% Hexagon-compatible flattened source; generated 2026-11-21 \documentclass[reqno]{aomart} % --- Font and encoding --- \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage{lmodern} %\usepackage[nopatch=footnote]{microtype} % --- Math packages --- \usepackage{amsmath,amssymb,amsfonts,amsthm,mathtools,bm,mathrsfs} \usepackage{geometry} \usepackage{graphicx} \usepackage[section]{placeins} \usepackage{booktabs,tabularx,array,multirow,makecell} \usepackage{enumitem} \usepackage{xcolor} \usepackage{hyperref} \usepackage[nameinlink,noabbrev,capitalise]{cleveref} \usepackage{url} \usepackage{longtable} \usepackage{ragged2e} \usepackage{setspace} % The original draft loaded tensor, but no tensor-specific command is used. % Removing it keeps the source portable on standard physics-journal TeX setups. % --- Page layout --- \geometry{margin=1in} \setlength{\headheight}{12pt} \allowdisplaybreaks \setlength{\emergencystretch}{3em} \setlength{\parindent}{1.25em} \setlength{\parskip}{0.25em} \raggedbottom % --- Manuscript identity --- \newcommand{\PaperTitle}{A Stochastic Molecular Mechanism for the Thermodynamic Arrow of Time: Least Action, Equilibrium, and Dissipative Variational Dynamics} % --- Hyperref setup --- \hypersetup{ colorlinks=true, linkcolor=blue!55!black, citecolor=blue!55!black, urlcolor=blue!55!black, pdftitle={A Stochastic Molecular Mechanism for the Thermodynamic Arrow of Time: Least Action, Equilibrium, and Dissipative Variational Dynamics}, pdfauthor={Shipei Hu} } % --- Theorem environments --- \theoremstyle{plain} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{conjecture}[theorem]{Conjecture} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{assumption}[theorem]{Assumption} \newtheorem{principle}[theorem]{Principle} \newtheorem{example}[theorem]{Example} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \crefname{assumption}{Assumption}{Assumptions} \Crefname{assumption}{Assumption}{Assumptions} \crefname{principle}{Principle}{Principles} \Crefname{principle}{Principle}{Principles} % --- Math operators and commands --- \DeclareMathOperator{\Ent}{Ent} \DeclareMathOperator{\BridgeCov}{Cov} \DeclareMathOperator{\BridgeLaw}{Law} \DeclareMathOperator{\Tr}{Tr} \DeclareMathOperator{\supp}{supp} \DeclareMathOperator{\Dom}{Dom} \DeclareMathOperator{\diver}{div} \DeclareMathOperator*{\argmin}{arg\,min} \newcommand{\dd}{\,\mathrm{d}} \newcommand{\norm}[1]{\left\lVert #1\right\rVert} \newcommand{\RR}{\mathbb{R}} \newcommand{\PP}{\mathbb{P}} \newcommand{\QQ}{\mathbb{Q}} \newcommand{\EE}{\mathbb{E}} \newcommand{\cX}{\mathcal{X}} \newcommand{\cH}{\mathcal{H}} \newcommand{\cE}{\mathcal{E}} \newcommand{\cF}{\mathcal{F}} \newcommand{\cI}{\mathcal{I}} \newcommand{\cA}{\mathcal{A}} \newcommand{\cB}{\mathcal{B}} \newcommand{\cQ}{\mathcal{Q}} \newcommand{\cS}{\mathcal{S}} \newcommand{\cD}{\mathcal{D}} \newcommand{\cP}{\mathcal{P}} \newcommand{\cL}{\mathcal{L}} \newcommand{\cW}{\mathcal{W}} % --- Bold symbols for matrices --- \newcommand{\Mob}{\mathsf K} \newcommand{\Gfr}{\mathsf G} \newcommand{\Mmass}{\mathsf M} \newcommand{\Diff}{\mathsf D} % --- Functional and operator notation --- \newcommand{\KL}[2]{D\!\left(#1\middle\Vert #2\right)} \newcommand{\ip}[2]{\left\langle #1,#2\right\rangle} \newcommand{\normH}[1]{\left\lVert #1\right\rVert_{\cH}} \newcommand{\normV}[1]{\left\lVert #1\right\rVert_{V}} \newcommand{\gradH}{D_H} \newcommand{\Iinf}{\mathcal{I}} \newcommand{\Eform}{\mathcal{E}} \newcommand{\Aop}[1]{A_{#1}} \newcommand{\Sop}[1]{S_{#1}} \newcommand{\Uop}[1]{U_{#1}} \newcommand{\GammL}{\Gamma_L} \newcommand{\GammTwo}{\Gamma_{2,L}} \newcommand{\cC}{\mathcal{C}} \newcommand{\cCcore}{\mathcal{C}_{\mathrm{core}}} \newcommand{\cChist}{\mathcal{C}_{\mathrm{hist}}} \newcommand{\HH}{\mathcal{H}} \newcommand{\RRay}{\mathcal{R}} \newcommand{\RRaykin}{\mathcal{R}^{\mathrm{kin}}} \newcommand{\muEq}{\mu_{\mathrm{eq}}} \newcommand{\piEq}{\pi_{\mathrm{eq}}} \title[Least action, equilibrium, and dissipation]{\texorpdfstring{\PaperTitle}{A Stochastic Molecular Mechanism for the Thermodynamic Arrow of Time: Least Action, Equilibrium, and Dissipative Variational Dynamics}} \author{Shipei Hu} \address{College of Finance and Mathematics, West Anhui University, Lu'an, Anhui 237012, China} \email{hushipei@126.com} \doinumber{}\volumenumber{}\issuenumber{}\publicationyear{}\copyrightnote{}\papernumber{} \fancypagestyle{firstpage}{\fancyhf{}\fancyfoot[C]{\thepage}} \date{} % Shared first-page typography; the body geometry is restored afterwards. \RequirePackage{etoolbox} \makeatletter \@ifpackageloaded{geometry}{}{\RequirePackage[pass]{geometry}} \newcommand{\BeginArticleFrontmatter}{% \newgeometry{top=20mm,bottom=20mm,left=25mm,right=25mm}% } \newcommand{\EndArticleFrontmatter}{\restoregeometry\clearpage} \patchcmd{\@maketitle}{\global\topskip42\p@\relax} {\global\topskip14\p@\relax}{}% {\PackageError{frontmatter}{Cannot set the title-page top spacing}{}} \def\@settitle{% \par\begingroup \centering\normalfont\bfseries\fontsize{15}{18}\selectfont \@title\par \ifpdf\hypersetup{pdftitle=\@title}\fi \endgroup\vskip6pt } \patchcmd{\@setauthors}{\@topsep30\p@\relax} {\@topsep16\p@\relax}{}% {\PackageError{frontmatter}{Cannot set the author spacing}{}} \long\def\@getabstract#1{% \begingroup \ifx\languagename\undefined\def\languagename{english}\fi \g@addto@abstract{}% \g@addto@macro\@aom@abstract{#1}% \g@addto@abstract{}% \endgroup \global\setbox\abstractbox=\vtop\bgroup \box\abstractbox\vskip6pt {\centering\normalfont\bfseries\fontsize{10}{12}\selectfont \abstractname\par\vskip6pt}% \list{}{\labelwidth\z@\leftmargin6pt\rightmargin6pt \listparindent\normalparindent\itemindent\normalparindent \parsep\z@\let\fullwidthdisplay\relax}% \item[]\normalfont\def\baselinestretch{1}% \fontsize{10}{12}\selectfont#1 \endlist\egroup } % Let an extended abstract break across pages without reducing its type size. \patchcmd{\@setabstracta}{\box\abstractbox}{\unvbox\abstractbox}{}% {\PackageError{frontmatter}{Cannot enable a page-breaking abstract}{}} \makeatother \begin{document} \BeginArticleFrontmatter \begin{abstract} How does macroscopic irreversibility emerge from time-reversible Hamiltonian dynamics? Microscopic equations preserve phase volume and are invariant under time reversal, yet a system coupled to a single isothermal reservoir shows a one-way arrow: entropy grows, or equivalently, nonequilibrium free energy decays monotonically. The puzzle is real. To resolve it, we shift from individual trajectories to probability laws and path ensembles; trajectory-level fluctuations remain possible. This shift is necessary because, under Poincaré recurrence, almost every state returns arbitrarily close to its initial state, and the full Hamiltonian ensemble preserves relative entropy against an invariant reference. Ensemble formation alone cannot select a direction. Directed relaxation therefore requires a specified preparation, observation, and effective limit. The first step links two usually separate variational principles. Least action supplies the reversible Hamiltonian backbone; equilibrium statistical mechanics supplies the free-energy reference geometry needed for relaxation. A gap remains: least action describes orbit extremals, while equilibrium theory characterizes a reference state whose attraction needs dynamical estimates. The missing link is a molecular heat bath with controlled coarse-graining. In the Euclidean construction, harmonic oscillators couple linearly to the core system, and the bath is conditionally equilibrated with respect to the core’s instantaneous initial configuration. Shifted bath modes become independent centred Gaussians. This removes the initial-slip force while retaining thermal fluctuations and memory. With this preparation in hand, reduction proceeds through a fixed order: exact bath elimination gives a generalized Langevin equation with nonlocal memory; a Lorentzian spectrum and continuum limit give exponential memory; a short-memory limit gives instantaneous friction; a small-mass limit gives overdamped diffusion. The order is part of the theorem, not a convention. Crucially, no entropy-production inequality is assumed; it emerges from coarse-graining. On a torus, standard linear coupling fails globally, so periodic paired couplings are used to realize constant mobility. Even then, the dissipative direction is not automatic: it is fixed jointly by conditional preparation and retarded response, not by the bare Hamiltonian. Having reduced the dynamics to overdamped diffusion, we ask what variational principle selects the probability current. The Onsager–Rayleigh principle minimizes current cost plus signed free-energy rate, equivalent to weighted Wasserstein gradient flow. Full kinetics adds hypocoercivity: reversible transport spreads local friction, while thermal diffusion prevents relative entropy from being preserved. Finally, the Markov data-processing inequality unifies the picture: relative entropy cannot increase. Equilibrium is not a transient endpoint but the persistent reference that organizes kinetic skeleton, dissipation geometry, and path asymmetry; it is the zero of relative entropy with respect to that reference. Thus the arrow is neither postulated nor hidden microscopically; it emerges from preparation, retarded response, and coarse-graining. \end{abstract} \maketitle {\small\noindent\textbf{Keywords:} second law of thermodynamics; stochastic dynamics; minimum dissipation; relative entropy; free energy; hypocoercivity; Bakry-Emery curvature; path space; Poincare recurrence; stochastic Navier-Stokes equations.\par} \medskip \noindent\textbf{2020 Mathematics Subject Classification:} 60H10; 60J60; 60J75; 58J65; 82B31; 82C31.\par \EndArticleFrontmatter \medskip \setcounter{tocdepth}{1} \hypersetup{bookmarksdepth=2} \tableofcontents \newpage \setstretch{1.23} % ===================================================================== \section{Introduction} \label{sec:introduction} An equilibrium law fixes occupation probabilities but does not fix the rate of relaxation. Section~\ref{sec:qcc-correlated} gives a spatially correlated coframe example with a nonlinear mode limit and uniform inverse bounds. Section~\ref{sec:qtc-calibration} separates thermal populations from the independent observation that calibrates an evolution rate. It shows why physical heat rates require a calibrated clock in addition to a stationary reference. The spatial reference is explicitly chosen, while the heat interpretation continues to use the stated reservoir and physical time. The second law of thermodynamics is among the most robust empirical regularities in physics. In autonomous isothermal relaxation, it states that the appropriate nonequilibrium free energy cannot increase; convergence to equilibrium requires additional dynamical conditions. Everyday experience offers familiar illustrations of relaxation: hot tea cools, a compressed gas expands, and ink diffuses through still water. Yet this macroscopic picture becomes conceptually delicate at the microscopic level. For a Hamiltonian with the appropriate time-reversal symmetry, reversing a valid microscopic trajectory, including the momenta and any time-odd external parameters, yields another admissible trajectory. If these equations have no preference for one temporal direction, where does the observed irreversibility enter? This tension between macroscopic relaxation and microscopic reversibility is a foundational problem of statistical physics. It imposes a precise logical requirement: the dynamical premises connecting the two descriptions must be stated without assuming the entropy inequality that is to be derived. We therefore ask under which microscopic, stochastic, and thermodynamic hypotheses an entropy-production inequality follows, and which parts of that implication survive observation and time reversal. To address this question, we place the probability current at the center of the analysis. A probability current is the local flux of probability through state space. It connects the motion specified by a stochastic equation to the transport of an ensemble. Our framework admits two complementary starting points. In the fundamental interpretation, a stochastic law is specified as the microscopic model, together with its physical time reversal; that law may or may not break stationary reversal symmetry. In the effective interpretation, conservative Hamiltonian mechanics, a specified reservoir preparation, and an ordered sequence of limits produce dissipative dynamics after unresolved variables are eliminated. The latter interpretation tests whether a proposed dissipative structure has a concrete mechanical realization. We organize the construction around three connected variational ideas. First, the principle of least action provides the reversible dynamical skeleton. Here ``least action'' means Hamilton's stationary-action principle, \(\delta S=0\); a physical trajectory need not minimize the action globally. Second, equilibrium supplies the free-energy reference for nonequilibrium evolution. Minimization of the calibrated free energy identifies the Gibbs law; attraction to that law is a separate dynamical result. Third, the Onsager--Rayleigh principle characterizes the dissipative current by balancing its quadratic cost against the associated free-energy change. This is the meaning of the minimum-dissipation principle used below. It minimizes a Rayleighian, not dissipation alone, and its physical sign is supplied by the prepared reservoir and retarded response in the mechanically realized branch. The guiding idea can therefore be stated directly: least action provides the variational foundation of microscopic motion, while equilibrium supplies the free-energy reference for nonequilibrium evolution. We connect these structures through a specified molecular reservoir and controlled reduction, obtaining dissipative variational dynamics and identifying when free-energy loss represents total entropy production. The equilibrium reference remains part of the description throughout relaxation; it is not used only to name a possible final state. The irreversible statement concerns evolving probability laws and forward--reverse path comparisons. It does not assert that every microscopic orbit is individually irreversible. A stochastic entropy production can be assigned to a trajectory once its forward and reverse path laws have been specified, and individual values can fluctuate in sign. Likewise, a dissipative generator can satisfy equilibrium detailed balance even while a nonequilibrium initial law relaxes. These distinctions allow reversible microscopic motion, equilibrium fluctuations, and irreversible relaxation to be treated within the same description. \subsection{From microscopic reversible motion to the thermodynamic arrow} The directional content of the second law has its thermodynamic origins in Clausius's work \cite{Clausius1850} and its statistical interpretation in Boltzmann's work \cite{Boltzmann1872}. For autonomous isothermal relaxation at a fixed reservoir temperature \(T>0\), one convenient form is \begin{equation} \frac{\dd F}{\dd t}\leq 0. \label{eq:intro_free_energy_decay} \end{equation} When the energy, entropy, and reservoir heat satisfy the balance specified below, this implies \begin{equation} \dot S_{\mathrm{tot}}\geq 0. \label{eq:intro_second_law} \end{equation} The connection is thermodynamic, not merely a consequence of calling an invariant density an equilibrium law. The isothermal free-energy statement also differs from the microcanonical formulation for an isolated system. We specify the reference measure, free energy, and reservoir conditions before applying either interpretation. The microscopic tension has a simple mathematical form. Let \(z=(q,p)\) be a phase-space point and \(\Phi_t\) an autonomous Hamiltonian flow, defined for the times in question. Its group property gives \begin{equation} \Phi_{t+s}=\Phi_t\circ\Phi_s,\qquad \Phi_t^{-1}=\Phi_{-t}. \label{eq:Hamilton_group} \end{equation} For a time-reversal-symmetric Hamiltonian without time-odd external parameters, the physical parity transformation is \begin{equation} \Theta(q,p)=(q,-p), \label{eq:phase_time_reversal} \end{equation} and reversibility reads \begin{equation} \Theta\Phi_t\Theta^{-1}=\Phi_{-t}. \label{eq:Hamilton_reversal} \end{equation} A time-reversal-even scalar cannot be strictly increasing along every such trajectory and its reverse. Loschmidt's objection \cite{Loschmidt1876} therefore concerns the identification of a macroscopic monotonic law with a pointwise property of every microscopic orbit. It does not contradict relaxation observed under a specified preparation. Poincar\'e's recurrence theorem \cite{Poincare1890} and Zermelo's thermodynamic objection \cite{Zermelo1896} raise a different issue. Recurrence applies to a measure-preserving evolution with a finite invariant measure. Liouville's theorem supplies volume preservation for Hamiltonian mechanics, but it does not make the volume of an unbounded phase space finite. A finite invariant probability or a finite-volume invariant region must be supplied separately. Changing from a finite state description to a history space does not, by itself, remove recurrence. Our nonrecurrence result concerns nonequilibrium probability laws under strict entropy contraction; individual trajectories may still revisit previously occupied regions. These observations lead to the central question: \begin{quote} \emph{Under which microscopic stochastic and thermodynamic hypotheses does probability evolution have the Lyapunov structure needed for the second law, and how does that structure coexist with conservative motion and well-defined reverse processes?} \end{quote} The relevant objects are the generator, invariant reference, variable parities, preparation, and path-probability comparison. Specifying them turns the apparent conflict into a set of distinct mathematical questions. The physical motivation is the tension between two descriptions of time. At the microscopic level, a Hamiltonian system can obey an exact reversal law; at the macroscopic level, prepared systems relax and produce entropy. The central task is therefore to identify the precise operation that connects these descriptions rather than to assume that equilibrium alone supplies a time direction. We eliminate unresolved variables in a declared order, fix the reference measure, match mobility with noise covariance, and then derive the free-energy and path-space statements for the observed law. The unified-field interface asks whether the same regulated state description can also account for the coefficients entering the thermodynamic reduction. Its physical identification requires the stated matching conditions in addition to the thermodynamic proof. The Lorentzian geometric sector supplies constraints and reversible transport; a positive dissipative mobility arises only after the declared reduction and projection. The finite four-dimensional branch developed below makes this distinction concrete. A single shared regulated functional produces the gauge, coframe, and order-parameter contributions, while the Gibbs and entropy identities follow from the positive reduced generator. Thus the interface is tested by an explicit calculation rather than by assigning the same symbol to two unrelated actions. The resulting architecture has a controlled scope. It proves a finite, regulated thermodynamic reduction under a fixed parameter registry and the stated positivity, coercivity, boundary, and local-detailed-balance assumptions. It does not silently promote that result to a theorem about the complete Lorentzian field or its quantum completion. This distinction is essential: a finite regulator can verify the mechanism while leaving the continuum constraint algebra, removal of the regulator, and construction of the physical Hilbert space open. The standalone stochastic statements are therefore read at their own level, whereas the field interpretation is conditional on the interface hypotheses made explicit below. Section~\ref{sec:executable-closure-certificates} develops the finite construction further. It gives a response-based inverse map, an explicit energy-scale ambiguity, and a constrained nonlinear example. An alternative symmetric temporal regulator then gives a positive interacting quantum model at fixed spatial resolution. The distinction between this model, physical gravitational constraints, and a spatial continuum limit determines the order of the remaining analysis. Two further questions make the interface physically testable. First, does a geometric probability law control the inverse coframes that enter observables? A finite covariant logarithmic model answers this question together with an explicit free-energy balance (Section~\ref{subsec:sbr-covariant}). Second, does preserving an equilibrium family identify the physical evolution? The time-change calculation shows why the generator and clock must also be matched (Section~\ref{subsec:sbr-clock}). Exact temporal normalization and a finite-time backreaction estimate then quantify two errors that would otherwise be hidden in the interface. Synthetic comparisons use a fixed parameter registry and distinguish transient trajectories from periodic frequency measurements. The resulting statements strengthen the finite thermodynamic connection while specifying the estimates still required for a field limit. \subsection{Fundamental dynamics and effective reduction} \label{sec:foundational_stance_expansion} The fundamental interpretation treats stochastic dynamics as primitive within the chosen model class. The effective interpretation derives a stochastic equation by eliminating unresolved microscopic variables. These routes can give the same reduced equation while assigning different physical meanings to its assumptions. The Mori--Zwanzig formalism \cite{Mori1965,Zwanzig1961} and analyses of memory in coarse-grained Hamiltonian systems \cite{Espanol2026} explain why elimination and a Markov limit are separate operations. When only a slow variable is observed, hidden variables can retain information and later return it to the observed system. Accordingly, the exact projected equation generally contains memory and fluctuating forces. Schematically, a resolved observable \(A_t\) satisfies \begin{equation} \frac{\dd A_t}{\dd t} =\mathcal R(A_t)+\int_0^t K(t-s,A_s)\dd s+F_t. \label{eq:Mori_Zwanzig_schematic} \end{equation} Here \(K\) denotes a memory kernel and \(F_t\) the fluctuating force defined by the chosen projection; these symbols are local to this schematic equation. A memoryless approximation requires an additional limit or closure argument. Section~\ref{sec:microscopic-bridge30} supplies a concrete effective realization. A harmonic reservoir is prepared in its canonical distribution conditional on the initial system coordinate. Exact Hamiltonian elimination first produces a memory equation and its fluctuation--dissipation relation. Ordered reservoir, short-memory, and small-mass limits then give the overdamped diffusion. Periodic paired couplings realize the constant-mobility model on a torus. Conditional equilibration here specifies the initial bath distribution; it does not assert that a finite isolated harmonic bath is itself an attractor. The preparation temperature is an input to the construction. Section~\ref{sec:scope-closure30} identifies the scope of this connection. A finite oscillator gives recurring observed relative entropy, so reversibility alone cannot imply monotone reduced relaxation. The positive bridge theorem states the reservoir and convergence assumptions that do yield the Rayleighian diffusion. A separate symmetry argument shows why time-reversal-invariant premises alone cannot select an absolute temporal orientation. The forward thermodynamic direction in the constructed branch belongs to the preparation and response protocol. Exact reversal retains system--reservoir correlations that are absent when a fresh equilibrium bath is prepared. The probability-current formulation also makes observation part of the argument. Section~\ref{sec:orientation-reduction} proves an exact quadratic contraction, calculates the conditional force left by hidden coordinates, and bounds the resulting defect in observed free-energy dissipation. A current minimum, an instantaneous marginal balance, and an autonomous Markov law are three different results. A hidden Gaussian example shows how information can return to a measured coordinate while the full relative entropy decreases. \subsection{Main results and their physical interpretation} The results are organized into four levels. Each adds a conclusion only after the assumptions needed for that conclusion have been supplied. The relative entropy \(D(\nu\Vert\mu)\) measures the difference between a probability law \(\nu\) and its reference \(\mu\): for \(\nu=f\mu\), it is \(\int f\log f\dd\mu\), and it is infinite when \(\nu\) is not absolutely continuous with respect to \(\mu\). It is nonnegative and vanishes precisely when the measures coincide. \paragraph{Level I: Markov contraction.} If \((P_t)_{t\geq0}\) is a Markov semigroup and \(\mu\) is invariant, then data processing gives \begin{equation} \KL{\nu P_t}{\mu}\leq\KL{\nu}{\mu}. \label{eq:intro_DPI} \end{equation} This statement requires neither detailed balance nor a differential generator formula. It applies in finite and infinite dimensions, but invariance alone does not identify \(\mu\) as a thermal equilibrium. \paragraph{Level II: Differential entropy dissipation.} For the diffusion generators considered here, the stated domain, integrability, and chain-rule assumptions refine contraction to \begin{equation} \frac{\dd}{\dd t}\KL{\nu_t}{\mu} =-\mathcal I_\mu(f_t),\qquad \mathcal I_\mu(f_t)\geq0, \label{eq:intro_entropy_identity} \end{equation} where \(\nu_t=f_t\mu\). The functional \(\mathcal I_\mu\) includes the diffusion coefficients and therefore has units of inverse time. It quantifies the rate at which the law loses relative entropy. Jump processes require their own logarithmic dissipation expression. \paragraph{Level III: Thermodynamic identification.} For a Gibbs equilibrium law \(\muEq\), with energy and temperature fixed independently of the entropy calculation, \begin{equation} F[\nu]-F[\muEq]=k_BT\KL{\nu}{\muEq}. \label{eq:intro_free_energy_KL} \end{equation} Thus the equilibrium reference fixes the free-energy minimum, and relative entropy measures the excess above that minimum. Combining \cref{eq:intro_entropy_identity,eq:intro_free_energy_KL} yields \begin{equation} \frac{\dd F[\nu_t]}{\dd t} =-k_BT\mathcal I_{\muEq}(f_t)\leq0. \label{eq:intro_F_decay} \end{equation} To identify the right-hand side with physical entropy production, we use a single ideal thermal bath, a time-independent energy function, and no external work, and compatible heat and local-detailed-balance conventions \cite{Seifert2012,EspositoVanDenBroeck2010I,VanDenBroeckEsposito2010II}. With heat positive into the bath, the first law gives \(\dot Q_{\rm bath}=-\dot E\), and hence \begin{equation} \dot S_{\mathrm{tot}} =\dot S_{\mathrm{sys}}+\frac{\dot Q_{\rm bath}}{T} =-\frac{\dot F}{T} =k_B\mathcal I_{\muEq}(f_t)\geq0. \label{eq:intro_total_entropy} \end{equation} Here \(S_{\rm sys}=-k_B\int\rho\log\rho\dd x\) uses a fixed reference volume, and kinetic energy is included when momenta are retained. These physical identifications require more than Gibbs invariance. A maintained stationary current can produce heat even though the one-time probability law and its relative entropy remain constant. \paragraph{Level IV: Convergence and its rate.} Free-energy decay does not by itself prove attraction to equilibrium. A separate coercive estimate must prevent entropy from remaining in undamped directions. For example, a logarithmic Sobolev inequality \cite{Gross1975}, normalized as \(\mathcal I_{\muEq}(f)\geq2\lambda_{\rm LS}D(f\muEq\Vert\muEq)\) with \(\lambda_{\rm LS}>0\), implies \begin{equation} \KL{\nu_t}{\muEq} \leq e^{-2\lambda_{\rm LS}t}\KL{\nu_0}{\muEq}. \label{eq:intro_LSI_convergence} \end{equation} In kinetic models, thermal noise usually acts directly on momenta but not on positions. Hypocoercivity uses the coupling between this diffusion and conservative transport to obtain relaxation of the full law. For the specified quadratic kinetic branch, an explicit Gaussian factorization gives the optimal finite-time entropy-contraction factor and a computable exponential bound. More general models require their own coercivity or mixing estimates. \paragraph{Variational characterization of the current.} For constant symmetric positive mobility \(\Mob\), the overdamped probability current minimizes the Onsager--Rayleigh functional \begin{equation} \mathscr R_\rho[j] =\frac12\int\frac{j^{\mathsf T}\Mob^{-1}j}{\rho}\dd x +\int j\cdot\nabla\frac{\delta F}{\delta\rho}\dd x. \label{eq:intro_Rayleighian} \end{equation} The first term is a quadratic cost for transporting probability; the second is the resulting free-energy rate under the continuity equation and compatible boundary conditions. At a fixed positive density, the unique minimizer in the stated current space is \begin{equation} j_\star=-\rho\Mob\nabla\frac{\delta F}{\delta\rho}. \label{eq:intro_min_current} \end{equation} Indeed, completing the square gives \[ \mathscr R_\rho[j]-\mathscr R_\rho[j_\star] =\frac12\int \frac{(j-j_\star)^{\mathsf T}\Mob^{-1}(j-j_\star)}{\rho}\dd x. \] This identity proves the variational statement without assuming entropy monotonicity. Minimizing the quadratic cost alone would give zero current; minimizing the free-energy rate alone, without a cost or a speed constraint, would generally have no finite minimum. The physical statement is the balance of these terms at the calibrated mobility. In the effective branch, it is an equivalent characterization of the derived diffusion. The constant-mobility formulation connects to Wasserstein gradient flow and the Jordan--Kinderlehrer--Otto minimizing-movement scheme \cite{JKO1998}, as well as Onsager--Rayleigh formulations \cite{Doi2011}. Conservative motion remains in a separate antisymmetric generator component. Hamilton's stationary action, Onsager reciprocal relations \cite{Onsager1931I,Onsager1931II}, Onsager--Machlup path actions \cite{OnsagerMachlup1953,MachlupOnsager1953}, and Freidlin--Wentzell rate functionals \cite{FreidlinWentzell2012} refer to different variational objects. Keeping them distinct preserves the physical meaning of the connection made here. \paragraph{Relation to previous work.