betaFeedback

WorksPhysics and Astronomy

A Principle of Maximum Entropy for the Navier-Stokes Equations

Abstract

A principle of maximum entropy is proposed in the context of viscous incompressible flow in Eulerian coordinates. The relative entropy functional, defined over the space of L2L^2 divergence-free velocity fields, is maximized relative to alternate measures supported over the energy–enstrophy surface. Since thermodynamic equilibrium distributions are characterized by maximum entropy, connections are drawn with stationary statistical solutions of the incompressible Navier-Stokes equations. Special emphasis is on the correspondence with the final statistics described by Kolmogorov's theory of fully developed turbulence.

Introduction

We are concerned with the reformulation of the principle of maximum entropy, as established in the context of statistical mechanics and information theory, and its applications to mathematical fluid dynamics. More specifically, the Boltzmann–Gibbs entropy under investigation (cf. [9, 21]):

S(f)=−∫f(u)log⁡f(u) dη(u)(1)S(f)=-\int f(u)\log f(u)\,d\eta(u) \tag*{(1)}

is maximized under appropriate constraints resulting from physical considerations1. From a mathematical standpoint, these considerations impose constraints on the class of admissible velocity fields uu and their associated probability distributions f(u)f(u).

The principle of maximum entropy, introduced by E. Jaynes [13] in 1957, revolves around making the most desirable choice, for given prior data. This choice aims to maximize the entropy functional (1). In other words, as expressed in [12], “[T]he [principle of maximum entropy] is the probability assignment that is consistent with the available information but is maximally noncommittal with regard to missing information.” From the standpoint of statistical mechanics, it is seen as a method to determine the correct probability distribution at the state of equilibrium, depending heavily on the constraints imposed on the system, i.e., our knowledge of prior or given information. As the principle of maximum entropy provides a rule for assigning probabilities to data, it proves particularly useful in “filling gaps” between otherwise scattered data. Therefore, maximum entropy modeling can be applied to a wide range of problems. We refer to [14] for applications to several areas including statistics, biology, and medicine.

The physics underlying our problem is fundamentally rooted in the first and second laws of thermodynamics. The former signifies the conservation of energy EE, while the latter imposes a constraint on admissible processes in terms of the production of entropy SS. At a micro-level, where each micro-state is equally probable, entropy can be expressed in terms of the Boltzmann factor kBk_B:

S=kBlog⁡(Ω)S = k_B \log(\Omega)

where Ω\Omega is the number of states present at a given energy EE. Equation (1) can be understood as a generalization of the above formula for entropy SS when the micro-states of the system may not be equally distributed.

This formulation of the second law in terms of a single entropy can be rewritten in a perhaps more conventional manner. Introduce two entropies, SES_E for the entropy of the energetic degrees of freedom which is related to the viscous and turbulent dissipation of energy, and SZS_Z for the entropy of the vortical and enstropic degrees of freedom which is related to the viscous and turbulent dissipation of enstrophy. We assume that both the energetic and the enstrophic degrees of freedom are governed by a common temperature, which can be fixed in the context of the second law of thermodynamics. Consequently, the second law takes the form:

dE=TdSE+TdSZ.\mathrm{d}E = T\mathrm{d}S_E + T\mathrm{d}S_Z.

The first of the terms on the right-hand side aligns with the familiar expression from standard thermodynamics, a scenario that typically excludes physics models featuring an enstropic degree of freedom. The inclusion of the second term is necessitated by the much richer physical model under consideration here.

In tandem with the modified or reformulated second law, we can express the Legendre transforms and the definition of entropy as the logarithm of the volume of a constant energy surface, now accounting for two energy sources. Consequently, we represent

dE=dEK+dEZ\mathrm{d}E = \mathrm{d}E_K + \mathrm{d}E_Z

as a summation of the kinetic and enstrophic energies, resulting in two terms on the left-hand side of the second law. The kinetic energy EKE_K aligns with the conventional description in thermodynamics, while EZE_Z is an additional energy source specific to the effect that the entropy considered is for the simplest of the fluid physics theories possible, namely incompressible with no mixtures.

From a mathematical standpoint, the Boltzmann–Gibbs entropy (or its discrete counterpart) is a special case of what is known as the relative entropy first introduced by Kullback and Leibler in 1951 [16] (consult [15] for more details). Consider a probability space (H,B(H))(\mathcal{H},\mathcal{B}(\mathcal{H})) and two measures η\eta and μ\mu defined on it, where B(H)\mathcal{B}(\mathcal{H}) denotes the space of Borel probability measures over space H\mathcal{H}. Then the relative entropy S(μ∣η)S(\mu|\eta) (of μ\mu with respect to η\eta) is defined by

S(μ∣η):={−∫Hdμdηlog⁡(dμdη) dη(u)if μ≪η,−∞otherwise.(2)S(\mu|\eta) := \begin{cases} -\int_{\mathcal{H}} \frac{\mathrm{d}\mu}{\mathrm{d}\eta}\log\left(\frac{\mathrm{d}\mu}{\mathrm{d}\eta}\right)\,\mathrm{d}\eta(u) & \text{if } \mu\ll\eta, \\ -\infty& \text{otherwise.} \end{cases} \tag*{(2)}

In the case where measure μ\mu is absolutely continuous with respect to η\eta, the Radon–Nikodym derivative dμdη\frac{\mathrm{d}\mu}{\mathrm{d}\eta} is in L1(dη)L^1(\mathrm{d}\eta). Once η\eta is chosen as the Lebesgue measure (when H=Rn\mathcal{H} = \mathbb{R}^n), then dμdη\frac{\mathrm{d}\mu}{\mathrm{d}\eta} can identified with the probability distribution f(u)f(u) in equation (1). The relative entropy turns into a functional that is seen to be maximized over the space of all probability measures, once we fix the reference measure η\eta. Therefore, the maximum entropy problem consists of finding a probability measure μ\mu on (H,B(H))(\mathcal{H},\mathcal{B}(\mathcal{H})), depending now on the choice of measure η\eta, such that the relative entropy is maximized under some given information. In the probability space, −S(μ∣η)-S(\mu|\eta) can be perceived as a measure of “distance” between the maximum entropy measure μ\mu and the reference measure η\eta, or any two measures for that matter.

