NOTE-420: Stochastic thermodynamics, fluctuation theorems and molecular machines

Contribution

A systematic review, with original technical parts, of the framework that extends work, heat and entropy production to single fluctuating trajectories. Its distinctive contributions as a review:

• the unification of all known fluctuation theorems for stochastic dynamics as cases of one “master” theorem (§4);
• a uniform treatment of continuous (Langevin) and discrete (master equation) dynamics (§§2, 6);
• a cycle-based theory of efficiency and efficiency at maximum power for isothermal machines (§10).

Key insight

Thermodynamics can be done on one trajectory once the system’s entropy is also made a trajectory quantity, s(τ) = −ln p(x(τ),τ), evaluated on the ensemble the trajectory came from (Eq. 28). Then medium entropy (heat/T) plus system entropy is a total entropy production whose statistics obey exact symmetries. Those symmetries come from one fact: the log-ratio of a path’s probability to its time-reversed probability is the heat dissipated along it (Eq. 73, “a deep physical interpretation”, p. 24).

Assumptions

• Time-scale separation. The observed degrees of freedom are slow. The unobserved ones, the bath and fast internal modes, are always in equilibrium constrained by the slow ones. Temperature is that of the embedding medium and stays well defined (§§1.2, 12).
• Markovian dynamics. Langevin with Stratonovich discretisation, or a master equation on a discrete state set (§§2.1, 6.1).
• Noise unaffected by driving. “The strength of the noise is not affected by the presence of a time-dependent force” (p. 11). The thermodynamic reading of every theorem “essentially rests on” this (§3.2.1). It “could be violated in experiments” (§5.1).
• Local detailed balance on rates or noise, for thermodynamic consistency (§§9.4.5, 12).
• Weak, state-independent coupling to the bath for the simple identification of heat. Biomolecules need an intrinsic-entropy term (§§2.3.3, 9.2).
• Coherences ignored for quantum systems. Then the open quantum dynamics is a classical stochastic one (§1.3).

Key results

• Trajectory first law and entropy (§§2.3–2.5).
  • Work increment: đw = (∂V/∂λ)dλ + f dx.
  • First law along a trajectory: w = q + ΔV (Eq. 19).
  • Medium entropy: Δs_m = q/T (Eq. 27).
  • Average total entropy production rate: Ṡ_tot = ⟨ν²⟩/D ≥ 0, with equality only in equilibrium (Eq. 35).
• Classification of fluctuation theorems (§3.1).
  • Integral: ⟨e^{−Ω}⟩ = 1, which forces negative-Ω trajectories.
  • Detailed: p(−Ω)/p(Ω) = e
  • Crooks-type: p†(−Ω) = p(Ω)e^{−Ω}, comparing two processes.
  • “Violations” of the second law are of order 1 in Ω and so exponentially rare in system size N (pp. 19–20).
• The named theorems (§§3.2–3.3).
  • Jarzynski: ⟨e^{−w/T}⟩ = e^{−ΔF/T} (Eq. 58).
  • Bochkov–Kuzovlev, which is distinct from Jarzynski for a different definition of work (Eq. 60).
  • Crooks: p̃(−w)/p(w) = e^{−(w−ΔF)/T} (Eq. 61).
  • Total entropy production: ⟨e^{−Δs_tot}⟩ = 1 for any initial state, driving and duration (Eq. 64).
  • Steady-state theorem, exact for finite times once system entropy is included (Eq. 65).
  • Hatano–Sasa for transitions between steady states (Eqs. 66–67).
• Master theorem (§4). Choose a conjugate dynamics: reversed, dual or dual-reversed. Then for functionals of definite parity the log-ratio functional R satisfies a generalised theorem (Eqs. 78–79). All of the above follow by choosing the conjugate, the initial distributions and the functional (§§4.3–4.4).
• Master-equation version (§6). The same results hold for discrete states. Work there is a formal quantity unless detailed balance gives an energy (§6.3.1). A footnote to §6.1 mentions approaches that derive rates from current constraints and points to Polettini (2011) “for a relation to the minimum entropy production principle”. There is also a current fluctuation theorem from the Schnakenberg cycle decomposition (§6.4).
• Optimal protocols (§7.1). Minimal-work protocols jump at the start and end. Total entropy production for a transition in time t is bounded, ΔS_tot ≥ C/t.
• Time’s arrow (§7.2), in units with T = 1:
  • ⟨w⟩ − ΔF = D[p(w−ΔF) ‖ p†(−(w−ΔF))] (Eq. 148);
  • ⟨w⟩ − ΔF ≥ D[p(x(t₁)) ‖ p†(x†(t₁))] (Eq. 149), which is an equality for Hamiltonian dynamics;
  • ⟨w⟩ − ΔF ≥ D[p(x(t)) ‖ p_eq(x(t))] (Eq. 150, Vaikuntanathan–Jarzynski).
• Feedback (§7.3).
  • Sagawa–Ueda: ⟨e^{−(w−ΔF+I)}⟩ = 1 (Eq. 154). So extracted work is at most −ΔF + T·I (Eq. 155).
  • A NESS analogue (Eqs. 156–157) and a stronger Hatano–Sasa-type bound, ⟨Δs_tot⟩ ≥ ⟨Δs_hk⟩ − I (Eq. 160).
  • The second law is restored by Landauer’s erasure cost. That cost is set aside “first” in these analyses (p. 46).
  • An experiment found that erasure heat approaches the Landauer bound for long cycles (§7.3.6).
• NESS fluctuation–dissipation theorem (§8). The response function is R_A = −⟨A ∂_h ṡ⟩_s, which splits into medium and total entropy production terms (Eqs. 191–192).
• Machines (§§10–11).
  • Isothermal efficiency is at most 1.
  • In linear response, efficiency at maximum power is at most 1/2. The cycle representation gives Onsager symmetry automatically (§10.4).
  • For unicyclic motors, efficiency at maximum power can exceed 1/2 beyond linear response (§10.5).

