WorksPhysics and Astronomy
The normal approximation to macroscopic dynamical fluctuations
Abstract
Under local detailed balance, the Lagrangian of macroscopic dynamical fluctuations has a canonical structure: its time-antisymmetric part is the entropy flux, and its time-symmetric part is a Legendre pair of kinetic and dissipation functions. We show what survives of that structure for normal (Gaussian) fluctuations around a nonequilibrium steady state. The quadratic Lagrangian keeps the canonical form, with the Hessian of the dissipation function as noise matrix. However, the force on the fluctuations differs from the linearized thermodynamic force by a term proportional to the driving and to the state dependence of the dynamical activity, so that local detailed balance is violated at the Gaussian level. Nevertheless, the Gaussian dynamics is gradient flow for the quadratic part of the nonequilibrium entropy, around an emergent nondissipative current. Finally, we show which parts of the Lagrangian can be reconstructed from measured Gaussian fluctuations.
Introduction
Macroscopic equations describe typical behavior; fluctuations around them carry the information about how that behavior comes about. Since Onsager [1, 2], the linear (Gaussian) processes around equilibrium have been the main bridge between fluctuations and response: their drift is fixed by the thermodynamic force, their noise by the Onsager matrix, and time-reversal symmetry produces reciprocity and the fluctuation–dissipation theorem.
Away from equilibrium, the large deviations of macroscopic trajectories are known to have a canonical structure as soon as local detailed balance holds [3]. That is very similar to and inspired by nonlinear extensions of the relation between fluctuations and gradient flow. using the (global) detailed balance condition [1, 4, 5, 6]. The time-antisymmetric part of the path-space Lagrangian can now be identified with the (irreversible) entropy flux in terms of a thermodynamic force. Everything else is encoded in a Legendre pair of time-symmetric functions, the kinetic function and the dissipation function . The zero-cost path gives the macroscopic relaxation equation toward nonequilibrium, and its structure extends those of gradient flows and GENERIC.
However in practice, what is most often measured or simulated are the small, normal fluctuations around the macroscopic evolution: linear-noise approximations of chemical kinetics [7, 8], fluctuating hydrodynamics and Gaussian current statistics. The present paper asks what of the canonical structure survives at that level, and what the Gaussian level can tell about the full structure as an inverse problem. The main results are:
The quadratic Lagrangian of normal fluctuations around a nonequilibrium steady state keeps the canonical form (Sec. IV). Its noise matrix is the Hessian of the dissipation function at the stationary force; its drift defines a linear force on fluctuations.
That force is not the linearized thermodynamic force. The difference is first order in the stationary driving, coming from the state dependence of the dynamical activity. Local detailed balance is thus violated at the Gaussian level (§ 5). It is restored with respect to the quadratic part of the nonequilibrium entropy, together with what we can call an emergent nondissipative current.
The associated Ornstein–Uhlenbeck dynamics for states and integrated currents (Sec. VI) allows one to reconstruct the quadratic Lagrangian from data. Current noise measures the frenetic Hessian at the operating point, information that the relaxation equation alone does not contain.
Linear response is expressed through the Gaussian correlations (Sec. VIII C). Violations of the fluctuation–dissipation and Green–Kubo relations are governed by the nondissipative current. Details are postponed to a future publication.
Two families of examples run through the text: chemical reaction networks with one free species, including the Schlögl model, and driven diffusion on a ring (Sec. III). Section VII gives connections to and distinctions from previous works. Section VIII discusses multistability and normal fluctuations around limit cycles. Detailed computations are in the Appendices.
Canonical structure of dynamical fluctuations
We recall the main results of [3, 9]. Let denote the macroscopic state (a density profile, a vector of concentrations, ) and a current, related by
with a linear operator (minus a divergence, a stoichiometric matrix, the identity). For a system with large parameter (number of particles, volume), we assume that the paths of states and currents satisfy a large deviation principle,
with convex in .
Spaces and pairings. At each state (element of a differentiable manifold) there are two pairs of dual spaces: the tangent space (velocities ) with its dual, the cotangent space; and the space of currents with its dual, the space of forces . The adjoint maps cotangent vectors to forces, . In a basis, currents and forces have components , , where labels the transport channels: the reactions of a network, the bonds of a graph, or positions in a continuum (where sums become integrals); tangent vectors have components , with labelling species or positions. The entropy flux is the pairing . We set , so entropies are dimensionless. Round brackets always denote such pairings (except when delimiting the argument of a function such as ); angular brackets denote expectations.