} \label{sec:prior_art_boundary} Variational descriptions of irreversible dynamics have several distinct origins. Onsager's reciprocal relations and variational ideas provide a near-equilibrium force--flux description \cite{Onsager1931I,Onsager1931II}. Prigogine's minimum-entropy-production principle concerns a particular near-equilibrium stationary regime \cite{Prigogine1945}. Onsager--Machlup and Freidlin--Wentzell theories describe stochastic trajectory likelihoods and small-noise large deviations \cite{OnsagerMachlup1953,MachlupOnsager1953,FreidlinWentzell2012,SantolinFreitasEspositoFalasco2025}. For gradient-drift Fokker--Planck equations, Jordan--Kinderlehrer--Otto identify the dynamics as free-energy gradient flow \cite{JKO1998}; related structures are available for discrete Markov chains and jump processes \cite{Maas2011,Erbar2014}. The GENERIC framework organizes the reversible and irreversible sectors of nonequilibrium dynamics \cite{GrmelaOttinger1997,OttingerGrmela1997,Ottinger1998}. Stochastic thermodynamics connects heat, work, local detailed balance, path-probability ratios, absolute irreversibility, and excess and housekeeping contributions \cite{Crooks1999,Jarzynski1997,LebowitzSpohn1999,Seifert2005,Seifert2012,EspositoVanDenBroeck2010I,VanDenBroeckEsposito2010II,MurashitaFunoUeda2014,KolchinskyDechantYoshimuraIto2026}. These connections require attention to the level of description: local detailed balance can fail after coarse graining even when the reduced dynamics is Markovian \cite{HartichGodec2023}. Results for one-dimensional response, stationary diffusions, and jump diffusions also distinguish transport response, path entropy production, and time reversal \cite{GaoChunHorowitz2022,DaCostaPavliotis2023,ZhangLu2025}. The question pursued here is whether a specified variational dynamics and equilibrium law imply an autonomous isothermal second law, and how that implication is compatible with time reversal and recurrence. The direction of this implication matters. In the variational construction of Vaquero del Pino et al. \cite{VaqueroDelPino2026}, nonnegative average total entropy production is imposed to obtain constitutive restrictions and generalized fluctuation--dissipation relations. In \cref{thm:minimum_dissipation_current}, the thermally matched mobility and current variational principle are the inputs, and the free-energy dissipation inequality is derived from them. These constructions have different hypotheses and model classes. Mohite and Rieger \cite{MohiteRieger2026} develop a minimum-action description of discrete-state transition dynamics, relating nonquadratic dissipation to entropy production and fluctuation and uncertainty relations. The diffusion current functional used here has a different domain and a quadratic mobility cost. The present argument connects it to Gibbs invariance, strict contraction of nonequilibrium laws, and the obstruction to an equilibrium-preserving Markov inverse. Separate hypotheses govern convergence, regulator limits, and physical interpretation. \subsubsection{Mechanical origins and observation} The mechanical origin of dissipation is a separate question from the variational description of a diffusion already specified. Zwanzig's nonlinear oscillator-reservoir construction gives conditional canonical preparation, a memory equation and its fluctuation--dissipation relation \cite{Zwanzig1973}. Our reservoir calculation belongs to that tradition. The paired trigonometric coupling used on the torus also has a one-dimensional antecedent in dissipative ring models \cite{EtzioniHorovitzLeDoussal2012}. The claims considered here concern the stated matrix realization and the specified sequence of limits, rather than the invention of an oscillator bath or of paired sine and cosine forces. Microscopic fluctuation costs have also been related to variational dissipation. Adams et al. connect a Brownian-particle large-deviation principle to Wasserstein gradient flow \cite{AdamsEtAl2011}. Mielke, Peletier and Renger identify conditions relating large deviations and generalized gradient structures \cite{MielkePeletierRenger2014}; Kraaij et al. derive GENERIC structure from a generalized fluctuation symmetry \cite{KraaijEtAl2018}. These works address the origin of thermodynamic structure, not merely an inequality after it has been postulated. They also involve different microscopic classes and do not supply an unrestricted derivation from every reversible mechanical system. Mielke, Peletier and Zimmer derive a GENERIC system by coupling a finite-dimensional resolved Hamiltonian system to an infinite-dimensional linear bath \cite{MielkePeletierZimmer2025}. Their special compression assumption produces closed projected dynamics and, at positive temperature, an exact multivariate Ornstein--Uhlenbeck bath process. Projection and closure are therefore central to their construction. The present observation calculation asks a complementary question: for a given joint thermal diffusion and a coordinate observation that need not close, which conditional terms remain in the marginal free-energy rate? The answer is the force defect of Theorem~\ref{thm:orientation-defect} and its extension including transport in Theorem~\ref{thm:complete-observation-defect}. The conditional-force estimate has a direct quantitative precursor in Legoll and Leli\`evre \cite{LegollLelievre2010}; related vector coarse-graining estimates appear in Duong et al. \cite{DuongEtAl2018}. Equation~\eqref{eq:mean-force-defect-correspondence} identifies the shared mean-force term exactly. Consequently, quantitative control of hidden forces is not presented as absent from the preceding literature. Ding and Cates propose a microscopic dynamical entropy and obtain a second-law implication when the resolved density evolves independently of hidden details, up to a model-dependent autonomy error \cite{DingCates2025}. Their condition concerns changes of hidden preparation over an evolution interval. Vanishing of an instantaneous conditional force does not establish that property. The distinction is demonstrated below by a stationary Gaussian observation with zero defect at every time and a non-Markov path law. The correspondence between their entropy and our marginal relative entropy is stated only when the temperature and effective energy defining both references agree. Earlier analyses already distinguish entropy changes under preparation, time reversal, and Markov approximation \cite{MaesNetocny2003}, as well as conserved fine-grained Hamiltonian entropy and irreversible macroscopic entropy \cite{GoldsteinLebowitz2004}. Xing's microcanonical construction supplies another route to stochastic thermodynamic relations under its declared assumptions, including a Markov coarse-grained description \cite{Xing2026}. Against this background, the physical question pursued here is how the specified reservoir, variational current, observation and reversal protocol fit together in one model. A general statement that the second law has previously only been assumed would be too broad. \subsection{Four aspects of time irreversibility} The probability-law formulation separates four complementary aspects of irreversibility. They address different questions rather than forming a chain in which every property implies the next. \paragraph{Generator structure and physical reversal.} The dissipative part of a generator and the breaking of stationary time-reversal symmetry are distinct. For time-odd variables such as momentum, generalized detailed balance reads \(L^{\dagger_\mu}=\Theta L\Theta^{-1}\), with the same boundary and domain conventions on both sides. An equilibrium kinetic process can satisfy this relation while containing friction and thermal noise. \paragraph{An inverse Markov evolution.} If a step strictly decreases finite relative entropy, no \(\mu\)-preserving positive Markov map can recover the initial law from the final law. Otherwise, applying data processing to that recovery would reverse the strict inequality. The restriction on the reference and recovery class is essential. \paragraph{Ordering and nonrecurrence of laws.} Under strict entropy contraction, a nonequilibrium law cannot return exactly to an earlier law. This concerns an ensemble distribution, not the recurrence of individual trajectories. A stationary law itself is unchanged and is excluded from the strict statement. \paragraph{Paths and observed histories.} Forward--reverse path relative entropy, relative entropy of a moving history, and one-time relative entropy compare different measures. For a fixed finite window of the same Markov process, once the prescribed prehistory has left the window, the history relative entropy equals that of its oldest endpoint. This identity transfers endpoint convergence without an additional history-space decay assumption. Infinite memory and delayed feedback require a separate construction. Together these distinctions explain the scope in which the classical reversal and recurrence objections are addressed. \subsection{Applications and physical tests} The current formulation can be tested wherever a physical model supplies a resolved state, thermal calibration, and an observation protocol. The applications below retain those inputs and ask which identities remain valid after the model is reduced or extended. \paragraph{Molecular and fluid dynamics.} The molecular examples compare relaxation rates with those predicted from the energy, mobility, and bath temperature. In a fluid regulator, only finitely many spatial modes are retained. Matching a projected drift does not suffice: the noise covariance, physical clock, and reservoir balance must also agree with the proposed thermal generator. Section~\ref{sec:NS} treats the stated two-dimensional stochastic Navier--Stokes setting, while Section~\ref{subsec:secondlaw13-ultraviolet} gives a thermal Galerkin counterexample. Its dimension-independent entropy-decay estimate coexists with failure of energy-space tightness. Thus convergence of probability laws at each resolution does not alone control the energy carried by increasingly fine spatial modes. \paragraph{The unified-field interface.} The common-energy programme \cite{HuUnified2026} asks whether one regulated state description can constrain both field dynamics and stochastic transport. In the notation used by that programme, the thermodynamic energy entering the Gibbs law is obtained only after the reference measure has been fixed. More precisely, on a regulated chart with reference density \(m_r\), the identification is \[ S_r=E_r-\theta s_r+C_r, \qquad \pi_r(x)=Z_r^{-1}m_r(x) \exp\!\left[s_r(x)-\frac{E_r(x)}{\theta}\right] =Z_r^{-1}m_r(x)e^{-S_r(x)/\theta}, \] where \(C_r\) is state independent. Thus \(S_r\) in the present paper is the energy-normalized reduced master functional, not the unscaled Lorentzian action. In an affine chart with \(m_r\equiv1\), this reduces to the simpler density used in the displayed finite-regulator theorem; otherwise the corresponding reference-volume drift correction must be retained. The function \(s_r\) is fixed as part of the regulated state description (for example as a degeneracy or entropy potential); it is not the evolving Shannon entropy of \(p_t\). This prevents the Gibbs reference from depending on the distribution whose dissipation is being evaluated. In the corresponding finite free-energy identity, \(S_r\) is the Hamiltonian-like function relative to \(m_r\,\mathrm dx\); \(E_r\) and \(s_r\) are not separately varied with the evolving law. The finite interface is therefore the following conditional map. The positive symmetric part of the master mobility supplies \(M_r\) and the fluctuation covariance \(2\theta M_r\); a reversible or maintained circulation supplied by the master transport sector is represented by \(c_r\), subject to \(\nabla\!\cdot(\pi_r c_r)=0\) and the stated boundary condition; and the declared projection, constraint surface, parity map, and physical clock transport these data to the observed process. Section~\ref{sec:secondlaw-mother-recovery} then proves the relative-free-energy decay for that finite generator. The proof uses these matching conditions and does not by itself derive the full gauge-constrained field mobility, its bath, or its continuum limit. A nonzero circulation has a separate stationary path-entropy rate. The Schwarzschild comparison \cite{HuSchwarzschild2026} makes the scope concrete: statistical equilibrium of field labels cannot replace local gravitational constraints or the quantum state needed for radiation. A rotating reservoir also requires energy--angular-momentum work accounting. These are requirements for transferring the thermodynamic statement to a physical field sector, rather than consequences of a shared symbol for the action or the Gibbs density. \paragraph{Calibration and independent observations.} The control and observation sections derive the HJB/Cole--Hopf relations, the stochastic conservation criterion, and the conditions for information projection onto a prescribed drift family. An explicit rank condition is needed for identifiable coefficients. The calibration and holdout observables must use the same physical parameters, reference volume, and clock; agreement after refitting is a different claim. Sampling, discretization, and finite-horizon errors are kept separate \cite{YongZhou1999,Girsanov1960,NewmanBarkema1999,Pavliotis2014}. \paragraph{Stochastic control and observation.} Section~\ref{sec:bk-common-origin} derives an exact cost--generator compatibility identity for a specified stochastic-control realization. It states when a chosen energy, drift, diffusion, and cost describe the same evolution. Compatibility does not uniquely determine a physical cost or eliminate unresolved memory. The observation results then calculate the conditional-force and transport terms omitted by a reduced description. They distinguish fitting a trajectory from predicting an independently observed one with frozen parameters. \paragraph{Solvable models and numerical comparisons.} The Ornstein--Uhlenbeck model, optical-trap quench, and driven ring in Section~\ref{sec:exact_diagnostics} connect the abstract quantities to an evolving variance, a heat balance, and a maintained current. The coupled realization in Section~\ref{subsec:secondlaw13-joint} uses the same mechanical coefficients at equilibrium and under heat transport. Section~\ref{subsec:positive24} proves boundary inaccessibility, invariance, and entropy convergence for a coupled positive-coordinate diffusion; an added circulation preserves its stationary law but changes path-reversal entropy. Numerical data generated from these models are synthetic. They test the stated calculations and finite-resolution predictions. \subsection{Organization of the paper} The remainder of the paper is organized as follows. We first establish the microscopic and observational setting. Section~\ref{sec:intrinsic-temporal-orientation} distinguishes the orientation of a dynamical law from reversal of a prepared experiment and relaxation of its probability distribution. Section~\ref{sec:orientation-reduction} then studies what an observer retains after hidden variables are eliminated: it derives the contracted current cost, the conditional-force and transport defects, and the additional condition needed for an autonomous observed process. With these distinctions in place, Section~\ref{sec:microscopic-bridge30} constructs the Hamiltonian system--reservoir model and follows its exact memory equation through the ordered bath-spectrum, short-memory, and small-mass limits. Section~\ref{sec:scope-closure30} collects the assumptions and states the resulting conditional bridge, including the periodic realization on the torus and the role of the preparation in selecting a forward description. The stochastic law obtained in this way becomes the starting point for the thermodynamic analysis. Section~\ref{sec:models_variation} develops the stochastic molecular models, their Gibbs references, and the Onsager--Rayleigh current principle. Section~\ref{sec:entropy_convergence} proves entropy dissipation and separates this monotonicity from the further estimates needed for convergence. Section~\ref{sec:time_arrow_unified} examines time reversal, path-space entropy, maintained stationary currents, and the recurrence objections. Section~\ref{sec:infinite_extensions} turns to moving histories and stochastic fluid dynamics, identifying the additional domain and noise conditions needed beyond finite dimensions. Section~\ref{sec:exact_diagnostics} makes these structures explicit in solvable models and numerical comparisons. Section~\ref{sec:control_interface} then relates the variational description to stochastic control and the associated conservation and balance laws. Section~\ref{sec:secondlaw-mother-recovery} states the compatibility conditions under which the mother process admits the thermodynamic reduction, including the \(S_r=E_r-\theta s_r+C_r\) reference-measure identification and the finite-regulator mapping of mobility and circulation. Section~\ref{sec:finite-4d-interacting-branch} then gives a concrete four-dimensional gauge--coframe--order-parameter regulator. It proves the finite Gibbs and entropy identities from one frozen interacting action and records exactly which Lorentzian, continuum, and quantum statements remain outside that theorem. This explicit branch is included to test the interface constructively; it is not counted as a proof of the complete unified field. The oriented coframe barrier branch immediately following it makes the nondegenerate-sector assumption explicit, proves finite inverse-coframe moments at fixed regulator, and quantifies the loss of uniformity when the barrier is removed. The finite extensions in Section~\ref{sec:executable-closure-certificates} then examine identification, local rotation covariance, temporal normalization, observation-scale nondegeneracy, and dynamical response. Section~\ref{subsec:sbr-backreaction} supplies the finite-time estimate and its synthetic comparisons. Section~\ref{sec:empirical-data-protocol} develops cross-sector response and independent validation with a fixed parameter registry. Section~\ref{sec:three-layer-invariant-regulator} separately treats existence, uniqueness, attraction, and passage to a regulator limit. Section~\ref{sec:conclusion} presents the conclusions and further questions. Appendices~\ref{app:notation_variation}--\ref{app:ring_expansion} provide the notation and the longer variational, entropy, operator, fluid, and driven-ring calculations. \subsection{Summary and outlook} The construction connects least-action mechanics, an equilibrium free-energy reference, and dissipative variational dynamics within a specified stochastic molecular description. Its physical content lies in the connection: the reservoir preparation determines the effective response, the current determines the free-energy rate, and compatible heat accounting gives the thermodynamic entropy production. Observation introduces additional conditions because hidden variables can carry both force and memory. The results also distinguish the questions that remain. In continuum fluid limits, uniform moment estimates, compact containment, generator convergence, and preservation of the reservoir balance are needed in addition to entropy decay. The thermal Galerkin example makes this requirement concrete. For non-Markovian dynamics, a reduced model must retain or control the memory produced by microscopic elimination; a finite-window observation of a Markov process does not solve that problem. A quantum extension would require a physical quantum generator, a suitable reference state, and a separately justified heat and work balance. In a gravitational application, this would also require a specified quantum state and observation map; equilibrium of classical field labels does not provide them. These questions identify the additional mechanisms and estimates needed to extend the present results while retaining their physical interpretation. \section{Microscopic Dynamics and Temporal Orientation} \label{sec:intrinsic-temporal-orientation} An evolution equation specifies how a preparation changes as a clock is advanced; it answers the dynamical question of how the system changes as physical time progresses. A thermodynamic statement adds a further claim: a particular ensemble quantity decreases along that evolution. Establishing this claim requires the energy, probability law, and dissipative response to be specified together. We first formulate their connection before studying a Hamiltonian reservoir that realizes one of its branches. A stochastic law may be either a microscopic modelling premise or an effective law obtained by eliminating a reservoir. When the generators, their domains, and the initial laws agree and the evolution is uniquely defined, these interpretations have the same probabilistic consequences but different physical premises. This distinction runs through the section and is developed in Section~\ref{subsec:orientation-two-starts}. \subsection{Orientation, Reversal, and Relaxation} \label{subsec:orientation-properties} A deterministic equation can contain a dissipative direction. For example, $\dot x=-M\nabla U(x)$ with $M=M^{\mathsf T}>0$ gives $\dot U=-|\nabla U|_M^2$. The gradient flow therefore moves toward decreasing $U$. Reversing a nonstationary solution changes the vector field: the reversed orbit does not satisfy the original equation. This observation does not directly determine the reversal of a stochastic equation, because reversed fluctuations contribute a density-dependent drift. The distinction matters even at equilibrium. Deterministic reversal cannot simply be carried over to the stochastic setting. Let $\Theta$ be a measurable involution preserving an invariant probability $\pi$, and let $\mathcal T f=f\circ\Theta$ be its action on observables. Then $\mathcal T^2=I$ and $\mathcal T$ is unitary in $L^2(\pi)$. For a stationary Markov process, the generator of physical reversal is \begin{equation} L^{\mathrm R}=\mathcal T L^{\dagger_\pi}\mathcal T, \qquad \Delta_\Theta L=L-L^{\mathrm R}. \label{eq:orientation-reversal} \end{equation} These identities concern the appropriate generator domains as well as the formal expressions on a core. They assume that external parameters are held fixed; reversing a magnetic field or another odd external parameter requires the corresponding family of generators. Generalized detailed balance is $L^{\mathrm R}=L$. A nonzero $\Delta_\Theta L$ describes stationary path asymmetry relative to the specified parity and external parameters. We call such asymmetry \emph{intrinsic to the declared stochastic model} when it belongs to its fixed generator rather than its nonequilibrium initial distribution. This terminology describes the fixed model structure rather than a transient effect of the initial preparation. It does not establish the stochastic model as a fundamental law of nature. The third property is relaxation of a nonequilibrium ensemble. Drift describes the local motion, stationary path asymmetry compares forward and reversed histories, and relaxation describes the evolution of the ensemble distribution. For example, the scalar Ornstein--Uhlenbeck equation \[ \dd X_t=-\gamma X_t\dd t+\sqrt{2\gamma\theta}\dd W_t, \qquad \gamma,\theta>0, \] has the reversible stationary law $N(0,\theta)$ and contracts relative entropy from nonequilibrium initial laws. Thus microscopic drift, stationary path asymmetry, and thermodynamic relaxation are distinct statements. The stochastic formulation below allows an asymmetry in the generator, but the relaxation theorem does not require it. This example explains why relaxation and stationary reversal breaking must be examined separately. \subsection{A Thermally Calibrated Microscopic Law} \label{subsec:orientation-thermal} To obtain a concrete theorem, we work on a flat torus $\mathbb T^d$ with its declared coordinate volume. The periodic domain removes boundary fluxes and permits the physical balance to be examined before boundary and nonconfinement issues are introduced. Let $U,M,c$ be smooth, $\theta=k_BT_{\mathrm{bath}}>0$, and \[ M=M^{\mathsf T},\qquad m_-I\preceq M(x)\preceq m_+I,\qquad m_->0. \] Here $M$ is a mobility, not the mass matrix used in kinetic equations. Put $\pi=Z^{-1}e^{-U/\theta}$ and require $\nabla\cdot(\pi c)=0$. In this subsection $\pi$ is a density relative to coordinate volume; integrals against it define the reference measure. Choose a smooth matrix $\sigma$ with $\sigma\sigma^{\mathsf T}=M$. The It\^o equation and its backward generator are \begin{align} \dd X_t&=[c-M\nabla U+\theta\operatorname{div}M](X_t)\dd t +\sqrt{2\theta}\,\sigma(X_t)\dd W_t, \label{eq:intrinsic-oriented-sde}\\ Lf&=[c-M\nabla U+\theta\operatorname{div}M]\cdot\nabla f +\theta M:D^2f. \label{eq:intrinsic-oriented-generator} \end{align} We use $(\operatorname{div}M)_i=\sum_j\partial_jM_{ij}$. All coefficients are independent of the evolving density. Smoothness and compactness give a unique conservative diffusion. On a noncompact space, nonexplosion, normalizability, and integration by parts need separate conditions. Compactness supplies a sufficient setting for the present argument; it is not necessary for every diffusion with the same entropy identity. The Euclidean example in Section~\ref{subsec:orientation-finite-model} verifies its own well-posedness and integrability conditions. The covariance is $a=2\theta M$. The current in the forward equation is $j_i=b_i\rho-\tfrac12\partial_j(a_{ij}\rho)$. Expanding each term gives \begin{align} j_i&=\rho c_i-\rho M_{ij}\partial_jU +\theta\rho\partial_jM_{ij} -\theta(\partial_jM_{ij})\rho-\theta M_{ij}\partial_j\rho, \notag\\ j_\rho&=\rho c-\theta\rho M\nabla\log(\rho/\pi), \qquad \partial_t\rho=-\nabla\cdot j_\rho. \label{eq:intrinsic-current} \end{align} The divergence correction is indispensable when $M$ varies in space. At $\rho=\pi$ the current is $\pi c$, so the well-posed forward equation establishes invariance. When $c\ne0$ and all variables are even, the same Gibbs density carries a stationary current and need not describe a thermal equilibrium experiment. This distinction will reappear when the entropy rate is separated into relaxation and maintained circulation. \begin{theorem}[Thermal relaxation from a specified microscopic law] \label{thm:intrinsic-microscopic-second-law} Under the compact hypotheses above, let $\rho_0$ be a smooth strictly positive probability density. The resulting densities satisfy \begin{align} \frac{\dd}{\dd t}D(\rho_t\Vert\pi) &=-\theta\int\rho_t|\nabla\log(\rho_t/\pi)|_M^2\dd x, \label{eq:intrinsic-relative-entropy}\\ \mathcal F[\rho]-\mathcal F[\pi]&=\theta D(\rho\Vert\pi), \qquad \mathcal F[\rho]=\int U\rho\dd x+\theta\int\rho\log\rho\dd x, \label{eq:intrinsic-free-energy}\\ \frac{\dd}{\dd t}\mathcal F[\rho_t] &=-\theta^2\int\rho_t|\nabla\log(\rho_t/\pi)|_M^2\dd x\le0. \label{eq:orientation-free-energy-rate} \end{align} The circulation drops out of these identities. For $c=0$, with an autonomous single thermal reservoir and the heat convention below, the mean total entropy-production rate is $-\dot{\mathcal F}/T_{\mathrm{bath}}$. For a maintained circulation, the relaxation rate and the total entropy-production rate must be distinguished. \end{theorem} \begin{proof} Uniform parabolicity on the compact torus preserves smoothness and positivity, and justifies differentiating the entropy on every finite interval. Set $f_t=\rho_t/\pi$. Mass conservation removes the derivative of the linear term in $\rho\log\rho$, giving \[ D'(t)=-\int(\nabla\cdot j_{\rho_t})\log f_t\dd x =\int j_{\rho_t}\cdot\nabla\log f_t\dd x. \] Substitution of the current yields \[ D'(t)=\int\pi c\cdot\nabla f_t\dd x -\theta\int\rho_t|\nabla\log f_t|_M^2\dd x. \] The first integral is $-\int f_t\nabla\cdot(\pi c)\dd x=0$. Next $\log\pi=-U/\theta-\log Z$ gives $\mathcal F[\rho]=\theta D(\rho\Vert\pi)-\theta\log Z$ and $\mathcal F[\pi]=-\theta\log Z$. This proves all three formulas. For $c=0$, define the system entropy $S_{\mathrm{sys}}=-k_B\int\rho\log\rho\dd x$ and the mean energy $E=\int U\rho\dd x$. In autonomous relaxation with no applied work, the first law reads $\dot Q_{\mathrm{bath}}=-\dot E$, with heat positive into the reservoir. Hence \[ \dot S_{\mathrm{tot}}=\dot S_{\mathrm{sys}} +\dot Q_{\mathrm{bath}}/T_{\mathrm{bath}} =-\dot{\mathcal F}/T_{\mathrm{bath}}. \] The thermal noise calibration and this energy balance supply the physical interpretation. They are independent of the entropy differentiation. \end{proof} \subsection{The Two Operator Decompositions} \label{subsec:orientation-operators} The Hilbert-space adjoint and physical reversal answer different questions. The first uses an inner product weighted by the invariant law; the second also uses the physical parity of the variables. On a common core define \[ L_{\mathrm s}=\tfrac12(L+L^{\dagger_\pi}),\qquad L_{\mathrm a}=\tfrac12(L-L^{\dagger_\pi}). \] These expressions are symmetric and skew symmetric on that core. For \eqref{eq:intrinsic-oriented-generator}, \[ L_{\mathrm s}f=\theta\pi^{-1}\nabla\cdot(\pi M\nabla f), \qquad L_{\mathrm a}f=c\cdot\nabla f. \] \[ -\langle g,L_{\mathrm s}h\rangle_\pi =\theta\int\nabla g^{\mathsf T}M\nabla h\,\pi\dd x. \] In contrast, $(L-L^{\mathrm R})/2$ and $(L+L^{\mathrm R})/2$ are not generally skew symmetric and symmetric in $L^2(\pi)$. For example, equilibrium kinetic Langevin dynamics has a nonzero Hamiltonian transport operator, yet $L^{\mathrm R}=L$ under momentum reversal. Its reversal defect vanishes while its Hilbert-space skew part does not. Physical reversibility and Hilbert-space symmetry therefore cannot be substituted for one another. They are not logically unrelated: for all-even variables the physical reverse is precisely the weighted adjoint. Even Hilbert-space skew symmetry alone does not imply $\langle Af,\log f\rangle_\pi=0$. The diffusion proof uses the additional fact that $A=c\cdot\nabla$ is a derivation and preserves $\pi$: \[ (c\cdot\nabla)(f\log f) =(1+\log f)c\cdot\nabla f. \] For the skew-symmetric part of a jump generator this chain rule fails. For instance, on three states with uniform $\pi$, let \[ A=\begin{pmatrix}0&1&-1\\-1&0&1\\1&-1&0\end{pmatrix}, \qquad f=\tfrac37(1,2,4)^{\mathsf T}. \] Then $\pi f=1$ and direct multiplication gives $\langle Af,\log f\rangle_\pi=(\log2)/7\ne0$. Here $A$ is not itself a Markov generator, because some off-diagonal entries are negative. It is the skew part of the valid generator \[ Q=\begin{pmatrix}-4&3&1\\1&-4&3\\3&1&-4\end{pmatrix}, \qquad A=\tfrac12(Q-Q^{\mathsf T}), \] whose off-diagonal rates are positive and whose row and column sums vanish. Thus the uniform reference is invariant. Jump-process entropy contraction remains valid by data processing, but its differential calculation cannot use this diffusion cancellation. \subsection{What the Variational Principle Selects} \label{subsec:orientation-variational} Let $\mu_\rho=U+\theta\log\rho$, with a fixed reference volume understood inside the logarithm. For smooth positive $\rho$, take currents in the Hilbert space with norm squared $\int j^{\mathsf T}M^{-1}j/\rho$. Completing the square yields \begin{align} \mathscr R_\rho[j] &=\tfrac12\int\frac{j^{\mathsf T}M^{-1}j}{\rho}\dd x +\int j\cdot\nabla\mu_\rho\dd x, \notag\\ &=\tfrac12\int\frac{|j+\rho M\nabla\mu_\rho|_{M^{-1}}^2}{\rho}\dd x -\tfrac12\int\rho|\nabla\mu_\rho|_M^2\dd x. \label{eq:intrinsic-rayleighian-square} \end{align} The unique minimizer is $j_{\mathrm d}=-\rho M\nabla\mu_\rho$. There is no prescribed endpoint or prescribed $\partial_t\rho$ in this instantaneous minimization. The transport $\rho c$ is added separately. In other words, the principle selects the current that minimizes the Rayleighian for the given density and mobility. It does not minimize the quadratic dissipation term alone, which would select zero current, and it does not prescribe an entire trajectory. Conservative transport requires its own operator, particularly when momentum is retained. Changing the sign of the entire linear term produces $j=+\rho M\nabla(U+\theta\log\rho)$. Its continuity equation has a negative diffusion tensor and cannot define a positive forward Markov semigroup on arbitrary smooth densities. To test the positive maximum principle, choose a smooth function with a negative-definite Hessian at its maximum. On the period-one torus, for example, \[ \psi(x)=\sum_{i=1}^d\cos(2\pi(x_i-x_{0,i})),\qquad L_-\psi(x_0)=4\pi^2\theta\operatorname{Tr}M(x_0)>0. \] The drift term vanishes at $x_0$ and the negative diffusion coefficient reverses the sign of the Hessian term. This contradicts the positive maximum principle. A strict maximum alone would not ensure a strictly positive generator value, since its Hessian could vanish. This observation concerns reversal of the \emph{entire} dissipative current. Merely replacing $-M\nabla U$ by $+M\nabla U$ while retaining positive noise still defines a diffusion on the torus, but generally changes its invariant law and fails to dissipate the original free energy. Reversing the force and reversing the entire dissipative current are therefore different operations: the first retains positive diffusion, whereas the second fails the maximum-principle test above. For all-even variables, the actual nonstationary reverse drift on a window $[0,\tau]$ is \begin{equation} b^{\leftarrow}_s=-b+2\theta\operatorname{div}M +2\theta M\nabla\log\rho_{\tau-s}. \label{eq:orientation-reverse-drift} \end{equation} It has positive covariance $2\theta M$. At stationarity this becomes $-c-M\nabla U+\theta\operatorname{div}M$; in particular the gradient part is unchanged when $c=0$. Thus choosing a forward stochastic model is not a proof that its stationary law breaks time reversal. Stationary asymmetry is characterized by the declared generator, parity, and path law; a relaxation direction can exist even when that asymmetry vanishes. \subsection{Relaxation and Maintained Entropy Production} \label{subsec:orientation-maintained} In the all-even convention, assume the usual finite path-energy and absolute-continuity conditions for a stationary reversal with external forcing held fixed. The drift difference is $2c$. Girsanov's quadratic cost per unit time is therefore \begin{equation} e_{\mathrm{path}}=\frac1\theta\int\pi c^{\mathsf T}M^{-1}c\dd x. \label{eq:orientation-path-rate} \end{equation} For a nonstationary density, expansion of the current square gives \begin{equation} \frac1\theta\int\frac{j_\rho^{\mathsf T}M^{-1}j_\rho}{\rho}\dd x =\theta\int\rho|\nabla\log(\rho/\pi)|_M^2\dd x +\frac1\theta\int\rho|c|_{M^{-1}}^2\dd x. \label{eq:orientation-rate-split} \end{equation} The cross term is proportional to $\int\pi c\cdot\nabla(\rho/\pi)=0$. The first term is $-D'(t)$. The second can remain positive at stationarity. Multiplication by $k_B$ gives physical entropy units only when a thermal force and reservoir balance have been supplied. With odd variables one must instead use the parity-correct reverse law; Hamiltonian transport must not be counted as dissipative by this all-even formula. At $\rho=\pi$ the relaxation term is zero, whereas a nonzero smooth circulation gives a positive maintained term. Mere weak convergence of densities to $\pi$ does not by itself imply convergence of the gradient-dependent dissipation. The displayed decomposition holds under the stated model and reversal conditions. \subsection{A Finite Model Separating the Statements} \label{subsec:orientation-finite-model} Take $U(x)=|x|^2/2$, $\theta=1$, $M=\operatorname{diag}(1,2)$, and \[ \Omega=\begin{pmatrix}0&1\\-1&0\end{pmatrix},\qquad B_\omega=-M+\omega\Omega,\qquad \dd X_t=B_\omega X_t\dd t+\sqrt{2M}\dd W_t. \] Although the theorem above was stated on the torus, this linear Euclidean example is globally well posed and all Gaussian integrations