}$ , which is exactly the exponent of the free-diffusion propagator $P(L,T)=\tfrac{1}{\sqrt{4\pi DT}}\,e^{-L^2/4DT}$ . Why: the minimum-action exponent reproduces the Gaussian heat kernel — the definitive consistency check that the OM measure is correct. Verified in verify_aq-sm-onsager-machlup-action.py: $S_{\min}=L^2/4DT$ symbolically, and the exponent matches the Gaussian propagator; the free-diffusion kernel normalizes to 1.

Go deeper · Derivation C · 1 minWeak-noise instanton and the Kramers exponent

Derivation CWeak-noise instanton and the Kramers exponent

In the small- $D$ limit the path integral is dominated by the single trajectory that minimizes the action, and for a gradient drift that minimum reduces to a pure boundary term — the Arrhenius factor of barrier crossing. The point is to show that the exponential in Kramers' escape rate is a state function $\Delta V/k_BT$ that falls out of the action's structure, independent of protocol, leaving only the prefactor to Gaussian fluctuations. We complete the square in the action, isolate the boundary term, and identify the optimal (uphill) path.

Step 1. With $a=-\mu V'$ write $S=\tfrac{1}{4D}\int(\dot x+\mu V')^2\,\mathrm{d}t = \tfrac{1}{4D}\int\big[(\dot x-\mu V')^2 + 4\mu V'\dot x\big]\mathrm{d}t$ . Why: completing the square splits the action into a non-negative perfect square plus a cross term that is a total derivative.

Step 2. The cross term integrates to a boundary: $\tfrac{1}{4D}\int 4\mu V'\dot x\,\mathrm{d}t = \tfrac{\mu}{D}\Delta V = \dfrac{\Delta V}{k_BT}$ , using the Einstein relation $D=\mu k_BT$ . Why: $V'\dot x = \tfrac{d}{dt}V$ , so the integral is path-independent — it is the free-energy barrier.

Step 3. The perfect square vanishes on the optimal (instanton) path $\dot x = +\mu V'$ — the time-reversed relaxation that climbs uphill — so $S_{\min}=\Delta V/k_BT$ and the escape rate carries the exponential $k\propto e^{-\Delta V/k_BT}$ . Why: minimizing the non-negative square sets it to zero; only the boundary survives, giving the Arrhenius exponent. Verified in verify_aq-sm-onsager-machlup-action.py: the Einstein-relation cancellation $\mu\Delta V/D=\Delta V/k_BT$ , and the instanton condition $\dot x=+\mu V'$ zeroing the square.

Limiting cases

The OM path integral interpolates between several familiar regimes as noise strength and driving change:

Case	Result	Meaning
Weak noise $D\to0$	$P[x]\sim e^{-S_{\min}}$	Saddle-point dominance: only the minimum-action instanton survives, giving Freidlin–Wentzell large deviations and Kramers' $e^{-\Delta V/k_BT}$ .
Linear (OU) drift $a=-\gamma x$	$S=\tfrac{1}{4D}\int(\dot x+\gamma x)^2\mathrm{d}t + \tfrac{\gamma T}{2}$	Purely Gaussian path integral; reproduces the Ornstein–Uhlenbeck propagator exactly. The Jacobian is a constant $\gamma T/2$ .
Equilibrium / detailed balance	$S_{\text{rev}}-S_{\text{fwd}}=\Delta V/k_BT$	Bulk terms cancel; only the state-function boundary survives, so the average entropy production vanishes in steady state.
Non-equilibrium (no potential)	$\Delta S_{\text{tot}}=k_B(S_{\text{rev}}-S_{\text{fwd}})$	The action asymmetry equals the total entropy produced — the basis of the detailed and integral fluctuation theorems.

Worked examples

Example 1 — the free-diffusion action reproduces the propagator. Take a colloidal particle with diffusion constant $D=0.20\ \mu\text{m}^2/\text{s}$ (a typical 1-µm bead in water). What action does the straight-line path from the origin to $L=1.0\ \mu\text{m}$ over $T=2.0\ \text{s}$ carry, and does its weight match the Gaussian propagator?

Step 1. With no drift, $S_{\min}=L^2/(4DT) = (1.0)^2 / (4\cdot 0.20\cdot 2.0) = 1.0/1.6 = \mathbf{0.625}$ (dimensionless).