Local detailed balance. There are a thermodynamic force and a Hamiltonian flow with such that for all ,
Remark: The force may contain a gradient part and a non-gradient (rotational, cycle) part ,
with an equilibrium entropy; then is the nonequilibrium driving.
Canonical form. The time-symmetric part of is free but given by a Legendre pair of convex functions, even in their first argument and vanishing at zero [3]:
The pair may also depend on the driving ; we suppress that dependence in the notation. We call the dissipation function: it determines how forces are turned into currents.
Relaxation equation. The zero-cost path gives the typical current and the macroscopic evolution,
Equilibrium corresponds to ; the entropy then increases along (II.6). Away from equilibrium there is another nonequilibrium entropy , the static fluctuation functional of the stationary state (aka quasipotential), which is a Lyapunov function [3].
Assumptions. We assume that (II.6) has a stable stationary point (Sec. VIII relaxes this), with stationary current , , and that is twice continuously differentiable near . For simplicity of notation we take from now on, as in both running examples; Remark IV.2 explains the changes for .
Running examples
Chemical networks with one free species
Consider a well-mixed volume with one free species (concentration ) and two reactions , each producing one in its forward direction, so that . All other species are chemostatted. With mass-action rates ,
so that and . The prefactor is the frenetic part of the channel; the corresponding traffic is
The cycle (reaction 1 forward, reaction 2 backward) has affinity . At a stationary state, is the stationary cycle current.
Catalytic network. As a first example [10], take
with chemostat concentrations . In units where , and with
The cycle affinity does not depend on ; the network is in equilibrium if and only if . The stationary concentration is
with cycle current .
Schlögl model. The second example is Schlögl’s model [11],
with , , , . The cycle converts , with affinity . The cubic rate equation is bistable in a range of chemostat concentrations, which requires a nonzero cycle affinity (Sec. VIII A).
Driven diffusion on a ring
Let , with periodic boundary conditions, be a density with local entropy ( concave), driven by a constant field along the ring. Then
with mobility . The current is with diffusivity (Einstein relation). For the weakly asymmetric exclusion process, , and , and the hydrodynamic equation is the viscous Burgers equation. The uniform profile is stationary, with current around the ring. The non-gradient part of the force is the field: .
Such examples, weakly driven or boundary-driven diffusions are the main models of the macroscopic fluctuation theory developed by [12, 13, 14], as limits of stochastic interacting particle systems.
A nonlocal mobility (kernel) and nonlocal entropy fit the same framework. The formulas below then extend with read as an operator and as a general functional of .
The quadratic Lagrangian
Scaling and definition
Normal fluctuations live on the scale . We fix a stable stationary point of (II.6). Writing
the rescaled state fluctuation and current fluctuation satisfy , and their path probabilities are governed by the quadratic Lagrangian
in the sense that, as ,
The passage from the large deviation principle (II.2) to this central-limit statement is not automatic; it holds under smoothness of the cumulant generating functions [15], and is a theorem for density-dependent Markov jump processes [8] and their linear-noise approximation [7]. We assume it throughout.
Main formula
Introduce the stationary force and three linear objects, all evaluated at the operation point :
where . is symmetric and strictly positive definite (by the assumed stability) Hessian of the dissipation function on the space of forces, evaluated at steady thermodynamic force . The outer round brackets in (IV.5) denote the pairing of the tangent space and its dual, where linearization allows us to identify state perturbations with tangent vectors at the stationary state . maps state fluctuations to current fluctuations. Define the force acting on state fluctuations
Proposition IV.1. The quadratic Lagrangian is
and has the canonical form
with the quadratic Legendre pair
The proof (Appendix A) uses only that and for all , together with by Legendre duality.
The force in (IV.7) has two parts. The first, , is the variation of the thermodynamic force. The second, , comes from the sensitivity of the dissipation function on the state. Since is even in , for all , we have that
Therefore, the second part of is a genuinely nonequilibrium term.
Two classes of dissipation functions. (i) For quadratic , one has , and . (ii) For channel-wise products , as in (III.2), with even and convex,
which for gives the factor , the ratio of current to traffic [16, 17].
Examples
One-species networks. See Example III A. Here and with (derivatives at ). Then per unit , and by (IV.12),
For the catalytic network, , and , so that
in terms of the cycle current .
Ring. See Example III B. At the uniform state, and
with at , and
The first term is ; the second is the “correction.” For exclusion, is the drift velocity in the linearized dynamics.
