Contribution
A unified framework that takes stochastic thermodynamics (ST) of occupation-number jump processes to a large-size limit V → ∞. The scaling conditions (extensive Massieu potential, rates ∝ V, nonconservative forces O(1); eqs. 26–29) yield at once a deterministic rate equation, an extensive thermodynamics, and a large-deviation theory of fluctuations, which the authors show to preserve the fluctuation theorem. Equilibrium thermodynamics, linear irreversible thermodynamics, Onsager–Machlup and macroscopic fluctuation theory are recovered as limits (Fig. 9). Far from equilibrium the free energy is replaced by the quasi-potential. The bounds of §IV.H and the emergent ST of §VI are the review’s own new results (§VIII.B.6).
Key insight
Far from equilibrium the potential that orders the macroscopic dynamics is dynamical, not thermodynamic: it is the rate function of the stationary distribution, I_ss, computed from the rates, a Lyapunov function of the deterministic flow (eq. 48), and equal to the Massieu potential only at detailed balance (eqs. 36–38). Thermodynamics still constrains it: near detailed balance I_ss = ϕ plus the nonconservative dissipation along the relaxation path (eq. 120), and in general the barrier ΔI_ss that sets an attractor’s lifetime is bracketed by the entropy produced in relaxing into it and escaping from it (eq. 160).
Assumptions
- Mesoscopic states with equilibrated interiors: rates obey local detailed balance, ln[R_ρ(n)/R_{−ρ}(n + Δ_ρ)] = −Φ(n + Δ_ρ) + Φ(n) + a_ρ (eq. 5); every transition has its reverse (eq. 1).
- Extensivity: Φ/V → ϕ(c), R_ρ/V → r_ρ(c), a_ρ = O(1) (eqs. 26–28).
- Isolated, nondegenerate fixed points for most results; limit cycles only in passing, chaos excluded (§§III.C, VIII.A).
- Thermal noise only; quenched disorder and extrinsic noise excluded (§VIII.A).
Key results (parts read)
Deterministic level (§III)
- Rate function I(c, t) solves a Hamilton–Jacobi equation with H(c, π) = Σ_ρ (e^{Δ_ρ·π} − 1) r_ρ(c) (eqs. 32–33); I_ss solves H(c, ∂I_ss) = 0 (eq. 34). The deterministic flow F = Σ Δ_ρ r_ρ (eq. 46) decreases I_ss (eq. 48).
- Keizer’s paradox from the spectrum (§III.C). Metastability is eigenvalues of the generator going to zero as V → ∞; the deterministic limit is multistable while the mesoscopic stationary law is unique, because “the long time limit and the macroscopic limit do not commute”. For nondetailed-balance dynamics I_ss is only piecewise differentiable, built from local quasi-potentials plus attractor weights fixed by a jump process between attractors (eq. 55).
- Macroscopic Shannon entropy vanishes (eq. 57); the macroscopic system entropy is the internal entropy only (eq. 58).
- Drift decomposition F = v_ss − M·∂I_ss, with v_ss orthogonal to ∂I_ss (eqs. 71–76): descent in the quasi-potential plus circulation on its level sets. At detailed balance v_ss = 0.
- Minimum entropy production is recovered only near detailed balance with linear dynamics (eqs. 101–102), not in general.
Fluctuations and escape (§IV)
- Instantons from an attractor are time-reversed trajectories of the dual (adjoint) dynamics (eq. 109); at detailed balance they are the reversed relaxations (eq. 154).
- Emergent second law ⟨σ̇⟩ ≥ −dₜI_ss ≥ 0 along relaxation (eq. 119); tight to first order in a_ρ, where I_ss = ϕ + σ_nc along the detailed-balanced relaxation (eqs. 120–122).
- Transition rates κ_ν ≍ e^{−V[I_ss(x_ν) − I_ss(x*_γ)]} (eq. 157); Arrhenius at detailed balance (eq. 158). Near detailed balance the inter-attractor jumps inherit local detailed balance (eq. 159).
- Bounds (eq. 160). −σ_{ν→γ} ≤ lim V⁻¹ ln κ_ν ≤ σ_{γ→ν}, the entropy produced along relaxation from the saddle and along the instanton. The lower bound is the emergent second law integrated; the upper one uses non-negativity of the adiabatic entropy production conditioned on instantons (appendix A, by Jensen). They coincide at detailed balance. Appendix A concedes that between fixed points these integrals diverge unless detailed balance holds, and the endpoints must be taken on balls of radius O(V^{−1/2}) around the fixed points. For the Schlögl model “σ_{γ→ν} and |σ_{ν→γ}| are very large, thus making the bound (160) very loose” (§VII.A.1).
- The Langevin truncation fails (§IV.D–E). Truncating the Kramers–Moyal expansion at second order (the chemical Fokker–Planck or nonlinear Langevin equation) gives dI_NLE = −F/D instead of the exact dI_ss = ln(Σr₋/Σr₊) in one dimension (eqs. 137–138). They agree only at fixed points, so escape rates are wrong by an exponential factor; and the Langevin entropy production vanishes at every fixed point (eq. 145), missing the dissipation that sustains an attractor.