Step 2. The path weight is $P\propto e^{-S_{\min}}=e^{-0.625}=\mathbf{0.535}$ , and the full normalized propagator is $P(L,T)=\tfrac{1}{\sqrt{4\pi DT}}e^{-L^2/4DT} = \tfrac{1}{\sqrt{4\pi\cdot0.4}}e^{-0.625}=0.446\times0.535=\mathbf{0.239\ \mu\text{m}^{-1}}$ .

Step 3. The exponent from the action exactly matches the exponent of the exact heat kernel — confirming the OM measure. The typical displacement scale $\sqrt{2DT}=\sqrt{0.8}=0.894\ \mu\text{m}$ is of order $L$ , so $S\sim1$ and this displacement is not rare, as expected.

Example 2 — the Kramers exponent from a real barrier. A protein-folding-like double well has a barrier $\Delta V = 5\,k_BT$ at $T=300\ \text{K}$ . Using the instanton result, what is the exponential suppression of the escape rate?

Step 1. The minimum action is $S_{\min}=\Delta V/k_BT = 5$ (the Einstein-relation cancellation makes it a pure state function).

Step 2. The exponential factor is $e^{-S_{\min}}=e^{-5}=\mathbf{6.74\times10^{-3}}$ , i.e. escape is suppressed by a factor of about 148 relative to the attempt frequency.

Step 3. In absolute terms, $\Delta V = 5\times(1.381\times10^{-23})(300) = \mathbf{2.07\times10^{-20}\ \text{J}} = 0.129\ \text{eV}$ — a modest barrier whose exponential nonetheless makes crossing rare, exactly why the least-action path dominates the statistics.

:::physicist[action functional|central object of the path-integral formulation|The action functional $S[x]$ assigns a single real number to an entire trajectory, and the Onsager–Machlup construction (Lars Onsager and Stefan Machlup, 1953) showed that for dissipative Langevin dynamics this number is a "least-dissipation" cost: $P[x]\propto e^{-S[x]}$ . It is the classical, real-weighted analogue of the quantum action in Feynman's path integral, with the diffusion constant playing the role of $\hbar$ . Its time-reversal asymmetry is the entropy produced, making the action functional the pivot between single-particle dynamics and the fluctuation theorems.]

Origins · 1 minhow this was discovered

History

1931

— Onsager states the reciprocal relations, forcing the linear-response matrix to be symmetric near equilibrium

1953

— Onsager and Machlup publish the "principle of least dissipation" for the path probability of Gaussian fluctuations (Phys. Rev. 91, 1505 and 1512)

1970

— Freidlin and Wentzell begin the rigorous large-deviation theory of small-noise diffusions, formalizing the instanton (Russian Math. Surveys)

1973

— Martin, Siggia and Rose introduce the response-field functional integral for classical dynamics

1976

— Janssen and De Dominicis recast Langevin dynamics as a field theory, generalizing the OM action to critical dynamics

Going further

Texts. Risken, The Fokker-Planck Equation, 2nd ed. (Springer), Ch. 4 · Kardar, Statistical Physics of Fields (dynamic field theory) · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity · Zinn-Justin, Quantum Field Theory and Critical Phenomena (path-integral methods for Langevin dynamics) · Wio, Path Integrals for Stochastic Processes (World Scientific).

Provenance. Action and Jacobian derived from the standard noise-integration procedure (Risken Ch. 4; Kardar; Sethna). Original result: Onsager L. & Machlup S., "Fluctuations and Irreversible Processes," Phys. Rev. 91, 1505 (1953). Verification reproduces (1) the OM action and the $\tfrac12 a'$ Jacobian via midpoint discretization; (2) the short-time kernel solving Fokker–Planck to $O(\Delta t)$ with moments $\langle\Delta x\rangle=a\Delta t$ , $\langle\Delta x^2\rangle=2D\Delta t$ ; (3) the instanton boundary term $\mu\Delta V/D=\Delta V/k_BT$ using the Einstein relation $D=\mu k_BT$ from the langevin-equation.

Related. Prerequisites: the langevin-equation · wiener-process-white-noise · ito-stratonovich-calculus · fokker-planck-equation. Leads to: kramers-escape-rate · fluctuation-dissipation-theorem · fluctuation-theorems (whose exact symmetry this action's time-reversal asymmetry generates) · ornstein-uhlenbeck-process.