Claims

id	claim	strength	support
C1	All known fluctuation theorems for stochastic dynamics follow from one master theorem with a suitable conjugate dynamics	strong for the theorems listed in §§3–4	§4 derivations; the “all known” is the author’s survey judgement
C2	The integral theorem for total entropy production holds for any initial state, driving and duration	strong	Eq. 64 via Eq. 80
C3	The integral theorems are not a proof of the second law, because a Langevin or master-equation dynamics builds irreversibility in	strong (conceptual)	§3.3.1, §12
C4	Dissipated work bounds the relative entropy between forward and reverse distributions, at the work level and at any intermediate time	strong	Eqs. 148–150; reported from Kawai–Parrondo–van den Broeck and Vaikuntanathan–Jarzynski
C5	With feedback, extracted work can exceed −ΔF by at most T·I, and erasure restores the second law	strong for the bound (Eqs. 154–155); the erasure balance is asserted, not modelled	§7.3; the author says a full integration of measurement and erasure is “still missing” (§12)
C6	In a NESS, the response of any observable is a correlation with the h-derivative of stochastic entropy	strong	§8.2, Eqs. 191–192
C7	Efficiency at maximum power is at most 1/2 in linear response but can exceed it beyond	strong for linear response (§10.4); shown for unicyclic models beyond it (§10.5)	the result depends on the choice of variational parameters, as the author stresses

Concepts

• Stochastic entropy s(τ) = −ln p(x(τ),τ). It depends on the ensemble the trajectory is drawn from, not on the trajectory alone (§2.4).
• Housekeeping and excess heat: the heat needed to maintain a NESS, and the heat from changing the control parameter (§2.3.2).
• Conjugate dynamics: reversed, dual or dual-reversed. A device for deriving fluctuation theorems. The dual dynamics has the same stationary state and reversed currents (§4.1).
• Efficiency at maximum power (EMP): efficiency at the parameter values that maximise output power. Its value depends on which parameters may vary (§10.3.2).