Remark IV.2. With a Hamiltonian flow , and consequently remain unchanged, and both formulas (IV.8) as well as (IV.9) hold with replaced by .
Local detailed balance at the Gaussian level
Formal structure and its mismatch
Reversing the current fluctuation at fixed in (IV.9) gives
The quadratic Lagrangian therefore has the form of local detailed balance, with in the role of the force. Local detailed balance proper would require that to be the entropy flux of the fluctuation. However, by (IV.7),
which vanishes only at equilibrium. It also vanishes if the dissipation function does not depend on the state at the operating point, . In conclusion, the antisymmetric part of need not be the physical entropy flux.
The origin of the mismatch is the time reversal itself. Physical time reversal maps to . When , the reversed path lies outside the Gaussian window:
is the stationary entropy production, and its probability is exponentially small in . What remains at the Gaussian level is the reversal around the stationary current, which is not a symmetry of the physics. The antisymmetric part (V.1) then also picks up how the current responds kinetically to the state, through .
Gaussian entropy and nondissipative current
Local detailed balance can be restored at the Gaussian level, but with respect to new objects. Let be the stationary covariance of (Sec. VI), which solves the Lyapunov equation (for derivation see Appendix B)
Define the Gaussian entropy and
Proposition V.1. (a) , so that
(b) is nondissipative: for all .
(c) with antisymmetric. In particular is conserved by the flow .
The proof is in Appendix C. The Gaussian dynamics is thus the gradient flow of with mobility , plus a nondissipative current , exactly the structure (II.5) with a Hamiltonian-like part. In other words, the zero-cost flow for is part of the GENERIC family [5, 6]. However, from the validity of the normal approximation we must have (shown in Proposition D.1, when is twice differentiable at ). Hence, (defined under (V.3)) is the quadratic part of the nonequilibrium entropy of [3], not of the equilibrium entropy . Proof is in Appendix D.
Comparing with physical local detailed balance, two violations appear, both vanishing at equilibrium:
the Gaussian force differs from whenever the stationary covariance differs from the equilibrium one;
the emergent current is nonzero. Its part outside moves the state while conserving (a circulation ); its part in is a cycle current driven by the state fluctuation.
At equilibrium , : , , and and the standard Onsager (linear) processes are recovered.
Examples
One-species networks. See Example III A. The state is one-dimensional, so , and we can solve (V.3),
See also the end of Appendix D. The emergent current is a pure cycle current,
The physical force has no fluctuating cycle component: is fixed by the chemostats, so . The Gaussian force does: is, by (IV.13), proportional to the stationary cycle current. For the catalytic network,
and, using and ,
At equilibrium (, , ) the Fano factor (variance of the number of molecules divided by its mean) is and equals one, which is the Poisson value you would predict from the equilibrium expression of the entropy .
Ring. See Example III B. On mean-zero fluctuations, and , and the Lyapunov equation is solved by the local-equilibrium covariance times the identity (Appendix E). Hence : the Gaussian entropy is the local-equilibrium one, as for the product invariant measure of exclusion. The “generalized local detailed balance” (V.5) holds with current,
which transports fluctuations at the drift velocity without changing . For exclusion at half filling, : there, normal fluctuations cannot be told apart from equilibrium ones.
Intuitive meaning
At equilibrium a fluctuation relaxes because the thermodynamic force pulls it back, and its reversal costs exactly the entropy it produces. Around a driven steady state, a fluctuation also changes how active the transport channels are, and the stationary current through those channels then changes even at fixed force. Seen from the fluctuation, that kinetic effect looks like an extra force, which is the term . It is proportional both to how strongly the activity depends on the state and to how hard the system is driven. In the chemical example, a fluctuation in changes the activity of the two reactions unequally and so drives a cycle current, even though the cycle affinity itself is fixed. On the ring, a density fluctuation changes the mobility and so the field-driven current, and the fluctuation is carried along. The Gaussian description absorbs all such effects into its own entropy and its own nondissipative current.
Normal dynamics
We come to the second point of discussion about the normal approximation, the Gaussian dynamics itself.
Langevin equations
The Gaussian path measure (IV.3) with Lagrangian (IV.8) is that of the linear stochastic dynamics
with white noise on the space of currents. For the state alone,
The integrated current fluctuation satisfies and
with the intertwining . The noise matrix is the same for all channels whose currents are measured: the Hessian has become the strength of the current noise in (VI.1).
