Contribution
The diffusion-process counterpart of the bounds in Falasco & Esposito’s review (LIT-1892, §IV.H), which the authors cite as previously derived for jump processes with a macroscopic limit ([17, 18]). Two things are added: a recursive expansion of the quasi-potential I_ss in powers of the nonconservative force, each order an integral along detailed-balanced relaxation trajectories (eq. 36); and a proof of the escape-rate bracket (eq. 76) from the orthogonal decomposition of the drift, with a numerical illustration in a sheared double well (§VIII).
Key insight
Split the weak-noise drift as u = −d·∇I_ss + v with v·∇I_ss = 0 (eq. 45). Then relaxation and escape share the circulating part v and differ only in the sign of the gradient part (eq. 61). The entropy flow along either path is ∓ΔI_ss plus the same non-negative adiabatic term v·d⁻¹·v (eqs. 68–70). So relaxation always produces at least ΔI_ss, and escape at least −ΔI_ss; the rate e^{−ΩΔI_ss} sits between, and the gap is twice the circulation’s dissipation.
Assumptions
- Overdamped diffusion with drift μ = −D·∇Φ + f and diffusion D ∝ 1/Ω (eqs. 3–6, §IV); autonomous; a unique steady state.
- Weak noise, Ω → ∞, with the large-deviation ansatz P ≍ e^{−ΩI} (eq. 23).
- Isolated stable fixed points separated by a saddle; limit cycles excluded (§VII).
- Differentiable quasi-potential along the paths used; caustics are not discussed.
Key results
- Emergent second law (eq. 32). dₜI_ss(xₜ) + σ̇(xₜ) ≥ 0 along the deterministic dynamics: the entropy produced while relaxing bounds the drop in quasi-potential. An earlier derivation by Gaveau, Moreau and Toth (1998) “apparently went unnoticed”.
- Nonequilibrium expansion (eqs. 34–38). I_ss = I⁽⁰⁾ + I⁽¹⁾ + …, I⁽⁰⁾ = ϕ, and u⁽⁰⁾·∇I⁽ᵏ⁾ is given by lower orders, so each order is a line integral along zeroth-order (detailed-balanced) trajectories. To first order the emergent second law is saturated (eq. 38). The linear response approximation works far from equilibrium wherever v·∇ϕ is small (eq. 48).
- Instanton (eqs. 54–61). ẋ = v + d·∇I_ss, against relaxation ẋ = v − d·∇I_ss; at f = 0 one is the time-reverse of the other.
- Escape rate and bounds (eqs. 63, 72–76). r ≍ e^{−Ω[I_ss(x_ν) − I_ss(x_i)]}, and −σ_{ν→i} ≤ lim Ω⁻¹ ln r ≤ σ_{i→ν}, with σ_{i→ν} = σ_a − ΔI_ss and σ_{ν→i} = σ_a + ΔI_ss.
- Example (§VIII). Φ = Ω(αx₁⁴ − βx₁² + x₂²) with shear f = γ(−x₂, x₁), α = 1, β = 1.5, Ω = 10–30. Relaxation entropy bounds the sampled rate function, tightly at γ = 0 (Fig. 4); first-order expansion departs from the sampled one as γ grows (Fig. 6); minimum-action instantons give the bracket of Fig. 8, which narrows as γ → 0. No numbers are given in the text for the bracket’s width.
My check
On dx = −Ax dt + √2 dW with A = [[1, g], [−g, 2]], I_ss = ½xᵀC⁻¹x from the Lyapunov equation, and the endpoint x = (1, 0.5) with ΔI = I(x):
- g = 0: σ_relax = 0.75 = ΔI, σ_inst = −0.75 = −ΔI; the bracket is closed.
- g = 0.5: ΔI = 0.694, σ_relax = 0.771, σ_inst = −0.617.
- g = 1.5: ΔI = 0.719, σ_relax = 1.438, σ_inst = 0.000.
- g = 3: ΔI = 0.800, σ_relax = 4.000, σ_inst = +2.400.
Both inequalities hold, and in each case σ_relax − ΔI = σ_inst + ΔI, the common adiabatic part, as eqs. (72)–(73) require. At g = 3 the escape path produces entropy, so the upper bound on ln r is positive and says nothing; the bracket’s width grows quickly with the rotational drive.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | ln r/Ω is bracketed by −σ_relax and σ_instanton | strong | eqs. 68–76, two lines from the orthogonal decomposition; checked on a linear case |
| C2 | The bracket closes at detailed balance and is tight near it | strong | eq. 61 with f = 0; Fig. 8; my g = 0 |
| C3 | The first-order expansion approximates I_ss well for weak force, and far from equilibrium where v·∇ϕ ≈ 0 | moderate | eq. 48, Fig. 6 |
| C4 | The bounds make “a reliable predictive tool” for weak forcing (§IX) | weak | one model, qualitative figure |
Concepts
- Quasi-potential — I_ss = lim −Ω⁻¹ ln P_ss, a Lyapunov function of the noiseless dynamics (§I).
- Orthogonal decomposition — u = −d∇I_ss + v, v ⊥ ∇I_ss (eq. 45).
- Adiabatic entropy production — σ̇_a = v·d⁻¹·v, the part that sustains the circulation, non-negative along relaxation and instanton alike (eq. 70).
Connections
- Freidlin & Wentzell; Bouchet & Reygner 2016 for the nonreversible Eyring–Kramers prefactor (cited, appendix A); Freitas & Esposito 2022 for the emergent second law; LIT-1892 for the jump-process version.
Bearing on the record
- CASE-189. For a diffusion model of a maintained regime, this is the theorem that converts “finite lifetime” into thermodynamic terms: the lifetime is e^{ΩΔI_ss}, and ΔI_ss lies between the dissipation of relaxing into the regime and minus that of escaping it. It does not say the more dissipative regime lives longer; that is true only where the bracket is tight (THEORY-1264).
- THEORY-671. The paper’s bound is a constraint on rates, not a selection principle among steady states.
- The A14 SLT analogy. The quasi-potential plays the role a free energy plays at detailed balance; the disanalogy is the circulation v, which a Bayesian posterior (a Gibbs measure) does not have.
- Produces THEORY-1264 with LIT-1892.
Limitations
- Diffusions only; overdamped; fixed-point attractors.
- The example is two-dimensional, and computing instantons in higher dimension needs minimum-action methods whose cost the paper notes but does not assess.
- Prefactors are left out; at Ω = 30 they may matter.