Connections

The review synthesises Evans–Cohen–Morriss, Gallavotti–Cohen, Kurchan and Lebowitz–Spohn for the steady-state theorem; Jarzynski (LIT-517) and Crooks (LIT-522) for the work relations; Sekimoto’s stochastic energetics; Maes on the time-antisymmetric action; and the author’s 2005 stochastic entropy. It names the Brussels school (van den Broeck; Mou, Luo and Nicolis 1986) as the earlier, ensemble-level use of “stochastic thermodynamics” for chemical systems (§1.1). The quantum thermodynamics review LIT-010 covers the quantum side this one leaves out (§1.3).

Bearing on the record

• THEORY-026: framework, not source.
  • LIT-327 defines dissipation as ⟨W⟩ − ΔF_neq, with F_neq = ⟨E⟩ + k_BT⟨ln p⟩ (LIT-327, Eqs. 8–9). For one protocol started in equilibrium this equals ⟨W⟩ − ΔF − k_BT·D_KL(p_τ‖p_eq). Eq. 150 here says that is ≥ 0. So the protocol-level nonnegativity of LIT-327’s dissipation is a reported result of this framework.
  • That bound is for continuous Langevin dynamics, stated for the discrete case by analogy (§7.2). It is a protocol total, not a per-step quantity.
  • It does not settle what NOTE-295 found open and THEORY-026’s promote_when asks for: the sign of the per-step nostalgia I[s_t;x_t] − I[s_t;x_{t+1}] under a non-Markov drive, with free energies conditioned on the current signal value only.
  • Recommendation: THEORY-026 need not cite this review as a source. It could name it in Connections as the framework, citing Eq. 150 as where protocol-level nonnegativity comes from.
• THEORY-030 and Landauer. §7.3 treats measurement and feedback first with the erasure cost set aside, then restores the second law through Landauer’s principle. It reports an experiment in which the mean erasure heat saturates the Landauer bound in the limit of long cycles. The review does not separate logically irreversible steps from others, which is THEORY-030’s distinction. It is consistent with THEORY-030 and adds the experiment, not an argument. The record’s Landauer 1961 entry (LIT-328) is still unread.
• LIT-308. Goldt and Seifert’s learning bound is a subsystem second-law result in this framework. This review predates it and gives the information-flow machinery only in its feedback form (§7.3).
• Prigogine’s minimum principle. The review does not assess it. It points to one paper on it (Polettini 2011) in a footnote, and its own steady-state results are exact symmetries of fluctuations, not extremum principles for the mean. That silence is itself informative: the framework that replaced the Brussels programme’s ensemble thermodynamics for small systems does not use a variational selection principle for steady states.
• Dissipative structures and life. §9 treats enzymes and molecular motors as nonequilibrium steady states driven by chemical potential differences, the small-scale version of the “life as dissipation” picture. It makes no claim about organisms or about selection of structures.
• No ML instruction. The anthology need not hold it.

Limitations

• A review. Its support for any reported result is the cited paper’s. Its own original material is in §4 (the unifying derivation) and §§9.2 and 10 (biopolymers, machines), as the author says (§1.4).
• Markovian and classical throughout. §12.2 surveys non-Markovian and coarse-grained cases as open: entropy production with hidden slow variables “is then difficult if not impossible” to identify.
• Measurement and erasure. Their integration into a full machine energy balance is “still missing” (§12).
• Experiments are scarce for discrete-state systems (§§6.3.2, 9.4.8).

Open questions

The author’s own (§12):

• Are there universality classes for the distributions of work, heat and entropy production?
• When can the additive NESS response–correlation relation be recast with an effective temperature?
• How should entropy production be identified under coarse-graining and memory?
• Is there a zeroth law for coupled nonequilibrium steady states (§12.3)?

Corrections

• none to a seeded skim (there was no seed)
• Title. The arXiv title has a serial comma (“fluctuation theorems, and molecular machines”). The Reports on Progress in Physics title does not. The LIT records the published title.
• Units in §7.2. Eqs. 148–150 are written without T, in effect T = 1. Elsewhere the review keeps T explicit and sets only k_B = 1. Read with T, Eq. 150 is ⟨w⟩ − ΔF ≥ T·D[p‖p_eq].