Stochastic Thermodynamics

We take from the SM10 article stochastic-thermodynamics the continuous-trajectory framework: Sekimoto's stochastic energetics, which identifies heat as the Stratonovich integral of the bath force along a Langevin path, and Seifert's fluctuating system entropy $s(t)=-k_B\ln p(x(t),t)$ . That article owns the overdamped-Langevin formulation and the continuous heat integral. Here we do not re-derive the Stratonovich heat; instead we rebuild the whole framework on a discrete Markov network, where thermodynamic bookkeeping becomes a sum over jumps and every statement can be checked by direct simulation. On a jump process governed by a master equation, a single trajectory is a sequence of transitions between states. Local detailed balance fixes each transition-rate ratio, $\ln(W_{n\to m}/W_{m\to n})=(E_n-E_m)/k_BT$ , so the log-ratio of a forward jump to its reverse is the heat released to the bath divided by temperature. Summing along the path and adding the change in Seifert's system entropy gives the total trajectory entropy production $\Sigma=\Delta s_{\mathrm{sys}}+\sum_{\mathrm{jumps}}\ln(W_{n\to m}/W_{m\to n})$ — a well-defined random variable, positive on average but negative on rare trajectories. The centerpiece is the integral fluctuation theorem $\langle e^{-\Sigma/k_B}\rangle=1$ , which we verify by Monte-Carlo on a driven three-state ring: averaging $e^{-\Sigma}$ over jump trajectories returns 1 to within a percent, and Jensen's inequality then delivers the second law $\langle\Sigma\rangle\ge0$ . We close on the frontier hook of the field, the thermodynamic uncertainty relation $\mathrm{Var}(J)/\langle J\rangle^2\ge2k_B/\langle\Sigma\rangle$ , which trades precision for dissipation and is saturated by a biased Poisson walk. The jump-process view is the natural home for molecular motors, chemical-reaction networks, and single-molecule experiments.

Key takeaways

On a discrete Markov network, trajectory entropy production is a sum over jumps: $\Sigma=\Delta s_{\mathrm{sys}}+\sum_{\mathrm{jumps}}\ln(W_{n\to m}/W_{m\to n})$ , with each jump term equal to the bath entropy by local detailed balance.
Local detailed balance $\ln(W_{n\to m}/W_{m\to n})=(E_n-E_m)/k_BT$ makes the medium term $-Q/T$ — the microscopic-reversibility input that ties rates to thermodynamics.
The integral fluctuation theorem $\langle e^{-\Sigma/k_B}\rangle=1$ holds for any initial condition and driving; the second law $\langle\Sigma\rangle\ge0$ is its Jensen corollary.
Seifert's stochastic system entropy $s=-k_B\ln p$ makes $\Delta s_{\mathrm{sys}}$ a genuine trajectory quantity; on a nonequilibrium ring the steady state carries a persistent current and positive housekeeping entropy production.
The thermodynamic uncertainty relation $\mathrm{Var}(J)/\langle J\rangle^2\ge2k_B/\langle\Sigma\rangle$ bounds the precision of any current by its entropy cost — the field's modern frontier.
Key idea: On a jump process, entropy production is literally the log-ratio of a forward trajectory to its time-reverse; summing the jump-by-jump ratios (fixed by local detailed balance) plus the change in Seifert's system entropy gives a random variable $\Sigma$ that obeys the exact identity $\langle e^{-\Sigma/k_B}\rangle=1$ — the second law upgraded from an inequality to an equality.

The core insight

On a discrete Markov network, entropy production is the log-ratio of a forward jump trajectory to its time-reverse. Local detailed balance fixes each jump's contribution to be the bath heat over $T$ ; adding the change in Seifert's system entropy $s=-k_B\ln p$ gives the total $\Sigma$ . The exact integral fluctuation theorem $\langle e^{-\Sigma/k_B}\rangle=1$ holds arbitrarily far from equilibrium, and the second law $\langle\Sigma\rangle\ge0$ is just Jensen's inequality applied to it.

