Abstract
It is shown that the long-time dynamics (the global attractor) of the 2D Navier–Stokes system is embedded in the long-time dynamics of an ordinary differential equation, called a determining form, in a space of trajectories which is isomorphic to for sufficiently large depending on the physical parameters of the Navier–Stokes equations. A unified approach is presented, based on interpolant operators constructed from various determining parameters for the Navier–Stokes equations, namely, determining nodal values, Fourier modes, finite volume elements, finite elements, and so on. There are two immediate and interesting consequences of this unified approach. The first is that the constructed determining form has a Lyapunov function, and thus its solutions converge to the set of steady states of the determining form as the time goes to infinity. The second is that these steady states of the determining form can be uniquely identified with the trajectories in the global attractor of the Navier–Stokes system. It should be added that this unified approach is general enough that it applies, in an almost straightforward manner, to a whole class of dissipative dynamical systems.
Bibliography: 23 titles.
Export citation and abstract BibTeX RIS
The work of the first author was supported by NSF (grant no. DMS-1109784), that of the second author by NSF (grant nos. DMS-1008661 and DMS-1109638), and that of the fourth author by NSF (grant nos. DMS-1009950, DMS-1109640, and DMS-1109645), as well as the Minerva Stiftung/Foundation.
Dedicated to the memory of Professor Mark Vishik
1. Introduction
The 2D system of Navier–Stokes equations (NSE), (2.1) and (2.2), in addition to being a fundamental component of many fluid models, is intriguing for several theoretical reasons. In featuring both a direct cascade of enstrophy and an inverse cascade of energy, it displays more complicated turbulence phenomena than does the 3D system of NSE [3], [19], [20]. Also unlike in 3D, the global existence theory for the 2D NSE is complete (see, for instance, [6], [22]). In fact, the long-time dynamics of the 2D NSE is entirely contained in the global attractor (see (2.5)), a compact finite-dimensional subset of the infinite-dimensional phase space of solenoidal finite-energy vector fields, (see, for instance, [6], [10], [16], [22]). Sharp estimates of the dimension of the global attractor in terms of the relevant physical parameters were first established in [7] (see also [6], [21], [22] and references therein). If there were an inertial manifold (that is, a Lipschitz, finite-dimensional, forward-invariant manifold which attracts each bounded set at an exponential rate), then , and the dynamics on would be captured by an ordinary differential equation (ODE), called an inertial form, in a finite-dimensional phase space [6], [12], [13], [22]. This can be attained through reduction of the original evolution equation to an equation on the inertial manifold . Yet the existence of an inertial manifold for the 2D NSE has been an open problem since the 1980s!!
This is very surprising since there are even stronger indicators of the finite-dimensional behavior of the 2D NSE. The solutions in are determined by the asymptotic behavior of a sufficient (finite) number of determining parameters. If in the limit as a sufficiently large number of low Fourier modes (or nodal values, or finite volume elements) for two solutions in converge to each other, then the solutions coincide (see, for instance, [5] for a unified theory of determining parameters and projections). This is equivalent, at least in the case of Fourier modes, to the following:
This notion of determining modes, which was introduced in [11] (see also, [18] for sharp estimates of the number of determining modes), was used in [8] to construct a system of ODEs in the Banach space that govern the evolution of trajectories in the space . We call the system of ODEs a determining form. Trajectories in the global attractor of the 2D NSE (2.2) are identified with travelling wave solutions of the determining form. There are conceptually two time variables in play for the determining form: the evolving time for the ODE and the original time variable of the NSE that now parameterizes complete trajectories in . Though the determining form has an infinite-dimensional phase space, the vector field that governs the evolution is globally Lipschitz, so the determining form is an ODE in the true sense. The key to constructing the determining form in [8] is to extend to the whole space the map provided by (1.1) on the set of complete trajectories in . The extended map is shown to be Lipschitz, and its image plays the role of recovering the higher modes, while the evolving trajectory in represents the lower modes.
The determining form in this paper has an entirely different character. It is a system which possesses a Lyapunov function and whose steady states are precisely the trajectories in the global attractor of the 2D NSE. It is more general in that it can be used with a variety of determining parameters, including nodal values as well as Fourier modes. Furthermore, it provides a general framework and strategy that can be implemented for other dissipative systems. Like the determining form in [8], the key to its construction is the extension of a map defined at first only for projections of trajectories in . This is done by using the feedback control term added to the NSE as suggested in [1] and [2], and this involves an interpolating operator approximating the identity map at the level (for instance, can be based on nodal values, where represents the grid size). This construction and the statements of our main results are presented in § 3. In § 2 we provide some preliminary background material and useful inequalities concerning the Navier–Stokes equations. Details of the proofs of our main results are given in §§ 4 and 5.