When has a nontrivial co-kernel (conservation laws, as for the ring), has a kernel and the fluctuations are restricted to the subspace compatible with the conservation laws, on which is assumed stable. All inverses below are taken on that subspace.
Stationary correlations
The stationary covariance solves (V.3), and for (Onsager regression). For current fluctuations, for ,
plus the white-noise part . The integrated current has a diverging variance, linear in time, along .
Proposition VI.1. The long-time covariance rate of the integrated current is
is the (oblique) projector onto the cycle space along the image space of state-induced currents. The limit as for in (VI.3). Equivalently, .
The proof is in Appendix B. The long-time current fluctuations are therefore the current noise projected onto cycles, along the directions in which currents follow the state.
Note: At equilibrium, the image space is the -orthogonal complement of , so that is the orthogonal projector in the noise metric . Out of equilibrium, tilts the image space and becomes oblique. This obliqueness is what breaks the Green–Kubo relation by adding a frenetic contribution; more on that in Sec. VIII C.
One-species networks. See Example III A. is spanned by . Writing , the cycle current has variance rate
At equilibrium, and : the two reactions act as resistors in series.
Ring. See Example III B. The spatially averaged current obeys , because by mass conservation. Hence , unchanged by the drive at the Gaussian level.
Gaussian reconstruction
The quadratic Lagrangian is fully identified from stationary time series of states and integrated currents. Starting from the Langevin equations (VI.1) and (VI.3) we find:
with ; that is, the noise matrix is the quadratic variation of the currents, and is the regression of current increments on the state. From states alone, one obtains only , as , and from (V.3): the cycle components of and remain invisible. From currents alone, short times give , long times give , and the regression of current correlations gives , as in (VI.3).
That identifies part of the original Lagrangian. The noise matrix is the Hessian at the stationary force, while the relaxation equation only gives the gradient at that force. If is known from thermodynamics, gives the mixed derivative .
For chemical networks that is enough to recover everything locally: the mean current gives and the current noise gives , hence both rates, hence , and at . Varying a control parameter, such as a chemostat concentration, moves the stationary state; repeating the measurement along the way reconstructs the rate functions, and hence , over the range of states that can be made stationary, provided the dependence of the rates on the control parameter is known, as it is for mass-action kinetics.
In general, if the driving can be varied at fixed state, integrating along the ray recovers there.
Connections with earlier Gaussian theories and fluctuating hydrodynamics
The algebra of the Gaussian level is standard: a linear stochastic dynamics, its stationary covariance from a Lyapunov equation, and a split of the drift into a symmetric and an antisymmetric part. What the canonical structure adds is the extension of the Langevin dynamics by including the current variables, and the physical identification of all objects: the current noise as the Hessian of the dissipation function at the operating point, and the linear force derived from (the Hessian of) the nonequilibrium entropy in combination with an emergent current from the state dependence of the dynamical activity.
In the following, we compare our results with previous studies and make the connections explicit.
Gaussian fluctuations around steady states
Lax [18] treated Gaussian fluctuations around nonequilibrium steady states as linear processes whose stationary covariance solves a Lyapunov equation, our (V.3), and stressed that is in general not the inverse Hessian of the equilibrium entropy. Keizer [19] built on this a statistical thermodynamics of nonequilibrium processes. Two of his ingredients reappear here. First, in his canonical theory of elementary processes the noise intensity of each process is fixed by the sum of its forward and backward rates. That is exactly our noise matrix for a channel-wise dissipation function: for , the Hessian at the operating force is , the traffic. Keizer’s postulate is thus the statement that the noise matrix is at the stationary force, not at zero force. Second, Keizer proposed a nonequilibrium entropy whose second derivative at the steady state is , and suggested measuring it through fluctuations. Proposition D.1 derives that identification from large deviations: is the Hessian of the nonequilibrium entropy , the quasipotential of the stationary distribution. Keizer’s entropy is therefore our Gaussian entropy , and it is the quadratic part of a globally defined object.
Of course, knowing does not determine the Gaussian dynamics: by Proposition V.1 the drift is the gradient flow of plus the emergent current , and it is that will control the violations of fluctuation–dissipation and Green–Kubo relations. Moreover, working with currents rather than states, we keep track of cycle components of and that a description in terms of states alone, as in [18, 19], cannot see. In the one-species networks the state is one-dimensional and the entire nonequilibrium signature at the Gaussian level is such a cycle current.