Where this lives today

The jump-process formulation is the working language of nonequilibrium biophysics. Molecular motors — kinesin, myosin, the $F_1$ -ATPase — are modeled as Markov networks whose transitions consume ATP; local detailed balance turns the measured stepping rates into entropy-production estimates, and the thermodynamic uncertainty relation sets a floor on how precisely a motor can step given its dissipation. Chemical-reaction-network thermodynamics uses exactly the same ledger to bound the accuracy of biochemical clocks and sensory adaptation. On the experimental side, single-molecule pulling with optical tweezers reconstructs equilibrium free energies from irreversible work via the fluctuation theorems, and feedback-controlled "information engines" extract work from thermal noise up to the Landauer bound $k_BT\ln2$ per bit — the second law generalized to include information.

The physical picture

Imagine a single enzyme or molecular motor hopping between a handful of conformational states, each hop coupled to a chemical reaction that releases or absorbs heat. There is no smooth trajectory to integrate — the state jumps discontinuously. Stochastic thermodynamics on a Markov network assigns a thermodynamic cost to each hop: the log-ratio of the rate for the hop to the rate for its reverse. Local detailed balance says this ratio is exactly the heat dumped into the surroundings, divided by temperature. Thermodynamics becomes a ledger with one line per jump.

Figure

A jump trajectory on a driven 3-state ring: each transition n→m carries entropy cost ln(W_{n→m}/W_{m→n}); the running Σ(t) climbs but dips below zero on a rare excursion

The deep statement is a symmetry between a trajectory and its time-reverse. A forward jump path is more probable than the reversed path by exactly $e^{\Sigma/k_B}$ , where $\Sigma$ is the entropy it produces. Average $e^{-\Sigma/k_B}$ over all forward paths and you are summing the probabilities of all reverse paths — which must total one. That is the integral fluctuation theorem, and the second law is just the observation that an average of $e^{-x}$ being one forces the average of $x$ to be non-negative.

Figure

Punchline: Monte-Carlo ⟨e^(−Σ)⟩ converges to 1 (the exact IFT) even though ⟨Σ⟩ grows — the rare Σ<0 outliers carry the average

Local detailed balance is the bridge from microscopic reversibility to thermodynamics. It is not full detailed balance — that holds only in equilibrium and gives zero current — but a constraint on the rates that survives out of equilibrium. A driven ring can sustain a net circulating current with no detailed balance among populations, yet each individual transition still obeys local detailed balance, and that is what lets us read the heat off the rate ratios and keep the fluctuation theorems exact.

The core equations

We build the jump-process ledger from three ingredients: Seifert's system entropy (recalled from SM10), local detailed balance, and the path-probability ratio. Each equation is stated with the physical reading it carries.

s(t)=-k_B\ln p(x(t),t),\qquad \Delta s_{\mathrm{sys}}=-k_B\ln\frac{p(x_f,t_f)}{p(x_0,t_0)} \qquad (1)

where: $p(x,t)$ — the instantaneous solution of the master equation evaluated at the realized state $x(t)$ ; $s(t)$ — Seifert's fluctuating trajectory entropy, whose ensemble average is the Gibbs–Shannon entropy. Recalled from the SM10 owner; here it supplies the system part of the total production. Local detailed balance ties each transition-rate ratio to the reservoir entropy, the single input that makes the ledger thermodynamically consistent.

\ln\frac{W_{n\to m}}{W_{m\to n}}=\frac{E_n-E_m}{k_BT}=\frac{\Delta s_{\mathrm{med}}^{(n\to m)}}{k_B} \qquad (2)

where: $W_{n\to m}$ — transition rate from state $n$ to $m$ at fixed protocol; $E_n$ — energy of state $n$ ; $\Delta s_{\mathrm{med}}=Q_{\mathrm{med}}/T$ — entropy released to the bath in the jump, with $Q_{\mathrm{med}}=E_n-E_m$ the heat dumped. The log-rate-ratio is the medium entropy production per jump. The total trajectory entropy production sums the system change and the jump-by-jump medium terms.