2. Functional setting and the Navier–Stokes equations
We consider the two-dimensional incompressible Navier–Stokes equations
subject to periodic boundary conditions with the basic domain . The velocity field and the pressure are the unknown functions, while is a given forcing term, and is a given constant viscosity.
Let
For any subset we define . We denote by and the closures of in and , respectively. The inner product and norm in the Hilbert spaces and will be denoted by and , respectively, and the corresponding inner product and norm in the Hilbert spaces and will be denoted by and . Specifically, for every we set
Let denote the dual space of the space .
Using the above functional notation, we can write the Navier–Stokes equations as an evolution equation in the Hilbert space (cf. [6], [22]):
The Stokes operator , the bilinear operator , and the force are defined as
where is the Helmholtz orthogonal projection from onto , and where and are smooth enough that makes sense. In this paper we will assume that .
We remark that . The operator is self-adjoint, with compact inverse. Therefore, the space possesses an orthonormal basis of eigenfunctions of , namely, , with (cf. [6], [22]). The powers are given by
Note that all the powers of commute with . Note also that and that
It is well known that the NSE (2.2) has a global attractor
that is, is the maximal bounded invariant subset of under the NSE dynamics, or equivalently it is the minimal compact subset of which uniformly attracts all bounded sets in under the dynamics of (2.2). In particular, it is also known that
is the Grashof number, a dimensionless physical parameter, and . For the above properties see, for instance, [6], [10], [16], [22].
Next, we introduce a number of identities satisfied by the bilinear term. This includes the orthogonality relations
(where denotes the dual action between and ), and
(see, for instance, [6], [10], [22]). The relation (2.7) implies that (cf. [6], [22])
From now on, , , , , , , , , , will denote universal dimensionless positive constants. Our estimates for the non-linear term will involve Agmon's inequality
and the Sobolev and Ladyzhenskaya inequalities
which imply that
We also use the versions of the Poincaré inequality
and Young's inequality
By (2.13)
and by (2.10)
In addition,
(see [23]). Using the Brézis-Gallouet inequality [4] (see also a different proof in [23]), we have
We also use the following modified Gronwall inequality from [17] (see also [10]).
Lemma 2.1. Let and be locally integrable real-valued functions on which for some satisfy
and
where and . Suppose that is an absolutely continuous non-negative function on such that
Then as .
Lemma 2.1 will be combined later with the following estimates for averaged solutions (see [17], [18]).
Proposition 2.2. Let be a solution of the NSE (2.2) and let . Then
If , then
Moreover, it follows from the Cauchy–Schwarz inequality that
Proposition 2.3. Let be a solution of the NSE (2.2). Then
In particular,
Moreover, the solutions in the global attractor are analytic with respect to the time variable in a strip about the real axis with width . In addition, by the Cauchy formula the above estimates imply that
The ideas of the proof of the above proposition can be found in [6]. However, the new sharp estimates in Proposition 2.3 are obtained in [9].
We now derive two bounds for .
and for all and
By virtue of (2.11) we have
It follows from (2.26), (2.27), (2.10), and (2.12) that
and
Inspired by the proof of the Brézis–Gallouet inequality [4], we establish below a bound for the -norm which we later use to optimize an estimate.
Lemma 2.5. Let . Then for every
where .
and for ,
Using the Cauchy–Schwarz inequality and Parseval's identity, we have
3. Determining form and statements of main results
3.1. Interpolant operators.
In this subsection we introduce a unified approach for using various determining parameters (modes, nodes, volume elements, and so on) by representing them in terms of interpolant operators that approximate the identity.
Let be a finite-rank linear operator approximating the identity in the following sense: for every we have , has zero spatial average, and
Here is a small parameter that determines the order of approximation. The rank of is of order . For example, such interpolant polynomials are induced by the determining parameters of the NSE, such as determining modes, nodes, volume elements, projections of finite elements, and so on (see, for instance, [5], [11], [14], [15], [17], [18] and references therein). The most straightforward example of such interpolant operators is the projection operator onto , where . Also, the appendix of [1] provides explicit examples of such interpolant operators that are based on nodal values and that satisfy (3.1) and (3.2). We remark that a general framework employing interpolant polynomials satisfying (3.1) was introduced in [5] for investigating the long-time dynamics of the NSE.