are finite. Since $B_\omega+B_\omega^{\mathsf T}+2M=0$, its invariant law is $N(0,I_2)$. For $X_0\sim N(m_0,I_2)$, independent of future increments, \begin{align} m_t&=e^{B_\omega t}m_0,\qquad D(\rho_t\Vert\pi)=\tfrac12|m_t|^2, \notag\\ D'(t)&=-m_t^{\mathsf T}Mm_t,\qquad e_{\mathrm{path}}=\omega^2\operatorname{Tr} (\Omega^{\mathsf T}M^{-1}\Omega)=\tfrac32\omega^2. \label{eq:orientation-ou} \end{align} The covariance stays $I_2$ by its Lyapunov equation. With $\omega=0$ the stationary law is reversible but nonequilibrium means relax. With $\omega\ne0$, declaring both coordinates even gives positive stationary path asymmetry. With the energy and mobility fixed, these cases show that relaxation does not require generator-level reversal breaking. Moreover, a process started at its invariant law has no change in its one-time density, even when the circulating model has positive stationary path asymmetry. The example separates these properties; it does not prove that every pair of dynamical properties is logically independent. \subsection{Physical Interpretation of the Two Starting Points} \label{subsec:orientation-two-starts} Returning to the distinction raised at the beginning of the section, we can now state the physical meaning of the two starting points. The primitive stochastic interpretation takes the complete generator, temperature, parity, and physical clock as model inputs. Its microscopic temporal asymmetry, when present, is tested by \eqref{eq:orientation-reversal}; the free-energy theorem follows from invariance and the positive diffusion form. The reservoir interpretation instead derives a specific diffusion from a Hamiltonian model, preparation, and ordered finite-time limits. Its scope is the specified reservoir class. Neither interpretation requires a monotonic entropy along each sample path. A relative-entropy inequality orders ensemble laws. Exact reversal of a prepared Hamiltonian system retains hidden correlations, whereas a new canonical preparation resets them. These facts explain why the probability-law theorem can coexist with reversible microscopic trajectories and with trajectory recurrence. A common local relaxation direction also requires a common physical clock if several sectors are to be compared. Positive calibrated clock changes preserve their orientation, subject to the corresponding change of stationary sampling law. Choosing such clocks is not a proof of a global cosmological time function. The finite theory fixes the thermodynamic implication once its physical inputs are declared; the origin of a universal microscopic clock or a cosmological boundary condition remains a separate physical question. The stochastic molecular framework therefore supplies conditional thermodynamic statements that can be tested within their specified physical preparations. \section{Reduction of the dissipative current} \label{sec:orientation-reduction} Macroscopic measuring instruments typically record only a fraction of the underlying microscopic motion. In molecular dynamics, for instance, we often track a few characteristic slow variables of a large molecule, while many atomic coordinates remain outside the observational window. For the coordinate observations considered here, this loss of information means that the observed probability current is obtained by projecting the microscopic current and integrating over the hidden coordinates. This raises a question that no static equilibrium density can answer on its own: does minimizing a dissipative Rayleighian before observation give the same current as minimizing its effective counterpart after the hidden coordinates have been eliminated? The question concerns whether coordinate reduction and variational current selection commute. If they do not, the current predicted by the proposed reduced constitutive law differs from the projection of the microscopic current. This is a statement about ensemble currents, rather than an assertion that an individual molecular trajectory minimizes a dissipation functional. To address this question, we proceed in three steps. First, we solve the contraction of a positive quadratic cost in finite dimensions. For a specified observation map, this determines the effective mobility and the unique minimizing lift. Second, we extend the calculation to current fields on a product torus and identify the conditional force left by hidden coordinates. Third, we derive the observed free-energy balance, including both force and transport defects. A separate generator condition then determines when the observed path law is autonomous and Markovian. The distinction between these questions is essential: an instantaneous variational identity does not determine the law of an entire observed history. \subsection{Quadratic cost contraction in finite dimensions} Before considering current fields, it is useful to build a clear algebraic picture. Let $\mathbb R^n$ be the microscopic current space, and let $H=H^{\mathsf T}>0$ be its mobility matrix. The quadratic cost of a current $J$ is $\tfrac12J^{\mathsf T}H^{-1}J$. An $m\times n$ observation matrix $C$, with $m0$. The proposed lift has $CJ_*=j$. Writing $J=J_*+K$ gives $CK=0$ and $K^{\mathsf T}H^{-1}J_*=K^{\mathsf T}C^{\mathsf T} H_{\mathrm{eff}}^{-1}j=0$. Expanding the square proves the identity and uniqueness. \end{proof} The decomposition separates the cost needed to realize the measured current from a nonnegative cost carried by unresolved directions. The latter is lost when one retains only the minimum over the fibre; it is not additional dissipation generated by observation. Nor is this quadratic cost automatically a physical entropy-production rate: that identification requires a thermally calibrated dynamical model. The separation of costs does not automatically extend to driving forces. If the microscopic force is the pullback $C^{\mathsf T}g$ of an observed force $g$, then $J^{\mathsf T}C^{\mathsf T}g=j^{\mathsf T}g$ on every constraint fibre. Consequently the contracted Rayleighian is \[ \inf_{CJ=j}\{\tfrac12J^{\mathsf T}H^{-1}J+ J^{\mathsf T}C^{\mathsf T}g\} =\tfrac12j^{\mathsf T}H_{\mathrm{eff}}^{-1}j+j^{\mathsf T}g. \] Its minimizer is $-H_{\mathrm{eff}}g$. For a general force $h$, the unconstrained microscopic minimizer projects instead to $-CHh$. Equality with the proposed thermodynamic current therefore requires the separate force-matching condition $CH(h-C^{\mathsf T}g)=0$. Positive quadratic contraction alone does not establish force compatibility. When $C$ selects a coordinate block, $CHC^{\mathsf T}$ is the corresponding mobility block. Its inverse is also the Schur complement obtained by eliminating the unobserved block in $H^{-1}$. A Schur complement of a mobility under a different constraint is a different operation; the constraint must be stated before either formula is used. \subsection{Contraction along hidden coordinates} The linear calculation extends to current fields when observation retains some coordinates and integrates out the others. Work on the product torus $\mathbb T^m_y\times\mathbb T^k_z$, and let $P=(I_m,0)$ select the observed coordinates. Let $\rho(y,z)>0$ be smooth, and let $M(y,z)=M(y,z)^{\mathsf T}$ be smooth and uniformly positive. Define \[ p(y)=\int\rho(y,z)\dd z,\qquad \overline M_\rho(y)=\frac1{p(y)} \int\rho(y,z)PM(y,z)P^{\mathsf T}\dd z. \] For a microscopic current $J(y,z)$, its observed current is $j(y)=\int PJ(y,z)\dd z$. Let \[ \Phi_\rho[J]=\tfrac12\int J^{\mathsf T}M^{-1}J/\rho\,\dd y\dd z. \] \begin{theorem}[Exact current contraction] \label{thm:orientation-fibre-contraction} For every smooth observed current $j$, \begin{align} \inf_{\int PJ\,\dd z=j}\Phi_\rho[J] &=\tfrac12\int j^{\mathsf T}\overline M_\rho^{-1}j/p\,\dd y, \label{eq:orientation-fibre-cost}\\ J_*(y,z)&=\frac{\rho(y,z)}{p(y)} M(y,z)P^{\mathsf T}\overline M_\rho(y)^{-1}j(y). \label{eq:orientation-fibre-lift} \end{align} The minimum is unique in the weighted current space. Moreover, \[ \Phi_\rho[J]-\Phi_\rho[J_*] =\tfrac12\int|J-J_*|_{M^{-1}}^2/\rho\,\dd y\dd z. \] \end{theorem} \begin{proof} Uniform positivity passes to $\overline M_\rho$. Integrating $PJ_*$ over $z$ gives $j$. For $K=J-J_*$, the constraint implies $\int PK\,\dd z=0$. The cross term is \[ \int K^{\mathsf T}M^{-1}J_*/\rho\,\dd y\dd z =\int\left(\int PK\,\dd z\right)^{\mathsf T} \overline M_\rho^{-1}j/p\,\dd y=0. \] The same integration gives the stated cost of $J_*$. Expansion and positivity finish the proof. \end{proof} This is an instantaneous contraction at a fixed density. Once $j$ is chosen, its observed density derivative is $-\nabla_y\cdot j$; no independently prescribed hidden density derivative is imposed. If a hidden derivative or an endpoint law is also fixed, the admissible currents form a smaller set and the displayed unconstrained lift need not be admissible. The theorem therefore gives an instantaneous optimum, rather than an optimum over a prescribed density history. Path autonomy and memory introduce further questions beyond this difference in admissible sets. \subsection{The thermodynamic force left by hidden variables} Instantaneous contraction determines the least quadratic cost at a given joint density. It does not yet determine whether hidden variables leave a residual force when only the observed density is specified. Let $\pi(y,z)>0$ be the Gibbs reference density, and write \[ \bar\pi(y)=\int\pi(y,z)\dd z,\quad r(z\mid y)=\rho(y,z)/p(y),\quad s(z\mid y)=\pi(y,z)/\bar\pi(y). \] The exact relative-entropy chain rule is \begin{equation} D(\rho\Vert\pi)=D(p\Vert\bar\pi) +\int p(y)D(r(\cdot\mid y)\Vert s(\cdot\mid y))\dd y. \label{eq:orientation-chain} \end{equation} It follows by inserting $\log(\rho/\pi)=\log(p/\bar\pi)+\log(r/s)$ into the full integral. Thus the full relative entropy equals the marginal mismatch plus an average of the conditional mismatch. This decomposition also separates the thermodynamic forces. Define the observed force and the residual hidden force by \begin{equation} g=\nabla_y\log(p/\bar\pi),\qquad R_\rho(y)=\int r(z\mid y)PM(y,z) \nabla_{y,z}\log(r/s)\dd z. \label{eq:orientation-force-defect} \end{equation} The term $R_\rho$ is determined by the conditional densities and the mobility. Given these quantities it can be computed directly; it is not an entropy-production coefficient fitted after the calculation. Estimating it from observations nevertheless requires information about the hidden conditional distribution. \begin{theorem}[Projected dissipation and its defect] \label{thm:orientation-defect} Suppose the smooth joint density evolves with current \[ J=\rho c-\theta\rho M\nabla\log(\rho/\pi), \qquad \nabla\cdot(\pi c)=0. \] Set $\bar c_\rho(y)=\int r(z\mid y)Pc(y,z)\dd z$. Then the marginal current and marginal free-energy rate satisfy \begin{align} j&=p\bar c_\rho-\theta p\overline M_\rho g-\theta pR_\rho, \label{eq:orientation-projected-current}\\ \frac{\dd}{\dd t}\{\theta D(p\Vert\bar\pi)\} &=\theta\int p\bar c_\rho\cdot g\dd y -\theta^2\int p|g|_{\overline M_\rho}^2\dd y -\theta^2\int p\,g\cdot R_\rho\dd y. \label{eq:orientation-projected-rate} \end{align} If $\nabla_y\cdot(\bar\pi\bar c_\rho)=0$, then for every $0<\delta<1$, \begin{equation} \frac{\dd}{\dd t}\{\theta D(p\Vert\bar\pi)\} \le-\theta^2(1-\delta)I_{\overline M_\rho} +\frac{\theta^2}{4\delta} \int p|R_\rho|_{\overline M_\rho^{-1}}^2\dd y, \label{eq:orientation-defect-bound} \end{equation} where $I_{\overline M_\rho}=\int p|g|_{\overline M_\rho}^2\dd y$. \end{theorem} \begin{proof} Differentiate the logarithmic factorization: \[ \nabla_{y,z}\log(\rho/\pi) =P^{\mathsf T}g+\nabla_{y,z}\log(r/s). \] Projecting and integrating the current gives \eqref{eq:orientation-projected-current}. Integration of the hidden divergence on the torus gives $\partial_t p=-\nabla_y\cdot j$. Differentiation of the marginal entropy therefore gives $\theta\int j\cdot g$, which is \eqref{eq:orientation-projected-rate}. Under the weighted-solenoidal condition the transport integral vanishes: $\int p\bar c_\rho\cdot g =\int\bar\pi\bar c_\rho\cdot\nabla(p/\bar\pi)=0$. Finally, \[ |g\cdot R_\rho| \le |g|_{\overline M_\rho}|R_\rho|_{\overline M_\rho^{-1}} \le\delta|g|_{\overline M_\rho}^2 +(4\delta)^{-1}|R_\rho|_{\overline M_\rho^{-1}}^2. \] Multiplying by $p\theta^2$ and integrating proves the estimate. \end{proof} Conditional equilibrium $r=s$ implies $R_\rho=0$. In that case $\overline M_\rho=\int sPMP^{\mathsf T}\dd z$ depends on the reference law rather than on the evolving marginal. Also $\bar\pi\bar c_\rho=\int\pi Pc\dd z$ is divergence-free after integration over hidden coordinates. The marginal instantaneous Rayleighian then agrees exactly with the contracted microscopic one. Conditional equilibrium is a sufficient condition, not a necessary one: the weaker projected force condition $R_\rho=0$ can also hold. Neither condition is automatically preserved by the joint dynamics. Conditional equilibrium is therefore not a one-time check that remains valid throughout the experiment; its persistence needs a dynamical argument. For $\delta=1/2$ and compatible transport, integration gives the useful finite-window bound \[ \theta D(p_\tau\Vert\bar\pi) +\frac{\theta^2}{2}\int_0^\tau I_{\overline M_{\rho_t}}\dd t \le\theta D(p_0\Vert\bar\pi) +\frac{\theta^2}{2}\int_0^\tau\!\int p_t|R_{\rho_t}|_{\overline M_{\rho_t}^{-1}}^2\dd y\dd t. \] This estimate quantifies the accuracy required of a proposed conditional-equilibration or scale-separation argument. Once the conditional force is controlled over the time window, its effect on the observed balance is bounded explicitly. \subsection{A balance including the transport defect} \label{subsec:complete-observation-defect} Theorem~\ref{thm:orientation-defect} isolates the conditional thermodynamic force when the averaged transport preserves the observed reference. That additional transport condition is automatic at conditional equilibrium, but it need not hold away from it. To treat both effects together, average the transport against the reference conditional density first. Define \begin{equation} \bar c_{\rm eq}=\int sPc\,\dd z,\qquad A_\rho=\bar c_\rho-\bar c_{\rm eq},\qquad B_\rho=R_\rho-\theta^{-1}A_\rho. \label{eq:complete-defect-definitions} \end{equation} Here $A_\rho$ is the transport change caused by replacing the reference conditional density $s$ with the actual conditional density $r$. The factor $\theta^{-1}$ converts this velocity contribution to the units of $R_\rho$, so that $B_\rho$ combines the two effects in a single current defect. The subscript ${\rm eq}$ denotes averaging with the reference conditional law. It does not assert that the joint process satisfies detailed balance: the reference may carry a stationary current. \begin{theorem}[Complete observed balance] \label{thm:complete-observation-defect} Under the smoothness, positivity and periodic hypotheses of Theorem~\ref{thm:orientation-defect}, define \[ F_{\rm obs}=\theta D(p\Vert\bar\pi),\qquad I=\int p|g|_{\overline M_\rho}^2\dd y,\qquad E_B=\int p|B_\rho|_{\overline M_\rho^{-1}}^2\dd y. \] Then $\nabla_y\cdot(\bar\pi\bar c_{\rm eq})=0$, and \begin{align} j&=p\bar c_{\rm eq}-\theta p\overline M_\rho g-\theta pB_\rho, \label{eq:complete-observed-current}\\ F_{\rm obs}'&=-\theta^2I-\theta^2\int p\,g\cdot B_\rho\dd y. \label{eq:complete-observed-rate} \end{align} No condition on $\nabla_y\cdot(\bar\pi\bar c_\rho)$ is needed. For every $0<\delta<1$, \begin{align} F_{\rm obs}'&\le\theta^2(-I+\sqrt{IE_B}), \label{eq:complete-CS-bound}\\ F_{\rm obs}'&\le-\theta^2(1-\delta)I+\frac{\theta^2}{4\delta}E_B, \label{eq:complete-Young-bound}\\ F_{\rm obs}'&\le\frac{\theta^2}{4}E_B. \label{eq:complete-recovery-bound} \end{align} These are instantaneous statements. If the hypotheses and integrability hold on $[0,\tau]$, then \begin{align} F_{\rm obs}(\tau)-F_{\rm obs}(0) &\le\frac{\theta^2}{4}\int_0^\tau E_B(t)\dd t, \label{eq:complete-window-recovery}\\ F_{\rm obs}(\tau)+\frac{\theta^2}{2}\int_0^\tau I(t)\dd t &\le F_{\rm obs}(0)+\frac{\theta^2}{2}\int_0^\tau E_B(t)\dd t. \label{eq:complete-window-dissipation} \end{align} \end{theorem} \begin{proof} Since $\bar\pi\bar c_{\rm eq}=\int\pi Pc\,\dd z$, integrate $\nabla\cdot(\pi c)=0$ over the hidden torus. The integral of its hidden divergence is zero. The remaining observed divergence is precisely $\nabla_y\cdot(\bar\pi\bar c_{\rm eq})$. Substitute $\bar c_\rho=\bar c_{\rm eq}+A_\rho$ in \eqref{eq:orientation-projected-current} to obtain \eqref{eq:complete-observed-current}. The reference is time independent and mass is conserved, so differentiation and integration by parts give \[ F_{\rm obs}'=\theta\int j\cdot g\dd y, \qquad \int p\bar c_{\rm eq}\cdot g\dd y =\int\bar\pi\bar c_{\rm eq}\cdot\nabla_y(p/\bar\pi)\dd y=0. \] This proves \eqref{eq:complete-observed-rate}. Weighted Cauchy--Schwarz gives \[ \left|\int p\,g\cdot B_\rho\dd y\right| \le\left(\int p|g|_{\overline M_\rho}^2\dd y\right)^{1/2} \left(\int p|B_\rho|_{\overline M_\rho^{-1}}^2\dd y\right)^{1/2}. \] Apply $uv\le\delta u^2+v^2/(4\delta)$ to get \eqref{eq:complete-Young-bound}. For the last pointwise bound use $-u^2+uv=v^2/4-(u-v/2)^2$ with $u=\sqrt I$ and $v=\sqrt{E_B}$. Integration gives \eqref{eq:complete-window-recovery}; choosing $\delta=1/2$ before integration gives \eqref{eq:complete-window-dissipation}. \end{proof} In particular $E_B\le I$ is sufficient for instantaneous observed monotonicity. The condition $B_\rho=0$ is stronger and is not necessary. If $E_B(t)\le\varepsilon^2$ throughout a time window, the possible increase is at most $\theta^2\varepsilon^2\tau/4$. This $\varepsilon$ measures a conditional current defect, with units fixed by $E_B$. It is not automatically a memory time, a path-space approximation error, or an error in a microscopic limit. Each of those errors requires its own definition and estimate. When $A_\rho=0$, $B_\rho=R_\rho$ and the previous force estimate is recovered. Under the weaker transport condition of Theorem~\ref{thm:orientation-defect}, its estimate in terms of $R_\rho$ continues to apply directly, whether or not $A_\rho$ vanishes. \begin{corollary}[Bounds from conditional information] \label{cor:complete-conditional-information} Suppose additionally $M\ge\lambda I$ and $|Pc|\le C_c$ uniformly, with $\lambda>0$. Put \[ H_{\rm cond}=\int pD(r\Vert s)\dd y,\qquad J_{\rm cond}=\int\rho|\nabla_{y,z}\log(r/s)|_M^2\dd y\dd z. \] Then for every $\eta>0$, \begin{equation} E_B\le(1+\eta)J_{\rm cond} +\frac{2(1+\eta^{-1})C_c^2}{\theta^2\lambda}H_{\rm cond}. \label{eq:complete-conditional-information-bound} \end{equation} \end{corollary} \begin{proof} Fix $y$ and write $h=\nabla_{y,z}\log(r/s)$. For any $v\in\mathbb R^m$, weighted Cauchy--Schwarz gives \[ |v\cdot R_\rho|^2 =\left|\int r(P^{\mathsf T}v)^{\mathsf T}Mh\dd z\right|^2 \le(v^{\mathsf T}\overline M_\rho v)\int r|h|_M^2\dd z. \] Take $v=\overline M_\rho^{-1}R_\rho$, treating $R_\rho=0$ separately, to obtain $|R_\rho|_{\overline M_\rho^{-1}}^2\le\int r|h|_M^2\dd z$. For the transport term, total variation and Pinsker's inequality give \[ |A_\rho|=\left|\int(r-s)Pc\dd z\right| \le C_c\int|r-s|\dd z\le C_c\sqrt{2D(r\Vert s)}. \] Since $\overline M_\rho\ge\lambda I$, integration yields $\int p|A_\rho|_{\overline M_\rho^{-1}}^2\dd y \le2C_c^2H_{\rm cond}/\lambda$. Finally apply the Hilbert-space inequality $|u-v|^2\le(1+\eta)|u|^2+(1+\eta^{-1})|v|^2$ to $B_\rho=R_\rho-A_\rho/\theta$ and integrate. \end{proof} The conditional Fisher term contains observed as well as hidden derivatives. Small conditional entropy alone does not bound those derivatives. A conditional logarithmic Sobolev inequality gives an entropy bound by a Fisher quantity, not the reverse estimate needed to discard $J_{\rm cond}$. The corollary therefore identifies the derivative control needed to turn a conditional-equilibration assertion into a quantitative observed balance. A small mismatch of conditional densities is not, by itself, enough to make their thermodynamic forces close. The factor $1/4$ is sharp for the stated information $I,E_B$. Equality in the two inequalities requires $B_\rho=-2\overline M_\rho g$ in the weighted space. An admissible diffusion realizes this equality at an instant, as the next example shows. Thus the bound measures possible observed recovery; it does not assert that a nonzero defect always causes recovery or that the complete system violates its entropy balance. \subsubsection{A transport defect with no conditional force defect} The product-torus theorem also holds on Euclidean space when all displayed integrals are finite and integration by parts has vanishing boundary terms. The following Gaussian example has these properties. Set $\theta=1$, $\pi=N(0,I_2)$, $M=I_2$, and $c(y,z)=(\omega z,-\omega y)$ with $\omega\ne0$. Then \[ \nabla\cdot(\pi c)=\omega z(-y\pi)-\omega y(-z\pi)=0. \] The joint diffusion has drift $(-y+\omega z,-z-\omega y)$ and noise $\sqrt2 I_2$. For $\rho=N((m_1,m_2)^{\mathsf T},I_2)$, $r=N(m_2,1)$ and $s=N(0,1)$ do not depend on $y$. Consequently \begin{equation} R_\rho=0,\quad A_\rho=\omega m_2,\quad B_\rho=-\omega m_2, \quad F_{\rm obs}'=-m_1^2+\omega m_1m_2. \label{eq:transport-only-defect} \end{equation} The covariance stays equal to $I_2$, since the skew transport cancels from the covariance equation. Meanwhile $m_1'=-m_1+\omega m_2$ and $m_2'=-m_2-\omega m_1$, so \[ \frac{\dd}{\dd t}\frac{|m|^2}{2}=-|m|^2, \qquad I=m_1^2,\qquad E_B=\omega^2m_2^2. \] At $m_1=1,m_2=2,\omega=1$, the observed rate is $1=E_B/4$. It is positive despite $R_\rho=0$, while the full rate is $-5$. This example does not contradict Theorem~\ref{thm:orientation-defect}: its additional condition fails because $\partial_y(\bar\pi\bar c_\rho)=-\omega m_2y\bar\pi$. \subsection{Connection with conditional mean-force estimates} \label{subsec:conditional-mean-force-comparison} The conditional-force term has a direct antecedent in effective molecular dynamics. To make the comparison explicit, take $M=I$ and $c=0$, and let $\pi=Z^{-1}e^{-U/\theta}$. Denote the potential of mean force by $\mathcal A(y)=-\theta\log\int e^{-U(y,z)/\theta}\dd z$. Differentiating its defining integral gives \[ \nabla_y\mathcal A=\int s\nabla_yU\dd z, \qquad \nabla_y\log s=-\theta^{-1}\nabla_yU+ \theta^{-1}\nabla_y\mathcal A. \] Normalization of $r$ gives $\int\nabla_y r\,\dd z=0$. Therefore \begin{equation} \theta R_\rho =\int r\nabla_yU\dd z-\int s\nabla_yU\dd z. \label{eq:mean-force-defect-correspondence} \end{equation} Thus in this special case the defect is the instantaneous conditional mean force minus its equilibrium counterpart, divided by $\theta$. Legoll and Leli\`evre derive this mean-force discrepancy and a quadratic entropy estimate in their conditional-expectation construction \cite{LegollLelievre2010}. Vector coarse-graining and quantitative relative-entropy and Wasserstein error estimates for Langevin and overdamped Langevin dynamics are developed by Duong et al. \cite{DuongEtAl2018}. The conditional-force square and Young estimate are accordingly used here with their prior context. The present formulation records the simultaneous current contraction, general mobility, transport mismatch and separate requirement of path autonomy. These distinctions specify the result used here, while the cited works provide the earlier conditional mean-force and coarse-graining estimates on which the comparison rests. \subsection{The additional requirement for a Markov law} A current identity alone does not identify the observed path law. For an autonomous projected diffusion, a sufficient generator condition is \begin{equation} L(f\circ P)=(\bar Lf)\circ P \quad\hbox{for all smooth }f, \label{eq:orientation-generator-closure} \end{equation} with a well-posed conservative martingale problem for $\bar L$. For a coordinate projection this requires $Pb(y,z)=\bar b(y)$ and $Pa(y,z)P^{\mathsf T}=\bar a(y)$, including the common physical clock. The full martingale problem then gives \[ f(Y_t)-f(Y_0)-\int_0^t\bar Lf(Y_s)\dd s \] as a martingale in the full filtration. Because this expression is adapted to the observed filtration, the tower property makes it a martingale there as well. Uniqueness identifies the law of $Y=P X$ with the claimed Markov diffusion. This proof establishes path-law closure, rather than only equality of one-time Fokker--Planck equations. The stationary example in Section~\ref{subsec:zero-defect-nonmarkov} explains why this extra step is necessary even when the instantaneous defect vanishes. \subsection{A hidden variable can restore observed relative entropy} There is a simple counterexample to inferring marginal monotonicity from full monotonicity. On $\mathbb R^2$, put \[ M_*=\begin{pmatrix}1&\alpha\\\alpha&1\end{pmatrix}, \quad 0<|\alpha|<1,\quad \dd X_t=-M_*X_t\dd t+\sqrt{2M_*}\dd W_t. \] The stationary covariance is $I_2$. Start from $N((0,a)^{\mathsf T},I_2)$, where $a\ne0$, and observe $Y=X_1$. Then \[ m_1(t)=-ae^{-t}\sinh(\alpha t),\qquad m_2(t)=ae^{-t}\cosh(\alpha t). \] The full entropy $D_{\mathrm{full}}=\tfrac12(m_1^2+m_2^2)$ decreases, whereas \begin{equation} D_Y(t)=\tfrac12a^2e^{-2t}\sinh^2(\alpha t) \label{eq:orientation-hidden-rebound} \end{equation} starts at zero and becomes positive. Here $R_\rho=\alpha m_2$, $g=m_1$, and $\overline M_\rho=1$, so \[ D_Y'(t)=-m_1^2-\alpha m_1m_2. \] This agrees exactly with the defect formula. It shows why the instantaneous marginal free energy can increase even though the complete process is a reversible dissipative diffusion. The same projection is not a scalar stationary Gaussian Markov process. Its stationary covariance at lag $t\ge0$ is $C(t)=e^{-t}\cosh(\alpha t)$. A centered scalar stationary Gaussian Markov process with $C(0)=1$ must satisfy $C(t+s)=C(t)C(s)$: Gaussian regression and conditional independence give this identity. Here the difference is $e^{-(t+s)}\sinh(\alpha t)\sinh(\alpha s)$, which is nonzero for positive $s,t$. Force defect and memory are therefore both visible in this exact example. The same model displays an increase of marginal relative entropy and a failure of the observed Markov property, with both effects traced to the hidden coordinate. \subsection{Zero defect does not establish autonomy} \label{subsec:zero-defect-nonmarkov} The hidden Gaussian model also distinguishes a statement about a density from a statement about paths. Start that model at its stationary law $N(0,I_2)$, rather than at the displaced law above. At every time $\rho=\pi$, $r=s$, $R_\rho=0$ and $B_\rho=0$. The observed density is constant and its relative entropy is zero. Nevertheless its stationary covariance is $C(t)=e^{-t}\cosh(\alpha t)$, and \begin{equation} C(2t)-C(t)^2=e^{-2t}\sinh^2(\alpha t)>0\qquad(t>0). \label{eq:zero-defect-nonmarkov-covariance} \end{equation} For a centered scalar stationary Gaussian Markov process, Gaussian regression gives $\mathbb E[Y_{s+t}\mid Y_s]=C(t)Y_s/C(0)$. Taking its covariance with $Y_0$ and using the Markov property yields $C(s+t)=C(s)C(t)/C(0)$. Equation \eqref{eq:zero-defect-nonmarkov-covariance} violates this necessary condition, since $C(0)=1$. Thus even zero defect throughout a stationary experiment does not imply a Markov observed law. Changing the hidden initial displacement while leaving the initial observed marginal fixed also changes its later mean. Independence from hidden preparation is therefore stronger than cancellation of a force at one evolving density. The preceding displaced example shows that a defect can restore marginal relative entropy; the stationary example shows that zero defect does not imply Markovianity. Together they establish the need to check force matching and path autonomy separately, rather than a general logical independence theorem for arbitrary observations. \subsection{Relation to microscopic dynamical entropy} \label{subsec:mde-correspondence} For comparison with Ding and Cates \cite{DingCates2025}, take the dimensionless entropy convention $k_B=1$. For readers unfamiliar with that construction, the comparison concerns the entropy functional on matching model data. Their microscopic dynamical entropy has the form \[ S^{\rm md}=S_X^G-\beta\langle H_X\rangle+S_Y^\Omega(E). \] Suppose $\beta=\theta^{-1}$ is fixed, the effective energy used in that definition is $H_X$, and the observed reference in the present calculation is exactly $\bar\pi=Z_X^{-1}e^{-\beta H_X}$. Expansion gives \[ D(p\Vert\bar\pi)=\int p\log p\dd y+ \beta\langle H_X\rangle+\log Z_X, \qquad S^{\rm md}=S_Y^\Omega(E)+\log Z_X-D(p\Vert\bar\pi). \] For fixed total energy and a fixed bath reference, the first two terms in the last expression are constants. On this common reference the complete-defect estimate implies, for the specified joint diffusion, \[ \frac{\dd S^{\rm md}}{\dd t} =\theta\left(I+\int p\,g\cdot B_\rho\dd y\right) \ge-\frac{\theta}{4}E_B. \] The correspondence is between functionals with the same energy and reference measure. It does not identify the diffusion with the finite Hamiltonian bath of Ding and Cates. Different strong-coupling energies, varying temperatures, or changing reference measures add terms and require a separate calculation. The autonomy error in their microscopic construction likewise cannot be identified with $E_B$ without a dynamical approximation theorem. \subsection{Finite baths, exact memory and the limiting balance} \label{subsec:finite-bath-entropy-scope} There is a simple consistency check on every microscopic interpretation. Let $\Phi_t$ be an invertible Hamiltonian flow preserving a reference probability $\pi_{\rm tot}$, and let $\rho_0=f_0\pi_{\rm tot}$ with finite relative entropy. The evolved density ratio is $f_t=f_0\circ\Phi_{-t}$. A change of variables gives \begin{equation} D((\Phi_t)_\#\rho_0\Vert\pi_{\rm tot}) =\int f_0\circ\Phi_{-t}\log(f_0\circ\Phi_{-t})\dd\pi_{\rm tot} =D(\rho_0\Vert\pi_{\rm tot}). \label{eq:finite-hamiltonian-entropy-conservation} \end{equation} Observation can lower this entropy by data processing, but it does not by itself make the observed entropy monotone in time. The reversible oscillator and hidden Gaussian examples illustrate the distinction. Disintegrating the reference over the observed coordinate identifies the unobserved contribution as conditional relative entropy. A product reference is a special case, not a requirement for this chain rule. If the full Hamiltonian relative entropy is constant, a rise in the observed relative entropy is balanced by a fall in that conditional term. The bath construction takes its continuum-spectrum, short-memory and small-mass limits in the stated order on finite time windows. The entropy identity is proved for the resulting diffusion. Weak path convergence alone does not imply convergence of relative entropy, its time derivative or its dissipation. A transfer of a quantitative entropy statement from a finite bath would additionally need convergence or uniform integrability of the density ratios, control of the relevant conditional gradients, and a stated time interval. Such estimates must precede an interchange of a reservoir limit with an infinite-time limit. Finite-system recurrence and effective-law contraction therefore refer to different evolutions; a limiting theorem must specify their relation. Within the declared diffusion class, the generator, Gibbs law, current minimization, and entropy balance are connected by the stated proofs. These results concern a specified preparation and effective evolution. Whether a broader class of microscopic systems has the same limit, and whether cosmological boundary data select a universal orientation, are separate physical questions. \subsection{Limits require an additional theorem} The exact contractions above do not assert a general $\Gamma$-limit. For a sequence of mobilities, density laws, or observation maps, a limiting variational statement needs a specified topology, a liminf inequality, a recovery sequence, and sufficient compactness to pass minimizers. A limiting diffusion additionally needs tightness and identification of its martingale problem. These obligations have different roles. A conditional limiting argument can transfer an integrated entropy balance when the required convergence is available. Suppose, for fixed times, $\rho_t^n\Rightarrow\rho_t$ and $\pi_n\Rightarrow\pi$, with convergent initial relative entropies, and \[ D(\rho_t^n\Vert\pi_n)+\int_0^t\mathcal D_n(\rho_s^n)\dd s \le D(\rho_0^n\Vert\pi_n). \] If the integrated dissipation has a liminf bound by a nonnegative limiting functional, joint lower semicontinuity of relative entropy yields \[ D(\rho_t\Vert\pi)+\int_0^t\mathcal D(\rho_s)\dd s \le D(\rho_0\Vert\pi). \] Indeed, apply the liminf to the nonnegative left-hand terms and use the convergence of the right-hand side. Equality of entropy derivatives requires stronger convergence of the dissipation, not merely weak path convergence. The Hamiltonian limits established later determine a specific limiting generator; their entropy identity is then proved for that limiting generator on its own regularity domain. \section{Microscopic implementation and the emergence of a variational time direction} \label{sec:microscopic-bridge30} This section develops the reversible-reservoir interpretation introduced in Section~\ref{sec:intrinsic-temporal-orientation}. The underlying Hamiltonian equations retain their physical time-reversal symmetry. The effective diffusion is obtained from a specified initial preparation and an ordered reduction of the unresolved variables. The appearance of a negative gradient drift does not by itself imply stationary time-reversal breaking. The current-reduction results of Section~\ref{sec:orientation-reduction} also show why a static equilibrium density does not determine an autonomous observed law. A microscopic realization must specify four related inputs: the system--reservoir coupling, the initial conditional bath distribution, the response spectrum, and the scales on which memory and momentum are removed. We use a system coupled to unresolved harmonic oscillators. Their initial coordinates and momenta are sampled from a conditional canonical law, while the full evolution remains Hamiltonian. Eliminating the oscillators produces a memory force that records past motion and a fluctuating force inherited from the bath preparation. Their relation is derived before a memoryless approximation is made. The mechanical starting point is Hamilton's stationary-action principle for the total system--reservoir coordinates, \[ \delta\!