\Sigma=\Delta s_{\mathrm{sys}}+\sum_{\mathrm{jumps}}k_B\ln\frac{W_{n\to m}}{W_{m\to n}} \qquad (3)

where: the sum runs over every jump along the path; the second term is the total medium entropy $\Delta s_{\mathrm{med}}$ . $\Sigma$ is a random variable — positive on average, negative on rare trajectories. The path-probability ratio identifies $\Sigma$ as the log-ratio of a forward path to its time-reverse, the origin of every fluctuation theorem.

\frac{P_F[\text{path}]}{P_R[\text{reverse}]}=\exp\!\left(\frac{\Sigma}{k_B}\right) \qquad (4)

where: $P_F$ — probability of the forward jump path (initial distribution times the product of forward rates); $P_R$ — probability of the reversed path under the reversed protocol. A path that produces more entropy is exponentially more probable than its reverse. Summing (4) over all paths gives the integral fluctuation theorem and, by Jensen, the second law.

\langle e^{-\Sigma/k_B}\rangle=1\;\Longrightarrow\;\langle\Sigma\rangle\ge0 \qquad (5)

where: $\langle\cdot\rangle$ — average over forward trajectories with any initial condition and any driving. The equality is exact; the inequality is the second law it contains. The thermodynamic uncertainty relation bounds any current's relative precision by its dissipation — the modern frontier of the field.

\frac{\mathrm{Var}(J)}{\langle J\rangle^2}\ge\frac{2k_B}{\langle\Sigma\rangle} \qquad (6)

where: $J$ — any time-integrated current (displacement, particle number, cycles); $\langle\Sigma\rangle$ — total entropy production over the observation window. A precise current is expensive: halving its relative variance at least doubles the dissipation.

Derivation ATrajectory entropy production on a Markov network

This derivation constructs the total entropy production for a discrete jump process from first principles, showing that it equals the log-ratio of a forward path to its time-reverse and that its jump-by-jump structure is fixed entirely by local detailed balance. We work on a Markov network with time-dependent rates $W_{n\to m}(\lambda(t))$ driven by a protocol $\lambda(t)$ , and we recall Seifert's system entropy from the SM10 owner rather than re-deriving it. The goal is a closed expression $\Sigma=\Delta s_{\mathrm{sys}}+\Delta s_{\mathrm{med}}$ whose exponential average we can then prove is one.

Step 1. Write the forward path probability as the initial distribution times the product of transition rates along the realized jump sequence, $P_F=p(x_0)\prod_j W_{x_j\to x_{j+1}}$ , and the reverse path probability as $P_R=p^R(x_f)\prod_j W_{x_{j+1}\to x_j}$ under the reversed protocol. Why: a Markov trajectory is memoryless, so its probability factorizes into the starting weight and independent per-jump factors — the reverse path visits the same states in the opposite order.

Step 2. Take the log-ratio: $\ln(P_F/P_R)=\ln[p(x_0)/p^R(x_f)]+\sum_j\ln(W_{x_j\to x_{j+1}}/W_{x_{j+1}\to x_j})$ . The first term is $\Delta s_{\mathrm{sys}}/k_B$ (Seifert), the second the sum of per-jump rate-ratios. Why: the initial/final weights carry the system entropy change; the rate ratios carry everything the bath did.

Step 3. Apply local detailed balance, equation (2): each rate ratio equals $(E_n-E_m)/k_BT$ , the heat released to the medium over $k_BT$ . Summing, the jump terms become $\Delta s_{\mathrm{med}}/k_B=-Q/k_BT$ where $Q$ is the total heat absorbed by the system. Why: this is the microscopic-reversibility input — without it the rate ratios would be arbitrary and no thermodynamic interpretation would exist.