In the context of fluid flow, a theory for entropy maximization was first established in [10], in which a well-defined probability measure acting as a linear functional on the configuration space of the Euler and Navier-Stokes equations, reduced to enforce particle interchange symmetries, was proved to maximize the Boltzmann–Gibbs entropy relative to alternate solutions, restricted to the constant energy surface. In this paper, we establish the existence of a probability measure on the configuration space satisfying a maximum entropy admissibility condition relative to alternate solutions of the Navier-Stokes equations, yet with a restriction to the energy-enstrophy surface. The significance of this restriction lies in its connection to the pursuit of establishing a rigorous mathematical foundation for the investigation of turbulent phenomena, a link that we briefly outline in the following section.

Motivation for entropy maximization

The recent nonuniqueness result for the Navier-Stokes equations [2] suggest that weak solutions may fall short from describing the complex nature of fluid flow. On this basis, a probabilistic description of the flow can serve as an alternate framework for studying the Navier-Stokes and Euler equations, as opposed to classical weak solutions and their Euler limits. In fact, any comprehensive exploration of turbulent flow phenomena must contend with some notion of averaging; especially as it relates to formulating the Kolmogorov theory of fully developed turbulence, widely considered to be valid by the practitioners of the field. We refer the reader to [8] for a modern comprehensive account on this topic.

In an endeavor to establish a mathematical theory of turbulence, Foias [5] (see also [6]), inspired by earlier work of G. Prodi, initiated a program to give, among other things, a rigorous formulation of the notion of statistical solutions for the Navier-Stokes equations and eventually derive the predictions of Kolmogorov for the energy spectrum and the structure functions [17, 18]. The Kolmogorov power laws are formulated in relation to the average (long time averages or ensemble averages) of certain physically measurable quantities; hence, by constructing a family of time-dependent and spatially homogeneous probability measures μt\mu_t (over an appropriate Hilbert space) that satisfy the Navier-Stokes equations in some averaged sense2. Therefore, Foias, later with Manley, Roas, and Temam [5, 6], was able to attach an unambiguous meaning to the notion of an ensemble average3. In the context of this program, one main assumption embedded in the derivation of the Kolmogorov power laws is that the global energy identity for the Navier-Stokes equations (with normalized mass density and viscosity ν\nu):

12∥u(t)∥2+ν∫0t∥∇u(s)∥2 ds=12∥u0∥2(3)\frac{1}{2}\lVert u(t)\rVert^2+\nu\int_0^t\lVert\nabla u(s)\rVert^2\,\mathrm{d}s=\frac{1}{2}\lVert u_0\rVert^2 \tag*{(3)}

is satisfied when integrated against a (spatially) homogeneous statistical solution μt\mu_t for all t>0t>0. Here u0u_0 is taken to be the (given) initial velocity of the fluid, i.e., u(x,0)=u0u(x,0)=u_0, and ∥⋅∥\lVert\cdot\rVert denotes the Lx2L_x^2-norm for simplicity. These solutions, which also remain invariant under re-scaling, are called self-similar homogeneous statistical solutions to the Navier-Stokes equations.

Leray–Hopf solutions are known to satisfy the global energy inequality (obtained by replacing “=” with “≤\leq” in relation (3)). Whether identity (3) holds in three space dimensions remains one of the major open problems for weak solutions of the Navier-Stokes equations. Incidentally, the same is true for statistical solutions: only an averaged version of the global energy inequality is known to hold.

In other words, these claims infer that a probabilistic description consistent with Kolmogorov’s spectrum needs to be supported over the set

Ge(t)≔{u:12∥u(t)∥2+ν∫0t∥∇u(s)∥2 ds=e(t)}(4)G_{e(t)}\coloneqq\left\{u:\frac{1}{2}\lVert u(t)\rVert^2+\nu\int_0^t\lVert\nabla u(s)\rVert^2\,\mathrm{d}s=e(t)\right\} \tag*{(4)}

where the prescribed energy e(t)e(t) is taken to be the energy profile for the initial velocity. However, since the existence and uniqueness of such solutions remain in question, energy e(t)e(t) can be prescribed a priori to ensure that Ge(t)G_{e(t)} is nonempty. Throughout this paper, we assume that, for each t≥0t\geq0, Ge(t)G_{e(t)} is defined for any prescribed positive energy e(t)e(t); such a set Ge(t)G_{e(t)} is called the energy-enstrophy surface.

A similar challenge arises when constructing numerical solutions. It is known that statistical solutions of the incompressible Euler equations and multidimensional systems of conservation laws can be computed. However, due to the lack of the well-posedness of these problems and the lack of convergence of their numerical approximations, it is not known whether these solutions are entropy or entropy maximizing solutions4. See [4, 19] and references therein for more details.

Next, we demonstrate how the principle of maximum entropy provides a systematic approach for determining probability distributions in the presence of constraints on the mean energy of the system.

Principle of maximum entropy: classical formulation

As indicated above, the principle of maximum entropy can be formulated in a number of different ways depending on modeling considerations. The formulation in the context of equilibrium statistical mechanics, which is presented here, most resembles the application of the principle for the purposes of this work. Consider the entropy functional (1) together with the two constraints:

∫g(u) dξ(u)=1,∫g(u)mu22 dξ(u)=ε,(5)\int g(u)\,\mathrm{d}\xi(u) = 1,\qquad\int g(u)\frac{m u^2}{2}\,\mathrm{d}\xi(u) = \varepsilon, \tag*{(5)}

where mu22\frac{m u^2}{2} is the kinetic energy of a single particle as measured in a fixed frame, ε\varepsilon is the mean energy per particle which is a priori known, and measure ξ\xi is assumed to exist over the configuration space consisting of all velocity fields5. The principle of maximum entropy, then, asserts that the physical entropy achieves its maximum, which we denote g∗(u)g^*(u), for the constrained system. Hence we may apply the methods of the calculus of variations to functional (1) subject to constraints (5) to conclude that there exist two Lagrange multipliers α\alpha and β\beta satisfying

∫δLδg(g(u);α,β) ϕ(u) dξ(u)∣g=g∗=0\left.\int\frac{\delta L}{\delta g}(g(u);\alpha,\beta)\,\phi(u)\,\mathrm{d}\xi(u)\right|_{g=g^*}=0

for any test function ϕ(u)\phi(u), where

L(g(u);α,β)=−g(u)log⁡g(u)+α(1−g(u))+β(ε−g(u)mu22).L(g(u);\alpha,\beta)=-g(u)\log g(u)+\alpha(1-g(u))+\beta\left(\varepsilon-g(u)\frac{m u^2}{2}\right).