\int_{t_0}^{t_1} \left[p\cdot\dot q+\sum_{a=1}^N P_a\dot x_a -H_N(q,p,x,P)\right]\dd t=0, \] where $(q,p)$ are the resolved variables, $(x_a,P_a)$ are the reservoir coordinates and momenta, and $H_N$ is the total Hamiltonian specified below. Configuration endpoints are fixed in the variation. This gives Hamiltonian equations, which are reversible for the Hamiltonian used here; the effective relaxation direction additionally uses the initial preparation and retarded response. We first identify why these inputs are needed, then construct the reservoir and take the limits in the declared order. Throughout, $\theta=k_BT$, $G=G^{\mathsf T}>0$ is friction, $K=G^{-1}$ is mobility, and $F$ denotes nonequilibrium free energy rather than mechanical action or control cost. \subsection{The logical target and the two variational signs} An arbitrary reversible system may be uncoupled, periodic, or retain persistent memory. Reversibility alone therefore does not guarantee a particular coarse-grained behavior. The following elementary calculation rules out, by an explicit counterexample, an unrestricted inference from microscopic reversibility to monotone reduced entropy. \begin{proposition}[A reversible system with recurring observed entropy] \label{b30:prop:oscillator} Consider $H(q,p)=(p^2+\omega^2q^2)/2$, $\omega>0$, and temperature parameter $\theta>0$. Let $q_0\sim N(a,\theta/\omega^2)$ and $p_0\sim N(0,\theta)$ be independent, with $a\ne0$. The observed position law satisfies \[ D\!\left(\BridgeLaw(q_t)\middle\Vert N(0,\theta/\omega^2)\right) =\frac{\omega^2a^2}{2\theta}\cos^2(\omega t). \] It reaches zero and subsequently returns to its initial positive value. The relative entropy of the full phase-space law against the canonical phase-space law is constant. \end{proposition} \begin{proof} Hamilton's equations give \[ q_t=q_0\cos(\omega t)+\frac{p_0}{\omega}\sin(\omega t),\qquad p_t=-\omega q_0\sin(\omega t)+p_0\cos(\omega t). \] The mean of $q_t$ is $a\cos(\omega t)$ and its variance is $\theta\omega^{-2}(\cos^2+\sin^2)=\theta/\omega^2$. For Gaussians with the same covariance, relative entropy is one half the squared mean in the inverse covariance. This proves the formula. The energy of the phase-space mean is $[\omega^2a^2\cos^2+\omega^2a^2\sin^2]/2=\omega^2a^2/2$. This is the excess energy over the centered canonical state; the total expected energy also includes the constant thermal contribution $\theta$. The covariance remains the canonical covariance, so the full relative entropy is $\omega^2a^2/(2\theta)$ and is constant. Equivalently, its constancy follows from the invertible Hamiltonian flow preserving the canonical reference. \end{proof} This example shows that the observed law can return to its initial state rather than follow a monotone relaxation. It motivates a closer look at the dynamical information supplied by the variational sign. \subsubsection{Two convex functionals of opposite sign and the selection of orientation} To make the sign question explicit, consider two convex functionals with opposite signs in their linear term. Let $K=K^T>0$ be a constant mobility matrix, $\theta=k_BT$, and define the free energy and its chemical potential, up to an irrelevant additive constant, by \[ F[\rho]=\int U\rho\dd q+\theta\int\rho\log\rho\dd q, \qquad \mu_\rho=U+\theta\log\rho. \] All logarithms in physical coordinates use a fixed reference volume; writing $\log\rho$ suppresses only the resulting additive constant. This constant does not change a functional derivative gradient or a free-energy difference. For smooth positive densities and currents with no boundary flux, \[ \partial_t\rho+\nabla\cdot j=0 \quad\Longrightarrow\quad \dot F=\int j\cdot\nabla\mu_\rho\dd q. \] Define \[ \Phi_\rho[j]=\frac12\int\frac{j^TK^{-1}j}{\rho}\dd q, \qquad \mathscr R_\rho^\pm[j]=\Phi_\rho[j]\pm\dot F. \] The quadratic cost is the same in the two functionals. At a fixed positive density it makes each functional strictly convex on the weighted current space, although their linear terms have opposite signs. Direct differentiation gives \begin{equation} j_+=-\rho K\nabla\mu_\rho,\qquad j_-=+\rho K\nabla\mu_\rho. \label{b30:eq:two-signs} \end{equation} Consequently, \[ \dot F[j_+]=-\int\rho|\nabla\mu_\rho|_K^2\dd q, \qquad \dot F[j_-]=+\int\rho|\nabla\mu_\rho|_K^2\dd q. \] Positive definiteness makes the quadratic cost nonnegative and ensures uniqueness. It does not select between these two signs. The free-energy functional supplies an equilibrium reference, but it is also unchanged when a history is read in the opposite direction. The sign in $\Phi+\dot F$ therefore specifies a dynamical orientation when adopted as a constitutive principle. To connect that algebraic choice to a physical process, we must also require a positive, probability-preserving evolution. The next proposition excludes the opposite constitutive branch as a forward diffusion; the subsequent reservoir construction identifies the preparation and response that produce the admissible branch. \begin{proposition}[Why the opposite sign is not a forward diffusion] \label{b30:prop:anti-diffusion} The current $j_-$ in \eqref{b30:eq:two-signs} would give \[ \partial_t\rho=-\nabla\cdot(\rho K\nabla U)-\theta\nabla\cdot(K\nabla\rho). \] Its formal backward generator is $L_-f=K\nabla U\cdot\nabla f-\theta K:D^2f$. It cannot generate a conservative positive forward Markov semigroup on an unrestricted smooth class of initial probability laws. \end{proposition} \begin{proof} A positive Markov semigroup has the positive maximum principle: at a global maximum $x_*$ of a function in its generator domain, $Lf(x_*)\le0$. This follows by differentiating $P_tf(x_*)\le f(x_*)$ at $t=0+$. Choose a smooth function with a strict maximum at $x_*$, zero gradient there, and negative-definite Hessian. Then $L_-f(x_*)=-\theta K:D^2f(x_*)>0$, a contradiction. The same problem is visible for a Fourier mode of the backward heat equation: its amplitude grows like $e^{\theta k^TKk\,t}$. \end{proof} This result explains why a probability-preserving diffusion with positive noise covariance uses the dissipative sign in a chosen forward time. It does not establish which temporal orientation the entire universe must adopt. A time-reversed diffusion still has positive covariance; its drift contains a density-dependent correction, computed in Section~\ref{b30:sec:time-information}. \subsubsection{Equilibrium does not uniquely determine the dynamics: circulation currents} Even after fixing a positive diffusion covariance and a Gibbs invariant law, the drift retains freedom. Equilibrium constrains the divergence of the stationary current, rather than necessarily making that current vanish. For constant $K=K^T>0$, all smooth drifts of the form \[ b=-K\nabla U+\ell,\qquad \nabla\cdot(\pi\ell)=0, \qquad \pi=Z^{-1}e^{-U/\theta}, \] preserve $\pi$, subject to a conservative realization and boundary conditions. Their stationary current is $\pi\ell$, which need not vanish. Its zero divergence preserves the density, but does not require every flow line to be a closed orbit. For example, in dimension at least two, let $A^T=-A$ be a constant matrix and choose $\ell=A\nabla U$. Then \[ \nabla\cdot(\pi A\nabla U) =-\frac\pi\theta\nabla U^TA\nabla U+\pi A:D^2U=0. \] For a regular density its current is $j=-\rho K\nabla\mu_\rho+\rho\ell$. The extra contribution to the free-energy rate also vanishes: \[ \int\rho\ell\cdot\nabla\mu_\rho\dd q =\theta\int\pi\ell\cdot\nabla(\rho/\pi)\dd q=0. \] Thus the same Gibbs reference law and even the same free-energy dissipation identity need not select the pure gradient current. The current variational principle specifies additional dynamical structure. For the position-only diffusion with fixed constant $K$, zero equilibrium current, equivalently detailed balance on the regular conservative realization, sets $\ell=0$. Detailed balance is stronger than invariance. In the proposed realization the microscopic model and its preparation, rather than the equilibrium density alone, supply that additional information. With momentum, the appropriate reversal includes odd parity and cannot be replaced by the position-only zero-current test. \subsection{Exact elimination of a reversible Hamiltonian reservoir and the origin of memory} \label{b30:sec:bath} Equilibrium and covariance alone leave the drift undetermined. We now specify the mechanical system whose reduction selects the required drift. The resolved particle is linearly coupled to a Hamiltonian bath of harmonic oscillators, an established construction in statistical mechanics \cite{FordKacMazur1965,Zwanzig1973}. This model gives an explicit mechanism for energy exchange through the coupling forces. It contains no molecular collision law. At finite bath size, the eliminated dynamics retains memory and can return energy or information to the resolved system. A Markov description will require the additional spectral limit and estimates developed below. \subsubsection{Hamiltonian, variational assumptions, and conditional preparation} Let $q,p\in\RR^d$, let the resolved mass be $m>0$, and let $U\in C^3(\RR^d)$ satisfy \begin{equation} \lambda I\le D^2U(q)\le\Lambda I, \qquad 0<\lambda\le\Lambda<\infty. \label{b30:eq:potential-class} \end{equation} Higher smoothness can be imposed for the later classical density calculations. Bounded Hessian makes $\nabla U$ globally Lipschitz; the lower bound supplies confinement and a finite canonical partition function. This is a specific nonlinear class, not an arbitrary molecular interaction with collisions or singular forces. For reservoir oscillators $(x_a,P_a)\in\RR^2$, with masses $M_a>0$, frequencies $\omega_a>0$, and coupling vectors $c_a\in\RR^d$, set \begin{equation} H_N=\frac{|p|^2}{2m}+U(q)+ \sum_{a=1}^N\left\{ \frac{P_a^2}{2M_a} +\frac{M_a\omega_a^2}{2} \left(x_a-\frac{c_a^Tq}{M_a\omega_a^2}\right)^2\right\}. \label{b30:eq:hamilton-bath} \end{equation} The quadratic counterterm inside the square is essential. Without it, integrating over the reservoir generally changes the resolved potential and may even destroy stability. Expanding the square gives \[ \frac{M_a\omega_a^2x_a^2}{2}-x_ac_a^Tq +\frac{(c_a^Tq)^2}{2M_a\omega_a^2}. \] The counterterm compensates the static potential shift produced by the linear coupling. Consequently the reduced equilibrium potential can be checked directly by integration, rather than inferred from a desired invariant density. Once the mechanical model is fixed, its statistical interpretation also requires a preparation protocol. Prepare the resolved state with law $\nu_0(\dd q\dd p)$ and prepare the reservoir conditionally on $q_0=q$. The shifted bath coordinates are \[ y_a=x_a-\frac{c_a^Tq}{M_a\omega_a^2}. \] The conditional reservoir law has density proportional to \begin{equation} \exp\left[-\frac1\theta\sum_a \left(\frac{P_a^2}{2M_a}+\frac{M_a\omega_a^2y_a^2}{2}\right)\right]. \label{b30:eq:bath-preparation} \end{equation} Equivalently, the $y_a(0)$ and $P_a(0)$ are independent centered Gaussians, independent of $(q_0,p_0)$ in these shifted variables, with \[ \EE y_a(0)^2=\frac{\theta}{M_a\omega_a^2},\qquad \EE P_a(0)^2=M_a\theta. \] This puts the bath in conditional equilibrium at the preparation time. The physical coordinates $x_a(0)$ are not independent of $q_0$. An unshifted product preparation changes the initial transient and adds the initial-slip force computed below. Conditional preparation removes that particular term; it does not remove fluctuations, eliminate all initial layers, or by itself establish a Markov limit. \subsubsection{A static canonical limit from a finite-energy reservoir} The canonical preparation has a further statistical-mechanical interpretation. Integrating a specified microcanonical energy-shell measure over a large harmonic bath yields a canonical resolved marginal in the limit. This geometric calculation requires no entropy monotonicity. It explains the limiting exponential weight for that ensemble, rather than proving that every initial bath state thermalizes. Introduce energy coordinates $a_i=\sqrt{M_i}\omega_i y_i$ and $b_i=P_i/\sqrt{M_i}$. The reservoir energy is $\tfrac12\sum_i(a_i^2+b_i^2)$. The density of states of $N$ oscillators is, up to a positive constant, \[ \Omega_N(E)=E^{N-1}\mathbf1_{\{E>0\}}. \] This follows by differentiating the volume of a ball of radius $\sqrt{2E}$ in $2N$ dimensions; its volume is proportional to $E^N$. The linear coordinate Jacobian does not depend on $q$ because of the counterterm and the shift used in \eqref{b30:eq:hamilton-bath}. For a full microcanonical equilibrium at energy $E_N=N\theta$, the resolved marginal is proportional to \[ \left(1-\frac{H_S(q,p)}{N\theta}\right)_+^{N-1}, \qquad H_S=U(q)+|p|^2/(2m), \] after shifting the minimum of $U$ to zero. At each fixed resolved state, this tends to $e^{-H_S/\theta}$. For $N\ge2$ its value is at most $e^{-H_S/(2\theta)}$, because $\log(1-x)\le-x$ on $[0,1)$. The upper bound is integrable under \eqref{b30:eq:potential-class}. Dominated convergence therefore proves convergence of the normalizing constants and of the marginal densities in $L^1$. For an isolated harmonic bath at energy $E_B$, its kinetic temperature parameter is $E_B/N$. At a fixed resolved state in the full microcanonical ensemble, $E_B=E_N-H_S$; thus the conditional bath temperature is $\theta-H_S/N$, rather than exactly $\theta$ at finite $N$. If the derivative of the surface density-of-states entropy is used instead, its temperature parameter is $E_B/(N-1)$ for $N>1$. Both tend to $\theta$ at each fixed resolved state. The separate fixed-bath-energy preparation used below has $E_B=N\theta$ exactly. This equilibrium calculation derives the canonical limiting weight from a specified uniform energy-shell ensemble; it does not derive the choice of that ensemble from Hamilton's principle. More generally, a reservoir entropy $S_B(E)=k_B\log\Omega_B(E)$ gives \[ \log\Omega_B(E-h)-\log\Omega_B(E) =-\frac h{k_BT}-\frac{h^2}{2k_BT^2C_B}+\text{higher terms}, \] when the required derivatives exist, $C_B>0$, and $h$ is small compared with the reservoir's energy scale. Here $S_B'=1/T$ and $S_B''=-1/(T^2C_B)$. For the harmonic example the preceding exact calculation replaces this expansion and its unquantified remainder. \subsubsection{Canonical marginal decoupling before dissipation} Before deriving the dynamical reduction, we check the static marginal directly. Integration over each $x_a$ uses a translation and gives \[ \int_{\RR}\exp\left[-\frac{M_a\omega_a^2}{2\theta} \left(x_a-\frac{c_a^Tq}{M_a\omega_a^2}\right)^2\right]\dd x_a =\sqrt{\frac{2\pi\theta}{M_a\omega_a^2}}. \] The result is independent of $q$. The momentum integral is $\sqrt{2\pi M_a\theta}$. Hence the full canonical law $\Pi_N$ has resolved marginal \begin{equation} \pi_m(\dd q\dd p)=Z_m^{-1} \exp\left[-\frac{U(q)+|p|^2/(2m)}\theta\right]\dd q\dd p. \label{b30:eq:canonical-marginal} \end{equation} Integrating out $p$ gives $\pi(\dd q)=Z^{-1}e^{-U/\theta}\dd q$. Thus the equilibrium marginal depends on $U$ and the temperature, with the bath normalization absorbed into the partition function. This does not make the physical bath coordinates independent of $q$, nor does it put an arbitrarily prepared resolved state in equilibrium. These are static integrations; they do not use a free-energy decay law. We next eliminate the bath from the equations of motion to determine how a nonequilibrium preparation evolves. \subsubsection{Hamilton equations and the retarded solution} Hamilton's equations from \eqref{b30:eq:hamilton-bath} are \begin{align} \dot q&=p/m,\notag\\ \dot p&=-\nabla U(q)+\sum_a c_a \left(x_a-\frac{c_a^Tq}{M_a\omega_a^2}\right),\label{b30:eq:hamilton-resolved}\\ \dot x_a&=P_a/M_a,\qquad \dot P_a=-M_a\omega_a^2x_a+c_a^Tq.\notag \end{align} For finite $N$ these are globally well posed under \eqref{b30:eq:potential-class}. The full flow preserves phase volume and $H_N$. Its physical reversal changes $(p,P)$ to $(-p,-P)$ and leaves $(q,x)$ unchanged. Solving the oscillator equation with data at $0$ gives \begin{equation} x_a(t)=x_a(0)\cos\omega_at +\frac{P_a(0)}{M_a\omega_a}\sin\omega_at +\int_0^t\frac{\sin\omega_a(t-s)}{M_a\omega_a}c_a^Tq_s\dd s. \label{b30:eq:oscillator-solution} \end{equation} The lower limit is a boundary condition: this is the solution in terms of the prepared initial state. It has not been selected by minimizing entropy production. For $f(s)=c_a^Tq_s$, integration by parts gives \[ \int_0^t\frac{\sin\omega_a(t-s)}{M_a\omega_a}f(s)\dd s =\frac{f(t)-f(0)\cos\omega_at}{M_a\omega_a^2} -\int_0^t\frac{\cos\omega_a(t-s)}{M_a\omega_a^2}\dot f(s)\dd s. \] The endpoint term at $s=t$ cancels the counterterm in \eqref{b30:eq:hamilton-resolved}. The endpoint at $0$ combines with $x_a(0)$ to produce the centered variable $y_a(0)$. \begin{theorem}[Exact generalized Langevin equation] \label{b30:thm:exact-gle} For the finite Hamiltonian model and conditional preparation above, the resolved motion satisfies \begin{equation} m\ddot q_t=-\nabla U(q_t)-\int_0^t\Gamma_N(t-s)\dot q_s\dd s+\eta_N(t), \label{b30:eq:gle} \end{equation} where \begin{align} \Gamma_N(t)&=\sum_a\frac{c_ac_a^T}{M_a\omega_a^2}\cos\omega_at, \label{b30:eq:finite-kernel}\\ \eta_N(t)&=\sum_a c_a\left[y_a(0)\cos\omega_at +\frac{P_a(0)}{M_a\omega_a}\sin\omega_at\right]. \label{b30:eq:finite-noise} \end{align} The force $\eta_N$ is a stationary centered Gaussian process, independent of the resolved initial state, with \begin{equation} \EE[\eta_N(t)\eta_N(s)^T]=\theta\Gamma_N(t-s). \label{b30:eq:exact-fdr} \end{equation} \end{theorem} \begin{proof} Substitution of \eqref{b30:eq:oscillator-solution} and the integration-by-parts identity proves \eqref{b30:eq:gle}. Independence removes the terms with different oscillator indices and the mixed position--momentum terms in the covariance. The remaining terms for oscillator $a$ are \[ \frac{\theta c_ac_a^T}{M_a\omega_a^2} [\cos\omega_at\cos\omega_as+\sin\omega_at\sin\omega_as] =\frac{\theta c_ac_a^T}{M_a\omega_a^2}\cos\omega_a(t-s). \] Summation gives \eqref{b30:eq:exact-fdr}. A centered Gaussian process is determined by its covariance, so dependence on $t-s$ proves stationarity. \end{proof} This derives the fluctuation--dissipation relation for this reservoir from its Hamiltonian and preparation. It is a more specific result than assuming an Einstein relation after writing a Langevin equation. General linear-response versions require their own hypotheses \cite{Kubo1966}. If instead $x_a(0)$ is centered at zero independently of $q_0$, the same calculation produces an additional force $-\Gamma_N(t)q_0$. This initial-slip term must either be retained or removed by changing the preparation. The proposed theorem removes it by the explicit conditional law \eqref{b30:eq:bath-preparation}. \subsubsection{What is positive, and what is not} The even matrix kernel need not be positive at every nonzero time: a cosine can be negative. Its positivity is covariance positivity. For any vector-valued test function $v$, \[ \int\!\!\int v(t)^T\Gamma_N(t-s)v(s)\dd t\dd s =\frac1\theta\EE\left(\int v(t)^T\eta_N(t)\dd t\right)^2\ge0. \] Replacing this positive-type statement by the stronger assertion $\Gamma_N(t)>0$ for every $t$ would be an error. Neither statement implies a memoryless evolution or monotone reduced entropy at finite $N$. For a prescribed deterministic velocity protocol $v$ on $[0,\tau]$, the mean work of the retarded reservoir friction on the resolved coordinate is \begin{align*} W_{\rm fric}&=-\int_0^\tau\!\int_0^t v(t)^T\Gamma_N(t-s)v(s)\dd s\dd t\\ &=-\frac12\int_0^\tau\!\int_0^\tau v(t)^T\Gamma_N(t-s)v(s)\dd s\dd t\le0. \end{align*} The second equality uses evenness and matrix symmetry of the kernel; the sign uses the covariance identity, not an entropy inequality. The centered random force does zero mean work against this prescribed deterministic protocol. This is an equilibrium passivity statement relative to an initially prepared retarded response. It does not assert that instantaneous frictional power has a fixed sign at finite memory, or that a randomly evolving coupled subsystem loses energy monotonically. For its random velocity, the force and velocity become correlated, and the fluctuating-force work cannot be dropped. \subsubsection{The finite model stores the information that a Markov limit loses} Let $f_t$ be the full evolved law and let $\nu_t$ be its resolved phase-space marginal. Since the Hamiltonian flow is invertible and preserves $\Pi_N$, \begin{equation} D(f_t\Vert\Pi_N)=D(f_0\Vert\Pi_N). \label{b30:eq:full-kl-conserved} \end{equation} This follows by changing variables under the flow in the KL integral; it holds whenever the initial relative entropy is finite. Disintegration against the conditional canonical reservoir gives the exact chain rule \begin{equation} D(f_t\Vert\Pi_N)=D(\nu_t\Vert\pi_m)+C_N(t), \label{b30:eq:conditional-information} \end{equation} where \[ C_N(t)=\int D\!\left(f_t(\dd B\mid s)\middle\Vert \Pi_N(\dd B\mid s)\right)\nu_t(\dd s)\ge0. \] Here $s=(q,p)$ and $B=(x,P)$. The initial preparation gives $C_N(0)=0$, so \begin{equation} D(\nu_t\Vert\pi_m)+C_N(t)=D(\nu_0\Vert\pi_m). \label{b30:eq:information-transfer} \end{equation} The conditional term includes reservoir disturbance and correlations relative to its equilibrium conditional law. Because the canonical reservoir coordinates depend on $q$, it should not automatically be identified with ordinary mutual information in the original coordinates. Equation~\eqref{b30:eq:information-transfer} bounds the marginal entropy by its initial value but does not make it monotone in time. $C_N(t)$ can increase and later decrease. A finite reservoir can return information to the system. This exact accounting is the appropriate starting point for discussing the relation between microscopic reversibility and effective dissipation. \subsection{Continuum limit of the bath spectrum and controlled Markov reduction} \label{b30:sec:memory-limit} The exact memory equation is not yet an overdamped Markov diffusion. The latter requires selecting a bath spectrum and supplying a convergence argument. A Lorentzian spectral weight is particularly useful because its force covariance is exponential and its memory has a finite Markovian realization. This choice is a model specification, not a consequence of equilibrium alone. \subsubsection{A bath spectrum with a computable memory kernel} Let $G=G^T>0$ be the target friction matrix and $\varepsilon>0$ the memory time. Specify the matrix-valued spectral weight \begin{equation} S_\varepsilon(\omega)=\frac2\pi\frac{G}{1+\varepsilon^2\omega^2}, \qquad \omega\ge0. \label{b30:eq:spectrum} \end{equation} The cosine transform identity gives \begin{equation} \Gamma_\varepsilon(t)=\int_0^\infty S_\varepsilon(\omega) \cos(\omega t)\dd\omega =\frac G\varepsilon e^{-|t|/\varepsilon}. \label{b30:eq:exponential-kernel} \end{equation} In particular, \[ \int_0^\infty\Gamma_\varepsilon(t)\dd t=G, \qquad \int_{-\infty}^{\infty}\Gamma_\varepsilon(t)\dd t=2G. \] The factor of two is important. The causal memory convolution converges to $G\dot q$, whereas the two-sided force covariance converges to $2\theta G\delta$. Treating these two integrals as the same would produce the wrong Einstein coefficient. To realize the spectrum with finite baths, diagonalize $G$ and use oscillators in each eigenvector direction. Approximate the positive spectral measure in each direction by positive point masses $w_a$ at frequencies $\omega_a>0$ and choose, for example, $M_a=1$ and $c_a=\omega_a\sqrt{w_a}\,e_a$. Then $c_ac_a^T/(M_a\omega_a^2)=w_ae_ae_a^T$, exactly the weight required in \eqref{b30:eq:finite-kernel}. First let the frequency mesh tend to zero on a bounded interval, then let its cutoff tend to infinity, at a fixed $\varepsilon$. \subsubsection{A sufficient Gaussian convergence argument} Covariance convergence at individual times establishes only convergence of finite-dimensional Gaussian distributions. It does not by itself establish convergence in a continuous-path space. Add the following explicit tightness condition for every fixed horizon $\tau$: \begin{equation} \EE|\eta_N(t)-\eta_N(s)|^{2r}\le C_{\varepsilon,\tau,r}|t-s|^r, \qquad r>1. \label{b30:eq:gaussian-tightness} \end{equation} For the continuous spectral truncations this bound follows from \[ \EE|\eta(t)-\eta(s)|^2 =2\theta\Tr[\Gamma_\varepsilon(0)-\Gamma_\varepsilon(t-s)] \le C_\varepsilon|t-s|. \] Truncation only removes a nonnegative contribution to the scalar increment variance in each eigenvector direction. Gaussian moment identities then give \eqref{b30:eq:gaussian-tightness}. For a discrete frequency approximation, take mesh size at most one and put each bin's exact positive spectral mass at a representative frequency. For $h=t-s$, the estimate \[ |e^{i\omega_ah}-1|^2 \le2|e^{i\omega h}-1|^2+2|\omega_a-\omega|^2h^2 \] gives the same bound, uniformly along the mesh/cutoff sequence, up to $C_\varepsilon h^2$. Together with convergence of the covariance this proves tightness and weak convergence in $C([0,\tau];\RR^d)$. The solution map of the Volterra equation is continuous under uniform force convergence and $L^1$ kernel convergence on this finite interval: subtract the two integral equations, use the Lipschitz constant of $\nabla U$, and apply Gronwall's inequality. Thus the resolved paths converge in law at fixed $m,\varepsilon$. No assertion is needed that the infinite reservoir has a finite total canonical energy. Typically it does not. The rigorous object is the finite-bath sequence and its resolved finite-time limit. A normalized infinite-dimensional Lebesgue density $e^{-H/\theta}$ must not be introduced to bypass this issue. \subsubsection{The same force limit from a microcanonical bath preparation} The Gaussian bath variables in \eqref{b30:eq:bath-preparation} can be replaced by a uniform conditional energy-shell preparation, keeping the resolved nonequilibrium initial law independent of the standardized bath variables. In the energy coordinates $(a,b)$, sample uniformly on the sphere of radius $\sqrt{2N\theta}$ in $2N$ dimensions. This preparation has fixed bath energy $N\theta$ and covariance $\theta I$, so the finite-bath fluctuation--dissipation covariance \eqref{b30:eq:exact-fdr} still holds exactly. Its force is not exactly Gaussian at finite $N$. Let $g_N$ be a standard Gaussian vector in $\RR^{2N}$. The uniform sphere vector has the representation \[ \sqrt{2N\theta}\,\frac{g_N}{|g_N|} =R_N\sqrt\theta\,g_N, \qquad R_N=\frac{\sqrt{2N}}{|g_N|}\longrightarrow1 \quad\hbox{in probability}. \] The last statement follows from $|g_N|^2/(2N)\to1$ in probability, for example by its mean and variance. Since the exact reservoir force is linear in $(a,b)$, this coupling gives \[ \eta_N^{\rm mc}=R_N\eta_N^{\rm can}. \] The Gaussian force sequence at fixed $\varepsilon$ is tight in $C([0,\tau])$. Therefore $|R_N-1|\norm{\eta_N^{\rm can}}_\infty\to0$ in probability; independence of the two factors is not required. The microcanonical and canonical preparations have the same limiting force process and, by continuity of the Volterra solution map, the same resolved finite-time limit. This is an explicit ensemble-equivalence argument for the dynamical force, not just for one equilibrium marginal. The energy-shell measure is singular relative to the finite canonical bath density. Consequently the finite relative-entropy identity \eqref{b30:eq:information-transfer}, which used a canonical conditional preparation and finite initial KL, cannot be applied to this singular ensemble with $C_N(0)=0$. One may use the canonical preparation for that identity, or a finite-width energy shell with its nonzero initial conditional relative entropy retained. The limiting diffusion's entropy theorem is unaffected by this finite-ensemble distinction. Also, an integrable harmonic bath is not claimed to drive every arbitrary initial bath state to a microcanonical ensemble. The uniform sphere law is a specified preparation. \subsubsection{An exact exponential-memory realization} Let $\eta^\varepsilon$ be the stationary Ornstein--Uhlenbeck force \begin{equation} \dd\eta_t^\varepsilon=-\varepsilon^{-1}\eta_t^\varepsilon\dd t +\varepsilon^{-1}\sqrt{2\theta G}\dd W_t, \qquad \eta_0^\varepsilon\sim N(0,\theta G/\varepsilon). \label{b30:eq:colored-force} \end{equation} It is independent of the resolved initial state and has covariance $\theta G\varepsilon^{-1}e^{-|t-s|/\varepsilon}$, equal to the limiting force covariance derived from the Hamiltonian bath. Define \[ z_t^\varepsilon=\eta_t^\varepsilon -\int_0^t\frac G\varepsilon e^{-(t-s)/\varepsilon} \frac{p_s^\varepsilon}{m}\dd s. \] Differentiation of the convolution gives the exact system \begin{align} \dd q_t^\varepsilon&=p_t^\varepsilon\dd t/m,\notag\\ \dd p_t^\varepsilon&=-\nabla U(q_t^\varepsilon)\dd t+z_t^\varepsilon\dd t, \label{b30:eq:drude}\tag{D}\\ \dd z_t^\varepsilon&=-z_t^\varepsilon\dd t/\varepsilon -G p_t^\varepsilon\dd t/(m\varepsilon) +\sqrt{2\theta G}\dd W_t/\varepsilon.