Nonequilibrium potentials
For a diffusion with diffusion matrix , Graham [20] and Graham and Tél [21, 22] defined the nonequilibrium potential as the quasipotential of the stationary distribution, . It solves the Hamilton–Jacobi equation , which is equivalent to the decomposition
of the drift into a part driven by the potential and a remainder orthogonal to its gradient. The remainder conserves , and the time-reversed (adjoint) dynamics has drift .
Proposition V.1(c) is the Gaussian instance of (VII.1). With and , the state drift splits as
and is orthogonal to . Furthermore, here, the decomposition is lifted from states to currents, Eq. (V.5), where it takes the form of a generalized local detailed balance with respect to .
Kwon, Ao and Thouless [23] showed that any linear stochastic dynamics , , can be rewritten exactly as
with and stationary distribution . The construction solves the Lyapunov equation: , and with antisymmetric, so that . In our notation, and , and the identity is our
the same algebra as Proposition V.1(c). The decomposition (VII.3) is thus a mathematical fact about linear processes, and our Proposition V.1 is in that sense its restatement.
The physics enters through three identifications that [23] does not make: (i) the potential is the Hessian of the nonequilibrium entropy of the underlying nonlinear dynamics (Proposition D.1); (ii) the symmetric part is , the projection of the Hessian at the operating point; and (iii) the antisymmetric part has a definite origin, the frenetic sensitivity in the force (IV.7), so that it is of first order in the stationary driving and proportional to the state dependence of the activity. In particular, is a measurable signature of driving with state-dependent kinetics, not an arbitrary gauge choice.
Fluctuating hydrodynamics and macroscopic fluctuation theory
Fluctuating hydrodynamics adds noise to the hydrodynamic equations, following Landau and Lifshitz [24]: random stress and heat-flux terms are added to the Navier–Stokes equations, with covariance fixed by fluctuation–dissipation. This theory was extended and rigorously proven by [12, 13, 14] for large class of weakly driven or boundary-driven diffusions as limits of stochastic interacting particle systems.
In our framework, fluctuating hydrodynamics is the special case of a quadratic dissipation function, , which is exactly Example III B, Eq. (III.9). The quadratic Lagrangian is then linearized fluctuating hydrodynamics, with , and macroscopic fluctuation theory (MFT) is the large-deviation level of the same theory. MFT contains the decomposition that our Gaussian results linearize. With the quasipotential, the hydrodynamic current splits as
which is equivalent to the Hamilton–Jacobi equation for , and the adjoint hydrodynamics, which governs the optimal fluctuation paths, reverses the first part and keeps [12].
Linearizing (VII.5) around gives Proposition V.1: becomes , the force becomes , and becomes . For the ring, linearizes to the transport current with , and the Gaussian entropy is the local-equilibrium one, in agreement with the product invariant measure of the weakly asymmetric exclusion process on a ring.
The comparison also shows what is specific to the present setting. In fluctuating hydrodynamics and MFT the noise is fixed by the near-equilibrium mobility. Our general result, that the noise matrix is the Hessian at the operating force, has no counterpart there.
Outlook
Multistability: the Schlögl model
When the relaxation equation has several stable stationary points, as the Schlögl model (III.7) does for a range of parameters, the normal approximation applies around each of them separately, on times short compared with the switching time, which grows exponentially in with the barrier in the nonequilibrium entropy . All formulas of Secs. IV–VI for one-species networks apply at each stable point, with
Bistability requires a nonzero cycle affinity, since with detailed balance the rate equation has a unique stable point [25]. Consequently at each of the two stable points, and both carry an emergent cycle current . As a stable point approaches a saddle-node bifurcation, , so and diverge as and . The normal approximation then breaks down: fluctuations become non-Gaussian, on a larger scale in .
Limit cycles
Write the relaxation equation (II.6) as , with , and suppose it has a stable periodic orbit of period , as in the Brusselator [26]. Limit cycles need at least two state variables, so they do not occur for one-species networks. Writing
the derivation of Appendix A, with the stationary point replaced by the point on the orbit, gives a time-periodic quadratic Lagrangian
with and . The associated Gaussian dynamics is that of (VI.1), now with periodic coefficients:
with . The state noise has covariance , as in (VI.2).