Therefore, we obtain

−∫ϕ(u)((1+log⁡g∗(u))+α+βmu22) dξ(u)=0.(6)-\int\phi(u)\left((1+\log g^*(u))+\alpha+\beta\frac{m u^2}{2}\right)\,\mathrm{d}\xi(u)=0. \tag*{(6)}

If the integrand is assumed to be sufficiently regular, then

(1+log⁡g∗(u))+α+βmu22=0(1+\log g^*(u))+\alpha+\beta\frac{m u^2}{2}=0

for all uu, so that

g∗(u)=cαexp⁡(−βmu22),(7)g^*(u)=c_\alpha\exp\left(-\beta\frac{m u^2}{2}\right), \tag*{(7)}

where cα=exp⁡(−(1+α))c_\alpha=\exp(-(1+\alpha)). The two constraints (5) can be used to find parameters α\alpha and β\beta in terms of the known mean energy ε\varepsilon. One can even go a step further and determine the Lagrange multipliers in terms of more fundamental thermodynamics quantities by employing the first law of thermodynamics [1]. In any case, we obtain the well-known Maxwell-Boltzmann distribution describing the thermal equilibrium velocity distribution through the principle of maximum entropy.

In a similar fashion, the maximum entropy probability measure μt\mu_t (whose support is Gε(t)G_\varepsilon(t)) obtained for the problem of fluid flow should be associated with a universal law of decay in the limit: t→∞t\to\infty. This is indeed the case for self-similar homogeneous statistical solutions constructed: the average kinetic energy scales like Mt\frac{M}{t} as t→∞t\to\infty, where constant M=M(ν)M=M(\nu) is independent of any initial conditions (cf. [6]).

As will be illustrated in § 2, a version of the energy constraint similar to (5) will be presented, but not explicitly: it is included in terms of the support of the maximum entropy probability measure, that is to say over set Ge(t)G_e(t). This choice is motivated by the role that the global (averaged) energy identity for self-similar homogeneous statistical solutions plays in producing the predictions of Kolmogorov’s theory of turbulence.

Entropy Maximization and the Physical Measure

In this section, we establish the principle of maximum entropy in the context of incompressible flow. We begin first by setting out the notation and formulating the problem.

Principle of maximum entropy: formulation for the Navier-Stokes equations

The fixed time particle configuration space is the Hilbert space H\mathcal{H} of L2L^2 divergence-free velocity fields defined on a cube V⊂R3V \subset\mathbb{R}^3 and satisfying periodic boundary conditions on the boundary of VV. For every t∈[0,T]t \in[0,T] (fixed), define the Banach space XtX_t to be the space Lt∞Lx2∩Lt2Hx1L_t^\infty L_x^2 \cap L_t^2 H_x^1 endowed with the norm:

∥u∥t2:=12∥u(t)∥2+ν∫0t∥∇u(s)∥2 ds,\lVert u\rVert_t^2 := \frac{1}{2}\lVert u(t)\rVert^2 + \nu\int_0^t \lVert\nabla u(s)\rVert^2\,ds,

where ∥⋅∥\lVert\cdot\rVert denotes the Lx2L_x^2-norm. Throughout this section, given a fixed t≥0t \ge0, we fix a scalar e=e(t)≥0e=e(t)\ge0 to define the energy-enstrophy surface:

Ge(t)={u∈Xt:∥u∥t2=e(t)}for each t∈[0,T],(8)G_e(t)=\left\{u\in X_t:\lVert u\rVert_t^2=e(t)\right\}\qquad\text{for each }t\in[0,T], \tag*{(8)}

and make the identification

Ge(0)={u0∈H(V):∥u∥0=12∥u0∥2=e(0)}.(9)G_e(0)=\left\{u_0\in\mathcal{H}(V):\lVert u\rVert_0=\frac{1}{2}\lVert u_0\rVert^2=e(0)\right\}. \tag*{(9)}

It is clear that Ge(t)G_e(t) and Ge(0)G_e(0) are non-empty. For each t≥0t\ge0 and non-negative e(t)e(t), a given physical probability measure ηe,t\eta_{e,t} is supported over Ge(t)G_e(t). For measures ηe,t\eta_{e,t} and μe,t\mu_{e,t} belonging the probability space M:=(H,B(H))\mathcal{M}:=(\mathcal{H},\mathcal{B}(\mathcal{H})), we define

S(μe,t∣ηe,t):={−∫Hdμe,tdηe,tlog⁡(dμe,tdηe,t)dηe,t(u)if μe,t≪ηe,t,−∞otherwise,(10)S(\mu_{e,t}\mid\eta_{e,t}):= \begin{cases} -\int_{\mathcal{H}}\frac{d\mu_{e,t}}{d\eta_{e,t}}\log\left(\frac{d\mu_{e,t}}{d\eta_{e,t}}\right)d\eta_{e,t}(u) & \text{if }\mu_{e,t}\ll\eta_{e,t},\\ -\infty& \text{otherwise}, \end{cases} \tag*{(10)}

which is known as the relative entropy functional of measure μe,t\mu_{e,t} with respect to the reference probability measure ηe,t\eta_{e,t}. Since log⁡(y)\log(y) is a concave function and fe,t(u):=dμe,t(u)dηe,t(u)∈L1(ηe,t)f_{e,t}(u):=\frac{d\mu_{e,t}(u)}{d\eta_{e,t}(u)}\in L^1(\eta_{e,t}), we employ Jensen’s inequality to obtain