\notag \end{align} Here $z_0^\varepsilon=\eta_0^\varepsilon$. This stochastic embedding is used because its law equals that of the Gaussian continuum reservoir force. Starting by postulating \eqref{b30:eq:drude} without the preceding Hamiltonian elimination would not constitute a microscopic derivation. Its equilibrium density is proportional to \begin{equation} \exp\left[-\frac1\theta\left(U(q)+\frac{|p|^2}{2m} +\frac\varepsilon2 z^TG^{-1}z\right)\right]. \label{b30:eq:extended-gibbs} \end{equation} The conservative coupling cancels in the energy derivative: $(p/m)^Tz+\varepsilon z^TG^{-1}[-Gp/(m\varepsilon)]=0$. The remaining $z$-OU part has covariance $\theta G/\varepsilon$. The correct parity is $(q,p,z)\mapsto(q,-p,z)$; the force variable is even. Thus the equilibrium embedding is reversible under this parity, despite the appearance of a friction and a noise term. \subsubsection{White-memory limit at a fixed positive mass} \begin{theorem}[Controlled white-memory limit] \label{b30:thm:white-memory} Assume \eqref{b30:eq:potential-class}, fix $m>0$ and a finite horizon $\tau$, and give $(q_0,p_0)$ finite second moments. Couple \eqref{b30:eq:drude} and the kinetic Langevin equation \begin{equation} \dd q_t=p_t\dd t/m,\qquad \dd p_t=-\nabla U(q_t)\dd t-Gp_t\dd t/m+\sqrt{2\theta G}\dd W_t \label{b30:eq:kinetic-limit} \end{equation} with the same initial resolved state and Brownian motion. Then \[ \EE\sup_{0\le t\le\tau} (|q_t^\varepsilon-q_t|^2+|p_t^\varepsilon-p_t|^2)\longrightarrow0. \] In fixed dimensionless coordinates a bound of order $C_{m,\tau}\varepsilon[1+\log(1+\tau/\varepsilon)]$ is sufficient. Its constant is not asserted to be uniform as $m\downarrow0$. \end{theorem} \begin{proof} Multiplying the last equation of \eqref{b30:eq:drude} by $\varepsilon$ and integrating gives \[ \int_0^t z_s^\varepsilon\dd s =-G(q_t^\varepsilon-q_0)+\sqrt{2\theta G}W_t -\varepsilon(z_t^\varepsilon-z_0^\varepsilon). \] Therefore \begin{equation} p_t^\varepsilon=p_0-\int_0^t\nabla U(q_s^\varepsilon)\dd s -G(q_t^\varepsilon-q_0)+\sqrt{2\theta G}W_t+r_t^\varepsilon, \quad r_t^\varepsilon=-\varepsilon(z_t^\varepsilon-z_0^\varepsilon). \label{b30:eq:memory-remainder} \end{equation} The same equation for \eqref{b30:eq:kinetic-limit} has zero remainder. We first control the force maximum rather than replacing colored noise by white noise pointwise. On $[0,\tau]$, a stationary OU process \eqref{b30:eq:colored-force} satisfies \begin{equation} \EE\sup_{t\le\tau}|\varepsilon\eta_t^\varepsilon|^2 \le C\theta\norm{G}\varepsilon [1+\log(1+\tau/\varepsilon)]. \label{b30:eq:ou-maximum} \end{equation} One direct proof partitions time into intervals of length $\varepsilon$. The values at the endpoints have Gaussian tails with variance $O(\varepsilon)$ after multiplication by $\varepsilon$. A union bound over at most $2+\tau/\varepsilon$ endpoints yields an expected squared maximum $O(\varepsilon\log(2+\tau/\varepsilon))$. On one interval the OU increment is a deterministic decay plus a Gaussian stochastic integral. After scaling time by $\varepsilon$, the latter has a maximal tail bounded by $C\exp[-u^2/(C\varepsilon)]$, obtained from the Brownian maximal inequality after its deterministic time change. A second union bound and integration of this tail give the same order for the interval maxima. Summation over the fixed number of coordinate directions proves \eqref{b30:eq:ou-maximum}. The convolution representation of $z^\varepsilon$ gives \begin{equation} \varepsilon\sup_{s\le t}|z_s^\varepsilon| \le\varepsilon\sup_{s\le t}|\eta_s^\varepsilon| +\frac{\varepsilon\norm{G}}m\sup_{s\le t}|p_s^\varepsilon|. \label{b30:eq:z-bound} \end{equation} Let $V(t)=\sup_{s\le t}(|q_s^\varepsilon|+|p_s^\varepsilon|)$. The integral equations, linear growth of $\nabla U$, and \eqref{b30:eq:z-bound} imply, with constants depending on $m,\tau,G,U$, \[ V(t)\le C\left(1+|q_0|+|p_0|+\sup_{s\le t}|W_s| +\varepsilon\sup_{s\le t}|\eta_s^\varepsilon|\right) +C\int_0^tV(s)\dd s+C\varepsilon V(t). \] For sufficiently small $\varepsilon$, absorb the last term. Gronwall's inequality, the Brownian maximal second moment, and \eqref{b30:eq:ou-maximum} give $\sup_{\varepsilon\le\varepsilon_0}\EE V(\tau)^2<\infty$. Equations \eqref{b30:eq:z-bound}--\eqref{b30:eq:memory-remainder} now imply \[ \EE\norm{r^\varepsilon}_\infty^2 \le C_{m,\tau}\varepsilon[1+\log(1+\tau/\varepsilon)]. \] Finally, subtract the zero-remainder equation, and use $\delta q_t=\int_0^t\delta p_s\dd s/m$ and the Lipschitz bound on $\nabla U$. A second Gronwall estimate gives $\norm{\delta q}_\infty+\norm{\delta p}_\infty \le C_{m,\tau}\norm{r^\varepsilon}_\infty$. Taking squares and expectations proves the assertion. \end{proof} This proof derives both the damping sign and the matching white-noise amplitude without assuming an entropy inequality. It also explains why $\eta_t^\varepsilon$ itself has no finite pointwise white noise limit: its stationary variance diverges like $1/\varepsilon$. Only its time integral, or the resolved dynamics, has the stated limit. \subsubsection{Physical meaning of friction positivity} For a stationary force with integrable covariance $C(t)$ and the required equilibrium symmetry, the Green--Kubo candidate is \[ G=\frac1\theta\int_0^\infty C(t)\dd t. \] For every $v\in\RR^d$, \[ v^TGv=\lim_{T\to\infty}\frac1{2\theta T} \EE\left(\int_0^T v^T\eta_s\dd s\right)^2\ge0. \] To verify the equality, expand the variance as $2\int_0^T(T-u)v^TC(u)v\dd u$ and use integrability to pass to the limit. Strict positivity requires nonzero zero-frequency fluctuation power in every retained direction; it is not automatic. In the chosen spectrum it is exactly the hypothesis $G>0$. This gives a physical covariance origin for the positive metric. It does not assign a temporal orientation to that metric by itself. \subsection{Overdamped reduction and variational characterization} \label{b30:sec:overdamped} The white-memory limit retains momentum, whereas the present Rayleighian describes an overdamped position law. Removing momentum is therefore a second limiting step, with its own preparation and estimate. Deleting $m\ddot q$ from the equation would miss the initial momentum layer and does not justify that limit. \subsubsection{Small mass with a thermally scaled initial momentum} Set $K=G^{-1}$. Let $q_0$ have finite second moment and take $p_0^m\sim N(0,m\theta I)$ independently. This momentum scaling keeps the mean kinetic energy $d\theta/2$ finite and gives $\EE|p_0^m|^2=dm\theta\to0$. \begin{theorem}[Small-mass realization of the overdamped diffusion] \label{b30:thm:small-mass} Under \eqref{b30:eq:potential-class}, the position in \eqref{b30:eq:kinetic-limit} converges, on every fixed $[0,\tau]$, to the unique strong solution \begin{equation} \dd Q_t=-K\nabla U(Q_t)\dd t+\sqrt{2\theta K}\dd W_t, \qquad Q_0=q_0. \label{b30:eq:overdamped-sde} \end{equation} With the coupling just specified, $\EE\sup_{t\le\tau}|q_t^m-Q_t|^2\to0$. \end{theorem} \begin{proof} Integrating the momentum equation gives the exact identity \begin{equation} q_t^m=q_0-\int_0^tK\nabla U(q_s^m)\dd s +\sqrt{2\theta K}W_t-K(p_t^m-p_0^m). \label{b30:eq:small-mass-integral} \end{equation} We used $K\sqrt G=\sqrt K$, valid because $G$ is symmetric positive definite. For a nonsymmetric friction, the noise matrix must be handled separately; it is not obtained by this identity. Variation of constants gives \[ p_t^m=\xi_t^m-\int_0^t e^{-G(t-s)/m}\nabla U(q_s^m)\dd s, \] where \[ \xi_t^m=e^{-Gt/m}p_0^m +\int_0^t e^{-G(t-s)/m}\sqrt{2\theta G}\dd W_s. \] The chosen $p_0^m$ makes $\xi^m$ stationary with covariance $m\theta I$. The interval-maximal argument of \eqref{b30:eq:ou-maximum}, now with correlation time proportional to $m$, gives in fixed dimensionless units \[ \EE\sup_{t\le\tau}|\xi_t^m|^2 \le Cm[1+\log(1+\tau/m)]. \] If $g_0>0$ is the smallest eigenvalue of $G$, the convolution satisfies \[ \sup_{t\le\tau}\left|\int_0^t e^{-G(t-s)/m}\nabla U(q_s^m)\dd s\right| \le\frac m{g_0}\sup_{t\le\tau}|\nabla U(q_t^m)|. \] Linear growth of $\nabla U$ and \eqref{b30:eq:small-mass-integral} yield a Gronwall bound on $\EE\sup|q^m|^2$, uniformly for sufficiently small $m$, after the $Cm\sup|q^m|$ term is absorbed. Thus \[ \EE\sup_{t\le\tau}|p_t^m|^2 \le C_\tau\{m[1+\log(1+\tau/m)]+m^2\}\longrightarrow0. \] Subtracting the integral equation for $Q$ from \eqref{b30:eq:small-mass-integral} gives \[ |q_t^m-Q_t|\le\norm K\Lambda\int_0^t|q_s^m-Q_s|\dd s +\norm K(|p_t^m|+|p_0^m|). \] Gronwall and the preceding maximal estimate finish the proof. \end{proof} If a fixed nonzero initial momentum is used instead, the last term in \eqref{b30:eq:small-mass-integral} can produce a nonvanishing initial displacement during the momentum boundary layer. Uniform convergence from the same initial position cannot then be asserted without a modified statement. General state-dependent-friction limits have additional drift terms \cite{Hottovy2015}; the constant-friction theorem above avoids that extra complication. \subsubsection{The order of limits is part of the theorem} The proven construction has the order \begin{equation} \lim_{m\downarrow0}\ \lim_{\varepsilon\downarrow0}\ \lim_{\text{bath mesh}\downarrow0,\,\text{cutoff}\uparrow\infty} \BridgeLaw(q^{N,\varepsilon,m}_{[0,\tau]}) =\BridgeLaw(Q_{[0,\tau]}). \label{b30:eq:ordered-limits} \end{equation} The inner bath limit is at fixed $\varepsilon,m$. The estimates for the next limit may depend on $m$. The last limit uses the thermal momentum scaling. These statements do not establish commutation of limits, a simultaneous scaling for arbitrary parameter ratios, or uniform convergence as $\tau\to\infty$. A diagonal sequence can be chosen for each fixed finite-time weak-convergence criterion, but no uniform physical rate follows merely from a diagonal argument. Finite-bath recurrence is therefore not contradicted: a finite-time reservoir limit is taken before an unbounded observation horizon. Long-time and reservoir-size limits are consequently not interchanged without an additional estimate. \subsubsection{Generator, current, and equilibrium obtained from the limits} Equation~\eqref{b30:eq:overdamped-sde} has generator \[ Lf=-K\nabla U\cdot\nabla f+\theta K:D^2f. \] Its forward equation and current are \begin{equation} \partial_t\rho=-\nabla\cdot j_\rho, \qquad j_\rho=-\rho K\nabla U-\theta K\nabla\rho. \label{b30:eq:derived-current} \end{equation} The positive covariance $2\theta K$ fixes the sign of the second-order term. The restoring drift was obtained from the Hamiltonian potential and the retarded reservoir response. Neither sign was inferred from a desired entropy-production inequality. For $\pi=Z^{-1}e^{-U/\theta}$, $\theta\nabla\pi=-\pi\nabla U$ gives $j_\pi=0$. The globally Lipschitz drift defines a unique conservative diffusion; the corresponding well-posed forward equation identifies $\pi$ as an invariant probability. Equivalently, integration by parts gives \[ \int fLg\,\pi\dd q=-\theta\int\nabla f^TK\nabla g\,\pi\dd q \] on a suitable core, followed by the conservative realization. The formal equation $L^*\pi=0$ alone is not substituted for that realization. \subsubsection{The Rayleighian becomes a derived characterization} For each fixed smooth positive density, \[ j_\rho=-\rho K\nabla(U+\theta\log\rho) =-\rho K\nabla\mu_\rho. \] Assume $\int\rho|\nabla\mu_\rho|_K^2\dd q<\infty$ and minimize over the weighted Hilbert space of measurable currents with $\int j^TK^{-1}j/\rho\dd q<\infty$. The linear term is continuous on this space by Cauchy--Schwarz. No prescribed density derivative or two-endpoint constraint is imposed in this instantaneous constitutive minimum. Completing the square gives \begin{equation} \mathscr R_\rho[j] =\frac12\int\frac{|j+\rho K\nabla\mu_\rho|_{K^{-1}}^2}{\rho}\dd q -\frac12\int\rho|\nabla\mu_\rho|_K^2\dd q. \label{b30:eq:derived-rayleighian} \end{equation} The Hamiltonian reservoir reduction already determined $j_\rho$. Equation~\eqref{b30:eq:derived-rayleighian} shows that the same current uniquely minimizes $\Phi+\dot F$. In this microscopic realization, the Rayleighian need not be an independent dynamical postulate. This complements the mesoscopic route in which it is postulated and its consequences are proved. Onsager--Rayleigh formulations and free-energy gradient flows have established antecedents \cite{Doi2011,JKO1998}. \subsubsection{Free-energy identity and physical entropy balance} Work on the class of smooth positive probability densities with finite absolute energy and entropy integrals and with integrable current boundary terms. One convenient initial class is a smooth bounded positive perturbation of $\pi$ with controlled derivatives; the corresponding diffusion and cutoff estimates justify the integrations. If only weak convergence is available, use the appropriate entropy inequality instead of silently asserting a differentiated identity. Since $\log\pi=-U/\theta-\log Z$, \[ \theta D(\rho\Vert\pi) =\theta\int\rho\log\rho\dd q+\int U\rho\dd q+\theta\log Z. \] Also $F[\pi]=-\theta\log Z$. Thus \begin{equation} F[\rho]-F[\pi]=\theta D(\rho\Vert\pi). \label{b30:eq:free-energy-identity} \end{equation} Using mass conservation and \eqref{b30:eq:derived-current}, \begin{align} \dot F &=\int\mu_\rho\partial_t\rho\dd q =\int\nabla\mu_\rho\cdot j_\rho\dd q\notag\\ &=-\int\rho|\nabla\mu_\rho|_K^2\dd q =-\theta^2\int\rho|\nabla\log(\rho/\pi)|_K^2\dd q. \label{b30:eq:derived-dissipation} \end{align} No entropy sign was used to obtain the diffusion in the preceding sections. Equation~\eqref{b30:eq:derived-dissipation} is now a consequence. Positivity of $K$ and connectedness imply that a vanishing integrand forces $\rho=\pi$. For an autonomous system in contact with one isothermal reservoir and no work input, declare heat into the reservoir by $\dot Q_{\rm bath}=-\dd\EE U/\dd t$ after the kinetic reduction. Let $S_{\rm sys}=-k_B\int\rho\log\rho$. Then direct algebra yields \begin{equation} \dot S_{\rm sys}+\frac{\dot Q_{\rm bath}}T =-\frac{\dot F}T =k_B\theta\int\rho|\nabla\log(\rho/\pi)|_K^2\dd q\ge0. \label{b30:eq:physical-balance} \end{equation} The use of $Q/T$ as reservoir entropy assumes the isothermal reservoir limit. It is not an exact assertion that a finite disturbed oscillator bath remains in equilibrium. Before the small-mass reduction, the resolved energy is $U+|p|^2/(2m)$ and the corresponding kinetic entropy dissipation acts in momentum. These two balances must not be mixed. For the kinetic SDE, the Stratonovich energy identity is \[ \dd\left(U(q)+\frac{|p|^2}{2m}\right) =\frac{p^T}m\left(-\frac{Gp}m\dd t+\sqrt{2\theta G}\circ\dd W_t\right). \] The equivalent It\^o mean energy rate contains $\theta\Tr(G)/m$ in addition to $-\EE[p^TGp]/m^2$. The fluctuation term compensates friction at equilibrium; friction alone must not be interpreted as positive stationary total entropy production. \subsubsection{Two extensions requiring different statements} \paragraph{Position-dependent mobility.} For an isothermal symmetric $K(q)$, the Gibbs-preserving It\^o drift is \begin{equation} b_i(q)=-K_{ij}(q)\partial_jU(q)+\theta\partial_jK_{ij}(q). \label{b30:eq:noise-induced-drift} \end{equation} Indeed the current is $J_i=b_i\rho-\theta\partial_j(K_{ij}\rho) =-K_{ij}(\rho\partial_jU+\theta\partial_j\rho)$. Leaving out the second term in \eqref{b30:eq:noise-induced-drift} leaves an unwanted current $-\theta\pi\,\partial_jK_{ij}$ at $\pi$. Thus a state-dependent-friction generalization needs an additional limit theorem, not a symbol replacement in the constant-matrix proof. \paragraph{A periodic configuration space.} The linear coupling $c_a^Tq$ in \eqref{b30:eq:hamilton-bath} is globally defined on $\RR^d$, not on $\mathbb T^d$. Therefore this microscopic realization directly supports the confined Euclidean model above. It does not establish a microscopic realization on a torus. Such a realization needs periodic coupling functions, or a separately defined lifted/quotient model, and its own limiting calculation. Section~\ref{b30:sec:torus-realization} supplies that calculation using periodic paired couplings and a separate fixed-mass convergence estimate. The Euclidean and periodic proofs therefore keep their distinct couplings and hypotheses. \subsection{Where the effective time direction enters} \label{b30:sec:time-information} The sign question can now be answered explicitly. The positive metric is generated by equilibrium force fluctuations. The free-energy reference is generated by the canonical marginal. Their positivity and static form do not alone orient time. The effective orientation arises when initial preparation is propagated by a retarded response and the reservoir memory is removed in a specified forward description. \subsubsection{Physical inputs and their consequences} The inputs in the microscopic realization are: \begin{enumerate} \item A reversible Hamiltonian, specified potential, and resolved system--reservoir split. \item A conditional canonical reservoir distribution at the preparation time and an independently specified nonequilibrium resolved distribution. \item A reservoir spectrum with a positive integrated response and a finite correlation time, followed by the stated scale limits. \item Evolution from the preparation time using the retarded oscillator solution, with no separately arranged future system--reservoir correlations. \item The choice of retained variables and the time resolution at which a Markov description is asserted. \end{enumerate} The drift, diffusion coefficient, current, Rayleighian characterization, and entropy inequality are consequences for this model. Temperature, spectral shape, preparation, and the resolved variables remain physical inputs. Describing these inputs as if they had all been deduced from equilibrium would create a new unsupported claim. \subsubsection{The reverse experiment has a different preparation} At a later time $\tau$, the full state is $f_\tau=\Phi_\tau{}_*f_0$. A reversed microscopic experiment starts from $\Theta_*f_\tau$. It contains the reservoir conditional disturbance measured by $C_N(\tau)$ in \eqref{b30:eq:conditional-information}. It generally does not have the fresh conditional canonical bath preparation used at $0$. Resetting the bath to that canonical conditional law discards the information needed for exact reversal. This is a physically specified change of preparation, not a contradiction in Hamiltonian mechanics. An equilibrium bath by itself is time-reversal symmetric. If the full initial preparation is invariant under $\Theta$, an even resolved observable has the same law at $t$ and $-t$ under the exact reversible flow. Thus a prepared nonequilibrium state at a distinguished time does not identify an absolute orientation of all microscopic time. It can support relaxation away from that preparation time in either temporal direction when the two experiments are posed with the corresponding boundary data. A cosmological initial-condition theory or a fundamental time-asymmetric law is an additional problem. \subsubsection{Why reverse diffusion is not anti-diffusion} For the derived overdamped SDE with $b=-K\nabla U$ and covariance $2\theta K$, use smooth positive densities on $[0,\tau]$ for which the score and the reversed SDE are well defined, nonexplosive, and in the uniqueness class of the forward equation. The nonstationary time reverse over $[0,\tau]$ then has drift \begin{equation} b_s^R(q)=-b(q)+2\theta K\nabla\log\rho_{\tau-s}(q). \label{b30:eq:reverse-drift} \end{equation} The reversed initial law is $\rho_\tau$. Its diffusion covariance remains $2\theta K$, and its current is \[ b_s^R\rho_{\tau-s}-\theta K\nabla\rho_{\tau-s} =-b\rho_{\tau-s}+\theta K\nabla\rho_{\tau-s} =-j_{\rho_{\tau-s}}. \] This verifies the reversal at the level of the continuity equation. For path-law reversal, the same formula follows from the reversed transition density $p(s,x;t,y)\rho_s(x)/\rho_t(y)$ by expanding its conditional first and second moments; regularity is needed for this expansion. A reversed one-time density equation alone would not determine a path law without the Markov and uniqueness hypotheses. At equilibrium, $\nabla\log\pi=-\nabla U/\theta$, so $b^R=b$. Away from equilibrium the reverse drift depends on the forward marginal family and generally differs from the original autonomous drift. Merely changing $K\nabla U$ to its negative while retaining the same score term would not describe this reversal. The equilibrium-preserving Markov-inverse obstruction has a precise meaning. If $D(\rho_\tau\Vert\pi)0$, and let $G=G^T>0$. Consider the finite reversible Hamiltonians \eqref{b30:eq:hamilton-bath}, prepared with a resolved position law $\rho_0$ of finite second moment, thermal momentum $N(0,m\theta I)$, and the conditional canonical bath \eqref{b30:eq:bath-preparation}. Choose their positive spectral weights to approximate \eqref{b30:eq:spectrum}, with the covariance convergence and path tightness described in Section~\ref{b30:sec:memory-limit}. Then, on every fixed finite time interval, the resolved position laws converge in the order \eqref{b30:eq:ordered-limits} to \[ \dd Q_t=-K\nabla U(Q_t)\dd t+\sqrt{2\theta K}\dd W_t, \qquad K=G^{-1}. \] The limiting invariant law is $\pi=Z^{-1}e^{-U/\theta}\dd q$. For every smooth positive density in the stated finite-energy, finite-entropy and integration-by-parts class, with $\int\rho|\nabla\mu_\rho|_K^2\dd q<\infty$, the limiting current uniquely minimizes \[ \mathscr R_\rho[j] =\frac12\int\frac{j^TK^{-1}j}{\rho}\dd q +\int j\cdot\nabla(U+\theta\log\rho)\dd q. \] For its regular density evolution, \[ F[\rho_t]-F[\pi]=\theta D(\rho_t\Vert\pi),\qquad \dot F[\rho_t]=-\theta^2\int\rho_t |\nabla\log(\rho_t/\pi)|_K^2\dd q. \] With the autonomous single-isothermal-reservoir heat convention, the associated mean total entropy-production rate is $-\dot F/T\ge0$. No entropy monotonicity assumption is used in the Hamiltonian elimination or either stochastic limit. The assertion concerns this preparation, reservoir class, order of limits, and density domain. \end{theorem} \begin{proof} The canonical integration gives \eqref{b30:eq:canonical-marginal}. The exact Hamiltonian elimination is Theorem~\ref{b30:thm:exact-gle}. The positive spectral approximation converges to the Gaussian exponential-memory force and its Volterra solution. Theorem~\ref{b30:thm:white-memory} then gives the kinetic SDE at each fixed mass, and Theorem~\ref{b30:thm:small-mass} gives the position diffusion. Its well-posed forward equation and zero Gibbs current establish invariance. Equation~\eqref{b30:eq:derived-rayleighian} gives the current minimizer; \eqref{b30:eq:free-energy-identity} and \eqref{b30:eq:derived-dissipation} give the entropy conclusions on the stated regularity class. Finally \eqref{b30:eq:physical-balance} gives the heat-bath interpretation. The roles of the finite-time limits and the density-domain hypotheses are distinct; weak path convergence alone has not been used to pass an entropy derivative through a limit. \end{proof} \section{A conditional microscopic bridge and the scope of the arrow of time} \label{sec:scope-closure30} The reservoir construction above has two logically distinct tasks. The first is constructive: under a specified preparation, spectrum, and order of limits, it produces the overdamped diffusion used in the variational calculation. The second concerns scope: whether the same conclusion follows from reversibility alone for every microscopic system. The oscillator example shows that it does not. We now collect the hypotheses and conclusions to make this distinction precise. \subsection{Why reversibility alone cannot imply the proposed diffusion} The following proposition is a logical obstruction, rather than a technical limitation of the harmonic construction. A finite, smooth, exactly reversible model already disproves a universal inference from reversibility to overdamped diffusion. \begin{proposition}[No universal bridge from reversibility alone] \label{prop:no-universal-microscopic-bridge} There is no implication of the form \[ \begin{aligned} &\text{every finite reversible Hamiltonian system}\\[-2pt] &\qquad\Longrightarrow\quad \dd Q_t=-K\nabla U(Q_t)\dd t+\sqrt{2\theta K}\dd W_t \end{aligned} \] for a fixed positive matrix $K$ and a strictly decreasing observed relative entropy. In particular, the harmonic oscillator of Proposition~\ref{b30:prop:oscillator} is a counterexample: its microscopic flow is reversible, its full relative entropy is constant, and its observed position relative entropy is periodic. \end{proposition} \begin{proof} The oscillator has \[ D(\operatorname{Law}(q_t)\Vert\pi_q)=\frac{\omega^2a^2}{2\theta}\cos^2(\omega t). \] At $t=\pi/(2\omega)$ this quantity is zero, whereas at $t=\pi/\omega$ it equals its strictly positive initial value. A strictly decreasing Lyapunov function cannot have this behaviour. The contradiction arises before any issue about differentiating an entropy through a limit; it is a finite, smooth, exactly reversible model. Therefore additional conditions on the system--reservoir split, preparation, response kernel, observation scale, and limiting procedure are logically necessary. \end{proof} The counterexample identifies the scope of the constructive result: the proposed diffusion follows from an admissible reservoir realization, with the additional hypotheses stated next. \subsection{The admissible class for the positive bridge} We collect the assumptions used by the positive construction so that each can be checked in a concrete model. They specify the dynamics, preparation, and limit procedure without using the entropy-production conclusion. \begin{assumption}[Admissible reversible-reservoir bridge] \label{ass:admissible-bridge30} Fix a finite horizon $\tau<\infty$. A sequence of finite Hamiltonian models of the form \eqref{b30:eq:hamilton-bath} is called admissible on $[0,\tau]$ if the following conditions hold. \begin{enumerate}[label=(A\arabic*)] \item The Hamiltonian flow $\Phi_t^{N}$ is globally well posed, preserves phase volume and the Hamiltonian, and has a time-reversal involution $\Theta$ with $\Theta\Phi_t^{N}\Theta^{-1}=\Phi_{-t}^{N}$. \item The resolved potential satisfies the regularity and confinement assumptions in \eqref{b30:eq:potential-class}; the resolved initial position has finite second moment, and the initial momentum has the thermal scaling $p_0^m\sim N(0,m\theta I)$. \item At the preparation time the reservoir is conditionally canonical in the shifted variables of \eqref{b30:eq:bath-preparation}. Consequently the eliminated force is centered, independent of the resolved initial state in those variables, and satisfies the finite-bath fluctuation--dissipation identity \eqref{b30:eq:exact-fdr}. \item At fixed memory time $\varepsilon$ the positive bath spectra converge to \eqref{b30:eq:spectrum}. The force laws are tight in $C([0,\tau];\mathbb R^d)$ and their bounded--Lipschitz distance from the exponential-memory force tends to zero. No rate is assumed for this bath-discretisation error. \item The white-memory and small-mass estimates are uniform on the fixed horizon in the form proved in Theorems~\ref{b30:thm:white-memory} and \ref{b30:thm:small-mass}. The limits are taken in the order \eqref{b30:eq:ordered-limits}. \end{enumerate} \end{assumption} Condition (A4) is deliberately stated in terms of path laws. Covariance convergence at finitely many times is insufficient for a path-space result, whereas the tightness estimate \eqref{b30:eq:gaussian-tightness} supplies the missing compactness for the present harmonic realization. Condition (A5) also prevents a finite-time estimate from being silently promoted to a long-time statement. For a non-Gaussian reservoir, verifying (A4) is the place where a central-limit, martingale, or homogenisation theorem must be proved; it is not supplied by the fluctuation--dissipation identity alone. \subsection{Conditional bridge theorem} Let $\mathsf d_{\mathrm{BL}}$ denote the bounded--Lipschitz distance on probability laws over $C([0,\tau];\mathbb R^d)$, with the supremum norm truncated at one. Write $q^{N,\varepsilon,m}$ for the resolved path of a finite bath, $q^{\varepsilon,m}$ for the exponential-memory realization, $q^m$ for the white-memory kinetic process, and $Q$ for the overdamped limit. \begin{theorem}[Conditional microscopic bridge to the Rayleighian diffusion] \label{thm:conditional-microscopic-bridge30} Under Assumption~\ref{ass:admissible-bridge30}, let $G=G^{\mathsf T}>0$ and $K=G^{-1}$. Then, for every fixed $\tau<\infty$, \[ \operatorname{Law}(q^{N,\varepsilon,m}_{[0,\tau]}) \Longrightarrow \operatorname{Law}(Q_{[0,\tau]}) \] in the ordered limits \[ N\to\infty,\qquad \varepsilon\downarrow0,\qquad m\downarrow0, \] where \begin{equation} \dd Q_t=-K\nabla U(Q_t)\dd t+\sqrt{2\theta K}\dd W_t. \label{eq:scope-overdamped-limit30} \end{equation} The limiting invariant law is \[ \pi(\dd q)=Z^{-1}e^{-U(q)/\theta}\dd q, \] and the forward current for every density in the regularity class of Theorem~\ref{b30:thm:combined-bridge} is \[ j_\rho=-\rho K\nabla U-\theta K\nabla\rho =-\rho K\nabla(U+\theta\log\rho). \] It is the unique minimizer of \[ \mathscr R_\rho[j] =\frac12\int\frac{j^{\mathsf T}K^{-1}j}{\rho}\dd q +\int j\cdot\nabla(U+\theta\log\rho)\dd q. \] Consequently, \begin{equation} \frac{\dd}{\dd t}\left[\theta D(\rho_t\Vert\pi)\right] =-\theta^2\int\rho_t \left|\nabla\log\frac{\rho_t}{\pi}\right|_K^2\dd q\le0. \label{eq:scope-free-energy-dissipation30} \end{equation} No entropy monotonicity assumption is used in the Hamiltonian elimination, the white-memory limit, or the small-mass limit. The theorem is conditional on (A1)--(A5) and does not assert the implication for systems outside this class. \end{theorem} \begin{proof} The shifted canonical preparation gives the exact generalized Langevin equation and the fluctuation--dissipation identity in Theorem~\ref{b30:thm:exact-gle}. By (A4), the finite-bath force and the exponential-memory force have the same path-law limit at fixed $(\varepsilon,m)$. The Volterra solution map is continuous on the fixed horizon, so the corresponding resolved paths have the same limit. For fixed $m>0$, Theorem~\ref{b30:thm:white-memory} gives the kinetic Langevin equation. Its remainder estimate is obtained directly from \[ \int_0^t z_s^\varepsilon\dd s =-G(q_t^\varepsilon-q_0)+\sqrt{2\theta G}W_t -\varepsilon(z_t^\varepsilon-z_0^\varepsilon), \] and the stationary OU maximum bound \eqref{b30:eq:ou-maximum}; no sign of an entropy derivative enters this estimate. Theorem~\ref{b30:thm:small-mass} then gives the overdamped equation \eqref{eq:scope-overdamped-limit30}, with the thermal momentum scaling removing the initial momentum layer. The limiting generator is \[ Lf=-K\nabla U\cdot\nabla f+\theta K:D^2f. \] Since $\theta\nabla\pi=-\pi\nabla U$, its equilibrium current is zero. For $\mu_\rho=U+\theta\log\rho$, completing the square gives \[ \mathscr R_\rho[j] =\frac12\int\frac{|j+\rho K\nabla\mu_\rho|_{K^{-1}}^2}{\rho}\dd q -\frac12\int\rho|\nabla\mu_\rho|_K^2\dd q. \] Thus the displayed current is the unique minimizer. Finally, \[ F[\rho]-F[\pi]=\theta D(\rho\Vert\pi), \qquad \dot F[\rho_t]=\int j_{\rho_t}\cdot\nabla\mu_{\rho_t}\dd q, \] and substitution of the current yields \eqref{eq:scope-free-energy-dissipation30}. The proof uses the Hamiltonian equations, covariance convergence, and positive quadratic forms; the entropy sign appears only in this last calculation. \end{proof} \subsection{Finite-scale error decomposition} The ordered-limit theorem is qualitative unless a bath approximation rate is also established. The following decomposition records exactly what is and is not available from the estimates above. Define \[ \delta_{N,\varepsilon,m,\tau} :=\mathsf d_{\mathrm{BL}} \bigl(\operatorname{Law}(q^{N,\varepsilon,m}_{[0,\tau]}), \operatorname{Law}(q^{\varepsilon,m}_{[0,\tau]})\bigr). \] Then $\delta_{N,\varepsilon,m,\tau}\to0$ for fixed $(\varepsilon,m,\tau)$ by (A4), but no numerical rate is assigned without a specified quadrature or spectral discretisation theorem. The proved white- memory and small-mass estimates imply \begin{align} \mathsf d_{\mathrm{BL}} \bigl(\operatorname{Law}(q^{N,\varepsilon,m}_{[0,\tau]}),\operatorname{Law}(Q_{[0,\tau]})\bigr) &\le \delta_{N,\varepsilon,m,\tau} +C_{m,\tau}\sqrt{\varepsilon\bigl[1+\log(1+\tau/\varepsilon)\bigr]} \notag\\ &\quad +C_{\tau}\sqrt{m\bigl[1+\log(1+\tau/m)\bigr]+m^2}. \label{eq:scope-error-decomposition30} \end{align} The first term is a model-dependent bath error; the second is the proved white-memory error at fixed mass; the third is the proved small-mass error. The constants may depend on the displayed horizon and coefficients. In particular, \eqref{eq:scope-error-decomposition30} does not claim a uniform rate in $m$, a simultaneous scaling, or convergence of relative entropies and Fisher informations. Weak path convergence does not by itself justify passing an entropy derivative through either singular limit. \subsection{No absolute cosmic orientation from symmetric premises} The preceding bridge selects a direction relative to a preparation time and a retarded response convention. It cannot select an absolute orientation from premises that are themselves invariant under time reversal. \begin{proposition}[No absolute arrow from time-reversal-symmetric premises] \label{prop:no-absolute-arrow30} Let $\mathcal P$ be a class of microscopic model descriptions invariant under the reversal operation \[ \mathcal R(\Phi)_t=\Theta\Phi_{-t}\Theta^{-1}. \] Suppose an orientation selector $a:\mathcal P\to\{-1,0,+1\}$ depends only on the reversal-invariant premises, so that $a(\mathcal R\Phi)=a(\Phi)$, while physical covariance of an oriented statement requires \[ a(\mathcal R\Phi)=-a(\Phi). \] Then $a(\Phi)=0$ for every $\Phi\in\mathcal P$. Hence a nonzero absolute time orientation requires additional time-asymmetric data, such as a boundary preparation, a retarded-response condition, an explicitly time-asymmetric microscopic law, or a cosmological boundary condition. \end{proposition} \begin{proof} The two transformation rules give \[ a(\Phi)=a(\mathcal R\Phi)=-a(\Phi). \] Therefore $2a(\Phi)=0$. Since the codomain is discrete, $a(\Phi)=0$. The conclusion is a symmetry statement and does not depend on the particular form of the entropy or on the existence of a Markov limit. \end{proof} This proposition does not deny a time arrow in a prepared experiment. It states that the arrow cannot be an output of premises that contain no datum capable of distinguishing the two orientations. Positive dissipation and a positive mobility are compatible with both orientations until the preparation and response convention are fixed. \subsection{Conditional thermodynamic orientation} We can now state precisely what the reservoir construction does select. Let $f_0$ be the full prepared Hamiltonian law and let $f_\tau=\Phi_\tau{}_*f_0$. The exact reverse experiment starts from $\Theta_*f_\tau$. A fresh forward experiment at the observed state instead uses \[ \widehat f_0^{\,R}(\dd s\dd B) =\nu_\tau(\dd s)\,\Pi_N(\dd B\mid s), \] whereas the exact reverse contains the hidden reservoir disturbance and correlations inherited from $\Theta_*f_\tau$. The KL chain rule gives \begin{equation} D(f_\tau\Vert\Pi_N) =D(\nu_\tau\Vert\pi_m)+C_N(\tau), \qquad C_N(\tau)\ge0. \label{eq:scope-hidden-correlation30} \end{equation} \begin{theorem}[Conditional selection of the thermodynamic orientation] \label{thm:conditional-orientation30} Assume the admissible bridge hypotheses and prepare the reservoir at time $0$ by the conditional canonical law. Evolve the full Hamiltonian system with the retarded solution, observe the resolved coordinate, and take the ordered finite-time limits of Theorem~\ref{thm:conditional-microscopic-bridge30}. Then the forward effective law satisfies \[ \frac{\dd}{\dd t}\left[\theta D(\rho_t\Vert\pi)\right] =-\theta^2\int\rho_t \left|\nabla\log(\rho_t/\pi)\right|_K^2\dd q\le0. \] The selected orientation is therefore the composite protocol \[ \boxed{\begin{gathered} \text{preparation}+\text{retarded response}\\ +\text{hidden correlations}+\text{ordered Markov limit} \end{gathered}} \] The exact microscopic reverse is obtained from $\Theta_*f_\tau$ and retains the conditional information in \eqref{eq:scope-hidden-correlation30}. If $C_N(\tau)>0$, a fresh canonical reverse preparation is a different experiment and cannot be identified with the exact reverse. The theorem does not assert an absolute cosmic time direction. \end{theorem} \begin{proof} The first assertion is exactly the conclusion of the conditional bridge theorem. For the second assertion, Hamiltonian reversibility gives \[ \Phi_t{}_*(\Theta_*f_\tau)=\Theta_*f_{\tau-t}, \] so the exact reverse must carry the hidden conditional law at the final time. Replacing it by $\widehat f_0^{\,R}$ resets that conditional law. By the chain rule \eqref{eq:scope-hidden-correlation30}, a positive value of $C_N(\tau)$ records a nonzero discrepancy in the full relative-entropy accounting. The replacement is therefore not a reversal of the same prepared Hamiltonian trajectory ensemble. The forward entropy sign follows only after the response and ordered limits have produced the autonomous diffusion; it is not used to justify the preparation or the limits. \end{proof} The construction therefore identifies the data that select a forward thermodynamic description: a preparation time, a retarded response, and the stated effective limit. It does not derive these data from time-reversal symmetry alone. Extending the result to other microscopic systems requires verifying their coupling, preparation, and convergence hypotheses; a cosmological orientation requires additional boundary or dynamical information. \subsection{A periodic microscopic realization of the compact model} \label{b30:sec:torus-realization} The linear coupling used above is not a globally defined function on a torus. A periodic coupling supplies a separate realization of the compact current model. This construction permits any constant symmetric positive friction matrix; no convexity of the periodic potential is required. Write $\mathbb T^d=\mathbb R^d/\mathbb Z^d$ in fixed period-one coordinates and use normalized Haar volume. Let $U\in C^\infty(\mathbb T^d)$, $\theta=k_BT>0$, and $G=G^T>0$. \subsubsection{Periodic couplings with a constant friction matrix} \begin{lemma}[Positive integer-direction decomposition] \label{b30:torus:integer-decomposition} There are finitely many $k_\ell\in\mathbb Z^d\setminus\{0\}$ and $a_\ell>0$ such that $G=\sum_{\ell=1}^L a_\ell k_\ell k_\ell^T$. Consequently there are $J=2L$ smooth periodic scalar functions $h_\alpha$ whose gradient matrix $B(q)=(\nabla h_1(q),\ldots,\nabla h_J(q))$ satisfies $B(q)B(q)^T=G$ at every $q$. \end{lemma} \begin{proof} Choose an invertible rational matrix $R$ sufficiently close to $G^{1/2}$. The matrix $C=R^{-1}GR^{-T}$ is then sufficiently close to $I$ that $C_{ii}>\sum_{j\ne i}|C_{ij}|$ for every $i$. This choice is possible by continuity of matrix inversion and density of rational matrices. The identity \[ C=\sum_i\left(C_{ii}-\sum_{j\ne i}|C_{ij}|\right)e_ie_i^T +\sum_{i0$, set \begin{equation} H_N=\frac{|p|^2}{2m}+U(q) +\sum_{\alpha=1}^J\sum_{a=1}^N \left[\frac{P_{\alpha a}^2}{2M_a} +\frac{M_a\omega_a^2}{2} \left(x_{\alpha a}-\frac{\kappa_a h_\alpha(q)}{M_a\omega_a^2}\right)^2\right]. \label{b30:torus:hamiltonian} \end{equation} This is globally defined on $T^*\mathbb T^d$ times the reservoir phase space. Its reversal changes $p,P$ to $-p,-P$. Smoothness, periodicity, and energy conservation give a complete finite-dimensional flow: on each energy surface $p$ and the shifted reservoir coordinates are bounded, and the position remains on a compact space. Prepare $q_0$ with an assigned probability law, $p_0$ thermally when taking the small-mass limit, and the bath conditionally on $q_0$ by independent centered Gaussian variables \[ y_{\alpha a}=x_{\alpha a} -\frac{\kappa_a h_\alpha(q_0)}{M_a\omega_a^2}, \qquad \mathbb E y_{\alpha a}^2=\frac\theta{M_a\omega_a^2}, \qquad \mathbb E P_{\alpha a}^2=M_a\theta. \] In these shifted coordinates the bath variables are independent of $(q_0,p_0)$. Translation in each $x_{\alpha a}$ shows directly that the finite canonical marginal is proportional to $\exp[-(U(q)+|p|^2/(2m))/\theta]$. Solving each driven oscillator and integrating its convolution by parts, now with $h_\alpha(q_s)$ in place of a linear coordinate, gives \begin{align} \dot q_t&=p_t/m,\notag\\ \dot p_t&=-\nabla U(q_t)+B(q_t)\eta_N(t) -B(q_t)\int_0^t\gamma_N(t-s)B(q_s)^T\frac{p_s}{m}\dd s, \label{b30:torus:exact-memory}\\ \gamma_N(t)&=\sum_a\frac{\kappa_a^2}{M_a\omega_a^2}\cos(\omega_at), \notag\\ (\eta_N)_\alpha(t)&=\sum_a\kappa_a \left[y_{\alpha a}(0)\cos(\omega_at) +\frac{P_{\alpha a}(0)}{M_a\omega_a}\sin(\omega_at)\right].