Step 4. Combine: $\ln(P_F/P_R)=(\Delta s_{\mathrm{sys}}+\Delta s_{\mathrm{med}})/k_B\equiv\Sigma/k_B$ . The total entropy production is exactly the log-ratio of forward to reverse path probability. Why: this single identity is the seed of every fluctuation theorem — detailed, integral, Jarzynski, Crooks — which differ only in what boundary condition they impose on the endpoints.

Step 5. For a nonequilibrium steady state (a driven ring), the system entropy change over a cycle vanishes but the medium term does not: the ring sustains a persistent probability current and a positive housekeeping entropy production per cycle, equal to the ring affinity $\ln[(W_{12}W_{23}W_{31})/(W_{21}W_{32}W_{13})]$ . Why: detailed balance would require this affinity to vanish; a nonzero affinity is precisely what drives the steady current and the constant dissipation. *Verified symbolically and by Monte-Carlo in verify_aq-sm-stochastic-thermodynamics-sm11.py: local detailed balance gives the medium term $Q/T$ ; the ring affinity is $\ln 8$ for a 2:1 bias and vanishes at detailed balance; the Schnakenberg integrand $(a-b)\ln(a/b)\ge0$ ; and $\langle\Sigma\rangle>0$ on the driven ring.*

Go deeper · Derivation B · 1 minThe integral fluctuation theorem by trajectory reweighting

Derivation BThe integral fluctuation theorem by trajectory reweighting

Here we prove the exact identity $\langle e^{-\Sigma/k_B}\rangle=1$ and extract the second law from it, using only the path-ratio result of the previous derivation. The proof is a one-line reweighting once $\Sigma$ is identified with $\ln(P_F/P_R)$ , and it holds for any initial distribution and any driving protocol — the feature that makes it so much stronger than a near-equilibrium statement. We then verify it numerically on the driven three-state ring, where naive intuition (that $\Sigma$ only grows) coexists with the exact conservation law.

Step 1. Write the average as a sum over forward paths, $\langle e^{-\Sigma/k_B}\rangle=\sum_{\mathrm{paths}}P_F\,e^{-\Sigma/k_B}$ . Why: the exponential average is an expectation over the forward ensemble, weighted by each path's forward probability.

Step 2. Substitute $e^{-\Sigma/k_B}=P_R/P_F$ from equation (4). The $P_F$ factors cancel: $\langle e^{-\Sigma/k_B}\rangle=\sum_{\mathrm{paths}}P_R$ . Why: the reweighting converts a forward average into a bare sum of reverse-path probabilities — the essential algebraic move.

Step 3. The reverse-path probabilities are normalized, $\sum_{\mathrm{paths}}P_R=1$ , so $\langle e^{-\Sigma/k_B}\rangle=1$ exactly. Why: $P_R$ is a genuine probability distribution over the reversed trajectories; summing it must give one, independent of how far from equilibrium the driving is.

Step 4. Apply Jensen's inequality. Since $e^{-x}$ is convex, $\langle e^{-\Sigma/k_B}\rangle\ge e^{-\langle\Sigma\rangle/k_B}$ , and the left side equals one, so $e^{-\langle\Sigma\rangle/k_B}\le1$ , giving $\langle\Sigma\rangle\ge0$ . Why: the second law is not an extra postulate — it is a mathematical consequence of the exact equality, which is why individual trajectories may have $\Sigma<0$ while the mean cannot.

Step 5. Verify by simulation on the driven three-state ring: Gillespie-sample jump trajectories from the stationary distribution, accumulate $\Sigma=\Delta s_{\mathrm{sys}}+\sum\ln(W/W')$ , and average $e^{-\Sigma}$ . Why: the rare $\Sigma<0$ excursions, exponentially weighted, exactly compensate the typical dissipative paths — the numeric convergence to one is the tangible content of the theorem. *Verified in verify_aq-sm-stochastic-thermodynamics-sm11.py: the Monte-Carlo estimate $\langle e^{-\Sigma}\rangle=0.984$ (target 1, within 5%), $\langle\Sigma\rangle=1.36>0$ on the driven ring, and the TUR product $\mathrm{Var}(J)/\langle J\rangle^2\cdot\langle\Sigma\rangle\ge2$ for the biased Poisson walk (saturating as the bias vanishes).*