S(μe,t∣ηe,t)=S(fe,t)=∫Hlog⁡(1fe,t(u))dμe,t(u)≤log⁡(∫H1fe,t(u)dμe,t(u))≤log⁡(∫Hdηe,t(u))=log⁡(ηe,t(Ge(t)))=0.\begin{aligned} S(\mu_{e,t}\mid\eta_{e,t})=S(f_{e,t})=\int_{\mathcal{H}}\log\left(\frac{1}{f_{e,t}(u)}\right)d\mu_{e,t}(u)\le\log\left(\int_{\mathcal{H}}\frac{1}{f_{e,t}(u)}d\mu_{e,t}(u)\right) \\ &\le\log\left(\int_{\mathcal{H}}d\eta_{e,t}(u)\right)=\log\left(\eta_{e,t}(G_e(t))\right)=0. \end{aligned}

We postpone the discussion on the measure ηe,t\eta_{e,t} over the energy-enstrophy to the next section.

Our problem consists of maximizing the relative entropy functional (10) subject to some appropriate constraints: the solution obtained is a probability distribution

∫fe,t(u) dηe,t(u)=∫dηe,t(u)=1,(11)\int f_{e,t}(u)\,d\eta_{e,t}(u)=\int d\eta_{e,t}(u)=1, \tag*{(11)}

and the support of the resulting measure must be supported on Ge(t)G_e(t). For the latter constraint, it will be shown that the fact that ηe,t\eta_{e,t} is supported on Ge(t)G_e(t) is sufficient to ensure the second constraint is satisfied by the maximum entropy measure. We hence require a solution to the following maximum problem:

sup⁡{S(fe,t):μe,t∈M satisfying (2.4)}(12)\sup\{\mathcal{S}(f_{e,t}):\mu_{e,t}\in\mathcal{M}\text{ satisfying }(2.4)\} \tag*{(12)}

for each fixed t∈[0,T]t\in[0,T] and a priori prescribed e(t)e(t).

Main theorem for the principle of maximum entropy

The following basic lemma is key in establishing our main result – Theorem 2.2 – below.

Lemma 2.1. Fix t∈[0,T]t\in[0,T] and e(t)≥0e(t)\geq0. Then the following statements hold:

(i) The relative entropy is non-positive: S(fe,t)≤0\mathcal{S}(f_{e,t})\leq0 for all μe,t∈M\mu_{e,t}\in\mathcal{M}, where fe,t=dμe,tdηe,tf_{e,t}=\frac{\mathrm{d}\mu_{e,t}}{\mathrm{d}\eta_{e,t}}.

(ii) The relative entropy is concave in μe,t\mu_{e,t}: for μe,t1,μe,t2∈M\mu^1_{e,t},\mu^2_{e,t}\in\mathcal{M} so that S(μe,t1∣ηe,t)\mathcal{S}(\mu^1_{e,t}\mid\eta_{e,t}) and S(μe,t2∣ηe,t)\mathcal{S}(\mu^2_{e,t}\mid\eta_{e,t}) are finite, then

S(μe,t∣ηe,t)≥αS(μe,t1∣ηe,t)+(1−α)S(μe,t2∣ηe,t),\mathcal{S}(\mu_{e,t}\mid\eta_{e,t})\geq\alpha\mathcal{S}(\mu^1_{e,t}\mid\eta_{e,t})+(1-\alpha)\mathcal{S}(\mu^2_{e,t}\mid\eta_{e,t}),

where μe,t=αμe,t1+(1−α)μe,t2\mu_{e,t}=\alpha\mu^1_{e,t}+(1-\alpha)\mu^2_{e,t} and α∈[0,1]\alpha\in[0,1].

(iii) The total variation of μe,t\mu_{e,t}, defined by

∣μe,t∣=sup⁡Π∑i=1m∫Aidμe,t(u),|\mu_{e,t}|=\sup_{\Pi}\sum_{i=1}^{m}\int_{A_i}\mathrm{d}\mu_{e,t}(u),

satisfies

∣μe,t−ηe,t∣2≤−2S(μe,t∣ηe,t),|\mu_{e,t}-\eta_{e,t}|^2\leq-2\mathcal{S}(\mu_{e,t}\mid\eta_{e,t}),

where the supremum is taken over all finite partitions Π={A1,A2,⋯Am}\Pi=\{A_1,A_2,\cdots A_m\} of H\mathcal{H}.

(iv) Functional S(fe,t)\mathcal{S}(f_{e,t}) is upper-semicontinuous: if ∣μe,tn−μe,t∣→0|\mu^n_{e,t}-\mu_{e,t}|\to0 as n→∞n\to\infty, then

S(μe,t∣ηe,t)≥lim sup⁡nS(μe,tn∣ηe,t).\mathcal{S}(\mu_{e,t}\mid\eta_{e,t})\geq\limsup_{n}\mathcal{S}(\mu^n_{e,t}\mid\eta_{e,t}).

Proof. Statement (i) has been proven in §2.1. The proofs of (iii)–(iv) can be found in Ch.1 of [12].

A slight modification of (ii) is also stated in [12] but without explicit proof, so we include the proof of (ii) for the sake of completeness. We begin by invoking the inequality:

∑i=1mailog⁡(aibi)≥alog⁡(ab),\sum_{i=1}^{m}a_i\log\left(\frac{a_i}{b_i}\right)\geq a\log\left(\frac{a}{b}\right),

where ai,bia_i,b_i are any non-negative numbers for all i=1,2,…,mi=1,2,\ldots,m, with a=∑iaia=\sum_i a_i and b=∑ibib=\sum_i b_i. The equality holds if and only if a1b1=a2b2=⋯=ambm\frac{a_1}{b_1}=\frac{a_2}{b_2}=\cdots=\frac{a_m}{b_m}.