\notag \end{align} The endpoint containing $h_\alpha(q_t)$ cancels the quadratic counterterm. The endpoint at zero combines with the prepared shift. Since $d h_\alpha(q_s)/ds=\nabla h_\alpha(q_s)^Tp_s/m$, the remaining convolution is exactly the displayed one. Independence and $\cos u\cos v+\sin u\sin v=\cos(u-v)$ give \[ \mathbb E[\eta_N(t)\eta_N(s)^T]=\theta\gamma_N(t-s)I_J. \] The memory force is nonlinear in the resolved path; it is not replaced by a constant kernel before taking the short-memory limit. Approximate the positive scalar spectrum $2/[\pi(1+\varepsilon^2\omega^2)]$ by positive point masses and set $\kappa_a^2/(M_a\omega_a^2)$ equal to those masses. At fixed $m,\varepsilon$, the force converges in $C([0,\tau];\mathbb R^J)$ to the stationary Gaussian force with covariance $\theta\varepsilon^{-1}e^{-|t-s|/\varepsilon}I_J$. Covariance convergence is accompanied by the Gaussian increment bound $\mathbb E|\eta_N(t)-\eta_N(s)|^{2r}\le C|t-s|^r$, $r>1$. Positive spectral truncation and frequency quantization give this bound as in the preceding scalar-kernel construction. For completeness, the nonlinear Volterra solution map remains continuous in the force and in the $L^1$ kernel on a fixed interval. Boundedness of $B$ and $\nabla U$ gives \[ |p_t|\le |p_0|+C\tau+C\int_0^\tau|\eta(s)|\dd s +\frac C m\|\gamma\|_{L^1(0,\tau)}\int_0^t|p_s|\dd s. \] Gronwall bounds $p$ for bounded force and kernel sets. Subtracting two equations on that bounded set, using the bounded derivatives of $B$ and $\nabla U$, and applying Gronwall again proves continuity. Thus continuous-path convergence of the Gaussian forces transfers to weak convergence of the resolved paths at fixed $m,\varepsilon$. \subsubsection{The short-memory limit with periodic coupling} The limiting memory equation has the exact embedding \begin{align} \dd q_t^\varepsilon&=p_t^\varepsilon\dd t/m,\notag\\ \dd p_t^\varepsilon&=-\nabla U(q_t^\varepsilon)\dd t +B(q_t^\varepsilon)z_t^\varepsilon\dd t, \label{b30:torus:embedding}\\ \dd z_t^\varepsilon&=-z_t^\varepsilon\dd t/\varepsilon -B(q_t^\varepsilon)^Tp_t^\varepsilon\dd t/(m\varepsilon) +\sqrt{2\theta}\dd W_t/\varepsilon,\notag \end{align} where $W$ is $J$ dimensional and $z_0^\varepsilon\sim N(0,\theta I_J/\varepsilon)$ is independent of the resolved initial state and future Brownian increments. The extended Gibbs energy is $U+|p|^2/(2m)+\varepsilon|z|^2/2$. The conservative coupling cancels in its derivative and has zero phase-space divergence. The remaining $z$ part is an OU thermostat. The equilibrium parity is $(q,p,z)\mapsto(q,-p,z)$. \begin{proposition}[Fixed-mass periodic short-memory limit] \label{b30:torus:fixed-mass} At fixed $m>0$ and finite horizon $\tau$, the resolved paths of \eqref{b30:torus:embedding} converge in probability, under the same Brownian motion and resolved initial state, to \begin{equation} \dd q_t=p_t\dd t/m,\qquad \dd p_t=-\nabla U(q_t)\dd t-Gp_t\dd t/m +\sqrt{2\theta}B(q_t)\dd W_t. \label{b30:torus:kinetic-limit} \end{equation} The initial momentum has finite second moment. The law of this kinetic limit is the law of constant-friction Langevin dynamics with momentum noise covariance $2\theta G$. \end{proposition} \begin{proof} Write $z^\varepsilon=\eta^\varepsilon- \int_0^t\varepsilon^{-1}e^{-(t-s)/\varepsilon} B(q_s^\varepsilon)^Tp_s^\varepsilon/m\dd s$, where $\eta^\varepsilon$ is stationary OU. Multiplication of the last equation in \eqref{b30:torus:embedding} by $\varepsilon B(q_t^\varepsilon)$, and the product rule, give \begin{align} p_t^\varepsilon &=p_0-\int_0^t\nabla U(q_s^\varepsilon)\dd s -G\int_0^t p_s^\varepsilon\dd s/m +\sqrt{2\theta}\int_0^t B(q_s^\varepsilon)\dd W_s +r_t^\varepsilon,\label{b30:torus:remainder-identity}\\ r_t^\varepsilon &=-\varepsilon B(q_t^\varepsilon)z_t^\varepsilon +\varepsilon B(q_0)z_0^\varepsilon +\varepsilon\int_0^t DB(q_s^\varepsilon)[p_s^\varepsilon/m]z_s^\varepsilon\dd s. \notag \end{align} Here $BB^T=G$. There is no quadratic-covariation correction in this product: at positive mass $q^\varepsilon$ is a finite-variation path. The derivative term in the remainder cannot simply be discarded without an estimate. Lift positions continuously to $\mathbb R^d$ from the same initial representative. Stop both equations when either momentum reaches radius $R$, and denote this time by $\tau_R$. With $B_*=\sup\|B\|$, \[ \sup_{t\le\tau\wedge\tau_R}\varepsilon|z_t^\varepsilon| \le \varepsilon\sup_{t\le\tau}|\eta_t^\varepsilon| +\varepsilon B_*R/m. \] The stationary OU maximal bound is \[ \mathbb E\sup_{t\le\tau}|\varepsilon\eta_t^\varepsilon|^2 \le C\varepsilon[1+\log(1+\tau/\varepsilon)]. \] It follows by dividing time into intervals of length $\varepsilon$, using Gaussian endpoint tails and time-changed Brownian maximal tails within each interval, and integrating the union bound over $O(1+\tau/\varepsilon)$ intervals. Therefore \[ \mathbb E\sup_{t\le\tau\wedge\tau_R}|r_t^\varepsilon|^2 \le C_{R,m,\tau}\varepsilon[1+\log(1+\tau/\varepsilon)] +C_{R,m,\tau}\varepsilon^2\longrightarrow0. \] All derivatives of $B$ are bounded, and $|p^\varepsilon|\le R$ on the stopped interval. Subtract \eqref{b30:torus:kinetic-limit} from \eqref{b30:torus:remainder-identity}. The drift terms are Lipschitz; the Burkholder--Davis--Gundy inequality bounds the stopped noise difference by a constant times $\int_0^\tau\mathbb E\sup_{u\le s\wedge\tau_R}|q_u^\varepsilon-q_u|^2\dd s$. Together with $q_t^\varepsilon-q_t=\int_0^t(p_s^\varepsilon-p_s)\dd s/m$, Gronwall yields \[ \mathbb E\sup_{t\le\tau\wedge\tau_R} (|q_t^\varepsilon-q_t|^2+|p_t^\varepsilon-p_t|^2)\longrightarrow0. \] The limiting equation has globally Lipschitz coefficients on the lift and is nonexplosive. On the event $\sup_{t\le\tau}|p_t|0$, and $\theta>0$. Use the periodic Hamiltonians \eqref{b30:torus:hamiltonian}, the conditional canonical bath preparation, the positive spectral approximation just stated, and an independent initial momentum $N(0,m\theta I_d)$. On every fixed finite interval, the resolved position laws converge, first in the bath mesh and cutoff, next as $\varepsilon\downarrow0$ at fixed $m$, and finally as $m\downarrow0$, to the law of \begin{equation} \dd Q_t=-K\nabla U(Q_t)\dd t+\sqrt{2\theta K}\dd\beta_t \quad\hbox{on }\mathbb T^d,\qquad K=G^{-1}. \label{b30:torus:overdamped} \end{equation} For a smooth positive initial position density, its invariant density is $\pi=Z^{-1}e^{-U/\theta}$ and its current uniquely minimizes $\Phi+\dot F$ in the finite weighted-current space. Moreover, \[ F[\rho_t]-F[\pi]=\theta D(\rho_t\Vert\pi),\qquad \dot F[\rho_t]=-\theta^2\int\rho_t |\nabla\log(\rho_t/\pi)|_K^2\dd q\le0. \] The corresponding mean total entropy-production rate is $-\dot F/T$ under the autonomous single-isothermal-reservoir heat convention. \end{theorem} \begin{proof} The finite-bath elimination and force convergence give the memory embedding, and Proposition~\ref{b30:torus:fixed-mass} gives its fixed-mass kinetic limit. That kinetic law equals the additive-noise Langevin law. It remains to justify the momentum reduction for a periodic potential; a confining Euclidean Gibbs density is unnecessary. Use the periodic lift of $U$, whose gradient is bounded and globally Lipschitz. In the additive kinetic representation the exact integral identity is \[ q_t^m=q_0-\int_0^tK\nabla U(q_s^m)\dd s +\sqrt{2\theta K}\beta_t-K(p_t^m-p_0^m). \] Variation of constants gives \[ p_t^m=\xi_t^m-\int_0^t e^{-G(t-s)/m}\nabla U(q_s^m)\dd s, \qquad \mathbb E[\xi_t^m(\xi_t^m)^T]=m\theta I_d, \] where $\xi^m$ is stationary because the initial momentum has the thermal law. If $g_*>0$ is the least eigenvalue of $G$, then \[ \sup_{t\le\tau}|p_t^m| \le\sup_{t\le\tau}|\xi_t^m|+\frac m{g_*}\|\nabla U\|_\infty. \] The OU maximal estimate, with its momentum correlation time of order $m$, gives in fixed dimensionless units \[ \mathbb E\sup_{t\le\tau}|p_t^m|^2 \le C_\tau m[1+\log(1+\tau/m)]+C_\tau m^2\longrightarrow0. \] Subtract the lifted integral equation for $Q$. Lipschitz continuity of $\nabla U$ and Gronwall give \[ \mathbb E\sup_{t\le\tau}|q_t^m-Q_t|^2 \le C_\tau\mathbb E\left[\sup_{t\le\tau}|p_t^m|^2+|p_0^m|^2\right] \longrightarrow0. \] Projection to $\mathbb T^d$ gives the claimed final path-law limit. The bath and short-memory limits are taken first; no commutation or infinite-horizon uniformity has been used. The forward current of \eqref{b30:torus:overdamped} is $j_\rho=-\rho K\nabla U-\theta K\nabla\rho$. Its value at $\pi$ is zero. Smooth coefficients and ellipticity on the compact torus give a unique conservative diffusion and a smooth positive density evolution. Completing the weighted square gives the unique Rayleighian minimizer, and integration by parts on the torus gives the displayed free-energy identity. For the heat convention $\dot Q_{\rm bath}=-d\mathbb E U/dt$ and $S_{\rm sys}=-k_B\int\rho\log\rho$, direct algebra gives $\dot S_{\rm sys}+\dot Q_{\rm bath}/T=-\dot F/T$. The stochastic limits were established before this entropy calculation. \end{proof} This result supplies a microscopic realization of the compact constant-mobility current model, using globally periodic couplings. Its preparation, bath spectrum and sequential scale limits remain physical assumptions. Exact reverse preparation retains the correlations created during the forward evolution. The construction neither selects an absolute cosmic orientation from equilibrium alone nor realizes arbitrary interacting reservoirs or constrained continuum fields. \section{Stochastic Dynamics and Variational Thermodynamics} \label{sec:models_variation} The reservoir calculation has supplied one effective stochastic law. We now ask which features of that law are needed for a thermodynamic description. The energy specifies the conservative forces, the mobility specifies the response to those forces, and the bath temperature fixes the fluctuation scale. Their compatibility determines whether the Gibbs law is preserved. We first state these inputs, then derive the probability current and its variational form. The overdamped and kinetic models are treated separately because removing momentum changes the way dissipation acts on the state. Appendix~\ref{app:notation_variation} collects the notation and units. The physical content of this chain is an accounting of exchange. Friction transfers energy from a resolved degree of freedom to the reservoir, while reservoir fluctuations return energy to it. At thermal equilibrium these exchanges continue, although the probability law is stationary. Away from equilibrium, their net effect is expressed by a probability current. The equilibrium reference identifies the free-energy excess carried by that current; the local heat balance determines whether its loss is the physical entropy gain of system plus reservoir. Thus the stationary-action skeleton specifies admissible reversible motion, and the prepared reservoir supplies the direction of the effective dissipative response. The next hypotheses make each part of this interpretation explicit before the variational identity is used. \subsection{Assumption Hierarchy} The paper does not use a single monolithic axiom. Instead, assumptions are modular and are invoked only where needed. \begin{assumption}[M1: Markov well-posedness] \label{ass:M1} The microscopic stochastic equation under consideration admits a conservative Markov solution and a transition semigroup $(P_t)_{t\ge 0}$. Standard finite-dimensional diffusion background is used in the form summarized by Pavliotis \cite{Pavliotis2014}; infinite-dimensional stochastic evolution and ergodic background are invoked only under hypotheses compatible with Da Prato--Zabczyk theory \cite{DaPratoZabczyk1992,DaPratoZabczyk1996}; and Harris-type Markov-chain stability statements are invoked only in the scope covered by Meyn--Tweedie theory \cite{MeynTweedie2009}. \end{assumption} \begin{assumption}[M2: thermodynamic calibration] \label{ass:M2} In equilibrium molecular models, conservative forces and stochastic dissipation satisfy the stated Hamiltonian and fluctuation-dissipation structure. In particular, the thermal covariance is matched to the friction or mobility at temperature $T$. \end{assumption} \begin{assumption}[M3: normalizability] \label{ass:M3} The candidate equilibrium Gibbs measure has finite nonzero partition function. \end{assumption} \begin{assumption}[M4: invariant equilibrium law] \label{ass:M4} The equilibrium law $\muEq$ satisfies $\muEq P_t=\muEq$. Whenever possible this property is proved from \cref{ass:M2} rather than assumed. \end{assumption} \begin{assumption}[M5: entropy regularity] \label{ass:M5} For the diffusion entropy identity, the dissipative form is strongly local and has the stated diffusion chain rule; any antisymmetric part acts as a first-order derivation on the common core. For regular solutions, the common-domain and integrability conditions below permit differentiation of entropy and integration in time. For a weak-solution passage, one must instead verify an approximation with convergent endpoint entropies and convergence in \(L^1\) in time of the Fisher-information terms. Lower semicontinuity alone gives an inequality, not necessarily an identity. Closed symmetric and quasi-regular Dirichlet-form issues are interpreted in the sense of \cite{FukushimaOshimaTakeda2011,RocknerSchmuland1995}; closedness alone does not establish these extra properties. \end{assumption} \begin{assumption}[M6: physical time reversal] \label{ass:M6} There exists a measurable involution $\Theta$ such that $\Theta^2=I$ and $\Theta_{\#}\muEq=\muEq$. \end{assumption} \begin{assumption}[M7: irreducibility or hypoelliptic mixing] \label{ass:M7} Whenever uniqueness or strict long-time convergence is asserted, the system satisfies the corresponding irreducibility, positivity-improving, Harris, Hormander, or hypocoercive hypothesis explicitly cited in the theorem. Elliptic positivity and support arguments are not assumed silently; when used, they must be traced to results of Aronson type or Stroock--Varadhan type \cite{Aronson1967,StroockVaradhan1972}. \end{assumption} \begin{assumption}[M8: coercivity] \label{ass:M8} A quantitative convergence theorem uses a separately stated logarithmic Sobolev, Bakry-Emery, Harris, spectral, or hypocoercive estimate. The existence of an energy or action minimum is never used as a substitute for coercivity. \end{assumption} \begin{assumption}[M9: path-space consistency] \label{ass:M9} When a family of finite-window trajectory measures is interpreted as restrictions of one stochastic process, the family is projectively consistent. Stationarity is separately established by time-shift invariance. \end{assumption} \begin{assumption}[M10: history-space generation] \label{ass:M10} The physical moving-history process is described by its actual shift-boundary generator. An abstract Gibbs sampler on path space is not identified with the physical history generator unless that equality is independently proved. \end{assumption} \begin{assumption}[M11: Navier-Stokes trace regularity] \label{ass:M11} The noise covariance in the $\cH$-valued 2D stochastic Navier-Stokes model satisfies the trace assumptions stated in \cref{sec:NS}; in particular, enstrophy estimates require $\Tr(A_{\mathrm{St}}Q)<\infty$. \end{assumption} \begin{assumption}[M12: experimental protocol] \label{ass:M12} Each experimental prediction is conditioned on the specified calibration, time-scale separation, measurement model, and post-driving protocol. Work performed by external parameter changes is not absorbed into autonomous relaxation entropy production. \end{assumption} \begin{assumption}[M13: local detailed balance when physical entropy is inferred from paths] \label{ass:M13} Whenever a path likelihood ratio is interpreted as physical medium or total entropy production, the microscopic transition mechanism satisfies a model-specific local-detailed-balance relation. For jump channels this includes a declared channel involution \(\nu\mapsto\bar\nu\), support equivalence of forward and reversed channels, and cancellation of the corresponding holding-time terms. For diffusions this includes the declared stochastic convention, parity map, covariance, endpoint initialization, and heat sign. \end{assumption} \subsection{Dependency Map} The proof structure is summarized by the following implications: \begin{align} \text{Markovity}+\text{invariance} &\Longrightarrow \text{relative-entropy contraction}, \label{eq:dependency_DPI} \\ \text{Dirichlet regularity} &\Longrightarrow \text{entropy-dissipation identity}, \label{eq:dependency_dissipation} \\ \text{Gibbs thermodynamic identification} &\Longrightarrow F-F_{\mathrm{eq}}=k_B T\KL{\nu}{\muEq}, \label{eq:dependency_free_energy} \\ \text{coercivity/hypocoercivity} &\Longrightarrow \text{quantitative convergence}, \label{eq:dependency_convergence} \\ \text{strict contraction} &\Longrightarrow \begin{cases} \text{no equilibrium-preserving positive Markov inverse},\\ \text{no exact recurrence of a non-equilibrium law}. \end{cases} \label{eq:dependency_irreversibility} \end{align} These implications are deliberately one-directional. In particular, neither invariance nor the existence of a free-energy minimum uniquely determines the physical generator. The complete proof chain used for the central theorem is \begin{equation} \begin{array}{c} \text{open-system stochastic model} \\ \Downarrow \\ \text{thermal calibration }(\Diff=k_BT\Mob) \text{ and ideal-bath interpretation} \\ \Downarrow \\ \text{Gibbs invariant law and free-energy identity} \\ \Downarrow \\ \text{Onsager--Rayleigh characterization of the dissipative current} \\ \Downarrow \\ \text{Markov relative-entropy contraction and class-dependent dissipation} \\ \Downarrow \\ \text{autonomous unforced isothermal second law} \\ \Downarrow \\ \text{convergence in law only after an additional coercive theorem.} \end{array} \label{eq:noncircular_proof_chain} \end{equation} Every arrow in \cref{eq:noncircular_proof_chain} is either proved in the paper, imported with explicit hypotheses, or stated as a constitutive modeling premise. The sign of \(\dot S_{\mathrm{tot}}\) is never used as a premise for the relative-entropy contraction that later proves it in the autonomous Gibbs class. For the confined constant-mobility branch, Section~\ref{sec:microscopic-bridge30} also supplies the preceding Hamiltonian-reservoir realization. Its finite-time path convergence is distinct from the entropy identity of the limiting diffusion; no convergence of entropy derivatives is inferred from weak path convergence. % ===================================================================== \subsection{Microscopic Stochastic Molecular Dynamics} \label{sec:molecular_dynamics} The choice of state is physical. If momentum relaxes on a much shorter time scale than position, an overdamped model records position alone. If that separation is unavailable, momentum remains part of the state: forces change momentum, momentum transports position, and the reservoir acts through friction and momentum fluctuations. The two models below therefore share a thermal interpretation but have different generators. \subsubsection{Overdamped and Underdamped Model Classes} The finite-dimensional molecular state may be represented either by configuration variables alone or by full phase-space variables. The distinction is essential because the associated entropy-dissipation mechanisms are different. For an overdamped system with configuration \(X_t\in\RR^n\), the prototype is \begin{equation} \dd X_t =b(X_t)\dd t +\sigma(X_t)\dd W_t. \label{eq:general_overdamped_sde} \end{equation} When the mobility is constant and the force is conservative, the thermally calibrated equilibrium model takes the form \begin{equation} \dd X_t =-\Mob\nabla U(X_t)\dd t +\sqrt{2k_BT\Mob}\,\dd W_t, \qquad \Mob=\Mob^{\mathsf T}>0. \label{eq:overdamped_equilibrium_sde} \end{equation} When momentum relaxation must be resolved, the molecular model retains momentum. Let \(q_t,p_t\in\RR^{dN}\), let \(\Mmass=\Mmass^{\mathsf T}>0\) be the mass matrix, and let \(\Gfr=\Gfr^{\mathsf T}>0\) denote the friction matrix. The underdamped Langevin equation is \begin{align} \dd q_t &=\Mmass^{-1}p_t\dd t, \label{eq:underdamped_q} \\ \dd p_t &=-\nabla U(q_t)\dd t -\Gfr\Mmass^{-1}p_t\dd t +\sqrt{2k_BT\Gfr}\,\dd W_t. \label{eq:underdamped_p} \end{align} The molecular Hamiltonian is \begin{equation} \mathcal{H}(q,p) =\frac12p^{\mathsf T}\Mmass^{-1}p+U(q). \label{eq:molecular_hamiltonian} \end{equation} The noise in \cref{eq:underdamped_p} acts only in the momentum directions. Thus the kinetic diffusion is degenerate in the full \((q,p)\) state space and cannot be treated by a naive uniformly elliptic argument. \subsubsection{Fluctuation-Dissipation Calibration} The noise covariance and friction are not independent parameters in the equilibrium model. This calibration is the stochastic analogue of the fluctuation-dissipation structure developed in linear-response theory \cite{CallenWelton1951,Kubo1966}. In the kinetic normalization of \cref{eq:underdamped_p}, \begin{equation} \sigma_p\sigma_p^{\mathsf T} =2k_BT\Gfr. \label{eq:FDT_kinetic} \end{equation} Similarly, for the overdamped model \cref{eq:overdamped_equilibrium_sde}, \begin{equation} \sigma\sigma^{\mathsf T} =2k_BT\Mob. \label{eq:FDT_overdamped} \end{equation} These identities are fluctuation-dissipation calibrations. They do not by themselves imply Gibbs equilibrium: the conservative form of the drift is also required. For example, if a nonconservative force \(F_{\mathrm{nc}}(q)\) is added to \cref{eq:underdamped_p}, then the canonical Gibbs density generally ceases to be invariant. \begin{proposition}[Conservative force plus fluctuation-dissipation calibration yields Gibbs invariance] \label{prop:gibbs_invariance_kinetic} Assume \(U\in C^2(\mathbb R^{dN})\) is bounded below, has compact sublevel sets, and has globally Lipschitz gradient. Keep the fixed positive mass and friction matrices specified above, and assume \begin{equation} Z =\int_{\RR^{2dN}} \exp[-\beta\mathcal{H}(q,p)]\dd q\dd p <\infty, \qquad \beta=(k_BT)^{-1}. \label{eq:kinetic_partition} \end{equation} Then the probability measure \begin{equation} \piEq(\dd q\dd p) =Z^{-1} \exp[-\beta\mathcal{H}(q,p)]\dd q\dd p \label{eq:kinetic_gibbs_measure} \end{equation} is invariant for \cref{eq:underdamped_q,eq:underdamped_p}. \end{proposition} \begin{proof} The backward generator is decomposed as \begin{equation} L=A+S, \label{eq:kinetic_generator_decomp} \end{equation} where \begin{equation} A =\Mmass^{-1}p\cdot\nabla_q -\nabla U(q)\cdot\nabla_p \label{eq:kinetic_A} \end{equation} and \begin{equation} S =-\Gfr\Mmass^{-1}p\cdot\nabla_p +k_BT\Gfr:\nabla_p^2. \label{eq:kinetic_S} \end{equation} Let \(\pi(q,p)\) denote the density in \cref{eq:kinetic_gibbs_measure}. Differentiating gives \begin{align} \nabla_q\pi &=-\beta\nabla U(q)\pi, \label{eq:gibbs_q_gradient} \\ \nabla_p\pi &=-\beta\Mmass^{-1}p\,\pi. \label{eq:gibbs_p_gradient} \end{align} The Lebesgue adjoint of \(A\) is \begin{equation} A^* =-\Mmass^{-1}p\cdot\nabla_q +\nabla U(q)\cdot\nabla_p, \label{eq:kinetic_A_lebesgue_adjoint} \end{equation} because the Hamiltonian vector field has zero Euclidean divergence. Applying \cref{eq:kinetic_A_lebesgue_adjoint} to \(\pi\) and using \cref{eq:gibbs_q_gradient,eq:gibbs_p_gradient}, \begin{align} A^*\pi &=-\Mmass^{-1}p\cdot(-\beta\nabla U\pi) +\nabla U\cdot(-\beta\Mmass^{-1}p\,\pi) \notag\\ &=\beta\Mmass^{-1}p\cdot\nabla U\,\pi -\beta\nabla U\cdot\Mmass^{-1}p\,\pi \notag\\ &=0. \label{eq:Astar_pi_zero} \end{align} The forward dissipative operator is \begin{equation} S^*\rho =\nabla_p\cdot \left( \Gfr\Mmass^{-1}p\,\rho +k_BT\Gfr\nabla_p\rho \right). \label{eq:kinetic_S_forward} \end{equation} Substituting \cref{eq:gibbs_p_gradient}, the vector field inside the divergence is \begin{align} \Gfr\Mmass^{-1}p\,\pi +k_BT\Gfr\nabla_p\pi &=\Gfr\Mmass^{-1}p\,\pi -k_BT\beta\Gfr\Mmass^{-1}p\,\pi \notag\\ &=0, \label{eq:dissipative_gibbs_current_zero} \end{align} since \(k_BT\beta=1\). Hence \(S^*\pi=0\). Combining this identity with \cref{eq:Astar_pi_zero} gives \begin{equation} L^*\pi=(A^*+S^*)\pi=0, \label{eq:full_gibbs_stationarity} \end{equation} This establishes infinitesimal stationarity. To pass from it to invariance of the actual semigroup, we also need a conservative well-posed realization; the global Lipschitz hypothesis supplies it. One direct argument avoids identifying a formal adjoint solution with an invariant probability without this step. Let \(H_t\) be the deterministic Hamiltonian transition and \(O_t\) the Ornstein--Uhlenbeck transition acting only on momentum. The Hamiltonian flow is complete, preserves \(\mathcal H\), and has unit Jacobian because its vector field is divergence free. Thus \(\piEq H_t=\piEq\). The momentum covariance \(C_p=k_BT\Mmass\) satisfies \[ (\Gfr\Mmass^{-1})C_p+ C_p(\Mmass^{-1}\Gfr)=2k_BT\Gfr. \] Consequently \(O_t\) preserves the Maxwell momentum factor while leaving the position factor unchanged, and \(\piEq O_t=\piEq\). Every product \((H_{t/n}O_{t/n})^n\) therefore preserves \(\piEq\) exactly. For globally Lipschitz drift, these split transitions converge to the Langevin transition \(P_t\) from each fixed initial state. To see the ingredients, expand each deterministic step to first order and use the exact Gaussian momentum step. Its conditional drift is the momentum drift times \(t/n+O(n^{-2})\), its conditional covariance is \(2k_BT\Gfr\,t/n+O(n^{-2})\), and the deterministic part supplies the Hamiltonian drift. Global linear-growth moment bounds give tightness of the interpolations. Their martingale limits have generator \(A+S\); global Lipschitz uniqueness identifies that limit with \(P_t\). Equivalently, the usual coupled time-discretization estimate proves convergence on a finite interval. For every bounded continuous \(f\), dominated convergence now gives \[ \int P_tf\,d\piEq =\lim_{n\to\infty}\int(H_{t/n}O_{t/n})^nf\,d\piEq =\int f\,d\piEq. \] No energy moment of the equilibrium is needed for this last passage, because the test is bounded and convergence holds pointwise in the initial state. This proves invariance. \end{proof} The global Lipschitz condition is a transparent sufficient condition, not a necessary one. Singular molecular forces or faster-growing smooth forces need a separate conservative-realization and convergence argument; normalizability of a formal Gibbs density alone does not supply that argument. \begin{remark}[Why fluctuation-dissipation calibration alone is insufficient] \label{rem:FDT_not_enough} Suppose a force \(F_{\mathrm{nc}}(q)\) is added to the momentum equation. The additional forward operator is \begin{equation} \mathcal N^*\rho =-\nabla_p\cdot(F_{\mathrm{nc}}\rho). \label{eq:nonconservative_forward} \end{equation} Because \(F_{\mathrm{nc}}\) is independent of \(p\), \begin{align} \mathcal N^*\pi &=-F_{\mathrm{nc}}\cdot\nabla_p\pi \notag\\ &=\beta F_{\mathrm{nc}}\cdot\Mmass^{-1}p\,\pi, \label{eq:nonconservative_breaks_gibbs} \end{align} which is generally nonzero. Thus Gibbs invariance is produced by the joint conservative-force and fluctuation-dissipation structure, not by stochasticity or the covariance relation alone. \end{remark} \subsubsection{The It\^o Convention and the Stratonovich Correction} Whenever an explicit finite-dimensional calculation is required, the manuscript uses It\^o form. If a model is originally written as \begin{equation} \dd X_i =b_i^{\mathrm S}(X)\dd t +\sigma_{ik}(X)\circ\dd W_k, \label{eq:stratonovich_general} \end{equation} then the corresponding It\^o drift is \begin{equation} b_i^{\mathrm I} =b_i^{\mathrm S} +\frac12 \sum_{j,k} \sigma_{jk}\partial_j\sigma_{ik}. \label{eq:correct_stratonovich_ito} \end{equation} In general this is not equal to \(b_i^{\mathrm S}+\frac12\partial_j(\sigma\sigma^{\mathsf T})_{ij}\) because \begin{align} \partial_j(\sigma\sigma^{\mathsf T})_{ij} &=\sum_k \partial_j(\sigma_{ik}\sigma_{jk}) \notag\\ &=\sum_k \sigma_{jk}\partial_j\sigma_{ik} +\sum_k \sigma_{ik}\partial_j\sigma_{jk}. \label{eq:stratonovich_difference} \end{align} The second term in \cref{eq:stratonovich_difference} is not part of the general Stratonovich-to-It\^o correction. \subsubsection{Why Stochasticity Is Not Identical to Time-Reversal Breaking} The equilibrium Ornstein-Uhlenbeck process already provides a decisive counterexample to the statement that noise by itself generates a physical arrow of time. The stationary process is stochastic and dissipative, but its stationary law is reversible. Physical time reversal must therefore be formulated relative to the parity involution and the stationary law, as developed in \cref{sec:four_levels}. % ===================================================================== \subsection{Stochastic