Is the Onsager–Machlup action the same as the Feynman action?

No, though the structure is analogous. Feynman's action is real for a Lagrangian but enters the amplitude as $e^{iS/\hbar}$ — a complex phase that produces interference. The OM action enters as $e^{-S}$ , a real Euclidean weight with no interference: it is a probability, not an amplitude. The role of $\hbar$ is played by the diffusion constant $D$ , and the "classical limit" $\hbar\to0$ becomes the weak-noise limit $D\to0$ , in which the path integral is dominated by the least-action instanton rather than by a classical trajectory of an oscillating phase.

Why does the discretization convention (Itô vs Stratonovich) change the action?

Because the continuum action is shorthand for a discretized limit, and the Jacobian of the change of variables from noise increments to path increments depends on where the drift is evaluated within each step. The Stratonovich (midpoint) rule evaluates the drift at $(x_n+x_{n+1})/2$ , which makes $\partial\eta_n/\partial x_{n+1}$ pick up $-\tfrac12 a'$ ; the Itô (prepoint) rule evaluates at $x_n$ and gives a different Jacobian. The physics — the Fokker–Planck propagator — is convention-independent, but the split between the "bare" action and the Jacobian term is not, so the two must be kept consistent.

What is an instanton in this context?

It is the single trajectory that minimizes the OM action between two states — the "optimal fluctuation path." In the weak-noise limit the entire path integral collapses onto this one path (a saddle point of the functional), so its action $S_{\min}$ sets the exponential rate of the transition: $k\sim e^{-S_{\min}}$ . For a particle escaping a potential well the instanton is the time-reversed relaxation trajectory that climbs up the barrier, and its action reduces to the Arrhenius exponent $\Delta V/k_BT$ . Instantons are how large-deviation theory extracts rare-event rates without simulating the exponentially rare events themselves.

How exactly does the action connect to entropy production?

For overdamped Langevin dynamics the log-ratio of the forward path weight to the time-reversed path weight is $\ln(P[x]/P[\tilde x]) = \tfrac{1}{k_BT}\int F\circ\dot x\,\mathrm{d}t$ , which is the heat dumped to the bath divided by $T$ — the medium entropy production $\Sigma_{\text{med}}/k_B$ . The bulk "kinetic" part of the action is time-reversal symmetric and cancels in the ratio; only the drift–velocity cross term survives, and that term is the dissipation. Adding the change in system (Shannon) entropy gives the total, whose exponential average is unity — the integral fluctuation theorem owned by the fluctuation-theorems article.

Does the OM action apply to non-gradient (non-equilibrium) drifts?

Yes, and this is where it earns its keep. For a gradient drift $a=-\mu V'$ the escape exponent collapses to the equilibrium barrier $\Delta V/k_BT$ , but for a genuinely non-conservative flow (no potential $V$ exists) there is no Boltzmann weight and no shortcut. The instanton must then be found by actually minimizing the action — the geometric minimum-action method — and the resulting "quasipotential" is the Freidlin–Wentzell generalization of $\Delta V$ . This is precisely why numerical minimum-action methods are indispensable in driven, non-equilibrium systems.

Why is the noise-free path the most probable one?

Because the action penalizes exactly the mismatch $(\dot x - a)^2$ between the real velocity and the drift. A trajectory that follows the drift precisely has $\dot x = a$ , so the kinetic cost is zero and the weight $e^{-S}$ is maximal (up to the boundary and Jacobian terms). Any deviation forces the noise to supply the difference, and since large noise excursions are Gaussian-suppressed, deviating paths carry exponentially smaller weight. Formally, $\delta S=0$ yields the Euler–Lagrange equation, which for the OM Lagrangian is the deterministic (noise-free) equation of motion.

Can the action be negative, and what would that mean?

The bulk term $(\dot x - a)^2/4D$ is manifestly non-negative, and the physically meaningful quantity is always a difference of actions relative to the least-action path. The Jacobian term $\tfrac12 a'$ can be negative (e.g. $a'=-\gamma<0$ for a restoring drift), but it appears as a path-independent constant that is absorbed into normalization. So an individual action value is not itself a probability — only $e^{-S}$ relative to the normalization is — and the relative action of any path exceeds that of the instanton, which is what makes the instanton the most probable history.