A Three-Layer Bound on Entropy Production: From the Hellinger Rate to the Jensen–Shannon Divergence
For a first-order, stationary, irreducible Markov chain we prove the three-layer inequality ΔS ≥ 4H(π) ≥ 4JS,H(π)=i<j∑(wij−wji)2, where wij=πiTij is the edge flux and JS=∑iπiJS(Pf(i)∥Pb(i)) is the π-weighted sum of the Jensen–Shannon divergences between the forward and backward conditional distributions at each state. Equality holds if and only if detailed balance holds; in that case all three quantities vanish simultaneously, and there is no nontrivial equality case. The middle layer is the structural part of this work. The quantity H(π) admits two characterizations. (a) A sum of squared Hellinger distances. H(π)=∑iπiH2(Pf(i),Pb(i)), which is an identity (Sec. 3.5). (b) The level-2.5 large-deviation rate at zero empirical flux (continuous-time formulation). In the level-2.5 description by the empirical measure and the empirical flux, the edge rate function at zero flux equals (w−w′)2, and summing over edges gives H(π) (Sec. 4.2). Its physical reading is the large-deviation cost of forcing the empirical flux to vanish. At equilibrium the zero flux is the typical flux and the cost vanishes; away from equilibrium, forcing the flux to zero is a rare event and the cost is H(π). Under a discrete-time formulation the corresponding quantity is the KL distance to the nearest reversible chain, which in general differs from H(π) (Sec. 4.5). The three-layer chain therefore reads entropy production ≥ 4×(Hellinger distinguishability of forward and backward) ≥ 4×(Jensen–Shannon distinguishability), while the middle layer simultaneously equals the cost of suppressing the flux to zero. The proof combines two known scalar inequalities, edge by edge and state by state. We do not claim that any single step is new. What we claim is the three-layer combination, its sharpness, the level-2.5 evaluation of the middle layer, the single-cycle degeneration law r(a), the relation to the Jeffreys-type inequality of Lin (Sec. 9.4), and an in-class comparison on three explicitly defined families of networks. With the Landauer correspondence ΔS=Q/(kBT) we obtain the quantitative strengthening of the second law Q ≥ 4kBTH(π) ≥ 4kBTJS. 本文证明:对一阶、平稳、不可约马尔可夫链,每步熵产生 ΔS 满足三层不等式 ΔS ≥ 4H(π) ≥ 4JS,H(π)=i<j∑(wij−wji)2, 其中 wij=πiTij 为边流,JS=∑iπiJS(Pf(i)∥Pb(i)) 是逐态前向与后向条件分布的 Jensen–Shannon 散度的 π-加权和。等号成立当且仅当细致平衡;此时三量同时为零,不存在非平凡的等号情形。 中间层是本文的结构性部分。H(π) 有两种刻画。 (a)平方 Hellinger 和。H(π)=∑iπiH2(Pf(i),Pb(i)),为恒等式(§3.5)。 (b)零经验流处的 level-2.5 大偏差率(连续时间表述)。在(经验测度,经验流)的 level-2.5 表述中,逐边率函数在零流处的取值为 (w−w′)2,逐边相加即 H(π)(§4.2)。其物理含义为:将经验流强制为零所需的大偏差率函数值。平衡时零流是典型流,代价为零;偏离平衡时,将流压制为零属于罕见事件,代价即 H(π)。该刻画在离散时间下的对应量为到最近可逆链的 KL 距离,与 H(π) 一般不等(§4.5)。 三层链因此读作 熵产生 ≥ 4×(前向与后向的 Hellinger 可区分性) ≥ 4×(JS 可区分性), 而中间层同时等于将流压制为零的代价。 证明是两条已知标量不等式的逐边与逐态组合。本文不主张任一单步为新;所主张的是三层组合、其尖锐性、中间层的 level-2.5 求值、单循环退化律 r(a)、与 Lin 不等式的 Jeffreys 型形式的关系(§9.4),以及三个明确界定网络类上的类内比较。 经 Landauer 对应关系 ΔS=Q/(kBT) 得第二定律的定量加强 Q ≥ 4kBTH(π) ≥ 4kBTJS.
Authors
- kaifa chen (ORCID: https://orcid.org/0009-0007-1428-711X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24421464
- Primary Topic
- Advanced Thermodynamics and Statistical Mechanics
- Type
- preprint