Actions and the Minimum-Dissipation Principle} \label{sec:minimum_dissipation} Least action and variational dissipation concern different objects. Hamilton's principle varies a mechanical trajectory with prescribed endpoint data. The principle used here varies an ensemble current at a fixed density. The free-energy gradient provides the thermodynamic force, and the mobility fixes how much current that force produces. The Rayleighian combines the free-energy rate with a positive transport cost, so that its minimizer reproduces the already specified dissipative dynamics. A comparison with path-action formulations is needed before writing this current principle mathematically. \subsubsection{Three Inequivalent Action Concepts} The word ``action'' is used in several branches of mechanics and probability, but the corresponding objects are not interchangeable. Confusing them creates both dimensional and logical errors. \paragraph{Hamilton Action} The classical real-time Hamilton action is \begin{equation} \mathcal S_H[q] =\int_0^T \left[ \frac12m|\dot q_t|^2-U(q_t) \right]\dd t. \label{eq:Hamilton_action} \end{equation} Its variational condition is stationarity, \begin{equation} \delta\mathcal S_H=0, \label{eq:Hamilton_stationarity} \end{equation} not a universal monotone descent law. The resulting Euler-Lagrange or Hamilton equations are conservative under the usual assumptions and do not by themselves imply entropy production. There is also a dimensional obstruction to using \(\exp(-\beta\mathcal S_H)\) as a classical thermal path weight. The dimensions are \begin{equation} [\mathcal S_H]=\mathrm{J\,s}, \qquad [\beta]=\mathrm{J}^{-1}, \label{eq:action_dimensions} \end{equation} so \(\beta\mathcal S_H\) has units of time rather than being dimensionless. Accordingly, the final theory never inserts the real-time Hamilton action directly into a Boltzmann factor. \paragraph{Girsanov Relative Path Action} Consider, as an elliptic prototype, two path measures generated by diffusions with the same constant invertible diffusion matrix \(\sigma\) and the same initial law. The reference process is \begin{equation} \dd X_t=\sigma\dd W_t, \label{eq:Girsanov_reference} \end{equation} and the drifted process is \begin{equation} \dd X_t=b(X_t)\dd t+\sigma\dd W_t. \label{eq:Girsanov_drifted} \end{equation} Set \begin{equation} u(x)=\sigma^{-1}b(x). \label{eq:Girsanov_u} \end{equation} Under Novikov's condition \begin{equation} \EE_{\PP_0} \exp\left[ \frac12 \int_0^T|u(X_t)|^2\dd t \right] <\infty, \label{eq:Novikov} \end{equation} Here the expectation is under the reference law and the stochastic integral below uses its Brownian motion \(W^0\). Girsanov's theorem gives \begin{equation} \frac{\dd\PP_b}{\dd\PP_0} =\exp\left[ \int_0^T u(X_t)\cdot\dd W_t^0 -\frac12\int_0^T|u(X_t)|^2\dd t \right]. \label{eq:Girsanov_RN} \end{equation} The relative path action is therefore defined by \begin{equation} \mathcal A_T^{G} =-\log\frac{\dd\PP_b}{\dd\PP_0}. \label{eq:Girsanov_action} \end{equation} This quantity is dimensionless by construction. It is a logarithmic likelihood ratio between path measures; it is not automatically a most-probable-path action. Under the drifted law, \(W^b=W^0-\int_0^\cdot u(X_s)\dd s\) is Brownian. If \(\EE_{\PP_b}\int_0^T|u(X_t)|^2\dd t<\infty\), taking the expectation of the log likelihood under that law gives \begin{equation} D(\PP_b\Vert\PP_0)=\frac12\EE_{\PP_b} \int_0^T|u(X_t)|^2\dd t. \end{equation} This expectation condition and the reference Novikov condition serve different purposes. The drifted-law expectation of \(\mathcal A_T^G\) is the negative of this KL divergence. If the initial distributions differ, \(\nu_0\ll\mu_0\), the path Radon-Nikodym derivative acquires the initial contribution \begin{equation} \log \frac{\dd\PP_b^{\nu_0}} {\dd\PP_0^{\mu_0}} =\log\frac{\dd\nu_0}{\dd\mu_0}(X_0) +\int_0^T u(X_t)\cdot\dd W_t^0 -\frac12\int_0^T|u(X_t)|^2\dd t. \label{eq:Girsanov_initial_term} \end{equation} \begin{remark}[Degenerate kinetic diffusions] \label{rem:Girsanov_degenerate} The full underdamped Langevin diffusion matrix is degenerate in the \((q,p)\) state space, so a full inverse \(\sigma^{-1}\) does not exist. The elementary derivation \cref{eq:Girsanov_u,eq:Girsanov_RN} is therefore only an elliptic prototype. A kinetic Girsanov comparison is legitimate only when the drift difference belongs to the range of the noise operator, for example when two kinetic processes share the same position kinematics and differ only in suitably integrable momentum-direction drifts. \end{remark} \paragraph{Onsager-Machlup and Freidlin-Wentzell Functionals} The Onsager-Machlup functional characterizes small-tube probabilities of smooth comparison paths for appropriate diffusion models, whereas the Freidlin-Wentzell functional is a small-noise large-deviation rate functional. They need not coincide with \cref{eq:Girsanov_action}; divergence, geometry, normalization, and limiting procedures matter. Accordingly, the manuscript uses separate notations \(\mathcal A^G\), \(\mathcal A^{OM}\), and \(I_{FW}\) and never replaces one by another without a theorem. \subsubsection{Overdamped Free-Energy Gradient Flow} For the constant-mobility equilibrium diffusion \cref{eq:overdamped_equilibrium_sde}, the Fokker-Planck equation is \begin{equation} \partial_t\rho =\nabla\cdot \left[ \Mob\left( \rho\nabla U +k_BT\nabla\rho \right) \right]. \label{eq:overdamped_FPE} \end{equation} Define the nonequilibrium free-energy functional relative to Lebesgue measure for the present finite-dimensional calculation by \begin{equation} F[\rho] =\int_{\RR^n}U(x)\rho(x)\dd x +k_BT\int_{\RR^n}\rho(x)\log\rho(x)\dd x. \label{eq:overdamped_free_energy} \end{equation} The reference-measure refinement needed for physical dimensional consistency is given in \cref{sec:thermodynamic_free_energy}. Its first variation is \begin{equation} \frac{\delta F}{\delta\rho} =U+k_BT(1+\log\rho). \label{eq:free_energy_variation} \end{equation} Therefore \begin{equation} \nabla\frac{\delta F}{\delta\rho} =\nabla U+k_BT\nabla\log\rho. \label{eq:free_energy_variation_gradient} \end{equation} Using \(\rho\nabla\log\rho=\nabla\rho\), we obtain \begin{equation} \rho\Mob\nabla\frac{\delta F}{\delta\rho} =\Mob(\rho\nabla U+k_BT\nabla\rho). \label{eq:gradient_flow_identity} \end{equation} Consequently \cref{eq:overdamped_FPE} can be written as \begin{equation} \partial_t\rho =\nabla\cdot \left[ \rho\Mob\nabla \frac{\delta F}{\delta\rho} \right]. \label{eq:overdamped_gradient_flow} \end{equation} \subsubsection{Rayleighian Minimization of the Dissipative Current} The sign of the linear free-energy term is part of this characterization. Both $\Phi+\dot F$ and $\Phi-\dot F$ are convex in the current, and their minimizers have opposite signs. In the microscopic realization of Section~\ref{sec:microscopic-bridge30}, equilibrium force covariance gives the positive metric, while the retarded response of the prepared reservoir gives the forward dissipative drift. The drift-change cost in Section~\ref{b30:sec:cost} recovers the same sign. In a model without that derivation, the variational sign remains a constitutive premise. Write the continuity equation as \begin{equation} \partial_t\rho+\nabla\cdot j=0. \label{eq:continuity_current} \end{equation} For a fixed density \(\rho\), define the extended Rayleighian \begin{equation} \mathscr R_{\rho}[j] =\frac12\int_{\RR^n} \frac{j^{\mathsf T}\Mob^{-1}j}{\rho}\dd x +\int_{\RR^n} j\cdot\nabla \frac{\delta F}{\delta\rho}\dd x, \label{eq:Rayleighian_full} \end{equation} with the convention \begin{equation} \frac{j^{\mathsf T}\Mob^{-1}j}{\rho} =\begin{cases} 0,&\rho=0\ \text{and}\ j=0,\\ +\infty,&\rho=0\ \text{and}\ j\ne0. \end{cases} \label{eq:Rayleighian_zero_density} \end{equation} Boundary conditions are assumed periodic, no-flux, or sufficiently decaying so that all integrations by parts below have no boundary contribution. \begin{theorem}[Minimum-dissipation current] \label{thm:minimum_dissipation_current} Assume \(\Mob=\Mob^{\mathsf T}>0\) is constant, \(\rho\) is positive almost everywhere, and \(\int\rho\,\nabla(\delta F/\delta\rho)^{\mathsf T}\Mob \nabla(\delta F/\delta\rho)\dd x<\infty\). Use the Hilbert space of currents with finite \(\int j^{\mathsf T}\Mob^{-1}j/\rho\dd x\), with no additional pointwise current constraints. If a physical boundary is present, assume the displayed minimizing current has the required zero normal trace. Then \cref{eq:Rayleighian_full} is strictly convex and attains its unique minimum at \begin{equation} j_{\star} =-\rho\Mob\nabla \frac{\delta F}{\delta\rho}. \label{eq:Rayleighian_minimizer} \end{equation} Substitution into \cref{eq:continuity_current} yields the Fokker-Planck equation \cref{eq:overdamped_gradient_flow}. \end{theorem} \begin{proof} Let \(h\) be any admissible variation and set \(j_\varepsilon=j+\varepsilon h\). The domain assumptions give the complete-square identity \begin{equation} \mathscr R_\rho[j]=\frac12\int \frac{\bigl(j+\rho\Mob\nabla(\delta F/\delta\rho)\bigr)^{\mathsf T}\Mob^{-1} \bigl(j+\rho\Mob\nabla(\delta F/\delta\rho)\bigr)}\rho\dd x -\frac12\int\rho\, \left(\nabla\frac{\delta F}{\delta\rho}\right)^{\mathsf T} \Mob\nabla\frac{\delta F}{\delta\rho}\dd x. \end{equation} It proves attainment as well as uniqueness; strict convexity alone would prove only the latter. Equivalently, the first variation is \begin{align} \frac{\dd}{\dd\varepsilon} \mathscr R_\rho[j_\varepsilon] \bigg|_{\varepsilon=0} &=\int h^{\mathsf T}(\rho\Mob)^{-1}j\dd x +\int h\cdot\nabla \frac{\delta F}{\delta\rho}\dd x \notag\\ &=\int h\cdot \left[ (\rho\Mob)^{-1}j +\nabla\frac{\delta F}{\delta\rho} \right]\dd x. \label{eq:Rayleighian_first_variation} \end{align} For this expression to vanish for every admissible \(h\), the bracket must vanish almost everywhere: \begin{equation} (\rho\Mob)^{-1}j +\nabla\frac{\delta F}{\delta\rho}=0. \label{eq:Rayleighian_EL} \end{equation} Multiplication by \(\rho\Mob\) gives \cref{eq:Rayleighian_minimizer}. The second variation is \begin{equation} \delta^2\mathscr R_\rho[j](h,h) =\int \frac{h^{\mathsf T}\Mob^{-1}h}{\rho}\dd x, \label{eq:Rayleighian_second_variation} \end{equation} which is strictly positive for nonzero admissible \(h\) on the support of \(\rho\). Therefore the stationary point is the unique minimizer. Finally, substituting \cref{eq:Rayleighian_minimizer} into \cref{eq:continuity_current} gives \cref{eq:overdamped_gradient_flow}. \end{proof} The same structure yields free-energy dissipation directly. Using \cref{eq:continuity_current} and integrating by parts, \begin{align} \frac{\dd F}{\dd t} &=\int \frac{\delta F}{\delta\rho}\partial_t\rho\dd x \notag\\ &=-\int \frac{\delta F}{\delta\rho}\nabla\cdot j_{\star}\dd x \notag\\ &=\int j_{\star}\cdot\nabla \frac{\delta F}{\delta\rho}\dd x \notag\\ &=-\int \rho \left( \nabla\frac{\delta F}{\delta\rho} \right)^{\mathsf T} \Mob \left( \nabla\frac{\delta F}{\delta\rho} \right)\dd x \le0. \label{eq:Rayleighian_free_energy_decay} \end{align} \subsubsection{Weighted Wasserstein Form of the JKO Step} For \(\Mob=I\), the Jordan-Kinderlehrer-Otto scheme uses the standard \(2\)-Wasserstein metric. When \(\Mob\ne I\) is a constant symmetric positive-definite matrix, the correct transport geometry is the metric induced by \(\Mob^{-1}\). Define \begin{equation} W_{\Mob}^2(\rho_0,\rho_1) =\inf_{\Pi\in\Gamma(\rho_0,\rho_1)} \int (x-y)^{\mathsf T}\Mob^{-1}(x-y) \dd\Pi(x,y), \label{eq:weighted_Wasserstein} \end{equation} where \(\Gamma(\rho_0,\rho_1)\) denotes the set of transport plans with the prescribed marginals. The corresponding minimizing-movement step is \begin{equation} \rho_{n+1} \in\argmin_{\rho} \left\{ F[\rho] +\frac{1}{2\tau} W_{\Mob}^2(\rho,\rho_n) \right\}. \label{eq:weighted_JKO} \end{equation} If \(y=\Mob^{-1/2}x\), then the cost in \cref{eq:weighted_Wasserstein} becomes the ordinary Euclidean quadratic cost. Thus \cref{eq:weighted_JKO} is the constant-mobility coordinate transform of the standard Wasserstein scheme. The JKO theorem is used here only for the gradient-drift Fokker-Planck class for which its hypotheses apply; it is not promoted to a theorem for arbitrary molecular SDEs. \subsubsection{State-Dependent Mobility and the It\^o Drift Correction} If the mobility depends on position, \(\Mob=\Mob(x)\), the It\^o equation must include a thermal drift correction to generate the desired free-energy gradient flow. Let \begin{equation} \sigma(x)\sigma(x)^{\mathsf T} =2k_BT\Mob(x). \label{eq:variable_mobility_covariance} \end{equation} Then the It\^o drift should be \begin{equation} b_i(x) =-\Mob_{ij}(x)\partial_jU(x) +k_BT\partial_j\Mob_{ij}(x). \label{eq:variable_mobility_drift} \end{equation} Indeed, the It\^o Fokker-Planck equation is \begin{equation} \partial_t\rho =-\partial_i(b_i\rho) +k_BT\partial_i\partial_j(\Mob_{ij}\rho). \label{eq:variable_mobility_FPE_start} \end{equation} Substituting \cref{eq:variable_mobility_drift} and expanding the second derivative gives cancellation of the terms containing \(\partial_j\Mob_{ij}\) and yields \begin{equation} \partial_t\rho =\nabla\cdot \left[ \Mob(x) \left( \rho\nabla U+k_BT\nabla\rho \right) \right]. \label{eq:variable_mobility_FPE_final} \end{equation} The constant-mobility theorem is retained as the clean foundational case; \cref{eq:variable_mobility_drift} records the precise modification required for state-dependent mobility. The consequence of omitting this term is visible without solving a time-dependent equation. On a periodic one-dimensional domain, write $\theta=k_BT$, take smooth $K(x)>0$, and use periodic $U$. The corrected drift has stationary density $\pi\propto e^{-U/\theta}$. If the It\^o drift is instead $-KU'$, its zero-current equation is $0=-KU'\widetilde\pi-\theta(K\widetilde\pi)'$, and therefore \[ \widetilde\pi(x)=\widetilde Z^{-1}K(x)^{-1}e^{-U(x)/\theta}. \] Spatially varying mobility then changes the equilibrium populations even though the intended potential and temperature have been held fixed. For the synthetic dimensionless choice $K=1+\tfrac12\cos x$, $U=0.6\sin x$, and $\theta=0.7$, the total-variation distance between the two normalized stationary laws is approximately $0.16397$. This is a change of model, not a time-discretization error. Figure~\ref{fig:mobility-reference-20261105} shows the resulting redistribution of probability. \begin{figure}[tbp] \centering \includegraphics[width=0.78\linewidth]{mobility_reference_20261105.pdf} \caption{Synthetic stationary-density comparison on a periodic domain. Both curves use the same potential, mobility, and temperature. The additional factor $K^{-1}$ is the consequence of omitting the thermal It\^o drift; no experimental data are used.} \label{fig:mobility-reference-20261105} \end{figure} \subsubsection{Conservative Transport Plus Dissipative Minimization} The preceding gradient-flow structure does not imply that every physical generator is symmetric. A more realistic representation is \begin{equation} L=A+S, \label{eq:AS_general} \end{equation} where \(A\) is conservative or circulatory transport and \(S\) is dissipative. The minimum-dissipation principle applies to the dissipative probability current generated by \(S\); it does not force \(A\) to vanish. This separation is closely related to Onsager's variational ideas and to the GENERIC framework of reversible-irreversible coupling. Accordingly, the decomposition itself is not claimed as novel. The contribution sought here lies in connecting this structure to the exact Markov relative-entropy theorem, the thermodynamic free-energy identity, the positive-Markov-inverse obstruction, law-level nonrecurrence, and history-space formulations. % ===================================================================== \subsection{Equilibrium Gibbs Measures and Thermodynamic Free Energy} \label{sec:thermodynamic_free_energy} Equilibrium supplies a reference throughout the motion, not just a name for its possible endpoint. For a time-independent energy function and bath temperature, the Gibbs law fixes the minimum of the nonequilibrium free energy. A different law has an excess free energy measured by relative entropy. This static comparison does not specify a relaxation time or guarantee that the dynamics reach the minimum; those questions also involve the mobility and the accessible states. We first fix the reference volume so that logarithms of densities and free-energy differences have unambiguous meanings. \subsubsection{Reference Measures and Dimensional Consistency} A continuous probability density generally has physical units, so an expression such as \(\log\rho\) should be interpreted relative to a fixed reference measure. Let \(m\) be a reference phase-space measure and assume \(\nu\ll m\) with density \begin{equation} r=\frac{\dd\nu}{\dd m}. \label{eq:reference_density} \end{equation} Define the entropy relative to \(m\) by \begin{equation} \Ent_m(\nu) =\int r\log r\dd m. \label{eq:relative_entropy_reference_measure} \end{equation} Any fixed dimensional normalization, such as a classical phase-space cell factor, changes the absolute free energy by an additive constant but does not change the dissipation identities proved below. Let \(\mathcal{H}\) be a measurable real-valued energy function and define the canonical equilibrium measure \begin{equation} \muEq(\dd x) =Z^{-1}e^{-\beta\mathcal{H}(x)}m(\dd x), \qquad Z=\int e^{-\beta\mathcal{H}}\dd m, \label{eq:canonical_gibbs_general} \end{equation} where \(00\), and \(\beta=(k_BT)^{-1}\). The nonequilibrium Helmholtz-type free energy is \begin{equation} F[\nu] =\int\mathcal{H}\dd\nu +k_BT\Ent_m(\nu). \label{eq:general_free_energy} \end{equation} \begin{theorem}[Exact Gibbs free-energy identity] \label{thm:gibbs_free_energy_identity} Assume additionally \(\int|\mathcal H|\,d\muEq<\infty\). Use the finite-term class \[ \mathcal A_F=\left\{\nu\ll\muEq: \int|\mathcal H|\,d\nu<\infty,\ \int r|\log r|\,dm<\infty,\quad r=d\nu/dm\right\}. \] For every probability \(\nu\in\mathcal A_F\), \begin{equation} F[\nu]-F[\muEq] =k_BT\KL{\nu}{\muEq}. \label{eq:gibbs_free_energy_identity} \end{equation} In particular, \(\muEq\in\mathcal A_F\) and is the unique minimizer of \(F\) in this class. \end{theorem} \begin{proof} Put \(w=e^{-\beta\mathcal H}\). Since \(w>0\) and \(\int w\,dm=Z<\infty\), the measurable sets \(B_n=\{w\geq1/n\}\), \(n\geq1\), cover the state space and satisfy \(m(B_n)\leq nZ\). Thus \(m\) is \(\sigma\)-finite; no separate volume assumption is needed. The Gibbs density is strictly positive, so \(m\) and \(\muEq\) have the same null sets. The finite-term assumptions make every logarithmic identity below integrable with respect to \(\nu\). Let \(r=\dd\nu/\dd m\). From \cref{eq:canonical_gibbs_general}, \begin{equation} \log\frac{\dd\muEq}{\dd m} =-\beta\mathcal{H}-\log Z. \label{eq:log_gibbs_density} \end{equation} Therefore \begin{align} \log\frac{\dd\nu}{\dd\muEq} &=\log r -\log\frac{\dd\muEq}{\dd m} \notag\\ &=\log r+\beta\mathcal{H}+\log Z. \label{eq:log_density_ratio_gibbs} \end{align} Integrating with respect to \(\nu\) yields \begin{align} \KL{\nu}{\muEq} &=\Ent_m(\nu) +\beta\int\mathcal{H}\dd\nu +\log Z. \label{eq:KL_expansion_gibbs} \end{align} Multiplying by \(k_BT\) gives \begin{equation} k_BT\KL{\nu}{\muEq} =F[\nu]+k_BT\log Z. \label{eq:KL_times_kT} \end{equation} For \(\nu=\muEq\), the relative entropy vanishes, so \begin{equation} F[\muEq]=-k_BT\log Z. \label{eq:equilibrium_free_energy_partition} \end{equation} The added energy moment and \(\log(d\muEq/dm)=-\beta\mathcal H-\log Z\) show that the equilibrium entropy term is absolutely integrable. Thus substituting \(\nu=\muEq\) above is legitimate. Substituting \cref{eq:equilibrium_free_energy_partition} into \cref{eq:KL_times_kT} proves \cref{eq:gibbs_free_energy_identity}. Relative entropy is nonnegative and vanishes only for equal measures, which proves uniqueness. \end{proof} There is a useful extension, but it must be defined explicitly: \[ F_{\rm rel}[\nu]=-k_BT\log Z+k_BT D(\nu\Vert\muEq). \] This expression is meaningful as an extended-real functional and agrees with \(F\) on \(\mathcal A_F\). Finite relative entropy alone need not make the separate energy and reference-entropy integrals finite. In particular, \(D(\muEq\Vert\muEq)=0\) even when the equilibrium energy moment is infinite. For example, use counting measure on \(n\ge2\), dimensionless \(k_BT=1\), and \(\mathcal H(n)=\log n+2\log\log n\). Then \(Z=\sum_{n\ge2}[n(\log n)^2]^{-1}<\infty\), whereas \(\sum_n\mathcal H(n)e^{-\mathcal H(n)}=\infty\), by comparison with \(\sum_{n\ge2}[n\log n]^{-1}\). The Gibbs relative entropy is zero, but the two separate terms in \eqref{eq:general_free_energy} do not define a finite sum. \subsubsection{Why a Minimum Does Not Imply Convergence} The existence of the minimizer in \cref{thm:gibbs_free_energy_identity} is not sufficient to prove that an arbitrary dynamics converges to that minimizer. A simple deterministic counterexample makes the logical point explicit. Let \begin{equation} V(x,y)=x^2+y^2, \label{eq:minimum_counterexample_potential} \end{equation} whose unique minimizer is \((0,0)\), and consider the rotational dynamics \begin{equation} \dot x=-y, \qquad \dot y=x. \label{eq:minimum_counterexample_dynamics} \end{equation} Then \begin{align} \frac{\dd}{\dd t}V(x_t,y_t) &=2x_t\dot x_t+2y_t\dot y_t \notag\\ &=-2x_ty_t+2y_tx_t =0. \label{eq:minimum_counterexample_derivative} \end{align} Thus the system moves forever on a level set of a function possessing a unique minimum. A quantitative sufficient condition for convergence is \begin{equation} \begin{gathered} H'(t)=-\mathcal I(t),\qquad \mathcal I(t)\ge2\rho H(t),\quad \rho>0,\quad H(0)<\infty\\ \Longrightarrow\quad H(t)\le e^{-2\rho t}H(0), \end{gathered} \label{eq:minimum_plus_coercivity} \end{equation} where \(H(t)=D(\nu_t\Vert\muEq)\). The dissipation inequality, rather than existence of a minimum alone, is the estimate that must be proved for the chosen dynamics. \subsubsection{Autonomous and Driven Thermodynamic Balances} Throughout this subsection heat is counted positive when it enters the system, \(T>0\) is fixed unless explicitly varied, and the mobility \(\Mob\) and diffusion \(\Diff=k_BT\Mob\) are interpreted in the same stochastic convention as the Fokker-Planck current. If an externally controlled potential \(U(x,t)\) is present, the protocol work rate performed on the system is \begin{equation} \dot W_t = \int\rho_t(x)\partial_tU(x,t)\dd x. \label{eq:external_work_rate} \end{equation} If no protocol is present, \(\dot W_t=0\). If the overdamped current has the force-driven form \begin{equation} j_t = \rho_t\Mob \left[ f_{\mathrm{nc}} -\nabla\frac{\delta F}{\delta\rho}(\rho_t) \right], \label{eq:force_driven_current_bridge} \end{equation} then the nonconservative work rate is \begin{equation} \dot W_{\mathrm{nc}}(t) = \int f_{\mathrm{nc}}(x,t)\cdot j_t(x)\dd x. \label{eq:nonconservative_work_rate} \end{equation} Combining \cref{eq:force_driven_current_bridge} with the free-energy rate induced by the continuity equation gives \begin{align} \dot W_t+\dot W_{\mathrm{nc}}(t)-\dot F_t &= \int \left[ f_{\mathrm{nc}} -\nabla\frac{\delta F}{\delta\rho}(\rho_t) \right]\cdot j_t\dd x \notag\\ &= \int \frac{j_t^{\mathsf T}\Mob^{-1}j_t}{\rho_t}\dd x \notag\\ &= k_BT \int \frac{j_t^{\mathsf T}\Diff^{-1}j_t}{\rho_t}\dd x \ge0. \label{eq:general_isothermal_force_bridge} \end{align} Thus the unforced Lyapunov identity is a special case of the driven balance, not a replacement for it. Time homogeneity alone is not enough: an autonomous nonequilibrium steady state can have \(\dot F=0\) and \(\dot W_{\mathrm{nc}}>0\). For autonomous unforced relaxation with a single ideal bath at fixed temperature, assume in addition that the effective dynamics obey local detailed balance with the heat convention stated above. Define $S_{\rm sys}=-k_B\Ent_m(\nu_t)$ relative to the fixed reference volume, and $\dot S_{\rm tot}=\dot S_{\rm sys}-\dot Q_{\rm in}/T$. The first law gives $\dot E=\dot Q_{\rm in}$, while $F=E-TS_{\rm sys}$. Consequently \begin{equation} T\dot S_{\mathrm{tot}} =-\dot F. \label{eq:autonomous_balance} \end{equation} If an externally controlled potential \(U(x,t)\) is present, the free-energy change contains the mechanical or parametric work defined in \cref{eq:external_work_rate}. The appropriate nonequilibrium balance becomes \begin{equation} T\dot S_{\mathrm{tot}} =\dot W_t+\dot W_{\mathrm{nc}}(t)-\dot F_t \ge0. \label{eq:driven_second_law_balance} \end{equation} The optical-trap quench in \cref{sec:experiments} is interpreted using this distinction: the rapid switching stage performs work, whereas the subsequent fixed-parameter relaxation is autonomous. A scalar harmonic calculation fixes the sign convention. Consider \begin{equation} \dd X_t = -K\kappa(t)X_t\dd t +\sqrt{2Kk_BT}\dd W_t, \label{eq:harmonic_sign_oracle_sde} \end{equation} where \(K>0\), \(\kappa(t)>0\), and \(y(t)=\EE[X_t^2]\). It\^o's formula gives \begin{equation} \dot y = -2K(\kappa y-k_BT). \label{eq:harmonic_sign_oracle_variance} \end{equation} Up to an additive constant independent of \(y\), the centered Gaussian free energy is \begin{equation} F(t) = \frac12\kappa(t)y(t) -\frac{k_BT}{2}\log y(t). \label{eq:harmonic_sign_oracle_free_energy} \end{equation} The work rate performed on the system is \(\dot W_t=\dot\kappa\,y/2\), and differentiation yields \begin{align} \dot F_t &= \frac12\dot\kappa\,y +\frac12 \left( \kappa-\frac{k_BT}{y} \right)\dot y \notag\\ &= \dot W_t -K\frac{(\kappa y-k_BT)^2}{y}. \label{eq:harmonic_sign_oracle_derivative} \end{align} Consequently, \begin{equation} T\dot S_{\mathrm{tot}} = \dot W_t-\dot F_t = K\frac{(\kappa y-k_BT)^2}{y} \ge0. \label{eq:harmonic_sign_oracle_entropy} \end{equation} At the dimensionless regression point \(K=k_BT=y=1\), \(\kappa=2\), and \(\dot\kappa=10\), this gives \begin{equation} \dot W_t=5, \qquad T\dot S_{\mathrm{tot}}=1, \qquad \dot F_t=4>0. \label{eq:harmonic_sign_oracle_numbers} \end{equation} This example shows explicitly why a driven process may increase nonequilibrium free energy even though the total entropy production is nonnegative. \subsubsection{Local Detailed Balance as an Additional Physical Bridge} \label{sec:local_detailed_balance_bridge} The free-energy identity compares probability laws; a heat measurement records energy transferred to a reservoir. To connect them, the forward and reversed transition weights must encode the same reservoir energy exchange. Local detailed balance supplies this relation. The sign convention is fixed below so that heat entering the system is the negative of heat entering the bath. This is the additional physical step that turns an information inequality into a thermodynamic balance. The equality between a path likelihood ratio and physical entropy production is not a consequence of Markovity or Jensen's inequality alone. It requires a model-specific local-detailed-balance relation connecting elementary forward and reversed transition weights to reservoir heat \cite{Seifert2005,Seifert2012,ZhangLu2025}. This premise is separate from relative-entropy contraction: data processing is information-theoretic, whereas local detailed balance assigns thermodynamic meaning to the logarithm of a path probability ratio. For a jump \(i\to j\) through reservoir channel \(\nu\), declare the dynamically reversed channel \(\bar\nu\). The support condition is \begin{equation} k_{ij}^{(\nu)}>0 \quad\Longleftrightarrow\quad k_{ji}^{(\bar\nu)}>0. \label{eq:jump_channel_support_LDB} \end{equation} When the same channel label is physically valid in both directions, \(\bar\nu=\nu\), the local detailed-balance convention reads \begin{equation} \log \frac{k_{ij}^{(\nu)}}{k_{ji}^{(\nu)}} = \frac{\Delta s_{\mathrm{med}}^{(\nu)}(i\to j)}{k_B} = -\beta_\nu Q_{\mathrm{in}}^{(\nu)}(i\to j), \label{eq:jump_LDB_heat} \end{equation} where \(Q_{\mathrm{in}}\) is heat entering the system. For one bath and no mechanical or chemical work, \cref{eq:jump_LDB_heat} reduces to \begin{equation} \frac{k_{ij}}{k_{ji}} = \exp[-\beta(E_j-E_i)], \label{eq:jump_Gibbs_detailed_balance} \end{equation} which is the Gibbs detailed-balance ratio. For a complete jump-path likelihood ratio, the rate ratio is not enough. Let \(r_i^F(t)=\sum_{j,\nu}k_{ij}^{F,(\nu)}(t)\) be the forward escape rate and let \(r_i^R(t)\) be the escape rate in the reversed protocol. The holding-time contribution cancels only under the declared convention, for example when \begin{equation} r_i^R(\tau-t)=r_i^F(t) \label{eq:jump_escape_rate_cancellation} \end{equation} along the reversed path. If \cref{eq:jump_channel_support_LDB} fails, the logarithm in \cref{eq:jump_LDB_heat} is an extended value and singular path components may appear. In that case the ordinary finite path-KL and ordinary integral fluctuation theorem must be replaced by an absolute-continuity statement with the appropriate singular correction \cite{MurashitaFunoUeda2014}. For diffusion paths in the Stratonovich convention, the corresponding work and heat increments are \begin{align} \delta W_{\mathrm{on}} &= \partial_tU_t(X_t)\dd t + f_{\mathrm{nc}}(X_t,t)\circ\dd X_t, \label{eq:diffusion_protocol_work_increment} \\ \delta Q_{\mathrm{bath}} &= [-\nabla U_t(X_t)+f_{\mathrm{nc}}(X_t,t)] \circ\dd X_t, \qquad \delta Q_{\mathrm{in}}=-\delta Q_{\mathrm{bath}}. \label{eq:diffusion_heat_increment} \end{align} The Stratonovich chain rule gives the first-law convention \begin{equation} \dd U = \delta W_{\mathrm{on}} -\delta Q_{\mathrm{bath}} = \delta W_{\mathrm{on}} +\delta Q_{\mathrm{in}}. \label{eq:diffusion_first_law_path} \end{equation} With a declared reverse protocol, parity map, endpoint initialization, and common covariance, the conditional path calculation may identify the medium entropy with a path likelihood ratio, \begin{equation} \log \frac{\dd P_F(\cdot\mid x_0)} {\dd P_R(\Theta\cdot\mid \vartheta x_\tau)} = \beta Q_{\mathrm{bath}}[\omega] = \frac{\Delta s_{\mathrm{med}}[\omega]}{k_B}. \label{eq:diffusion_medium_entropy_path_ratio} \end{equation} Only after the endpoint density term is restored does this become total entropy production. Local detailed balance fixes the time-antisymmetric ratio of forward and backward transition weights. It does not fix the time-symmetric kinetic activity and therefore does not uniquely determine the generator; moreover, coarse graining can destroy local detailed balance even when the reduced process remains Markovian \cite{HartichGodec2023}. \subsubsection{Optional Microcanonical Ensemble Formulation} For a genuinely isolated ensemble, suppose the stochastic dynamics preserves the energy shell \begin{equation} \Sigma_E=\{x:\mathcal{H}(x)=E\} \label{eq:energy_shell} \end{equation} and admits a normalized invariant microcanonical reference measure \(\mu_E\) on that shell. For ensemble laws \(\nu\ll\mu_E\) define the information-theoretic microcanonical entropy \begin{equation} S_E[\nu] =S_E^{\mathrm{eq}} -k_B\KL{\nu}{\mu_E}. \label{eq:microcanonical_information_entropy} \end{equation} If \(\mu_EP_t=\mu_E\), Markov data processing implies \begin{equation} S_E[\nu P_t] \ge S_E[\nu]. \label{eq:microcanonical_entropy_monotonicity} \end{equation} This is an ensemble theorem relative to the chosen microcanonical measure. It is not assigned to singular point masses for which the relative entropy may be infinite, and it is not used as a substitute for the canonical theorem. % ===================================================================== % ===================================================================== \section{Entropy Dissipation and Convergence to Equilibrium} \label{sec:entropy_convergence} The preceding current calculation gives an instantaneous free-energy balance. We now follow an entire probability law through time. Three questions must be distinguished: does relative entropy decrease, can its rate be computed, and must it eventually reach zero? Markov data processing answers the first, diffusion calculus answers the second under its domain assumptions, and coercive or hypocoercive estimates answer the third. For a particle cloud in a trap, these correspond to the direction of relaxation, its instantaneous rate, and the time needed to approach the equilibrium cloud. The equilibrium reference is the same in all three questions, but the required dynamical information is different. \subsection{Markov Relative-Entropy Contraction and Differential Dissipation} \label{sec:markov_entropy} \subsubsection{Relative Entropy} Let \(\nu\) and \(\mu\) be probability measures on a measurable state space \((\cX,\mathcal B)\) and assume \(\nu\ll\mu\). Write \begin{equation} f=\frac{\dd\nu}{\dd\mu}. \label{eq:relative_density_definition} \end{equation} The Kullback-Leibler relative entropy is \begin{equation} \KL{\nu}{\mu} =\int f\log f\dd\mu. \label{eq:KL_definition} \end{equation} If \(\nu\not\ll\mu\), the relative entropy is defined to be \(+\infty\). \subsubsection{Dimension-Free Markov Data Processing} \begin{theorem}[Invariant-measure relative-entropy contraction] \label{thm:Markov_DPI} Let \(P\) be a Markov kernel and let \(\mu\) be a probability measure satisfying \begin{equation} \mu P=\mu. \label{eq:kernel_invariance} \end{equation} Then for every probability