Contribution
The paper offers a statistical-mechanical derivation of the maximum-entropy-production principle (MaxEP). Until then MaxEP had been proposed empirically, from Paltridge’s climate models onward, and argued for qualitatively. Dewar applies Jaynes’s maximum-entropy (MaxEnt) procedure to microscopic paths instead of states, and reads off three consequences:
- the fluctuation theorem;
- MaxEP as the selection principle for stationary states;
- self-organized criticality (SOC) in slowly driven flux systems.
The aim is to make Jaynes’s formalism a “common predictive framework” for equilibrium and nonequilibrium statistical mechanics (p. 2).
Key insight
Count paths, not states. Choose the path distribution that is broadest given the macroscopic constraints, as Gibbs chose the broadest state distribution. Then entropy production appears in the exponent of the path weight, and the most probable macroscopic history is the one realised by the most paths. If more paths go with more irreversibility, the most probable stationary state is the one that produces entropy fastest. The whole argument turns on that “if”.
Assumptions
- MaxEnt as the method. The macroscopic behaviour is reproducible, so it is characteristic of nearly all compatible paths. Maximizing path entropy under the imposed constraints discards irrelevant microscopic information (§2.2). “When applications of the Jaynes procedure fail … [it] signals the presence of new constraints” (p. 6).
- Step 1 constraints (Eqs. 2–4): normalisation, a fixed initial configuration ⟨d(x,0)⟩ of energy and mass densities in the volume V, and fixed time-averaged boundary fluxes on the boundary Ω. Local energy and mass conservation are imposed through Eq. 9.
- Redundancy of the surface term. After integration by parts, the surface multiplier “can be set to zero” because the flux information is “now also contained” in the volume term (p. 10). This is stated, not shown.
- Step 2 logic. The quantities fixed in Step 1 are, for a reproducible state, “ultimately determined only by the remaining constraints”. So maximizing S_I over the multipliers λ(x) is the right second step (§2.3).
- MaxEP needs two further assumptions (§4.2):
- a mean-field approximation that ignores fluctuations of the action about its mean (Eq. 23);
- W(A^irr), the number of paths with irreversible action A^irr, is “an increasing function of A^irr” (p. 14).
- Time-reversal pairing. Every path has a reverse with the same end-point action and opposite irreversible action (§4.1).
- SOC (§4.3): a Landau–Ginzburg form H(F|F_ext) = rF² + gF⁴ with r > 0 and g < 0 for the output-flux distribution (Eqs. 26–27), and a mean-field (quadratic) treatment of fluctuations.
Key results
- Eq. 5. p_Γ = exp(A_Γ)/Z, with A_Γ = ∫_V λ·d(x,0)_Γ + ∫_Ω η·F^n_Γ (Eq. 6).
- Eq. 14. In terms of a local T and μ_i defined from the multipliers (Eq. 13), A_Γ = −½∫(H_Γ(0) + H_Γ(τ))/k_BT + τσ_Γ/2k_B. Here σ_Γ is the time-averaged entropy production rate of the path, the usual flux-times-force sum for a multicomponent fluid (Eq. 15). T, μ and σ are defined without local equilibrium (p. 12).
- Eq. 18. p_Γ/p_{Γ_R} = exp(τσ_Γ/k_B). From it, ⟨exp(−τσ/k_B)⟩ = 1 and ⟨σ⟩ ≥ 0 (Eq. 20). The fluctuation theorem is p(σ) = exp(τσ/k_B) p(−σ) (Eq. 21). These are exact given Eq. 14.
- Eqs. 22–24, MaxEP. S_I,max = ln Z − ⟨A⟩ ≈ ln W(⟨A^irr⟩), under the mean-field approximation. If W increases with A^irr, maximizing S_I over λ is maximizing ⟨σ⟩.
- Two-box climate toy (p. 15). With T_1 ∝ (f_SW − h)^{1/4} and T_2 ∝ h^{1/4}, σ = h(1/T_2 − 1/T_1) peaks at h_opt ≈ 0.199 f_SW. I re-derived this numerically: the maximum is at h = 0.1991 f_SW.
- Which σ is maximized. For climate, it is the material entropy production (Eq. 15), horizontal and vertical. Radiative entropy production does not count (pp. 15–16). This answers Essex’s objection to Paltridge.