Fix α∈[0,1]\alpha\in[0,1], and let μe,t=αμe,t1+(1−α)μe,t2\mu_{e,t}=\alpha\mu^1_{e,t}+(1-\alpha)\mu^2_{e,t}. Consider

\begin{aligned} \alpha\frac{\mathrm{d}\mu^1_{e,t}}{\mathrm{d}\eta_{e,t}}\log\left(\frac{\mathrm{d}\mu^1_{e,t}}{\mathrm{d}\eta_{e,t}}\right)+(1-\alpha)\frac{\mathrm{d}\mu^2_{e,t}}{\mathrm{d}\eta_{e,t}}\log\left(\frac{\mathrm{d}\mu^2_{e,t}}{\mathrm{d}\eta_{e,t}}\right) \\ &=\alpha\frac{\mathrm{d}\mu^1_{e,t}}{\mathrm{d}\eta_{e,t}}\log\left(\frac{\alpha\frac{\mathrm{d}\mu^1_{e,t}}{\mathrm{d}\eta_{e,t}}}{\alpha\frac{\mathrm{d}\mu^1_{e,t}}{\mathrm{d}\eta_{e,t}}}\right)+(1-\alpha)\frac{\mathrm{d}\mu^2_{e,t}}{\mathrm{d}\eta_{e,t}}\log\left(\frac{(1-\alpha)\frac{\mathrm{d}\mu^2_{e,t}}{\mathrm{d}\eta_{e,t}}}{(1-\alpha)\frac{\mathrm{d}\mu^2_{e,t}}{\mathrm{d}\eta_{e,t}}}\right) \\ &\ge\left(\alpha\frac{\mathrm{d}\mu_{e,t}^{1}}{\mathrm{d}\eta_{e,t}} + (1-\alpha)\frac{\mathrm{d}\mu_{e,t}^{2}}{\mathrm{d}\eta_{e,t}}\right)\log\left(\frac{\alpha\frac{\mathrm{d}\mu_{e,t}^{1}}{\mathrm{d}\eta_{e,t}} + (1-\alpha)\frac{\mathrm{d}\mu_{e,t}^{2}}{\mathrm{d}\eta_{e,t}}}{\alpha\frac{\mathrm{d}\eta_{e,t}}{\mathrm{d}\eta_{e,t}} + (1-\alpha)\frac{\mathrm{d}\eta_{e,t}}{\mathrm{d}\eta_{e,t}}\right)} \\ &= \frac{\mathrm{d}\mu_{e,t}}{\mathrm{d}\eta_{e,t}}\log\left(\frac{\mathrm{d}\mu_{e,t}}{\mathrm{d}\eta_{e,t}}\right). \end{aligned}

Integrating with respect to ηe,t\eta_{e,t} gives the desired conclusion.

We now state and prove the main result of this paper.

Theorem 2.2. The physical measure ηe,t\eta_{e,t} is a solution to the maximization problem (12). That is, the physical measure maximizes the entropy functional relative to all the comparison probability measures μe,t\mu_{e,t}, both restricted to the global energy-enstrophy surface Ge(t)G_e(t).

Proof. We divide the proof into two steps.

  1. Existence. We follow the argument in Ch.3 of [12]. Fix t∈[0,T]t \in[0,T] and a non-negative function e(t)e(t). Consider the problem

Isup=sup⁡μe,t∈AS(fe,t)≤0,(13)I_{\mathrm{sup}}=\sup_{\mu_{e,t}\in\mathcal{A}} S(f_{e,t})\le0, \tag*{(13)}

where A:={μe,t∈M satisfying \mathcal{A}:=\{\mu_{e,t}\in\mathcal{M}\text{ satisfying }(11)}\}. Let μe,tn\mu_{e,t}^{n} be a maximizing sequence for our problem, i.e., μe,tn∈A\mu_{e,t}^{n}\in\mathcal{A}, and S(fe,tn)→IsupS(f_{e,t}^{n})\to I_{\mathrm{sup}} as n→∞n\to\infty. It can be directly checked that

S(fe,tm)+S(fe,tn)=S(μe,tm∣ηe,t)+S(μe,tn∣ηe,t)=2S(μe,tm,n∣ηe,t)+S(μe,tm∣μe,tm,n)+S(μe,tn∣μe,tm,n),(14)\begin{aligned} S(f_{e,t}^{m})+S(f_{e,t}^{n})=S(\mu_{e,t}^{m}|\eta_{e,t})+S(\mu_{e,t}^{n}|\eta_{e,t}) \\ &=2S(\mu_{e,t}^{m,n}|\eta_{e,t})+S(\mu_{e,t}^{m}|\mu_{e,t}^{m,n})+S(\mu_{e,t}^{n}|\mu_{e,t}^{m,n}), \tag*{(14)} \end{aligned}

where μe,tm,n=12(μe,tm+μe,tn)\mu_{e,t}^{m,n}=\frac{1}{2}(\mu_{e,t}^{m}+\mu_{e,t}^{n}). Because of the linear nature of the constraint, we see that μe,tm,n∈A\mu_{e,t}^{m,n}\in\mathcal{A}. The concavity of the relative entropy, Lemma 2.1, gives

2S(μe,tm,n∣ηe,t)≥S(μe,tm∣ηe,t)+S(μe,tn∣ηe,t)2S(\mu_{e,t}^{m,n}|\eta_{e,t})\ge S(\mu_{e,t}^{m}|\eta_{e,t})+S(\mu_{e,t}^{n}|\eta_{e,t})

which, together with the non-positivity of S(⋅)S(\cdot), implies that lim⁡m,nS(μe,tm,n∣ηe,t)=2Isup\lim_{m,n}S(\mu_{e,t}^{m,n}|\eta_{e,t})=2I_{\mathrm{sup}} so that lim⁡m,nS(μe,tm∣μe,tm,n)=0=lim⁡m,nS(μe,tn∣μe,tm,n)\lim_{m,n}S(\mu_{e,t}^{m}|\mu_{e,t}^{m,n})=0=\lim_{m,n}S(\mu_{e,t}^{n}|\mu_{e,t}^{m,n}). By Lemma 2.1(iii),