It is possible to record the time along the limit cycle, called the asymptotic phase : each in the basin of attraction of the orbit has an asymptotic phase : the solution starting at converges to . Hence and along solutions. The level sets of are the isochrons. To first order, defining the phase fluctuation . It turns out that accumulates noise without any restoring force: it is a time-changed Brownian motion. For further details, we refer to [27]. Bounds on the number of coherent oscillations in terms of the cycle affinities have been conjectured [28].
Local detailed balance is violated here in an essential way. On a limit cycle, the reversed orbit does not solve the relaxation equation, and by (II.3) its cost per period is with
the entropy produced per period. Time reversal thus maps the Gaussian window around into a window around an exponentially improbable path, and no reversal symmetry, not even a modified one, acts on the phase.
Linear response
In Section VI all expectations are taken in the Gaussian dynamics (VI.1), that is, in the normal approximation. The variables and are the rescaled fluctuations of (IV.1), their stationary law is with from (V.3), and the time correlations are and (VI.4). They are the limits as of times the connected (cumulant) correlations of the state and current of the original process, at fixed times. Higher connected correlations (third, fourth,...) are smaller still, of order and beyond. At leading order only the covariance survives, so the process is Gaussian. The correlations are those within a single well of a stable stationary point.
At the macroscopic level the linear response is the same as the Gaussian one. Consider a perturbation at fixed dissipation function, i.e., unchanged. The typical current becomes , and expanding around to first order in and ,
where the -term is the Hessian acting on . On the other hand, in the Gaussian dynamics, and . The latter coincides with (VIII.6) upon identifying and taking the perturbation of order at the Gaussian level, , so that the shifted state remains in the Gaussian window. Hence, starting from the steady state,
and for constant the stationarity condition shifts the state by , giving .
The results on linear response will be part of subsequent work.
Ballistic regime
Our scaling, with drift and noise both of order in the action as for density-dependent jump processes [8], is the diffusive or reactive one. For strongly asymmetric lattice gases and for integrable systems the relevant scaling is Euler (ballistic), . The hydrodynamics is then hyperbolic, for instance the inviscid Burgers equation for the totally asymmetric exclusion process, or generalized hydrodynamics for integrable models, with dissipation only at shocks.
Fluctuation theories exist at that scale, but none with the canonical structure of Sec. II. In ballistic macroscopic fluctuation theory [29] there is no dynamical noise at leading order: fluctuations come from the randomness of the initial state and are transported deterministically by the Euler flow, and the action is the initial-state large-deviation functional pushed along characteristics. For the totally asymmetric exclusion process, the large deviations of Jensen and Varadhan at Euler scale [30, 31] have a rate function given by the entropy production of non-entropic weak solutions, so that the only time-antisymmetric cost sits at shocks. Beyond the Gaussian window, asymmetric systems show Kardar–Parisi–Zhang fluctuations with scaling [32], so a Gaussian theory can only describe short times or weak asymmetry; on the ring this is the window in which our drift velocity transports fluctuations without distorting them.
In the language of Sec. II, the ballistic limit is the one in which the Hamiltonian current , the Euler current that conserves the entropy of smooth solutions, is of order one, while the force and the noise are suppressed by . The canonical structure then degenerates: the zero-cost flow is purely Hamiltonian, the dynamical Legendre pair disappears at leading order, and local detailed balance reduces to the time-reversal invariance of the Euler flow, supplemented by entropy production at shocks. There is therefore no canonical formalism for dynamical fluctuations at the ballistic scale in the present sense. Ballistic MFT takes its place, and the diffusive corrections, at Navier–Stokes order, are where and re-enter as weak dissipation on top of an Euler . That “weakly dissipative” setting, with the Euler current and the canonical pair at order , is a natural extension of the present work.
Conclusion
Most often measured or simulated around a stable nonequilibrium steady state are the normal fluctuations. In this paper we asked 1) what survives of local detailed balance and of the canonical structure of macroscopic dynamical fluctuations, 2) whether and how the normal fluctuations and perhaps the larger dynamical fluctuations as well, can be reconstructed.
The quadratic Lagrangian of the fluctuations keeps the canonical form with a quadratic dissipation function and a linear force acting on the state fluctuation. The noise matrix is the Hessian of at the actual stationary force is not the linearized thermodynamic force : it contains an additional term expressing the sensitivity of the dynamical activity to the state, proportional to the stationary driving. It vanishes when the dissipation function does not depend on the state. As a result, normal fluctuations around steady nonequilibrium are Onsager processes with respect to a modified entropy plus a nondissipative flow, that is of GENERIC-type. Secondly, the Gaussian path measure is that of a linear (Ornstein–Uhlenbeck) dynamics, for the state and for the integrated current. The quadratic Lagrangian can be identified from stationary time series.