Limiting cases

Case	Result	Meaning
Detailed balance (equilibrium)	$W_{n\to m}p^{\mathrm{eq}}_n=W_{m\to n}p^{\mathrm{eq}}_m$ , $\langle\Sigma\rangle=0$	Every jump balanced; zero current, zero entropy production, no time-reversal asymmetry.
Nonequilibrium steady state	$\dot\Sigma_{\mathrm{hk}}=k_B\times$ ring affinity $>0$	Persistent current, constant housekeeping dissipation even though populations are stationary.
Macroscopic limit	$\Sigma\sim Nk_B$ ; $P(\Sigma<0)\sim e^{-N}$	Violating trajectories are exponentially rare; the deterministic second law $\langle\Sigma\rangle\ge0$ becomes effectively strict.
Near equilibrium (Gaussian)	$\sigma_\Sigma^2=2k_B\langle\Sigma\rangle$	Gaussian entropy production automatically satisfies the fluctuation theorem; the relation is the fluctuation–dissipation theorem in disguise.

Common trapLocal detailed balance is not detailed balance

The most common confusion is between detailed balance (an equilibrium property of the populations: $W_{n\to m}p_n=W_{m\to n}p_m$ , giving zero current) and local detailed balance (a constraint on the rates alone: $\ln(W_{n\to m}/W_{m\to n})=(E_n-E_m)/k_BT$ , valid arbitrarily far from equilibrium). A driven system violates the former while always obeying the latter — that is exactly what lets a nonequilibrium steady state carry a current and produce entropy while the jump-by-jump heat is still well defined. Also: work and heat are path functionals, not state functions; only the energy change $\Delta E$ is exact.

Worked examples

Example 1 — medium entropy per jump on a two-state driven system (a real number).

Step 1. Take a two-state system with forward rate $W_{\uparrow}=2\,\mathrm{s}^{-1}$ and backward rate $W_{\downarrow}=1\,\mathrm{s}^{-1}$ at fixed protocol. Local detailed balance says the medium entropy produced in one up-jump is $\Delta s_{\mathrm{med}}=k_B\ln(W_{\uparrow}/W_{\downarrow})$ .

Step 2. Compute $\Delta s_{\mathrm{med}}=k_B\ln 2=\mathbf{0.693\,k_B}$ per up-jump, and $-0.693\,k_B$ for the reverse down-jump. Over one full forward cycle of a symmetric two-state loop driven at this ratio, the net medium entropy is the cycle affinity.

Step 3. On the three-state ring with each forward-to-backward ratio $2{:}1$ , the affinity per forward cycle is $\ln[(2\cdot2\cdot2)/(1\cdot1\cdot1)]=\ln 8=\mathbf{2.079\,k_B}$ — the guaranteed positive entropy production driving the persistent current. At equal rates the affinity vanishes and detailed balance is restored.

Example 2 — the integral fluctuation theorem holds despite growing dissipation.

Step 1. Simulate jump trajectories on the driven ring with forward rates $(2.0,1.5,1.8)$ and backward rates $(1.0,1.0,0.9)$ , starting from the stationary distribution. For each trajectory accumulate $\Sigma=\Delta s_{\mathrm{sys}}+\sum_{\mathrm{jumps}}\ln(W/W')$ .

Step 2. The mean entropy production is positive and grows with the number of jumps: over eight jumps, $\langle\Sigma\rangle\approx\mathbf{1.36\,k_B}$ . Naively one might expect $e^{-\Sigma}$ to be tiny.

Step 3. Yet the exponential average is $\langle e^{-\Sigma}\rangle\approx\mathbf{0.98}$ , converging to the exact value 1. The rare trajectories with $\Sigma<0$ (thermal fluctuations transiently "un-producing" entropy), weighted by $e^{-\Sigma}$ , exactly compensate the typical dissipative paths. This is the integral fluctuation theorem made concrete — the second law $\langle\Sigma\rangle\ge0$ is its Jensen shadow.