∣μe,tm−μe,tn∣≤∣μe,tm−μe,tm,n∣+∣μe,tn−μe,tm,n∣≤−2S(μe,tm∣μe,tm,n)+−2S(μe,tn∣μe,tm,n)→0as m,n→∞.(15)\begin{aligned} |\mu_{e,t}^{m}-\mu_{e,t}^{n}|\le|\mu_{e,t}^{m}-\mu_{e,t}^{m,n}|+|\mu_{e,t}^{n}-\mu_{e,t}^{m,n}| \\ &\le\sqrt{-2S(\mu_{e,t}^{m}|\mu_{e,t}^{m,n})}+\sqrt{-2S(\mu_{e,t}^{n}|\mu_{e,t}^{m,n})}\to0\quad\text{as }m,n\to\infty. \tag*{(15)} \end{aligned}

Then we conclude that there exists μe,t∗∈M\mu_{e,t}^{*}\in\mathcal{M} satisfying

lim⁡n∣μe,tn−μe,t∗∣=0.(16)\lim_{n}|\mu_{e,t}^{n}-\mu_{e,t}^{*}|=0. \tag*{(16)}

Moreover, due again to the linearity of constraint (11), it readily follows that μe,t∗∈A\mu_{e,t}^{*}\in\mathcal{A}. Now, by the upper-semicontinuity of SS, Lemma 2.1(iv), we conclude

Isup=S(μe,t∗∣ηe,t).I_{\mathrm{sup}}=S(\mu_{e,t}^{*}|\eta_{e,t}).
  1. Euler-Lagrange equations. Since the constraint is linear in ff, then, by the standard methods of the Calculus of Variations, there exists a constant αt\alpha_t (independent of $u)

satisfying the Euler-Lagrange equations:

−∫Ψt(u)(1+log⁡(fe,t∗(u))) dηe,t(u)−αt∫Ψt(u) dηe,t(u)=0(17)-\int\Psi_t(u)\left(1+\log(f^*_{e,t}(u))\right)\,\mathrm{d}\eta_{e,t}(u)-\alpha_t\int\Psi_t(u)\,\mathrm{d}\eta_{e,t}(u)=0 \tag*{(17)}

for any test function Ψt(u)\Psi_t(u). Then we have

−(1+log⁡(fe,t∗(u)))−αt=0.(18)-\left(1+\log(f^*_{e,t}(u))\right)-\alpha_t=0. \tag*{(18)}

for ηe,t\eta_{e,t}-almost every u∈Ge(t)u\in G_e(t). Thus,

log⁡(fe,t∗(u))=−(1+αt).(19)\log(f^*_{e,t}(u))=-(1+\alpha_t). \tag*{(19)}

or equivalently

fe,t∗(u)=exp⁡{−(1+αt)}.(20)f^*_{e,t}(u)=\exp\{-(1+\alpha_t)\}. \tag*{(20)}

That is, fe,t∗f^*_{e,t} is constant in uu. Since both μe,t∗\mu^*_{e,t} and ηe,t\eta_{e,t} are probability measures, constraint (11) yields αt≡−1\alpha_t\equiv-1, implying that fe,t∗≡1f^*_{e,t}\equiv1 or equivalently μe,t∗=ηe,t\mu^*_{e,t}=\eta_{e,t}. In other words, the physical measure ηe,t\eta_{e,t} achieves the maximum entropy under constraint (11).

Remark 2.1. The global energy-enstrophy surface Ge(t)G_e(t) in (8) for Theorem 2.2 can be replaced by the local energy-enstrophy surface: For any given e(x,t)>0e(x,t)>0,

12∫V∣u(x,t)∣2ψ(x,t) dx+ν∫0t∫V∣∇u∣2ψ dx dτ−12∫0t∫V{∣u∣2(∂tψ+νΔψ)+(∣u∣2+2p)u⋅∇ψ} dx dτ=∫Ve(x,t)ψ(x,t) dx(21)\frac{1}{2}\int_V |u(x,t)|^2\psi(x,t)\,\mathrm{d}x+\nu\int_0^t\int_V|\nabla u|^2\psi\,\mathrm{d}x\,\mathrm{d}\tau -\frac{1}{2}\int_0^t\int_V\left\{|u|^2(\partial_t\psi+\nu\Delta\psi)+(|u|^2+2p)u\cdot\nabla\psi\right\}\,\mathrm{d}x\,\mathrm{d}\tau =\int_V e(x,t)\psi(x,t)\,\mathrm{d}x \tag*{(21)}

for any ψ(x,τ)∈Cc∞(V×[0,t])\psi(x,\tau)\in C^\infty_c(V\times[0,t]). Then Theorem 2.2 can state as follows: The physical measure maximizes the entropy production rate relative to all the comparison probability measures μe,t\mu_{e,t}, both restricted to the local entropy-enstrophy surface (21).

Remark 2.2. A connection can be drawn to the classical results highlighted in §1.2: at t=0t=0, we set ηe,0=g(u) ξ\eta_{e,0}=g(u)\,\xi, where g(u)g(u) is given by equation (7); this implies that

dμe,0∗dξ=g(u)∈L1(dξ)\frac{\mathrm{d}\mu^*_{e,0}}{\mathrm{d}\xi}=g(u)\in L^1(\mathrm{d}\xi)

in agreement with the classical principle of maximum entropy. Furthermore, at t=0t=0, the energy condition in (9) gives the classical energy constraint (1.5)$_2$, once we identify the mean energy with ∫e(0)g(u) dξ\int e(0)g(u)\,\mathrm{d}\xi.

Connections and Remarks

In view of the previous work [10] on the principle of maximum entropy applied to the problem of inviscid fluid flow, the current results outline a general approach for studying the case of viscous flow under a common framework. As such, it is clear that the standard energy class associated with the Navier-Stokes equations provides the natural setting for expressing the total energy ∥u∥22\|u\|_2^2 of the fluid at each time tt, and hence defining the energy-enstrophy surface Ge(t)G_e(t).