- SOC (Eq. 28). In mean field the output-flux variance is ⟨(F − ⟨F⟩)²⟩ ≈ 1/(8|g|F_ext²), which diverges as F_ext → 0. I checked this from Eqs. 26–27 and it follows given them. The author calls it “likely to be only qualitatively correct” (p. 18).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | MaxEnt over paths under the stated constraints gives p_Γ ∝ exp(A_Γ), with τσ_Γ/2k_B in the action | strong given the constraints | Eqs. 2–14; the surface-term step (p. 10) is asserted |
| C2 | The fluctuation theorem and the second law on average follow | strong for the inferred distribution | Eqs. 18–21; exact given Eq. 14 |
| C3 | MaxEP is Step 2 of the Jaynes procedure | weak | needs mean field (Eq. 23) and the unargued assumption that W(A^irr) increases (p. 14) |
| C4 | In climate applications it is the material, not the radiative, entropy production that is maximized | moderate | follows from which balance Jaynes’s procedure is applied to (pp. 15–16); argued, not tested |
| C5 | SOC emerges from MaxEP in the slow-driving limit | weak, and refuted as stated | Eq. 28 is a mean-field divergence; LIT-525 shows keeping the quartic term gives a finite variance |
| C6 | Accumulating empirical evidence for MaxEP, FT and SOC supports Jaynes’s formalism | assertion | §5; the evidence cited is for MaxEP fits in climate and elsewhere, not tests of this derivation |
Concepts
- Path information entropy S_I = −Σ_Γ p_Γ ln p_Γ over microscopic phase-space paths Γ of length τ (Eq. 1).
- Reversible and irreversible action. The end-point part of A_Γ is symmetric under path reversal. The part τσ_Γ/2k_B is antisymmetric (p. 14).
- MaxEP: the selection principle that the stationary state realised is the one of maximum mean entropy production under the constraints. The paper’s name for what later papers call MEP.
Connections
- Lineage. Jaynes (1957, 1979) for MaxEnt and its near-equilibrium success with linear transport theory. Paltridge (1975–81) and successors for the empirical MaxEP in climate. Ziegler’s maximum-dissipation theorem as a phenomenological route. Maes (1999) for the Gibbs-type path measure that gives the fluctuation theorem. Bak, Tang and Wiesenfeld for SOC.
- Prigogine. Jaynes, commenting in 1980 on Prigogine’s minimum principle, conjectured that Gibbs’s method “might be closer in spirit to a principle of maximum entropy production” (p. 4). Dewar takes that as the programme. The paper does not engage the minimum principle further.
- The dynamical fluctuation theorems. Eq. 21 has the form of the theorems Crooks (LIT-522) and Seifert’s review (LIT-523) derive from Markovian, locally-detailed-balanced dynamics. Here it follows from an inferred path measure. So the paper explains why the fluctuation theorem holds if real path statistics are MaxEnt-distributed. It does not show that they are.
- Critique and reception. Grinstein and Linsker (LIT-525) refute the SOC corollary. Kleidon (LIT-519) relies on this paper for MEP’s foundation and calls it work in progress. Martyushev and Seleznev (LIT-520) survey it, but could not be read.
Bearing on the record
- MaxEP vs Prigogine. The paper is the record’s text for the claim that MEP has a statistical foundation. On this reading the foundation is incomplete. The derivation of MaxEP is conditional on an assumption about path counting that is not argued and on a mean-field step. The 2005 sequel, by Grinstein and Linsker’s account, tried to remove the near-equilibrium restriction and did not.
- Upstream: Jaynes’s MaxEnt. LIT-114 and NOTE-103 record that the cogency of MaxEnt as a foundation of statistical mechanics is unsettled. Every doubt there transfers here, since this paper is MaxEnt applied to paths. NOTE-103’s “MEP” means maximum entropy principle, not maximum entropy production. A reader should not merge the two acronyms.
- THEORY-026 and the stochastic-thermodynamics line: no bearing. The paper concerns stationary states of continuum open systems, not driven finite systems.
- Life and dissipative structures. Dewar’s conclusion extends the procedure to “economies and biological populations” (p. 19). That is the hook the life-as-dissipation literature uses, and Kleidon (LIT-519) uses it for life and Gaia. The paper itself gives no biological argument.
- No ML instruction. The anthology does not hold it.
Limitations
- The MaxEP step is approximate and conditional. The paper says so (“within the mean-field approximation”, p. 15). The monotonicity of W(A^irr) is assumed with no argument or example (p. 14).
- Reproducibility does the work. The Step 2 justification, that fixed averages are “ultimately determined only by the remaining constraints”, is the substantive physical assumption. It is stated as a methodological rule.
- SOC. The result depends on a mean-field quadratic expansion extended to all F. The author concedes it is “likely to be only qualitatively correct” (p. 18).
- No new test. The paper offers no new prediction checked against data. The empirical support it cites is for MaxEP fits, not for this derivation.
Open questions
- Is W(A^irr) increasing in A^irr for any model where it can be computed? The paper gives none.
- Does the derivation survive outside mean field, and outside the linear regime? The 2005 sequel addressed the second. Grinstein and Linsker (LIT-525) show it did not succeed.
- Which constraints make a system “have enough degrees of freedom” for MEP to apply? Kleidon (LIT-519) states this condition. The paper’s answer is only that a failure signals a missing constraint.
Corrections
- none to a seeded skim (there was no seed)
- Abstract vs Eq. 18. The abstract states p_Γ ∝ exp(τσ_Γ/2k_B). The forward/backward ratio (Eq. 18) is exp(τσ_Γ/k_B). These agree, since the end-point terms cancel in the ratio, but readers comparing with Crooks’s e^{ω} should use Eq. 18.
- Extraction. Text-extracted copies of this preprint lose every displayed equation (Symbol font). Quotations of equations in this note are from rendered pages.