Origins · 1 minudo seifert and how this was discovered

udo seifert

b. 1958, German physicist (University of Stuttgart) · Udo Seifert reformulated nonequilibrium thermodynamics at the level of single stochastic trajectories. In a 2005 Physical Review Letter he introduced the fluctuating system entropy

s(t)=-k_B\ln p(x(t),t)

, making the entropy of an individual trajectory a well-defined quantity and proving the integral fluctuation theorem

\langle e^{-\Delta s_{\mathrm{tot}}/k_B}\rangle=1

for total entropy production under arbitrary driving. His 2012 review in Reports on Progress in Physics unified stochastic energetics, fluctuation theorems, and molecular machines into the modern framework of stochastic thermodynamics, and his later work with Andre Barato established the thermodynamic uncertainty relation.

History

1931

— Onsager formulates the regression hypothesis, the near-equilibrium ancestor of trajectory thermodynamics

1976

— Schnakenberg gives the network entropy-production formula for master equations (Rev. Mod. Phys. 48, 571)

1997

— Jarzynski proves the nonequilibrium work equality; Sekimoto formulates stochastic energetics for Langevin systems

1998

— Sekimoto defines trajectory heat as the Stratonovich integral of the bath force

2005

— Seifert introduces the fluctuating system entropy

s=-k_B\ln p

and proves the integral fluctuation theorem for total entropy (Phys. Rev. Lett. 95, 040602)

2010

— Toyabe et al. realize an information-to-energy conversion (Maxwell's demon) obeying the generalized second law

2012

— Seifert's review consolidates stochastic thermodynamics (Rep. Prog. Phys. 75, 126001)

2015

— Barato and Seifert establish the thermodynamic uncertainty relation (Phys. Rev. Lett. 114, 158101)

Going further

References. U. Seifert, 'Entropy production along a stochastic trajectory and an integral fluctuation theorem', Phys. Rev. Lett. 95, 040602 (2005) · U. Seifert, 'Stochastic thermodynamics, fluctuation theorems and molecular machines', Rep. Prog. Phys. 75, 126001 (2012) · K. Sekimoto, Stochastic Energetics (Springer LNP 799, 2010) · J. Schnakenberg, Rev. Mod. Phys. 48, 571 (1976) · A. C. Barato & U. Seifert, Phys. Rev. Lett. 114, 158101 (2015) · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, 2nd ed.

Provenance. The jump-process trajectory entropy, the path-ratio identity, and the integral fluctuation theorem follow Seifert (2005, 2012) and Schnakenberg (1976); the thermodynamic uncertainty relation is from Barato & Seifert (2015). The SM10 companion stochastic-thermodynamics owns Sekimoto's continuous Stratonovich heat and Seifert's continuous system entropy; this article recalls those and rebuilds the framework on a discrete Markov network. A verifier should recompute: (1) local detailed balance $\Rightarrow$ medium entropy $Q/T$ ; (2) the ring affinity ( $\ln 8$ for a 2:1 bias) and its vanishing at detailed balance; (3) the Monte-Carlo integral fluctuation theorem $\langle e^{-\Sigma}\rangle\to1$ ; (4) the TUR product $\ge2$ on a biased Poisson walk. All checks are in verify_aq-sm-stochastic-thermodynamics-sm11.py.

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

How is the SM11 jump-process treatment different from the SM10 stochastic-thermodynamics article?

The SM10 article stochastic-thermodynamics owns the continuous formulation: Sekimoto's heat as a Stratonovich integral of the bath force along a Langevin path, and Seifert's continuous system entropy from the Fokker–Planck solution. This SM11 article rebuilds the same framework on a discrete Markov network, where a trajectory is a sequence of jumps and the entropy production becomes an explicit sum $\Delta s_{\mathrm{sys}}+\sum_{\mathrm{jumps}}\ln(W/W')$ . The discrete view is the natural home for chemical reactions and molecular motors, and it makes the integral fluctuation theorem directly checkable by Gillespie simulation — which is exactly what the verify script does.