Equally clear is that having a complete theory of entropy maximization for fluid flow is tied with specifying selection criteria for the reference measure ηe,t\eta_{e,t}. Whether or not such criteria are model dependent is open to question. For instance, in the finite-dimensional setting, a natural candidate for the reference measure is the Lebesgue measure (see [3, 10]) as it is the only measure (up to a multiplicative constant) that possesses the property of translation invariance; that is, in the velocity space, it remains invariant under the Galilean transformation. Yet another candidate for the reference measure is given by considering the statistics of fluid flows at a long time; this is related to stationary statistical solutions of the Navier-Stokes equations and Kolmogorov’s theory of turbulence (cf. [6]).

We first recall the notion of the stationary statistical solutions of the Navier-Stokes equations. Assume that μ\mu is a Borel probability measure over space H\mathcal{H} with finite enstrophy, that is,

∫H∥∇u∥2 dμ(u)<∞.\int_{\mathcal{H}} \lVert\nabla u\rVert^{2}\,d\mu(u)<\infty.

Then the measure μ\mu is said to be a stationary statistical solution of the Navier-Stokes equations if μ\mu satisfies the following (stationary) Liouville-type equation:

∫H∫VQ(u) Φ′(u) dx dμ(u)=0(22)\int_{\mathcal{H}}\int_{V} Q(u)\,\Phi'(u)\,dx\,d\mu(u)=0 \tag*{(22)}

for all suitable test functions Φ\Phi, and the energy inequality:

∫H∫V(ν∥∇u∥2−F⋅u) dx dμ(u)≤0,(23)\int_{\mathcal{H}}\int_{V}\left(\nu\lVert\nabla u\rVert^{2}-F\cdot u\right)\,dx\,d\mu(u)\leq0, \tag*{(23)}

where Q(u)=−νΔu+(u⋅∇)u−FQ(u)=-\nu\Delta u+(u\cdot\nabla)u-F.

Now from the perspective of thermodynamics, the flow relaxes, after sufficiently long enough time, to its equilibrium state characterized by maximum entropy; so we may assume that the statistics of this (final) state is encoded by the stationary solution μ≐μ∞\mu\doteq\mu_\infty. On the other hand, one of the fundamental postulates of Kolmogorov’s theory for fully developed isotropic turbulence is the similarity hypothesis that the statistical behavior of the flow must be determined by only two parameters: the (average) energy dissipation rate ϵˉν(t)\bar{\epsilon}_\nu(t) and viscosity ν\nu (see also [11]). Therefore, the average energy balance, in the absence of body forces, for the case of non-zero energy dissipation rate (i.e., the average of equation (3)) reads

12∫H∥u∥2 dμt(u)+ν∫H∫0t∥∇u(s)∥2 ds dμt(u)⏟ϵˉν(t)=12∫H∥u∥2 dμ0(u),(24)\frac{1}{2}\int_{\mathcal{H}}\lVert u\rVert^{2}\,d\mu_t(u)+\nu\underbrace{\int_{\mathcal{H}}\int_{0}^{t}\lVert\nabla u(s)\rVert^{2}\,ds\,d\mu_t(u)}_{\bar{\epsilon}_\nu(t)}=\frac{1}{2}\int_{\mathcal{H}}\lVert u\rVert^{2}\,d\mu_0(u), \tag*{(24)}

where the probability measure μt\mu_t is such that equation (24) is well-defined (e.g. a time-dependent statistical solution). However, if we set μt=μ∞\mu_t=\mu_\infty, then the non-zero energy (finite) dissipation rate can be maintained only if the kinetic energy is infinite, as seen by equation (23). This nonphysical conclusion was pointed out first by E. Hopf [11] (see also [6]). The resolution of this paradox, according to Foias, Manley, Rosa, and Temam [5, 6, 7], lies instead in introducing a family of probability measures μt\mu_t (parameterized by time t∈[0,∞)t\in[0,\infty))6, a feature which the physical measure ηe,t\eta_{e,t} enjoys through the energy-enstrophy surface Ge(t)G_e(t). These solutions, called self-similar homogeneous statistical solutions, were originally postulated for describing the decay of fully developed turbulence according to Kolmogorov [17, 18], though to date the question of their existence has not been resolved yet.

Meanwhile, our results suggest replacing the right-hand side of equation (24) with eˉ(t)\bar{e}(t), a quantity that can ostensibly be identified with ∫Ge(t)e(t) dηe,t\int_{G_e(t)}e(t)\,d\eta_{e,t}. Such a choice adds a degree of freedom to the final statistics of the flow, now depending upon the initial prescribed energy e(t)e(t). In keeping with the above similarity hypothesis, however, we can in principle prescribe the total energy e(t)e(t) at a high Reynolds number in terms of parameters (ϵˉν,ν)(\bar{\epsilon}_{\nu},\nu) and wavenumber κ\kappa, removing the ambiguity associated with the term: ∫Ge(t)e(t) dηe,t\int_{\mathcal{G}_{e(t)}} e(t)\,\mathrm{d}\eta_{e,t}. Let E(κ,t)\mathcal{E}(\kappa,t) be the energy spectrum (in the Fourier space with ∣κ∣=κ|\boldsymbol{\kappa}|=\kappa) associated with the average kinetic energy. Then the average total energy can be expressed in terms of the energy spectrum:

∫0∞E(κ,t)(12+νκ2) dκ.(25)\int_{0}^{\infty}\mathcal{E}(\kappa,t)\left(\frac{1}{2}+\nu\kappa^{2}\right)\,\mathrm{d}\kappa. \tag*{(25)}

Equation (25) provides a natural alternative for our (average) total energy eˉ(t)\bar{e}(t). In fact, the amount of total energy contained between wavenumbers κ1\kappa_{1} and κ2\kappa_{2} at a given time tt, i.e.,

∫κ1κ2E(κ,t)(12+νκ2) dκ\int_{\kappa_{1}}^{\kappa_{2}}\mathcal{E}(\kappa,t)\left(\frac{1}{2}+\nu\kappa^{2}\right)\,\mathrm{d}\kappa

can now be identified with the change in the total energy eˉκ2(t)−eˉκ1(t)\bar{e}_{\kappa_{2}}(t)-\bar{e}_{\kappa_{1}}(t).