3.2. Determining form.
In this subsection we present a determining form that is induced by the interpolant operators . The connection between the long-time dynamics of the NSE (2.2) and the determining form is explained in the following result.
Proposition 3.1. Let and let , , be a solution of the NSE (2.2) that lies in the global attractor . Suppose that with satisfies the equation
Then if
and is small enough that
and and , , , are as in (2.18), (2.23), and (3.1).
The proof of Proposition 3.1 is given in § 4. We remark that the existence of solutions of (3.3), as specified in Proposition 3.1, follows from Theorem 3.2 below.
Next, we introduce the phase space of the dynamics of our determining form. Let
with the two norms
Now we let be a given element of and consider the equation
We show in the following result that under certain conditions on the parameters and , which depend on , (3.6) has a unique bounded global solution for .
Theorem 3.2. Let , and let for some . Fix and so that
and
where and . Choose small enough so that
Then for every equation (3.6) has a unique solution that exists globally for all and has the following properties:
Moreover, suppose that and and are the corresponding solutions of (3.6). Let and . Then
where for some universal constant .
The proof of Theorem 3.2 will be presented in § 5.
The following corollary is an immediate consequence of Theorem 3.2.
Corollary 3.3. Assume the conditions of Theorem 3.2. Then there exists a Lipschitz continuous map
with the following properties:
- (i)
- (ii)for every
This map plays a crucial role in the definition of our determining form. To be more specific, let be a steady state of the NSE (2.2). Our determining form is the equation
The precise properties of (3.10) are stated in Theorem 3.5, below. But first we need the following result.
Proposition 3.4. Suppose that . Then for every
Consequently,
Proof. Since , we apply (3.2) and use the fact that by (3.9) we have , together with Proposition 2.3. □
Theorem 3.5. Let and suppose that the conditions of Theorem 3.2 hold for , where as in Proposition 3.4. Then the following hold.
- (i)
- (ii)
- (iii)
- (iv)
Proof. To show short-time existence it is sufficient to show that the vector field (3.10) is Lipschitz. Let , where . Since
it suffices to show that the map is Lipschitz. Note that for we have
Next, by (3.2) and the triangle inequality,
By virtue of Corollary 3.3 and the fact that ,
where satisfies (3.7) with . This completes the proof of
By Proposition 3.4, . Thus, we have short-time existence of a solution of (3.10) with initial data in . The proof of
where . This property implies that the ball is forward invariant for all , which simultaneously proves
To prove part
4. Proof of Proposition 3.1
In view of (2.2) and (3.3) the difference satisfies
Suppose that for some . Since is a continuous function with values in , there is a maximal interval containing such that for all . Taking the scalar product with and using (2.8), (2.9), (2.18), we find that for all
By (3.5) we have
that is,
where
We now seek a lower estimate of for . Note that
and that is decreasing for and increasing for . Thus,
Now note that for we have . Indeed, it is easy to check that satisfies , , and . We conclude that
Applying (4.2) and then (2.22) to (4.1), we find that since ,
It follows that
where and . If , then , so we take and conclude that for . Otherwise it follows from (2.22) that
for large enough . Letting , we find by (3.4) that . Since is arbitrary, in particular, a contradiction proving the proposition.
5. Proof of Theorem 3.2
In this section we give a formal proof of each estimate stated in Theorem 3.2. However, we describe below how to give a rigorous justification for the existence of a solution of (3.6) which satisfies, together with , these estimates. First, one considers the following Galerkin approximation system for (3.6):
where is the orthogonal projection from onto , the space of the first eigenfunctions of the Stokes operator .
Proposition 5.1. Equation (5.1) has a solution for all which satisfies, together with , all the estimates in Theorem 3.2.
Proof. For we supplement (5.1) with an initial value in order to obtain a finite system of ODEs with a quadratic polynomial non-linearity. Therefore, (5.1) with this initial data possesses a unique solution on a small time interval symmetric about the initial time . Furthermore, on this small interval is the unique solution of the Cauchy problem
For the interval one can follow the same steps as used below in establishing estimates of to show that the same estimates are also valid for when . Thus, remains bounded for , and as a result it solves (5.1) for . And since is finite and independent of , one can show, by following steps similar to those used below for estimating and , that is bounded uniformly independent of for all .