Emergent ST between attractors (§VI)
- At long times the system is a jump process on attractors (eq. 217) with three departures from ordinary ST (eq. 224): each attractor has its own steady dissipation ⟨σ̇_γ⟩ > 0 “needed to sustain the attractors” (eq. 225); the entropy of a transition is not the log ratio of the rates (local detailed balance fails away from equilibrium); and σ_{−ν} ≠ −σ_ν.
Schlögl model (§VII.A.1)
- Exact quasi-potential in closed form (eq. 243), checked against ∫ ln(r₋/r₊). Bistability for k₋₂ ≥ √3 and Δμ ≥ ln 9; at k₋₂ = 3.5 from Δμ ≈ 3.85 (Fig. 4), checked (3.848). The Fokker–Planck barrier falls short by δ, which I find 0.19 at Δμ = 5 and 0.50 at Δμ = 5.3; the text’s “δ ≃ 0.2, for V ≃ 30, would result in a transition rate around 400 times larger” is e⁶ ≈ 403, checked.
- In one dimension the probability velocity vanishes, relaxation and instanton retrace each other in configuration space, and yet the reactions are out of equilibrium; the bound (160) becomes an equality only if the thermodynamic entropy production is replaced by an information-theoretic one (eqs. 245–248).
Discussion (§VIII)
- Two maximum entropy production principles. The review’s own (eq. 160 read as “the attractor with the largest life-time is the one with the largest relaxation (or smallest instantonic) entropy production”) holds only where the bracket is tight, near detailed balance. The usual one, that the attractor with the largest dissipation ⟨σ̇_γ⟩ lives longest, “can agree only under some peculiar conditions” (eqs. 290–292).
- Lifetime is not global stability: “a local criterion of stability, e.g. the life-time of an attractor, does not determine the global stability, e.g. the stationary probability” (§VIII.B.6, citing Maes & Netočný 2013). With only two attractors, as in the Schlögl model, the stationary weights are proportional to the lifetimes. The divergence needs three or more.
- TUR (§VIII.B.1): for observation shorter than the escape time, the TUR involves only the occupied attractor’s dissipation (eq. 286).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Extensive rates, potentials and O(1) forces give a deterministic limit with the quasi-potential as Lyapunov function | strong | §III, standard large deviations |
| C2 | Escape rates go as e^{−VΔI_ss} | strong | eq. 157, Freidlin–Wentzell, cited |
| C3 | Relaxation and instanton entropy productions bracket ln κ_ν / V | moderate | appendix A’s Jensen argument; endpoint cutoff heuristic; loose far from equilibrium, as the authors show |
| C4 | Langevin truncations mis-estimate barriers and miss the housekeeping dissipation | strong | eqs. 137–138, 145; Schlögl checked |
| C5 | Longest-lived attractor = most dissipative only under special conditions | moderate | eqs. 290–292, argument not example |
Concepts
- Quasi-potential I_ss — the stationary rate function, p_ss ≍ e
- Instanton — the most likely path out of an attractor; here the reverse of the dual dynamics.
- Adiabatic (housekeeping) / nonadiabatic entropy production — the parts that sustain a steady state and that drive relaxation (eqs. 18–20).
- Emergent states — attractors treated as the states of a coarse jump process, each “with internal dissipation” (Fig. 1).
Connections
- Vellela & Qian (LIT-1890) on the same model, cited for the two-state reduction and the diffusion failure.
- Santolin et al. (LIT-1881) proves the escape bounds for diffusions, citing this review for jump processes.
- Barato & Seifert (LIT-1885) for the TUR; Seifert’s review (LIT-523) for the ST it starts from.
- Langer 1969 (LIT-1664, unread) and Eyring–Kramers supply the prefactors this review leaves out.
Bearing on the record
- CASE-189 and CASE-111: maintained regimes as dissipative attractors. The review supplies the frame in which their folds have a finite-size meaning. A maintained regime is an attractor with its own housekeeping dissipation ⟨σ̇_γ⟩ (eq. 225). Its lifetime is e^{VΔI_ss}, set by a quasi-potential, not by any energy. A saddle-node is where the barrier closes (Fig. 10a–b). It also gives the condition the cases’ production terms would have to meet to be “paid for”: every production step reversible with local detailed balance.
- THEORY-671 (maximum entropy production selects the steady state). Both sides of the review count against the strong form: the dissipation-ranked principle holds only under “peculiar conditions”, the lifetime-ranked one only near detailed balance, and lifetime need not fix stationary probability when there are more than two attractors.
- The A14 analogy with SLT (THEORY-087). The formal parallel is exact as far as it goes: V here and n there are large parameters, and I_ss and K(w) are rate functions whose minima carry the mass. The disanalogy is what this review adds. I_ss is a property of a dynamics with currents (v_ss) and housekeeping dissipation, whereas a posterior is a static Gibbs measure, the detailed-balance case. Dissipative features, such as circulation on level sets, instanton asymmetry and ⟨σ̇_γ⟩ > 0, have no counterpart in the evidence.
- QUESTION-281’s nucleation route: the escape-rate theory here is its nonequilibrium general case.
- Produces THEORY-1264 and THEORY-1265; co-source of THEORY-1273 and THEORY-1270.
Limitations
- A review: §§IV.H and VI are new, the rest systematizes known results.
- The bounds of §IV.H need an endpoint cutoff that appendix A supplies heuristically, and are loose in the one worked far-from-equilibrium example.
- Prefactors are left out (§VIII.B.3).
- My reading omits §V, §§VII.A.2–VII.C and appendix B.