We mention that for linear response, the Gaussian description reproduces the macroscopic response, and the emergent nondissipative current is expected to control the violation of fluctuation–dissipation and Green–Kubo relations.
Acknowledgements
KM is supported by Charles University, project GA UK No. 414225.
References
Derivation of the quadratic Lagrangian
Write in the form (II.5) with . For every , minimizes with , so . By Legendre duality, and
Taylor’s theorem in at fixed gives . With (IV.1), , so , and . The derivative of in the direction is , which is (IV.6). Multiplying by and letting gives (IV.8). Expanding (IV.8) and using gives (IV.9).
Gaussian process
From (VI.2), , which gives . Taking the equal-time covariance and imposing stationarity yields the Lyapunov equation (V.3) (see, e.g., [33]). For , , which gives (VI.4).
Proof of Proposition VI.1. Integrating (VI.2) gives . Hence
This can be read as a decomposition to the space of cycle currents and state-induced currents. The first term is bounded in time, so .
Properties of : a) , so , b) for , c) . Thus is the projector onto along . Moreover .
The equivalence with follows from expanding and using , which is (V.3).
Proof of Proposition V.1
Since , expanding (IV.8) around gives (a) exactly as in Appendix A. For (c), define . The Lyapunov equation (V.3) gives . Then
For (b), by antisymmetry of . A part of in is orthogonal to automatically.
At equilibrium, and solve (V.3), so .
The Hessian of the nonequilibrium entropy
Hamiltonian on force space. For a force (an element of the dual of the space of currents at ) define
the Legendre transform of in . It is the cumulant generating function of the current at fixed state. With the canonical form (II.5), the substitution and Legendre duality give
For , local detailed balance (II.3) is equivalent to the symmetry
since turns into after . The same follows directly from (D.2) and the evenness of .
Hamilton–Jacobi equation. The quasipotential is the rate function of the stationary measure, ,
Wherever is differentiable,
Proposition D.1. Let be near , solve (D.5), and have a strict local minimum at , . Let be stable and be positive definite (both on the subspace compatible with the conservation laws, if any). Then
where is the unique solution of the Lyapunov equation (V.3).
Proof. Put . Since and is a minimum, , hence . Expanding (D.2) in at fixed , using that and ,
Now and , so
Inserting gives, to second order in ,
A quadratic form vanishes for all if and only if its symmetric part does. Using we get
Since is invertible, multiplying (D.10) from the left and from the right by gives , i.e. solves (V.3). For stable the Lyapunov equation has a unique solution, so .
For Example III A, , with the sums of the forward and reverse rates, so with . That equals of (V.6). For the Schlögl example with , the rate equation is , with stable points and . At : , , and .
Driven diffusion on a ring
In Fourier modes , , : and , all at .
The Lyapunov equation gives , independent of . Thus , and .
Each Fourier mode decays at rate and rotates at frequency ; the rotation is the circulation .
References
- [1]A. Mielke, D. R. M. Renger, and M. A. Peletier. A Generalization of Onsager’s Reciprocity Relations to Gradient Flows with Nonlinear Mobility. J. Non-Equilib. Thermodyn., 41(2):141–149, 2016.
- [2]L. Onsager. Reciprocal Relations in Irreversible Processes. I. Phys. Rev., 37(4):405–426, 1931.DOI
- [3]C. Maes and K. Netočný. Relaxation to nonequilibrium. J. Non-Equilib. Thermodyn., 51(3):279–291, 2026.
- [4]A. Mielke, M. A. Peletier, and D. R. M. Renger. On the Relation between Gradient Flows and the Large-Deviation Principle, with Applications to Markov Chains and Diffusion. Potential Analysis, 41(4):1293–1327, 2014.DOI
- [5]R. Kraaij, A. Lazarescu, C. Maes, and M. Peletier. Deriving GENERIC from a Generalized Fluctuation Symmetry. J. Stat. Phys., 170(3):492–508, 2017.DOI
- [6]R. C. Kraaij, A. Lazarescu, C. Maes, and M. Peletier. Fluctuation symmetry leads to GENERIC equations with non-quadratic dissipation. Stoch. Proc. Appl., 130(1):139–170, 2020.DOI
- [7]N. G. van Kampen. Stochastic Processes in Physics and Chemistry. Elsevier, Amsterdam, 3rd edition, 2007.