Now let . Then by employing the Arzelà–Ascoli compactness theorem one can extract a subsequence of which converges as to a solution of (5.1) on the interval . Moreover, and satisfy the estimates in Theorem 3.2 for all . Now by the Cantor diagonal process one can show that converges to as , and that has the properties stated in the proposition. □
We continue with our proof of the theorem. Based on Proposition 5.1, we use the Aubin compactness theorem (see, for instance, [6] or [22]) to show that for every there exists a subsequence of which converges to in the relevant spaces on the interval as . Moreover, by passing to the limit and following arguments similar to those for the 2D NSE, one infers that is a solution of (3.6) on the interval , and in addition, and satisfy the estimates in Theorem 3.2 on the interval . Now we again use the Cantor diagonal process to show that the diagonal subsequence converges as to a solution of (3.6). Moreover, and satisfy the estimates in Theorem 3.2 for all . This in turn concludes the formal justification of the estimates that will be established below.
5.1. Bound for .
Taking the inner product of (3.6) with and using (2.8), we have
From (3.1) we get that
and thus
Therefore, if we assume that is small enough that
then by Gronwall's inequality and the assumption that is bounded we have the estimate
Next, we consider the evolution equation for :
5.2. Bound for .
Taking the inner product of (5.6) with and using (2.9) and (3.1) (after following steps similar to those above), one obtains
Now by applying (2.28) to we get that
Hence by (5.4)
where
We get from (5.3) and (5.4) that
and in view of (5.5) integration gives us
Applying the Cauchy–Schwarz inequality, we have
and hence
We want to make sure that
If we choose , then (5.9) follows from requiring that
which automatically holds if
which is equivalent to
Since
it suffices that
Since , , it follows that , and we set and , so that if
then it suffices that
Thus, to ensure that (5.9) holds, we take
By a similar calculation, again taking , we get that for
Therefore, if
then for some we have
for some absolute constant . Ultimately, we will choose so that (5.12) is compatible with (5.11) (see (5.20)).
Lemma 5.2. Let be a constant and let be an absolutely continuous bounded function satisfying the inequality for all . Suppose also that for all . Then
Proof. Note that since , we have . Multiplying by the integrating factor and integrating from to , we find that
From Lemma 5.2, (5.9), and (5.13),
5.3. Bounds for and .
From (5.8) we get that
and hence by (5.14) and (5.5) we have
From (3.6) we find that
and thus by (2.24)
Since , we get from (3.2) that
Thus, by (5.12) and (5.15) we have
5.4. Bound for .
Taking the inner product of with the equation
we get that
We bound the first non-linear term using (2.15) and (2.25) to get that
For the second non-linear term we integrate by parts, using the fact that under periodic boundary conditions, so that by (2.7), (2.8), (2.13), (2.11), and (2.15) we have
Using (2.25), (3.2), and the bounds for and , we now obtain by Gronwall's inequality the uniform bound
provided that (which, in turn follows from (5.11)).
5.5. Lipschitz property of and in the norm.
In this section we show that the bounded solutions of (3.6) are unique and depend continuously on the input trajectory , in a sense that will be specified below. In particular, these properties are instrumental for introducing a well-defined map from the space to a space of trajectories, which is defined by for all .
To obtain these properties we consider the difference between two trajectories and and establish estimates for and similar to those for and . Indeed, by linearity the only complication is in the non-linear term. Let , and . Then
5.6. Bound for .
Taking the scalar product of (5.17) with , we find as in § 5.1 that
where for the non-linear terms we used (2.18) and (2.19). Applying Young's inequality to the contribution from the non-linear terms, we find that if (3.5) holds, then by (5.5)
At this point we can use (4.2) with and to obtain
Therefore, if
then
To ensure compatibility of (5.11) and (5.18) we take, as in (3.8),
where and .
5.7. Bound for .
The equation for is
Taking the scalar product with , we get that
Note that the difference from (5.7) (when one changes to and to ) is in the addition of the two terms and , so that if we can show that they are bounded by a constant multiple of , then we obtain the Lipschitz property of by using the same methods as in § 5.2.
We begin with
Consequently, we have
and (by (2.16))
along with (by (2.17))
Thus,
By (3.5) we can drop the term containing . Then applying (5.16) to and using (5.19), we have
where
and
Proceeding as in § 5.2, we get that
where for some universal constant . Note that once again (5.9) suffices to ensure that (5.12) holds, so there is no need to modify the range for in (5.12).