measure \(\nu\), \begin{equation} \KL{\nu P}{\mu} \le \KL{\nu}{\mu}. \label{eq:Markov_DPI} \end{equation} Consequently, if \((P_t)_{t\ge0}\) is a Markov semigroup preserving \(\mu\), then \(t\mapsto\KL{\nu P_t}{\mu}\) is nonincreasing. \end{theorem} \begin{proof} If \(\KL{\nu}{\mu}=+\infty\), the inequality is immediate. Assume therefore that \(\nu\ll\mu\) and define \(f\) by \cref{eq:relative_density_definition}. Introduce two probability measures on \(\cX\times\cX\): \begin{align} Q(\dd x,\dd y) &=\nu(\dd x)P(x,\dd y), \label{eq:joint_Q} \\ R(\dd x,\dd y) &=\mu(\dd x)P(x,\dd y). \label{eq:joint_R} \end{align} The two joint laws possess the same conditional transition kernel, hence \begin{equation} \frac{\dd Q}{\dd R}(x,y)=f(x). \label{eq:joint_density_ratio} \end{equation} Therefore \begin{align} \KL{Q}{R} &=\int\log f(x)\,Q(\dd x,\dd y) \notag\\ &=\int f(x)\log f(x)\mu(\dd x) \notag\\ &=\KL{\nu}{\mu}. \label{eq:joint_KL_equals_input} \end{align} Project the pair \((x,y)\) onto its second coordinate. The second marginal of \(Q\) is \(\nu P\), while the second marginal of \(R\) is \(\mu P=\mu\). Relative entropy cannot increase under a measurable map, so \begin{equation} \KL{\nu P}{\mu} \le \KL{Q}{R} =\KL{\nu}{\mu}, \label{eq:DPI_projection_proof} \end{equation} which proves the theorem. \end{proof} \subsubsection{Exact Equality Condition} The equality case has an exact interpretation as recoverability of the initial likelihood ratio from the output. \begin{proposition}[Equality in the Markov data-processing inequality] \label{prop:DPI_equality} Assume \(0\le\KL{\nu}{\mu}<\infty\) and use the stationary joint law \(R\) from \cref{eq:joint_R}. Let \(X\) and \(Y\) denote the first and second coordinates. Equality in \cref{eq:Markov_DPI} holds if and only if \begin{equation} f(X) =\EE_R[f(X)\mid Y] \qquad R\text{-almost surely}. \label{eq:DPI_equality_condition} \end{equation} Equivalently, there exists a measurable function \(g\) such that \(f(X)=g(Y)\) \(R\)-almost surely. \end{proposition} \begin{proof} Let \begin{equation} g(Y)=\EE_R[f(X)\mid Y]. \label{eq:output_density_conditional_expectation} \end{equation} For every bounded measurable \(\varphi\), \begin{align} \int\varphi(y)g(y)\mu(\dd y) &=\EE_R[\varphi(Y)g(Y)] \notag\\ &=\EE_R[\varphi(Y)f(X)] \notag\\ &=\int\varphi(y)(\nu P)(\dd y). \label{eq:output_density_test} \end{align} Hence \(g=\dd(\nu P)/\dd\mu\). Let \(\Psi(r)=r\log r\). Since \begin{equation} \Psi''(r)=\frac1r>0 \qquad (r>0), \label{eq:strict_convexity_rlogr} \end{equation} conditional Jensen gives \begin{equation} \Psi(g(Y)) =\Psi(\EE[f(X)\mid Y]) \le \EE[\Psi(f(X))\mid Y]. \label{eq:conditional_Jensen} \end{equation} After taking expectations, \cref{eq:conditional_Jensen} is precisely the data-processing inequality. Strict convexity gives equality if and only if \(f(X)\) is almost surely constant conditional on \(Y\), which is exactly \cref{eq:DPI_equality_condition}. The value \(\Psi(0)=0\) is defined by continuity. Strict convexity extends to \([0,\infty)\): when \(s>0\) and \(00\), \begin{equation} f_t\in\Dom(L)\cap\Dom(L^{\dagger_\mu}), \qquad \log f_t\in\Dom(L), \label{eq:strong_domain_assumption} \end{equation} and that every pairing below is integrable. Define \begin{equation} H(t)=\KL{\nu_t}{\mu}. \label{eq:H_definition} \end{equation} Then normalization gives \begin{equation} \int\partial_tf_t\dd\mu=0. \label{eq:density_mass_derivative_zero} \end{equation} Hence \begin{align} H'(t) &=\int (\partial_tf_t)\log f_t\dd\mu \notag\\ &=\int (L^{\dagger_\mu}f_t)\log f_t\dd\mu \notag\\ &=\int f_tL\log f_t\dd\mu. \label{eq:entropy_derivative_adjoint} \end{align} Using \cref{eq:log_chain_rule}, \begin{align} H'(t) &=\int Lf_t\dd\mu -\int f_t\Gamma_L(\log f_t)\dd\mu. \label{eq:entropy_chain_expanded} \end{align} Invariance gives \begin{equation} \int Lf_t\dd\mu=0, \label{eq:invariance_generator_integral} \end{equation} so \begin{equation} H'(t) =-\int f_t\Gamma_L(\log f_t)\dd\mu. \label{eq:strong_entropy_identity} \end{equation} \subsubsection{Why the Antisymmetric Part Does Not Directly Dissipate State Relative Entropy} Suppose \(A_\mu\) is a first-order derivation on a common algebra core. Let \begin{equation} \Psi(r)=r\log r-r, \label{eq:Psi_entropy_primitive} \end{equation} so that \(\Psi'(r)=\log r\). The derivation property gives \begin{equation} A_\mu\Psi(f)=(A_\mu f)\log f. \label{eq:A_derivation_entropy} \end{equation} Since \(A_\mu 1=0\) and \(A_\mu^{\dagger_\mu}=-A_\mu\), \begin{equation} \int A_\mu g\dd\mu =\ip{1}{A_\mu g}_{L^2(\mu)} =-\ip{A_\mu 1}{g}_{L^2(\mu)} =0. \label{eq:A_integral_zero} \end{equation} Applying this identity to \(g=\Psi(f)\) yields \begin{equation} \int(A_\mu f)\log f\dd\mu=0. \label{eq:A_entropy_zero} \end{equation} Thus the antisymmetric derivation transports probability without directly contributing to the state-relative-entropy dissipation. This statement is an entropic statement; it is not by itself a statement of physical reversibility. \subsubsection{Jump Generators and the Scope of the Antisymmetric Cancellation} \label{sec:jump_entropy_scope} The conclusion of \cref{eq:A_entropy_zero} is local. It relies on the first-order derivation property \(A_\mu\Psi(f)=\Psi'(f)A_\mu f\). A nonlocal jump generator does not satisfy this chain rule, and its antisymmetric part may contribute nontrivially to the derivative of state relative entropy. Discrete entropy decay therefore uses jump-process Dirichlet forms and modified logarithmic Sobolev inequalities rather than the local diffusion square-root identity \cite{CaputoDaiPraPosta2009,BobkovTetali2006,Erbar2014}. For a finite-state generator \begin{equation} (Lg)_i = \sum_{j\ne i}q_{ij}(g_j-g_i), \qquad \mu L=0, \label{eq:finite_jump_generator} \end{equation} assume \(\mu_i>0\), \(q_{ij}<\infty\), and \(f_i(t)>0\). Then \begin{align} \frac{\dd}{\dd t}\Ent_\mu(f_t) &= \ip{f_t}{L\log f_t}_{L^2(\mu)} \notag\\ &= \sum_i\mu_i f_i \sum_{j\ne i}q_{ij} \log\frac{f_j}{f_i}. \label{eq:finite_jump_entropy_derivative} \end{align} Equivalently, using stationarity of \(\mu\), \begin{equation} -\frac{\dd}{\dd t}\Ent_\mu(f_t) = \sum_{i\ne j} \mu_iq_{ij} \left[ f_i\log\frac{f_i}{f_j} -f_i+f_j \right]. \label{eq:finite_jump_entropy_bregman} \end{equation} This directed Bregman sum is nonnegative under the Markov contraction theorem, but it is not the same object as the physical housekeeping entropy production of a stationary nonequilibrium jump process. Even in the reversible two-state case, the square-root expression is generally only a lower bound. For \begin{equation} L= \begin{pmatrix} -1&1\\ 1&-1 \end{pmatrix}, \qquad \mu=\left(\frac12,\frac12\right), \qquad f=(1.6,0.4), \label{eq:two_state_jump_fixture} \end{equation} one obtains \begin{equation} \mathcal E(f,\log f) = -\ip{f}{L\log f}_{L^2(\mu)} = 0.6\log4 = 0.8317766166719344\ldots, \label{eq:two_state_exact_entropy_value} \end{equation} whereas \begin{equation} 4\mathcal E(\sqrt f,\sqrt f) = 4(\sqrt{1.6}-\sqrt{0.4})^2 \frac12 = 0.8. \label{eq:two_state_square_root_value} \end{equation} Thus the inequality \begin{equation} (a-b)(\log a-\log b) \ge 4(\sqrt a-\sqrt b)^2 \label{eq:jump_log_sqrt_inequality} \end{equation} must not be silently replaced by equality. For a nonreversible three-state cycle, take \begin{equation} L= \begin{pmatrix} -2.3&2&0.3\\ 0.3&-2.3&2\\ 2&0.3&-2.3 \end{pmatrix}, \qquad \mu=\left(\frac13,\frac13,\frac13\right), \qquad f=(1.5,1,0.5). \label{eq:three_state_cycle_fixture} \end{equation} Since \(\mu\) is uniform, \(L^{\dagger_\mu}=L^{\mathsf T}\). With \begin{equation} S_\mu=\frac12(L+L^{\mathsf T}), \qquad A_\mu=\frac12(L-L^{\mathsf T}), \label{eq:three_state_SA} \end{equation} direct multiplication gives \begin{align} \ip{f}{L\log f}_{L^2(\mu)} &= -0.5909471057201607\ldots, \label{eq:three_state_total_value} \\ \ip{f}{S_\mu\log f}_{L^2(\mu)} &= -0.6317020659841630\ldots, \label{eq:three_state_symmetric_value} \\ \ip{f}{A_\mu\log f}_{L^2(\mu)} &= 0.04075496026400233\ldots. \label{eq:three_state_antisymmetric_value} \end{align} The final number is nonzero. It does not contradict \cref{eq:A_entropy_zero}, because the finite-state \(A_\mu\) is not a first-order derivation on a local algebra. It only rules out the unqualified statement that every antisymmetric generator component makes zero direct contribution to the state-entropy derivative. % ===================================================================== \subsection{Coercivity, Hypocoercivity, and Convergence to Equilibrium} \label{sec:coercivity} An entropy identity gives a rate of loss, but the rate could become small while the law remains far from equilibrium. A coercive estimate rules this out by bounding the entropy itself in terms of its dissipation. In kinetic dynamics the noise acts only on momentum, so the missing control of position must come from transport between the two variables. This transfer is the physical role of hypocoercivity. The curvature and functional inequalities below quantify these two routes to relaxation. \subsubsection{Bakry-Emery Curvature in the Elliptic Reversible Prototype} For a diffusion generator \(L\) with carre du champ \(\Gamma_L\), define \begin{equation} \Gamma_{2,L}(g) =\frac12L\Gamma_L(g)-\Gamma_L(g,Lg). \label{eq:Gamma2_definition} \end{equation} The Bakry-Emery curvature condition \(BE(\rho,\infty)\) \cite{BakryEmery1985,BakryGentilLedoux2014} is \begin{equation} \Gamma_{2,L}(g) \ge\rho\Gamma_L(g) \qquad \text{for all admissible }g, \label{eq:BE_condition} \end{equation} with \(\rho>0\). For the weighted Riemannian diffusion \begin{equation} L=\Delta-\nabla V\cdot\nabla, \label{eq:weighted_laplacian} \end{equation} the Bochner identity gives \begin{equation} \Gamma_{2,L}(g) =\|\nabla^2g\|_{\mathrm{HS}}^2 +\left( \operatorname{Ric}+\nabla^2V \right)(\nabla g,\nabla g). \label{eq:Bochner_BE} \end{equation} Consequently, \begin{equation} \operatorname{Ric}+\nabla^2V \ge\rho I \label{eq:BE_geometric_bound} \end{equation} is a sufficient curvature condition. \subsubsection{A Minimum Is Not a Curvature Bound} The distinction between minimization and positive curvature is elementary but foundational. Consider \begin{equation} V(x)=x^4-x^2. \label{eq:double_well_example} \end{equation} The function possesses global minima, but \begin{equation} V''(x)=12x^2-2, \label{eq:double_well_second_derivative} \end{equation} and hence \begin{equation} V''(0)=-2. \label{eq:double_well_negative_curvature} \end{equation} Therefore the statement ``the action has a minimum'' cannot be used to infer a positive Bakry-Emery constant. \subsubsection{Logarithmic Sobolev Inequality and Exponential Entropy Decay} Assume the invariant measure \(\mu\) satisfies the logarithmic Sobolev inequality \begin{equation} \Ent_\mu(g^2) \le \frac{2}{\rho} \mathcal{E}_{S,\mu}(g,g), \qquad \rho>0, \label{eq:LSI_general} \end{equation} where \begin{equation} \Ent_\mu(g^2) =\int g^2\log g^2\dd\mu -\left(\int g^2\dd\mu\right) \log\left(\int g^2\dd\mu\right). \label{eq:entropy_functional_g2} \end{equation} Set \(g=\sqrt f\) for a normalized density \(f\). Then \begin{equation} \Ent_\mu(f)=\KL{f\mu}{\mu}, \label{eq:LSI_density_entropy} \end{equation} and by \cref{eq:Fisher_Dirichlet_definition}, \begin{equation} \KL{f\mu}{\mu} \le \frac{1}{2\rho}\mathcal{I}_\mu(f). \label{eq:LSI_Fisher_form} \end{equation} Equivalently, \begin{equation} \mathcal{I}_\mu(f) \ge 2\rho\KL{f\mu}{\mu}. \label{eq:Fisher_lower_bound} \end{equation} Combining \cref{eq:Fisher_lower_bound} with the entropy-dissipation identity yields \begin{equation} H'(t) \le-2\rho H(t). \label{eq:entropy_Gronwall_differential} \end{equation} Gronwall's inequality therefore gives \begin{equation} H(t) \le e^{-2\rho(t-s)}H(s) \qquad t\ge s. \label{eq:entropy_exponential_decay} \end{equation} Pinsker's inequality then implies \begin{equation} \|\nu_t-\mu\|_{\mathrm{TV}} \le \sqrt{\frac{H(s)}{2}} e^{-\rho(t-s)}. \label{eq:Pinsker_exponential} \end{equation} \subsubsection{The Second Law in the Autonomous Isothermal Class} Combining \cref{thm:gibbs_free_energy_identity,thm:integrated_entropy_dissipation} gives the central thermodynamic statement. \begin{theorem}[Autonomous unforced isothermal stochastic second law] \label{thm:isothermal_second_law} Assume \cref{ass:M1,ass:M2,ass:M3,ass:M4,ass:M5}. Suppose \(\muEq\) is the thermodynamic Gibbs law, \(\nu_s\in\mathcal A_F\), and the finite-term free-energy identity holds at the times considered. Require the regularity or convergent approximation hypotheses of \cref{thm:integrated_entropy_dissipation}; a weak-limit inequality alone does not imply the equality below. Throughout this theorem the reservoir temperature \(T>0\), the mobility/noise calibration, and the potential are time independent, and no nonconservative work source is present. Then \begin{equation} F[\nu_t] +k_BT\int_s^t\mathcal{I}_{\muEq}(f_r)\dd r =F[\nu_s], \label{eq:free_energy_integrated_balance} \end{equation} and therefore \begin{equation} F[\nu_t]\le F[\nu_s] \qquad (t\ge s). \label{eq:free_energy_monotonicity} \end{equation} Whenever the differential identity is valid and the physical heat-bath interpretation in \cref{ass:M13} is available, \begin{equation} \dot S_{\mathrm{tot}}(t) =k_B\mathcal{I}_{\muEq}(f_t) \ge0. \label{eq:second_law_entropy_production} \end{equation} If, in addition, \cref{ass:M8} supplies an LSI of constant \(\rho>0\), then \begin{equation} F[\nu_t]-F[\muEq] \le e^{-2\rho(t-s)} \left[ F[\nu_s]-F[\muEq] \right]. \label{eq:free_energy_exponential_convergence} \end{equation} \end{theorem} \begin{proof} By \cref{eq:gibbs_free_energy_identity}, \begin{equation} F[\nu_r]-F[\muEq] =k_BT\KL{\nu_r}{\muEq} \label{eq:second_law_proof_identity} \end{equation} for every admissible time \(r\). Apply \cref{eq:integrated_entropy_dissipation} with \(\mu=\muEq\): \begin{equation} \KL{\nu_t}{\muEq} +\int_s^t\mathcal{I}_{\muEq}(f_r)\dd r =\KL{\nu_s}{\muEq}. \label{eq:second_law_proof_KL_balance} \end{equation} Multiply by \(k_BT\) and use \cref{eq:second_law_proof_identity} at \(r=s,t\). The equilibrium constant cancels, giving \cref{eq:free_energy_integrated_balance}. Since the integral is nonnegative, \cref{eq:free_energy_monotonicity} follows. For autonomous isothermal relaxation, the thermodynamic balance \cref{eq:autonomous_balance} gives \begin{equation} \dot S_{\mathrm{tot}} =-T^{-1}\dot F =k_B\mathcal{I}_{\muEq}(f_t), \label{eq:second_law_proof_entropy_rate} \end{equation} which proves \cref{eq:second_law_entropy_production}. Finally, the LSI estimate \cref{eq:entropy_exponential_decay} multiplied by \(k_BT\) gives \cref{eq:free_energy_exponential_convergence}. \end{proof} \subsubsection{Kinetic Entropy Dissipation and the Failure of Instantaneous Strictness} For the underdamped generator \cref{eq:kinetic_generator_decomp}, the carre du champ is generated entirely by the momentum diffusion: \begin{equation} \Gamma_L(g) =k_BT (\nabla_pg)^{\mathsf T} \Gfr \nabla_pg. \label{eq:kinetic_carre_du_champ} \end{equation} Hence, under the corresponding entropy regularity, \begin{equation} \frac{\dd}{\dd t} \KL{f_t\piEq}{\piEq} =-k_BT \int f_t (\nabla_p\log f_t)^{\mathsf T} \Gfr \nabla_p\log f_t \dd\piEq. \label{eq:kinetic_entropy_dissipation} \end{equation} Now let \begin{equation} f(q,p)=g(q) \label{eq:kinetic_zero_instant_diss_density} \end{equation} with \(g\) nonconstant and normalized. Then \begin{equation} \nabla_p\log f=0, \label{eq:kinetic_zero_p_gradient} \end{equation} so the right-hand side of \cref{eq:kinetic_entropy_dissipation} vanishes at that instant even though \(f\ne1\). Therefore the implication \begin{equation} \mathcal{I}(f)=0 \Longrightarrow f=1 \label{eq:invalid_kinetic_strictness} \end{equation} is not valid as an instantaneous statement for the degenerate kinetic diffusion. \subsubsection{Hypocoercive Transfer of Information} The missing mechanism is coupling. Hamiltonian transport differentiates in both \(q\) and \(p\), while the noise differentiates only in \(p\). Commutators transfer regularity and decay between these directions. This is the essence of hypocoercivity, for which Villani \cite{Villani2009} develops a systematic framework. For a quadratic potential the transition can be computed exactly; the following theorem derives explicit entropy bounds without assuming a kinetic decay estimate. \begin{theorem}[Explicit entropy convergence for a quadratic kinetic branch] \label{thm:conditional_hypocoercivity} Fix dimensionless coordinates and time, and let \(k,\gamma,\theta>0\). Consider the one-coordinate Langevin system \[ dq_t=p_t\,dt,\qquad dp_t=(-kq_t-\gamma p_t)\,dt+\sqrt{2\gamma\theta}\,dW_t. \] Its unique invariant probability is \(\piEq=N(0,\operatorname{diag}(\theta/k,\theta))\). For every initial law \(\nu_0\) with finite relative entropy, put \begin{align*} B&=\begin{pmatrix}0&\sqrt{k}\\-\sqrt{k}&-\gamma\end{pmatrix}, &R_t&=e^{tB},&r_t&=\|R_t\|_{\mathrm{op}},\\ \mathsf P&= \begin{pmatrix} \gamma^{-1}+\gamma/(2k)&1/(2\sqrt{k})\\ 1/(2\sqrt{k})&\gamma^{-1} \end{pmatrix}. \end{align*} Let \(p_-\) and \(p_+\) be the smallest and largest eigenvalues of \(\mathsf P\). Then \(00\), and \begin{equation} \KL{\nu_t}{\piEq} \le r_t^2\KL{\nu_0}{\piEq} \le \min\{1,(p_+/p_-)e^{-t/p_+}\}\KL{\nu_0}{\piEq}. \label{eq:conditional_hypocoercive_rate} \end{equation} Thus relative entropy converges to zero with explicit constants, and \(\|\nu_t-\piEq\|_{\mathrm{TV}}\le \sqrt{\KL{\nu_t}{\piEq}/2}\). \end{theorem} \begin{proof} In the normalized variables \(x=(\sqrt{k/\theta}\,q,p/\sqrt{\theta})\), the invariant covariance is the identity and the diffusion covariance is \(A=\operatorname{diag}(0,2\gamma)\). Direct substitution gives \(B+B^{\mathsf T}+A=0\). The characteristic polynomial of \(B\) is \(z^2+\gamma z+k\), whose roots have negative real part. The exact transition is therefore the Gaussian channel \begin{align*} x_t&=R_tx_0+\xi_t,\qquad \xi_t\sim N(0,Q_t),\\ Q_t&=\int_0^t e^{sB}Ae^{sB^{\mathsf T}}\,ds =I-R_tR_t^{\mathsf T}, \end{align*} with \(\xi_t\) independent of \(x_0\). The two columns \(e_2\) and \(Be_2=(\sqrt{k},-\gamma)^{\mathsf T}\) span \(\mathbb R^2\). More explicitly, if \(v^{\mathsf T}Q_tv=0\), continuity and nonnegativity of the integrand give \(v^{\mathsf T}e^{sB}e_2=0\) on \([0,t]\). Its value and derivative at zero imply \(v_2=0\) and \(\sqrt{k}v_1=0\). Thus \(Q_t>0\), and \(r_t<1\). Invertibility of the matrix exponential gives \(r_t>0\). The same transition shows convergence to \(N(0,I)\) and uniqueness: in the stationary characteristic-function identity, let \(t\) tend to infinity to obtain the Gaussian characteristic function. For any matrix \(U\) with \(\|U\|_{\mathrm{op}}\le1\), the channel \[ \mathcal K_U(x,\cdot)=N(Ux,I-UU^{\mathsf T}) \] preserves \(N(0,I)\), including when its noise covariance is singular. At a fixed positive time set \(U_t=R_t/r_t\). Applying first \(\mathcal K_{U_t}\) and then the scalar channel \(\mathcal K_{r_tI}\) has mean \(R_tx\) and covariance \(r_t^2(I-U_tU_t^{\mathsf T})+(1-r_t^2)I=Q_t\). Hence it is exactly the transition above. Data processing bounds entropy after the first channel by the initial entropy. The second is the standard Gaussian Ornstein--Uhlenbeck semigroup at time \(s=-\log r_t\); Gross's logarithmic Sobolev inequality \cite{Gross1975}, with generator \(\Delta-x\cdot\nabla\), gives the further factor \(e^{-2s}=r_t^2\). The conclusion extends from regular densities to every finite-entropy law by entropy approximation and the semigroup contraction. Relative entropy is unchanged by the invertible coordinate normalization. Finally, \(\mathsf P>0\), since its first diagonal entry is positive and \(\det\mathsf P=\gamma^{-2}+(4k)^{-1}>0\). Matrix multiplication gives \[ B^{\mathsf T}\mathsf P+\mathsf PB=-I. \] Along \(\dot z=Bz\), \[ \frac d{dt}(z^{\mathsf T}\mathsf Pz) =-|z|^2\le-p_+^{-1}z^{\mathsf T}\mathsf Pz. \] Norm equivalence and Gronwall yield \(\|R_t\|_{\mathrm{op}}^2\le(p_+/p_-)e^{-t/p_+}\). Combining this with \(r_t\le1\) proves the estimate. Pinsker's inequality gives the total-variation bound in the convention \(\|\nu-\pi\|_{\mathrm{TV}}=\sup_A|\nu(A)-\pi(A)|\). \end{proof} \begin{remark} The theorem derives a rate for the specified quadratic branch using the established Gaussian logarithmic Sobolev inequality after an explicit channel factorization. The finite-time factor \(r_t^2\) is optimal among bounds of this form: a translated \(N(m,I)\) has input entropy \(|m|^2/2\) and output entropy \(|R_tm|^2/2\), with equality when \(m\) is a maximizing right singular vector of \(R_t\). For \(k=\gamma=1\), \(p_\pm=(5\pm\sqrt5)/4\), so the coarse prefactor is \((3+\sqrt5)/2\) and the exponential rate is \(4/(5+\sqrt5)\). For an initial position displacement, the entropy derivative at zero still vanishes; the prefactor cannot in general be replaced by one while retaining a positive uniform exponential rate from time zero. Nonquadratic confinement requires a separate theorem with its actual Hessian, regularity, and functional-inequality assumptions. \end{remark} The corresponding numerical diagnostic is shown in Figure~\ref{fig:kinetic-entropy14}. \begin{figure}[htbp] \centering \includegraphics[width=.93\linewidth]{kinetic_entropy14.pdf} \caption{Exact entropy relaxation in the normalized quadratic kinetic model with \(k=\gamma=\theta=1\). The two translated Gaussian initial laws have mean \(e_1\) or \(e_2\) and identity covariance. The solid curves show their entropy divided by its initial value; the dashed curve is the optimal all-input factor \(\|e^{tB}\|_{\mathrm{op}}^2\). The dotted curve is the explicit Lyapunov bound. These are analytic model evaluations, not experimental data.} \label{fig:kinetic-entropy14} \end{figure} \begin{remark} For a nonquadratic kinetic model, quantitative decay depends on confinement, Hessian growth, and the functional inequalities available for its invariant law. A rate for the quadratic model cannot be transferred to that setting without verifying the hypotheses of an appropriate theorem. \end{remark} % ===================================================================== % ===================================================================== \section{Time Reversal, Path Entropy, and Recurrence} \label{sec:time_arrow_unified} An experiment that lets a prepared system relax and an experiment that reverses its complete microscopic motion are different preparations. Reversing velocities is only part of the second operation: reservoir correlations, driving, and time-odd parameters must also be specified. We therefore compare explicit forward and reverse laws before assigning an entropy production to their difference. This explains how a nonequilibrium ensemble can relax even when its equilibrium dynamics obey detailed balance, and why recurrence of an individual trajectory does not imply recurrence of the relaxing probability law. \subsection{Four Complementary Manifestations of Time Irreversibility} \label{sec:four_levels} \subsubsection{Level I: Generator Structure and Generalized Detailed Balance} Physical time reversal must account for odd variables. Under \cref{ass:M6}, define generalized equilibrium reversibility by \begin{equation} L^{\dagger_\mu} =\Theta L\Theta^{-1}. \label{eq:generalized_detailed_balance} \end{equation} The generalized reversal defect is \begin{equation} \mathfrak R =L^{\dagger_\mu} -\Theta L\Theta^{-1}. \label{eq:reversal_defect} \end{equation} Thus \(\mathfrak R=0\) characterizes generalized detailed balance on the relevant domain. For the kinetic equilibrium model, \(\Theta(q,p)=(q,-p)\). The Hamiltonian transport obeys \begin{equation} \Theta A\Theta^{-1}=-A, \label{eq:Theta_A} \end{equation} and the dissipative component obeys \begin{equation} \Theta S\Theta^{-1}=S. \label{eq:Theta_S} \end{equation} Since \(A^{\dagger_{\piEq}}=-A\) and \(S^{\dagger_{\piEq}}=S\), \begin{equation} L^{\dagger_{\piEq}} =S-A =\Theta(S+A)\Theta^{-1}. \label{eq:kinetic_generalized_reversibility} \end{equation} Hence an equilibrium stochastic process can possess both nonzero dissipation and nonzero antisymmetric transport while remaining reversible in the physically correct generalized sense. This observation prevents the false identification \begin{equation} S\ne0 \quad\Longleftrightarrow\quad \text{stationary time-reversal breaking}. \label{eq:dissipation_not_reversal_breaking} \end{equation} \subsubsection{Level II: Strict Contraction Excludes a Physical Positive Markov Inverse} \begin{theorem}[No invariant-measure-preserving positive Markov inverse] \label{thm:no_Markov_inverse} Let \(P\) be a Markov kernel preserving \(\mu\), and suppose \begin{equation} \KL{\nu P}{\mu} <\KL{\nu}{\mu} <\infty. \label{eq:strict_DPI_assumption} \end{equation} Then there is no Markov kernel \(Q\) satisfying simultaneously \begin{align} (\nu P)Q&=\nu, \label{eq:Markov_inverse_recovery} \\ \mu Q&=\mu. \label{eq:Markov_inverse_invariance} \end{align} \end{theorem} \begin{proof} If such \(Q\) existed, data processing under \(Q\) would give \begin{align} \KL{\nu}{\mu} &=\KL{(\nu P)Q}{\mu Q} \notag\\ &\le\KL{\nu P}{\mu}, \label{eq:inverse_DPI_contradiction_step} \end{align} which contradicts \cref{eq:strict_DPI_assumption}. Therefore no such positive Markov recovery map exists. \end{proof} \begin{remark}[Reverse process versus inverse semigroup] \label{rem:reverse_not_inverse} \Cref{thm:no_Markov_inverse} does not say that a time-reversed diffusion cannot be mathematically constructed. A reverse process may exist and may itself be Markov under suitable regularity. The theorem excludes a much stronger object: a positive \(\mu\)-preserving Markov recovery map that undoes a strictly information-losing forward step on the specified law. \end{remark} \subsubsection{Level III: Thermodynamic-Law Ordering and Nonrecurrence} \begin{theorem}[No exact recurrence of a non-equilibrium law under strict exponential contraction] \label{thm:no_law_recurrence} Suppose that \(\KL{\nu_s}{\mu}<\infty\) and that for \(t>s\), \begin{equation} \KL{\nu_t}{\mu} \le e^{-2\rho(t-s)} \KL{\nu_s}{\mu}, \qquad \rho>0. \label{eq:strict_exponential_contraction_recurrence} \end{equation} If \begin{equation} \nu_t=\nu_s, \label{eq:law_recurrence_assumption} \end{equation} then \(\nu_s=\mu\). \end{theorem} \begin{proof} Let \begin{equation} H=\KL{\nu_s}{\mu}\ge0. \label{eq:recurrence_H} \end{equation} By \cref{eq:law_recurrence_assumption}, the left-hand side of \cref{eq:strict_exponential_contraction_recurrence} equals \(H\). Hence \begin{equation} H\le e^{-2\rho(t-s)}H. \label{eq:recurrence_contraction_inequality} \end{equation} Since \(t>s\) and \(\rho>0\), \begin{equation} 00\), the mean is \(+\infty\). For \(a=1-\lambda>0\), the claimed lower bound is immediate when the positive part has infinite integral. In the remaining case \(q>0\) almost surely and \(-\log q\) is integrable. The pointwise inequality \(\log(q/a)\le q/a-1\), after integration, yields \(\mathbb E_P[-\log q]\ge-\log a\). Finally, \(e^{-\Sigma}=q\), with \(e^{-\infty}=0\), so Markov's inequality gives \(P(\Sigma\le-s)\le e^{-s}\mathbb E_Pq =(1-\lambda)e^{-s}\). This also covers \(\lambda=1\), for which \(q=0\) almost surely. Setting \(\lambda=0\) recovers the equivalent-law bound. \end{proof} Nonnegative mean production permits negative production on individual finite histories. Identifying \(k_B\Sigma\) with physical total entropy requires local detailed balance and a specified reservoir; the measure identity alone does not supply a thermometer, a heat variable, or a microscopic bath. \subsubsection{A reference-density correction in the heat formula} To display every endpoint term, use even variables on a torus, a constant positive diffusion matrix \(D\), a smooth time-independent drift \(b\), and the equation \(dX=b(X)dt+\sqrt{2D}\,dW\). Let \(p_0,p_T\) denote positive densities relative to coordinate volume. The reverse experiment uses the same drift and starts from \(p_T\). Under the finite-window absolute-continuity assumptions, the path ratio is \begin{equation} \Sigma=\log\frac{p_0(X_0)}{p_T(X_T)} +\int_0^T(D^{-1}b)(X_t)\circ dX_t. \label{eq:ta-density-heat} \end{equation} One way to verify the sign is to compare midpoint transition weights. The difference of the two quadratic actions for increments \(dX\) and \(-dX\) is \(b^TD^{-1}\circ dX\); their divergence terms cancel. Multiplication by the initial densities gives the two endpoint terms. Equivalently, the diffusion time-reversal theorem and Girsanov's formula give the same expression. In a thermal realization \(b=D F/\theta\), \(\theta=k_BT_{\rm bath}\), the integral is \(Q_{\rm bath}/\theta\), with \(Q_{\rm bath}=\int F\circ dX\). Now introduce a Gibbs reference \(\pi=Z^{-1}e^{-U/\theta}\) and write \(p_t=f_t\pi\). Merely changing the reference density gives \begin{align} \log\frac{p_0(X_0)}{p_T(X_T)} &=\log\frac{f_0(X_0)}{f_T(X_T)} +\frac{U(X_T)-U(X_0)}{\theta},\notag\\ \Sigma&=\log\frac{f_0(X_0)}{f_T(X_T)} +\frac{\Delta U+Q_{\rm bath}}{\theta}. \label{eq:ta-gibbs-endpoint} \end{align} Consequently a Gibbs-relative endpoint logarithm cannot simply be added to the full bath heat. For a conservative undriven force \(F=-\nabla U\), \(Q_{\rm bath}=-\Delta U\) and those energy terms cancel. With work \(W=\Delta U+Q_{\rm bath}\), the last term is \(W/\theta\). Time-dependent reference potentials also require their partition-function change and the prescribed protocol reversal; they are not covered by the fixed-reference simplification above. \subsubsection{Relaxation, stationary circulation, and information loss} \begin{proposition}[Endpoint-matched relaxation identity] \label{prop:ta-relaxation} Let a Markov diffusion have invariant probability \(\pi\) and be reversible under the declared measurable parity involution \(\Theta\), with \(\Theta_\#\pi=\pi\). Write \(R=P_\pi\) for its stationary path law, so \(\mathcal R_\#R=R\). Let \(p_T=(X_T)_\#P_{p_0}\) and set \(Q=\mathcal R_\#P_{\Theta_\#p_T}\). Assume \(p_0=f_0\pi\) and \(p_T=f_T\pi\), where \(f_0,f_T\) are measurable, finite and positive \(\pi\)-almost everywhere, and \(\int|\log f_i|\,dp_i<\infty\) for \(i=0,T\). Then the path laws are equivalent, all three relative entropies below are finite, and \begin{equation} \log\frac{dP_{p_0}}{dQ} =\log f_0(X_0)-\log f_T(X_T),\qquad D(P_{p_0}\Vert Q)=D(p_0\Vert\pi)-D(p_T\Vert\pi). \label{eq:ta-relaxation-path} \end{equation} \end{proposition} \begin{proof} Let \(K_x\) denote the conditional path law started at \(x\). It is concentrated on \(\{X_0=x\}\), and \(P_p=\int K_x\,p(dx)\). For every nonnegative measurable path function \(g\), \[ \int g\,dP_{p_0} =\int f_0(x)\!\left(\int g\,dK_x\right)\pi(dx) =\int g\,f_0(X_0)\,dR. \] Thus the initial-density change is derived from the path kernel. Parity invariance gives \(d(\Theta_\#p_T)/d\pi=f_T\circ\Theta\). Reversing this path law and using \(\Theta X_0=X_T\circ\mathcal R\) and \(\mathcal R_\#R=R\) gives \(dQ/dR=f_T(X_T)\). Both densities are positive \(R\)-almost everywhere because the two stationary endpoint marginals are \(\pi\). Their Radon--Nikodym ratio therefore gives the first identity. The prescribed endpoint logarithmic moments make both terms integrable under \(P_{p_0}\), since its endpoint laws are exactly \(p_0,p_T\). Integrating their difference proves the second identity with finite terms; no subtraction of two infinite entropies is used. \end{proof} Thus stationary equilibrium has zero path production while relaxation from a nonequilibrium law can have positive endpoint-matched path production. There is no contradiction between reversible equilibrium dynamics and a strictly decreasing nonequilibrium free energy. For a complementary maintained process, retain even variables and set \begin{equation} b=D\nabla\log\pi+c,\qquad \nabla\cdot(\pi c)=0, \qquad f=p/\pi,\qquad J=pc-Dp\nabla\log f. \end{equation} Integration by parts, with periodic conditions or vanishing boundary terms, gives \begin{align} -\frac d{dt}D(p\Vert\pi) &=\int p\,\nabla\log f^TD\nabla\log f=:\mathcal I_D(p),\notag\\ \int\frac{J^TD^{-1}J}{p} &=\mathcal I_D(p)+\int p\,c^TD^{-1}c. \label{eq:ta-split} \end{align} Indeed the transport contribution is \(\int\pi c\cdot\nabla f=0\); the same expression removes the cross term in the second line. The first formula expresses free-energy decay since \(F(p)-F(\pi)=\theta D(p\Vert\pi)\). Under the thermal interpretation of \eqref{eq:ta-density-heat}, the second formula is the expected rate of total entropy production in units of \(k_B\): the mean medium rate is \(\int J\cdot D^{-1}b\), and the Shannon entropy rate is \(-\int J\cdot\nabla\log p\). At \(p=\pi\), the state KL is constant and the stationary path rate is \(\int\pi c^TD^{-1}c\). Directly, the reversed stationary drift is \(b^R=-b+2D\nabla\log\pi\); Girsanov's quadratic cost is \(\tfrac12(2c)^T(2D)^{-1}(2c)=c^TD^{-1}c\). This calculation requires the stated all-even fixed-drive reversal. An odd Hamiltonian transport current must not be counted as dissipative by this formula without its parity decomposition and noise-range test. Finally, suppose a Markov evolution \(K_t\) preserves \(\pi\) and strictly decreases \(D(p\Vert\pi)\) for a particular nonequilibrium \(p\). There is no positive Markov kernel \(R_t\) preserving \(\pi\) with \(R_tK_tp=p\): data processing would give \begin{equation} D(p\Vert\pi)=D(R_tK_tp\Vert R_t\pi) \leq D(K_tp\Vert\pi)