What is the difference between local detailed balance and ordinary detailed balance?

Ordinary detailed balance is an equilibrium property of the populations: $W_{n\to m}p_n=W_{m\to n}p_m$ , which forces zero net current in every transition. Local detailed balance is weaker and always holds: it constrains only the rates, $\ln(W_{n\to m}/W_{m\to n})=(E_n-E_m)/k_BT$ , and encodes the microscopic reversibility of the system–bath coupling. A driven system violates ordinary detailed balance (it carries currents) but obeys local detailed balance at every jump, which is precisely what lets us assign a well-defined heat and keep the fluctuation theorems exact out of equilibrium.

Why does the integral fluctuation theorem hold arbitrarily far from equilibrium?

Because its proof uses only the path-ratio identity $e^{-\Sigma/k_B}=P_R/P_F$ and the normalization of the reverse-path distribution — neither of which assumes small driving. Reweighting the forward average by $e^{-\Sigma}$ converts it into a bare sum of reverse-path probabilities, which is one. No linearization, no near-equilibrium expansion, no restriction on the protocol speed enters. This is what distinguishes fluctuation theorems from older near-equilibrium results like the fluctuation–dissipation theorem, which they contain as a limiting case.

What exactly is Seifert's stochastic system entropy, and why is it needed?

Seifert's stochastic entropy is $s(t)=-k_B\ln p(x(t),t)$ : you take the ensemble solution $p(x,t)$ of the master (or Fokker–Planck) equation and evaluate it at the state the individual trajectory happens to occupy. Its ensemble average is the ordinary Gibbs–Shannon entropy, but as a trajectory quantity it fluctuates. It is needed because the total entropy production is the sum of the medium part (heat over temperature, from the jump rates) and the system part (the change in $s$ ); without the system term the integral fluctuation theorem would not close to exactly one for arbitrary initial conditions.

What is the thermodynamic uncertainty relation telling us physically?

It says precision is not free. For any time-integrated current $J$ — a motor's displacement, the number of catalyzed reactions, cycles of a clock — the relative variance is bounded below by the dissipation: $\mathrm{Var}(J)/\langle J\rangle^2\ge2k_B/\langle\Sigma\rangle$ . To make a current twice as precise (halve its relative variance) you must at least double the entropy you produce. It sets a fundamental thermodynamic cost on biological accuracy, and it is saturated by a biased Poisson process in the near-equilibrium limit, where the bound becomes an equality.

Do fluctuation theorems violate the second law when a trajectory has Σ<0?

No. The second law is a statement about the average entropy production, $\langle\Sigma\rangle\ge0$ , and that is guaranteed by the integral fluctuation theorem via Jensen's inequality. Individual trajectories may have $\Sigma<0$ — a colloidal bead momentarily extracting heat to do work, say — but these are exponentially suppressed relative to their entropy-producing time-reverse, by the factor $e^{-\Sigma/k_B}$ . In the macroscopic limit $\Sigma\sim Nk_B$ and such trajectories have probability $\sim e^{-N}$ , so the deterministic second law re-emerges.

How does this connect to Jarzynski and Crooks?

Jarzynski's equality and Crooks' theorem are special cases of the same path-ratio identity with particular boundary conditions. If the system starts (and the reverse ends) in equilibrium, the endpoint system-entropy terms combine into $\beta\Delta F$ , and the total entropy production reduces to the dissipated work $\beta(W-\Delta F)$ . The integral fluctuation theorem then becomes Jarzynski's $\langle e^{-\beta W}\rangle=e^{-\beta\Delta F}$ , and the detailed version becomes Crooks' $P_F(W)/P_R(-W)=e^{\beta(W-\Delta F)}$ . The dedicated SM11 articles jarzynski-equality and crooks-fluctuation-theorem develop these in full.