- [8]T. G. Kurtz. Limit theorems for sequences of jump Markov processes approximating ordinary differential processes. J. Appl. Probab., 8:344–356, 1971.DOI
- [9]C. Maes and K. Netočný. Canonical structure of dynamical fluctuations in mesoscopic nonequilibrium steady states. EPL, 82(3):30003, 2008.
- [10]G. Falasco and M. Esposito. Macroscopic stochastic thermodynamics. Rev. Mod. Phys., 97:015002, 2025.DOI
- [11]F. Schlögl. Chemical reaction models for non-equilibrium phase transitions. Z. Phys., 253(2):147–161, 1972.
- [12]L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim. Macroscopic fluctuation theory. Rev. Mod. Phys., 87:593–636, 2015.DOI
- [13]L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim. Fluctuations in Stationary Nonequilibrium States of Irreversible Processes. Phys. Rev. Lett., 87:040601, 2001.pubmed.ncbi.nlm.nih.gov/11461605
- [14]L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim. Macroscopic fluctuation theory for stationary non equilibrium states. J. Stat. Phys., 107(3/4):635–675, 2002.
- [15]W. Bryc. A remark on the connection between the large deviation principle and the central limit theorem. Statist. Probab. Lett., 18:253–256, 1993.DOI
- [16]C. Maes, K. Netočný, and B. Wynants. Steady state statistics of driven diffusions. Physica A, 387(12):2675–2689, 2008.DOI
- [17]C. Maes, K. Netočný, and B. Wynants. On and beyond entropy production: the case of Markov jump processes. Markov Proc. Relat. Fields, 14:445–464, 2008.arxiv.org/abs/0709.4327
- [18]M. Lax. Fluctuations from the nonequilibrium steady state. Rev. Mod. Phys., 32:25–64, 1960.DOI
- [19]J. Keizer. Statistical Thermodynamics of Nonequilibrium Processes. Springer, New York, 1987.DOI
- [20]R. Graham. Covariant formulation of non-equilibrium statistical thermodynamics. Z. Phys. B, 26:397–405, 1977.DOI
- [21]R. Graham and T. Tél. Existence of a potential for dissipative dynamical systems. Phys. Rev. Lett., 52:9–12, 1984.DOI
- [22]R. Graham and T. Tél. Nonequilibrium potential for coexisting attractors. Phys. Rev. A, 33:1322–1337, 1986.pubmed.ncbi.nlm.nih.gov/9896751
- [23]C. Kwon, P. Ao, and D. J. Thouless. Structure of stochastic dynamics near fixed points. Proc. Natl. Acad. Sci. USA, 102:13029–13033, 2005.DOI
- [24]L. D. Landau and E. M. Lifshitz. Fluid Mechanics. Pergamon, Oxford, 2nd edition, 1987. Chap. XVII.
- [25]G. Nicolis and I. Prigogine. Self-Organization in Nonequilibrium Systems. Wiley, New York, 1977.
- [26]I. Prigogine and R. Lefever. Symmetry breaking instabilities in dissipative systems. II. J. Chem. Phys., 48:1695–1700, 1968.DOI
- [27]P. Gaspard. Correlation time of mesoscopic chemical clocks. J. Chem. Phys., 117:8905–8916, 2002.DOI
- [28]A. C. Barato and U. Seifert. Coherence of biochemical oscillations is bounded by driving force and network topology. Phys. Rev. E, 95:062409, 2017.DOI
- [29]B. Doyon, G. Perfetto, T. Sasamoto, and T. Yoshimura. Ballistic macroscopic fluctuation theory. SciPost Phys., 15:136, 2023.DOI
- [30]L. H. Jensen. Large deviations of the asymmetric simple exclusion process in one dimension. PhD thesis, New York University, 2000.
- [31]S. R. S. Varadhan. Large deviations for the asymmetric simple exclusion process. In Tadahisa Funaki and Hirofumi Osada, editors, Stochastic Analysis on Large Scale Interacting Systems, volume 39 of Advanced Studies in Pure Mathematics, pages 1–27. Mathematical Society of Japan, Tokyo, 2004.
- [32]B. Derrida and J. L. Lebowitz. Exact large deviation function in the asymmetric exclusion process. Phys. Rev. Lett., 80:209–213, 1998.DOI
- [33]C. W. Gardiner. Handbook of stochastic methods for physics, chemistry, and the natural sciences. Springer, 1985.