As ν→0\nu\to0, or equivalently at infinite Reynolds number, the implications of Kolmogorov’s theory give

E(κ,t)∼C(t)ϵˉ(t)23κ−53(26)\mathcal{E}(\kappa,t)\sim C(t)\bar{\epsilon}(t)^{\frac{2}{3}}\kappa^{-\frac{5}{3}} \tag*{(26)}

for large enough ∣κ∣=κ|\boldsymbol{\kappa}|=\kappa but still belonging to the inertial range, where C(t)C(t) is a positive dimensionless constant independent of viscosity ν\nu, and lim⁡ν→0ϵˉν(t)=ϵˉ(t)>0\lim_{\nu\to0}\bar{\epsilon}_{\nu}(t)=\bar{\epsilon}(t)>0. In both cases, nevertheless, the explicit dependence on e(t)e(t) can be removed in the final statistics in favor of the energy dissipation rate and viscosity.

In view of interpreting the total energy e(t)e(t) in terms of the energy spectrum, one can conceive of the maximum entropy probability measure, that is the reference probability measure, as assigning a probability for finding the total energy carried by a certain wavenumber κ\kappa to be eˉκ(t)\bar{e}_{\kappa}(t). When the wavenumber lies in the inertial range and in the limit of infinite Reynolds number, this probability can be calculated via the Kolmogorov energy spectrum (25).

Acknowledgments. Gui-Qiang G. Chen’s research is partially supported by the UK Engineering and Physical Sciences Research Council Awards EP/L015811/1, EP/V008854/1, and EP/V051121/1.

Declarations

The authors have no competing interests to declare that are relevant to the content of this article. In addition, our manuscript has no associated data.

References

Mathematical Institute, University of Oxford, Oxford, OX2 6GG, UK

Email address: [email protected]

Stony Brook University, Stony Brook, NY 11794, and GlimmAnalytics LLC, USA

Email address: [email protected]

Department of Mathematics, College of Science, Kuwait University, Safat 13060, Kuwait

Email address: [email protected]

References

  1. [1]F. A. BAIS AND J. D. FARMER, The physics of information, In: The Handbook on the Philosophy of Information, edited by J. van Benthem and P. Adriaan, Elsevier, (2007).
  2. [2]T. BACKMASTER AND V. VICOL, Nonuniqueness of weak solutions to the Navier-Stokes equations, Ann. Math., 189 (2019), pp. 101–144.
  3. [3]J. DUNKEL, P. TALKNER, AND P. HÄNGGI, Relative entropy, Haar measures and relativistic canonical velocity distributions, New Journal of Physics, 9 (2007), p. 144.DOI
  4. [4]U. S. FOJORDHOLM, K. LYE, S. MISHRA, AND F. WEBER, Statistical solutions of hyperbolic systems of conservation laws: numerical approximation, Mathematical Models and Methods in Applied Sciences, 30 (2020), pp. 539–609.
  5. [5]C. FOIAS, Statistical study of Navier-Stokes equations, I, Rendiconti del Seminario Matematico della Universita’ di Padova, 48 (1972), pp. 219–348.
  6. [6]C. FOIAS, O. MANLEY, R. ROSA, AND R. TEMAM, Navier-Stokes Equations and Turbulence, Vol. 83, Cambridge University Press, 2001.DOI
  7. [7]C. FOIAS, R. ROSA, AND R. TEMAM, Properties of time-dependent statistical solutions of the three-dimensional Navier-Stokes equations, Annales de L’Institut Fourier, 63 (2013), pp. 2515–2573.
  8. [8]U. FRISCH AND R. J. DONNELLY, Turbulence: the legacy of A. N. Kolmogorov, Physics Today, 49 (1996), pp. 82–84.
  9. [9]M. GAVRILOV, R. CHÉTRITE AND J. BECHHOEFER, Direct measurement of weakly nonequilibrium system entropy is consistent with Gibbs–Shannon form. Proceedings of the National Academy of Sciences, 114(42)(2017), pp. 11097–11102.DOI
  10. [10]J. GLIMM, D. LAZAREV, AND G.-Q. G. CHEN, Maximum entropy production as a necessary admissibility condition for the fluid Navier–Stokes and Euler equations, SN Applied Sciences, 2 (2020), pp. 1–9.DOI
  11. [11]E. HOPF, Statistical hydromechanics and functional calculus, Journal of Rational Mechanics and Analysis, 1 (1952), pp. 87–123.DOI
  12. [12]S. IHARA, Information Theory for Continuous Systems, Vol. 2, World Scientific, 1993.
  13. [13]E. T. JAYNES, Information theory and statistical mechanics, Physical review, 106 (1957), p. 620.DOI
  14. [14]J. N. KAPUR, Maximum-Entropy Models in Science and Engineering, John Wiley & Sons, 1989.
  15. [15]S. KULLBACK, Information Theory and Statistics, Courier Corporation, 1997.
  16. [16]S. KULLBACK, AND R. A. LEIBLER, On information and sufficiency, Annals of Mathematical Statistics, 22(1)(1951), pp. 79–86.DOI
  17. [17]A. N. KOLMOGOROV, Dissipation of energy in the locally isotropic turbulence, Dokl. Akad. Nauk. SSSR, 32 (1941), pp. 19–21.
  18. [18]A. N. KOLMOGOROV, The local structure of turbulence in incompressible viscous fluid for very large Reynolds numbers, Proc. Royal Soc. London. Series A: Mathematical and Physical Sciences, 434 (1991), pp. 9–13.
  19. [19]S. LANTHALER, S. MISHRA, AND C. PARÉS-PULIDO, Statistical solutions of the incompressible euler equations, Mathematical Models and Methods in Applied Sciences, 31 (2021), pp. 223–292.DOI
  20. [20]M. I. VISHIK AND A. V. FURKISOV, Mathematical Problems of Statistical Hydromechanics, vol. 9, Springer Science & Business Media, 2012.
  21. [21]A. WHERL, The many facets of entropy. Rep Math Phys, 30(1)(1991), pp. 119–129.

Paper details

vlnc:W_gy8mc94y
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents