KdV equation under periodic boundary conditions and its perturbations
aa r X i v : . [ m a t h . D S ] O c t KdV equation under periodic boundary conditionsand its perturbations
HUANG Guan and Sergei KUKSIN C.M.L.S, ´Ecole Polytechnique, Palaiseau, France I.M.J, Universit´e Paris Diderot-Paris 7, Paris, France
Abstract.
In this paper we discuss properties of the KdV equation underperiodic boundary conditions, especially those which are important to studyperturbations of the equation. Next we review what is known now about long-timebehaviour of solutions for perturbed KdV equations.
Contents0 Introduction 21 KdV under periodic boundary conditions as a hamiltonian system 32 Finite dimensional integrable systems 4 H K ( I ) . . . . . . . . . . . . . . . . . . . . . 15 ǫ -quasi-invariant measures . . . . . . . . . . . . . . . . . . 254.6 Nekhoroshev type results (long-time stability)? . . . . . . . . . . . . . 27 ONTENTS
0. Introduction
The famous Korteweg-de Vries (KdV) equation u t = − u xxx + 6 uu x , x ∈ R , was first proposed by Joseph Boussinesq [15] as a model for shallow water wavepropagation. It became famous later when two Dutch mathematicians, DiederikKorteweg and Gustav De Vries [38], used it to explain the existence of a solitonwater wave, previously observed by John Russel in physical experiments. Their workwas so successful that this equation is now named after them. Since the mid-sixtiesof 20th century the KdV equation received a lot of attention from mathematical andphysical communities after the numerical results of Kruskal and Zabusky [40] led tothe discovery that its solitary wave solutions interact in an integrable way. It turnsout that in some suitable setting, the KdV equation can be viewed as an integrableinfinite dimensional hamiltonian system.In his “New Methods of Celestial Mechanics”, Poincar´e calls the task to studyperturbations of integrable systems the “General Problem of Dynamics”. The greatscientist was motivated by the celestial mechanics, where perturbed integrable systemsplay a very important role ‡ . For a similar reason his maxim is true for mathematicalphysics, where many important processes are described by suitable perturbations of anintegrable PDE, while the unperturbed integrable equations correspond to idealizationof physical reality. In particular, no physical process is exactly described by the KdVequation.In this paper § , we focus on the KdV equation with zero mean value periodicboundary condition. It is known since the works of Novikov, Lax, Marcˆenko, Its-Matveev and McKean-Trubowitz that this system is integrable ([57, 59, 52, 67]). Allof its solutions are periodic, quasi-periodic or almost periodic in time. In Section 1we discuss the KdV equation in the framework of infinite-dimensional hamiltoniansystems, in Section 2 we present some normal form results for finite-dimensionalintegrable hamiltonian systems and in Section 3 — recent results on KdV which maybe regarded as infinite-dimensional versions of those in Section 2. Finally in Section 4we discuss long-time behavior of solutions for the perturbed KdV equations, underhamiltonian and non-hamiltonian perturbations. Results presented there are heavilybased on theorems from Section 3.As indicated in the title of our work, we restrict our study to the periodicboundary conditions. In this case the KdV equation behaves as a hamiltonian systemwith countablely-many degrees of freedom and the method of Dynamical Systemsmay be used for its study (same is true for other hamiltonian PDEs in finite volume,e.g. see [45]). The KdV is a good example of an integrable PDE in the sense thatproperties of many other integrable equations with self-adjoint Lax operators, e.g. ofthe defocusing Zakharov-Shabat equation (see [25]), and of their perturbations arevery similar to those of KdV and its perturbations, while the equations with non-selfadjoint Lax operators, e.g. the Sine-Gordon equation, are similar to KdV whenwe study their small-amplitude solutions (and the KAM-theory for such equation issimilar to the KAM theory for KdV without the smallness assumption, see [44]). ‡ For example, the solar system, regarded as a system of 8 interacting planets rotating around theSun, is a small perturbation of the Kaplerian system. The latter is integrable § Based on the courses, given by the second author in Saint-Etienne de Tin´ee in February 2012 andin the High School for Economics (Moscow) in April 2013.
ONTENTS
1. KdV under periodic boundary conditions as a hamiltonian system
Consider the KdV equation under zero mean value periodic boundary condition: u t + u xxx − uu x = 0 , x ∈ T = R / Z , Z T udx = 0 . (1.1)(Note that the mean-value R T udx of a space-periodic solution u is a time-independentquantity, to simplify presentation we choose it to be zero.) To fix the setup, for anyinteger p >
0, we introduce the Sobolev space of real valued functions on T with zeromean-value: H p = n u ∈ L ( T , R ) : || u || p < + ∞ , Z T = 0 o , || u || p = X k ∈ N | πk | p ( | ˆ u k | + | ˆ u − k | ) . Here ˆ u k , ˆ u − k , k ∈ N , are the Fourier coefficients of u with respect to the trigonometricbase e k = √ πkx, k > e k = √ πkx, k < , (1.2)i.e. u = X k ∈ N ˆ u k e k + ˆ u − k e − k . (1.3)In particular, H is the space of L -functions on T with zero mean-value. By h· , ·i wedenote the scalar product in H (i.e. the L -scalar product).For a C -smooth functional F on some space H p , we denote by ∇ F its gradientwith respect to h· , ·i , i.e. dF ( u )( v ) = h∇ F ( u ) , v i , if u and v are sufficiently smooth. So ∇ F ( u ) = δFδu ( x ) + const , where δFδu is thevariational derivative, and the constant is chosen in such a way that the mean-valueof the r.h.s vanishes. See [44, 31] for details. The initial value problem for KdV on thecircle T is well posed on every Sobolev space H p with p >
1, see [70]. The regularityof KdV in function spaces of lower smoothness was studied intensively, see [16, 34]and references in these works; also see [16] for some qualitative results concerning theKdV flow in these spaces. We avoid this topic.It was observed by Gardner [23] that if we introduce the Poisson bracket whichassigns to any two functionals F ( u ) and G ( u ) the new functional { F, G } , n F, G o ( u ) = Z T ddx ∇ F ( u ( x )) ∇ G ( u ( x )) dx (1.4)(we assume that the r.h.s is well defined, see [44, 45, 31] for details), then KdV becomesa hamiltonian PDE. Indeed, this bracket corresponds to a differentiable hamiltonianfunction F a vector filed V F , such that hV F ( u ) , ∇ G ( u ) i = { F, G } ( u ) ONTENTS G . From this relation we see that V F ( u ) = ∂∂x ∇ F ( u ) . So the KdV equation takes the hamiltonian form u t = ∂∂x ∇H ( u ) , (1.5)with the KdV Hamiltonian H ( u ) = Z T ( u x u ) dx. (1.6)The Gardner bracket (1.4) corresponds to the symplectic structure, defined in H (aswell as in any space H p , p >
0) by the 2-form ω G ( ξ, η ) = (cid:10) ( − ∂∂x ) − ξ, η (cid:11) for ξ, η ∈ H . (1.7)Indeed, since ω G ( V F ( u ) , ξ ) ≡ −h∇ F ( u ) , ξ i , then the 2-form ω G also assigns to aHamiltonian F the vector field V F (see [4, 31, 44, 45]).We note that the bracket (1.4) is well defined on the whole Sobolev spaces H p ( T ) = H p ⊕ R , while the symplectic form ω G is not, and the affine subspaces { u ∈ H p ( T ) : R T udx = const } ≃ H p are symplectic leaves for this Poisson system.We study the equation only on the leaf R T udx = 0, but on other leaves it may bestudied similarly.Writing a function u ( x ) ∈ H as in (1.3) we see that ω G = P ∞ k =1 k − d ˆ u k ∧ ˆ u − k and that H ( u ) = H (ˆ u ) := Λ(ˆ u ) + G (ˆ u ) withΛ(ˆ u ) = + ∞ X k =1 (2 πk ) (cid:0)
12 ˆ u k + 12 ˆ u − k (cid:1) , G (ˆ u ) = X k,l,m =0 ,k + l + m =0 ˆ u k ˆ u l ˆ u m . Accordingly, the KdV equation may be written as the infinite chain of hamiltonianequations ddt ˆ u j = − πj ∂H (ˆ u ) ∂ ˆ u − j , j = ± , ± , . . . .
2. Finite dimensional integrable systems
Classically, integrable systems are particular hamiltonian systems that can beintegrated in quadratures. It was observed by Liouville that for a hamiltonian systemwith n degrees of freedom to be integrable, it has to possess n independent integralsin involution. This assertion can be understood globally and locally. Now we recallcorresponding finite-dimensional definitions and results. Let Q ⊂ R n ( p,q ) be a 2 n -dimensional domain. We provide it with the standardsymplectic form ω = dp ∧ dq and the corresponding Poisson bracket { f, g } = ∇ p f · ∇ q g − ∇ q f · ∇ p g, where g, f ∈ C ( Q ) and “ · ” stands for the Euclidean scalar product in R n (see [4]).If { f, g } = 0, the functions f and g are called commuting , or in involution . If h ( p, q )is a C -function on Q , then the hamiltonian system with the Hamiltonian h is˙ p = −∇ q h, ˙ q = ∇ p h. (2.1) ONTENTS Definition 2.1 (Liouville-integrability). The hamiltonian system (2.1) is calledintegrable in the sense of Liouville if its Hamiltonian h admits n independent integralsin involution h , . . . , h n . That is, { h, h i } = 0 for i n ; { h i , h j } = 0 for i, j n , and dh ∧ · · · ∧ dh n = 0 . A nice structure of an Liouville-integrable system is assured by the celebratedLiouville-Arnold-Jost theorem (see [4, 62]) It claims that if an integrable systemsis such that the level sets T c = { ( p, q ) ∈ Q : h ( p, q ) = c , . . . , h n ( p, q ) = c n } , c = ( c , . . . , c n ) ∈ R n are compact, then each non-empty set T c is an embedded n -dimensional torus. Moreover for a suitable neighborhood O T c of T c in Q there existsa symplectomorphismΘ : O T c → O × T n = { ( I, ϕ ) } , O ⊂ R n , where the symplectic structure in O × T n is given by the 2-form dI ∧ dϕ . Finally,there exists a function ¯ h ( I ) such that h ( p, q ) = ¯ h (Θ( p, q )) . This result is true both inthe smooth and analytic categories.The coordinates (
I, ϕ ) are called the action-angle variables for (2.1). Using themthe hamiltonian system may be written as˙ I = 0 , ˙ ϕ = ∇ I ¯ h ( I ) . (2.2)Accordingly, in the original coordinates ( p, q ) solutions of the system are( p, q )( t ) = Θ − ( I , ϕ + ∇ I ¯ h ( I ) t ) . On O × T n , consider the 1-form Idϕ = P nj =1 I j dϕ j , then d ( Idϕ ) = dI ∧ dϕ .For any vector I ∈ O , and for j = 1 , . . . , n , denote by C j ( I ) the cycle { ( I, ϕ ) ∈ O × T n : ϕ j ∈ [0 , π ] and ϕ i = const, if i = j } . Then12 π Z C j Idϕ = 12 π Z C j I j dϕ j = I j . Consider a disc D j ⊂ O × T n such that ∂D j = C j . For any 1-form ω , satisfying dω = dI ∧ dϕ , we have12 π Z C j ( Idϕ − ω ) = 12 π Z D j d ( Idϕ − ω ) = 0 . So I j = 12 π Z C j ( I ) ω , if dω = dI ∧ dϕ. (2.3)This is the Arnold formula for actions . We denote by J the standard symplectic matrix J = diag n (cid:18) −
11 0 (cid:19) o , operating inany R n (e.g. in R ). Assume that the origin is an elliptic critical point of a smoothHamiltonian h , i.e. ∇ h (0) = 0 and that the matrix J ∇ h (0) has only pure imagineeigenvalues. Then there exists a linear symplectic change of coordinates which puts h to the form h = n X i =1 λ i ( p i + q i ) + h.o.t, λ j ∈ R ∀ j. If the frequencies ( λ , . . . , λ n ) satisfy some non-resonance conditions, then thisnormalization process can be carried out to higher order terms. The result of thisnormalization is known as the Birkhoff normal form for the Hamiltonian h .ONTENTS Definition 2.2
The frequencies λ , . . . , λ n are non-resonant up to order m > if P ni =1 k i λ i = 0 for each k ∈ Z n such that P ni =1 | k i | m . They are called non-resonant if k λ + . . . + k n λ n = 0 with integers k , . . . , k n only when all k j ’s vanish. Theorem 2.3 (Birkhoff normal form, see [61, 62]) Let H = N + · · · be a real analyticHamiltonian in the vicinity of the origin in ( R n ( p,q ) , dp ∧ dq ) with the quadratic part N = P ni =1 λ i ( q i + p i ) . If the (real) frequencies λ , . . . , λ n are non-resonant up toorder m > , then there exists a real analytic symplectic trasformation Ψ m = Id + · · · ,such that H ◦ Ψ m = N + N + . . . + N m + h.o.t. Here N i are homogeneous polynomials of order i , which are actually smooth functionsof variables p + q , . . . , p n + q n . If the frequencies are non-resonant, then there existsa formal symplectic transformation Ψ = Id + · · · , represented by a formal power series,such that H ◦ Ψ = N + N + · · · (this equality holds in the sense of formal series). If the transformation, converting H to the Birkhoff normal form, was convergent,then the resulting Hamiltonian would be integrable in a neighborhood of the originwith the integrals p + q , . . . , p n + q n . These functions are not independent when p i = q i = 0 for some i . So the system is not integrable in the sense of Liouville. Butit is integrable in a weaker sense: Definition 2.4
Functions f , . . . , f k are functionally independent if their differentials df , . . . , df k are linearly independent on a dense open set. A n -dimensionalHamiltonian is called Birkhoff integrable near an equilibrium m ∈ R n , if it admits n functionally independent integrals in involution in the vicinity of m . Birkhoff normal form provides a powerful tool to study the dynamics ofhamiltonian PDEs, e.g. see [49, 6] and references in [6] . However, in this paperwe shall not discuss its version for KdV, since for that equation there exists a strongernormal form. Now we pass to its counterpart in finite dimension.
The results of this section hold both in the C ∞ -smooth and analytic categories. Definition 2.5
Consider a Birkhoff integrable system, defined near an equilib-rium m ∈ R n , with independent commuting integrals F = ( F , . . . , F n ) . Its Poissonalgebra is the linear space A ( F ) = n G : { G, F i } = 0 , i = 1 , . . . , n o . Note that although the integrals of an integrable system are not defined in a uniqueway, the corresponding algebra A ( F ) is. Definition 2.6
A Poisson algebra A ( F ) is said to be non-resonant at a point m ∈ R n , if it contains a Hamiltonian with a non-resonant elliptic critical point at m (i.e.,around m one can introduce symplectic coordinates ( p, q ) such that the quadratic part ofthat Hamiltonian at m is P λ j ( p j + q j ) , where the real numbers λ j are non-resonant). It is easy to verify that if some F ∈ A ( F ) is elliptic and non-resonant at theequilibrium m , then all other functions in A ( F ) are elliptic at m as well. ONTENTS Theorem 2.7 (Vey’s theorem). Let F = ( F , . . . , F n ) be n functionally independentfunctions in involution in a neighbourhood of a point m ∈ R n . If the Poissonalgebra A ( F ) is non-resonant at m , then one can introduce around m symplecticcoordinates ( p, q ) so that A ( F ) consists of all functions, which are actually functionsof p + q , . . . , p n + q n . Example.
Let F = ( f , . . . , f n ) be a system of smooth commuting Hamiltonians,defined in the vicinity of their joint equilibrium m ∈ R n , such that the hessians ∇ f i ( m ), 1 i n , are linear independent. Then the theorem above applies to thePoisson algebra A ( F ).In [71] Vey proved the theorem in the analytic case with an additional non-degeneracy condition, which was later removed by Ito in [28]. The results in [71, 28]also apply to non-elliptic cases. The smooth version of Theorem 2.7 is due to Eliasson[21]. There exists an infinite dimensional extension of the theorem, see [47].
3. Integrability of KdV
The KdV equation (1.1) admits infinitely many integrals in involution, and there aredifferent ways to obtain them, see [23, 60, 63, 52, 73]. Below we present an elegant wayto construct a set of Poisson commuting integrals by considering the spectrum of anassociated Schr¨odinger operator, due to Piter Lax [52] (see [53] for a nice presentationof the theory).
Let u ( x ) be a L -function on T . Consider the differential operators L u and B u , actingon 2-periodic functions k L u = − d dx + u ( x ) , B u = − d dx + 3 u ( x ) ddx + 3 ddx u ( x ) , where we view u ( x ) as a multiplication operator f u ( x ) f . The operators B u and L u are called the Lax pair for KdV. Calculating the commutator [ B u , L u ] = B u L u − L u B u ,we see that most of the terms cancel and the only term left is − u xxx +6 uu x . Thereforeif u ( t, x ) is a solution of (1.1), then the operators L ( t ) = L u ( t, · ) and B ( t ) = B u ( t, · ) satisfy the operator equation ddt L ( t ) = [ B ( t ) , L ( t )] . (3.1)Note that the operator B ( t ) are skew-symmetric, B ( t ) ∗ = − B ( t ). Let U ( t ) be theone-parameter family of unitary operators, defined by the differential equation ddt U = B ( t ) U, U (0) = Id . Then L ( t ) = U − ( t ) L (0) U ( t ). Therefore, the operator L ( t ) is unitary conjugated to L (0). Consequently, its spectrum is independent of t . That is, the spectral data of theoperator L u provide a set of conserved quantities for the KdV equation (1.1). Since L u is the strurm-Liouville operator with a potential u ( x ), then in the context of thistheory functions u ( x ) are called potentials . k note the doubling of the period. ONTENTS L -potential u the spectrum of the Sturm-Liouvilleoperator L u , regarded as an unbounded operator in L ( R / Z ), is a sequence of simpleor double eigenvalues { λ j : j > } , tending to infinity:spec( u ) = { λ < λ λ < · · · ր ∞} . Equality or inequality may occur in every place with a ” ” sign (see [57, 31]). Thesegment [ λ j − , λ j ] is called the n -th spectral gap . The asymptotic behaviour of theperiodic eigenvalues is λ n − ( u ) , λ n ( u ) = n π + [ u ] + l ( n ) , where [ u ] is the mean value of u , and l ( n ) is the n -th number of an l sequence.Let g n ( u ) = λ n ( u ) − λ n − ( u ) > n >
1. These quantities are conserved underthe flow of KdV. We call g n the n -th gap-length of the spectrum. The n -th gap iscalled open if g n >
0, otherwise it is closed . However, from the analytic point of viewthe periodic eigenvalues and the gap-lengths are not satisfactory integrals, since λ n is not a smooth function of the potential u when g n = 0. Fortunately, the squaredgap lengths g n ( u ), n >
1, are real analytic functions on L , which Poisson commutewith each other (see [58, 53, 31]). Moreover, together with the mean value, the gaplengths determine uniquely the periodic spectrum of a potential, and their asymptoticbehavior characterizes the regularity of a potential in exactly the same way as itsFourier coefficients [57, 24].This method applies to integrate other hamiltonian systems in finite or infinitedimension. It is remarkably general and is referred to as the method of Lax pair.3.2. Action-angle coordinates We denote by Iso( u ) the isospectral set of a potential u ∈ H :Iso( u ) = n u ∈ H : spec( u ) = spec( u ) o . It is invariant under the flow of KdV and may be characterized by the gap lengthsIso( u ) = n u ∈ H : g n ( u ) = g n ( u ) , n > o . Moreover, for any n > u ∈ H n if and only if Iso( u ) ⊂ H n .In [58], McKean and Trobwitz showed that the Iso( u ) is homemorphic to acompact torus, whose dimension equals the number of open gaps. So the phase space H is foliated by a collection of KdV-invariant tori of different dimensions, finite orinfinite. A potential u ∈ H is called finite-gap if only a finite number of its spectralgaps are open. The finite-dimensional KdV-invariant torus Iso( u ) is called a finite-gaptorus . For any n ∈ N let us set J n = n u ∈ H : g j ( u ) = 0 if j > n o . (3.2)We call the sets J n , n ∈ N , the finite-gap manifolds. Theorem 3.1
For any n ∈ N , the finite gap manifold ( J n , ω G ) is a smooth symplectic n -manifold, invariant under the flow of KdV (1.1), and T J n = n u ∈ H : ˆ u k = 0 if | k | > n + 1 o , (see (1.3)). Moreover, the square gap lengths g k ( u ) , k = 1 , . . . , n , form n commutinganalytic integrals of motions, non-degenerated everywhere on the dense domain J n = { u ∈ J n : g ( u ) , . . . , g n ( u ) > } .ONTENTS J n , n ∈ N .Furthermore, the union of the finite gap manifolds ∪ n ∈ N J n is dense in each space H s (see [57]). This hints that on the spaces H s , s >
0, it may be possible to constructglobal action-angle coordinates for KdV. In [22], Flaschka and McLaughlin used theArnold formula (2.3) to get an explicit formula for action variables of KdV in terms ofthe 2-period spectral data of L u . To explain their construction, denote by y ( x, λ, u )and y ( x, λ, u ) the standard fundamental solutions of the equation − y ′′ + uy = λy ,defined by the initial conditions y (0 , λ, u ) = 1 , y (0 , λ, u ) = 0 ,y ′ (0 , λ, u ) = 0 , y ′ (0 , λ, u ) = 1 . The quantity △ ( λ, u ) = y (1 , λ, u ) + y ′ (1 , λ, u ) is called the discriminant , associatedwith this pair of solutions. The periodic spectrum of u is precisely the zero set of theentire function △ ( λ, u ) −
4, for which we have the explicit representation (see e.g.[73, 58]) △ ( λ, u ) − λ − λ ) Y n > ( λ n − λ )( λ n − − λ ) n π . This function is a spectral invariant. We also need the spectrum of the differentialoperator L u = − d dx + u under Dirichlet boundary conditions on the interval [0 , µ ( u ) < µ ( u ) < . . . ր ∞ , which satisfy λ n − ( u ) µ n ( u ) λ n ( u ), for all n ∈ N . Thus, the n -th Dirichleteigenvalue µ n is always contained in the n -th spectral gap. The Dirichlet spectrumprovides coordinates on the isospectral sets (see [58, 57, 31]). For any z ∈ T , denote by { µ j ( u, z ) , j > } the spectrum of the operator L u under the shifted Dirichlet boundaryconditions y ( z ) = y ( z + 1) = 0 (so µ j ( u,
0) = µ j ( u )); still λ n − µ n ( u, z ) λ n ( u ).Jointly with the spectrum { λ j } , it defines the potential u ( x ) via the remarkable traceformula (see [73, 20, 31, 58]): u ( z ) = λ ( u ) + ∞ X j =1 ( λ j − ( u ) + λ j ( u ) − µ j ( u, z )) . Define f n ( u ) = 2 log( − n y ′ (1 , µ n ( u ) , u ) , ∀ n ∈ N . Flashka and McLaughlin [22] observed that the quantities { µ n , f n } n ∈ N form canonicalcoordinates of H , i.e. { µ n , µ m } = { f n , f m } = 0 , { µ n , f m } = δ n,m , ∀ n, m ∈ N . Accordingly, the symplectic form ω G (see (1.7)) equals dω , where ω is the 1-form P n ∈ N f n dµ n . Now the KdV action variables are given by the Arnold formula (2.3),where C n is a circle on the invariant torus Iso( u ), corresponding to µ n ( u ). It is shownin [22] that I n = 2 π Z λ n λ n − λ ˙ △ ( λ ) p △ ( λ ) − dλ, ∀ n ∈ N . The analytic properties of the functions u I n and of the mapping u I =( I , I , . . . ) were studied later by Kappeler and Korotyaev (see references in [31, 36] ONTENTS I n ( u ), n ∈ N , are real analytic functionson H of the form I n = g n + higher order terms , and I n = 0 if and only if g n = 0, seein [31]. For any vector I = ( I , I , . . . ) with non-negative components we will denote T I = { u ( x ) ∈ H : I n ( u ) = I n ∀ n } . (3.3)The angle-variables ϕ n on the finite-gap manifolds J n were found in 1970’s bySoviet mathematicians, who constructed them from the Dirichlet spectrum { µ j ( u ) } by means of the Abel transform, associated with the Riemann surface of the function p △ −
4, see [20, 57, 73], and see [29, 19, 39, 7] for the celebrated explicit formulas forangle-variables ϕ n and for finite-gap solutions of KdV in terms of the theta-functions.In [42] and [44], Section 7, the action-angle variables ( I n , ϕ n ) on a finite-gapmanifold J n and the explicit formulas for solutions of KdV on manifolds J N , N ≥ n ,from the works [19, 39, 7] were used to obtain an analytic symplectic coordinate system( I n , ϕ n , y ) in the vicinity of J n in H p . The variable y belongs to a ball in a subspace Y ⊂ H p of co-dimention 2 n , and in the new coordinates the KdV Hamiltonian (1.6)reads H = const + h n ( I n ) + h A ( I n ) y, y i + O ( y ) . (3.4)The selfadjoint operator A ( I n ) is diagonal in some fixed symplectic basis of Y . Thenonlinearity O ( y ) defines a hamiltonian operator of order one. That is, the KdV’slinear operator, which is an operator of order three, mostly transforms to the linearpart of the new hamiltonian operator and “does not spread much” to its nonlinearpart. This is the crucial property of (3.4). The normal form (3.4) is instrumental forthe purposes of the KAM-theory, see below Section 4.1.McKean and Trubowitz in [58, 59] extended the construction of angles onfinite-gap manifolds to the set of all potentials, thus obtaining angle variables ϕ =( ϕ , ϕ , . . . ) on the whole space H p , p >
0. The angles ( ϕ k ( u ) , k ≥
1) are well definedGato-analytic functions of u outside the locus a = { u ( x ) : g j ( u ) = 0 for some j } , (3.5)which is dense in each space H p . The action-map u I was not considered in[58, 59], but it may be shown that outside a , in a certain weak sense, the variables( I, ϕ ) are KdV’s action-angles (see the next section for a stronger statement). Thisresult is nice and elegant, but it is insufficient to study perturbations of KdV sincethe transformation to the variables (
I, ϕ ) is singular at the dense locus a . In a number of publications (see in [31]), Kappeler with collaborators proved thatthe Birkhoff coordinates v = { v n , n = ± , ± , . . . } , associated with the action-anglesvariables ( I, ϕ ), v n = p I n cos( ϕ n ) , v − n = p I n sin( ϕ n ) , ∀ n ∈ N , (3.6)are analytic on the whole of H and define there a global coordinate system, in whichthe KdV Hamiltonian (1.6) is a function of the actions only. This remarkable resultsignificantly specifies the normal form (3.4). To state it exactly, for any p ∈ R , weintroduce the Hilbert space h p , h p := n v = ( v , v , · · · ) : | v | p = + ∞ X j =1 (2 πj ) p +1 | v j | < ∞ , v j = ( v j , v − j ) t ∈ R , j ∈ N o , ONTENTS l -space h pI , h pI := n I = ( I , . . . ) ∈ R ∞ : | I | ∼ p = 2 + ∞ X j =1 (2 πj ) p +1 | I j | < + ∞ o . Define the mappings π I : h p → h pI , v I = ( I , I , . . . ) , where I k = 12 | v k | ∀ k,π ϕ : h p → T ∞ , v ϕ = ( ϕ , ϕ , . . . ) , where ϕ k = arctan( v − k v k )if v k = 0 , and ϕ k = 0 if v k = 0 . Since | π I ( v ) | ∼ p = | v | p , then π I is continuous. Its image h pI + = π I ( h p ) is the positiveoctant in h pI . When there is no ambiguity, we write I ( v ) = π I ( v ).Consider the mappingΨ : u ( x ) v = ( v , v , . . . ) , v n = ( v n , v − n ) t ∈ R , where v ± n are defined by (3.6) and { I n ( u ) } , { ϕ n ( u ) } are the actions and angles as inSection 3.2. Clearly π I ◦ Ψ( u ) = I ( u ) and π ϕ ◦ Ψ( u ) = ϕ ( u ). Below we refer to Ψ asto the nonlinear Fourier transform . Theorem 3.2 (see [31, 30]) The mapping Ψ defines an analytical symplectomorphism Ψ : ( H , ω G ) → ( h , P ∞ k =1 dv k ∧ dv − k ) with the following properties:(i) For any p ∈ [ − , + ∞ ) , it defines an analytic diffeomorphism Ψ : H p h p .(ii) (Percival’s identity) If v = Ψ( u ) , then | v | = || u || .(iii) (Normalisation) The differential d Ψ(0) is the operator P u s e s v , where v s = | πs | − / u s for each s .(iv) The function ˆ H ( v ) = H (Ψ − ( v )) has the form ˆ H ( v ) = H K ( I ( v )) , where thefunction H K ( I ) is analytic in a suitable neighborhood of the octant h I + in h I , suchthat a curve u ∈ C (0 , T ; H ) is a solution of KdV if and only if v ( t ) = Ψ( u ( t )) satisfies the equations ˙ v j = J ∂H K ∂I j ( I ) v j , v j = ( v j , v − j ) t ∈ R , j ∈ N . (3.7)The assertion (iii) normalizes Ψ in the following sense. For any θ = ( θ , θ , . . . ) ∈ T ∞ denote by Φ θ the operatorΦ θ v = v ′ , v ′ j = ¯Φ θ j v j , ∀ j ∈ N , (3.8)where ¯Φ α is the rotation of the plane R by the angle α . Then Φ θ ◦ Ψ satisfies allassertions of the theorem except (iii). But the properties (i)-(iv) jointly determine Ψin a unique way.The theorem above can be viewed as a global infinite dimensional version ofthe Vey Theorem 2.7 for KdV, and eq. (3.7) – as a global Birkhoff normal form forKdV. Note that in finite dimension a global Birkhoff normal form exists only for veryexceptional integrable equations, which were found during the boom of activity inintegrable systems, provoked by the discovery of the method of Lax pair.
ONTENTS Remark 3.3
The map Ψ simultaneously transforms all Hamiltonians of the KdVhierarchy to the Birkhoff normal form. The KdV hierarchy is a collection ofhamiltonian functions J l , l > , commuting with the KdV Hamiltonian, and havingthe form J l ( u ) = Z (cid:16)
12 ( u ( l ) ) + J l − ( u ) (cid:17) dx. Here J − = 0 and J l − ( u ) , l ≥ , is a polynomial of u, . . . , u ( l − . The functions fromthe KdV hierarchy form another complete set of KdV integrals. E.g. see [20, 31, 53]. Properties of the nonlinear Fourier transform Ψ may be specified in two importantrespects. One of this specifications – the quasilinearity of Ψ – is presented in thetheorem below. Another one – its behaviour at infinity – is discussed in the nextsection.The nonlinear Fourier transform Ψ is quasi-linear. Precisely,
Theorem 3.4 If m > , then the map Ψ − d Ψ(0) : H m → h m +1 is analytic. That is, the non-linear part of Ψ is 1-smoother than its linearisation at the origin.See [47] for a local version of this theorem, applicable as well to other integrableinfinite-dimensional systems, and see [32, 33] for the global result. We note that thetransformation to the normal form (3.4) also is known to be quasi-linear, see [42, 44].
Problem 3.5
Does the mapping Ψ − d Ψ(0) analytically maps H m to h m +1+ γ with γ > ? It is proved in [47] that for γ > γ > Ψ near infinity and large solutions of KdV By the assertion (ii) of Theorem 3.2, | Ψ( u ) | = || u || . It was established by Korotyayevin [36] that higher order norms of u and v = Ψ( u ) are related by both-sides polynomialestimates: Theorem 3.6
For any m ∈ N , there are polynomials P m ( y ) and Q m ( y ) such that if u ∈ H m and v = Ψ( u ) , then | v | m P m ( || u || m ) , || u || m Q m ( | v | m ) . The polynomials P m and Q m are constructed in [36] inductively. From a personalcommunication of Korotyayev we know that one can take P m ( y ) = C m y (1 + y ) m +2)3 . (3.9)Estimating a potential u ( x ) via its actions ¶ is more complicated. Correspondingpolynomials Q m may be chosen of the form Q m ( y ) = C ′ m y (1 + y ) a m , (3.10)where a = , a = 3 and a m has a factorial growth as m → ∞ . ¶ Note that any | v | m is a linear combinations of the actions I j , j > ONTENTS Problem 3.7
Prove that there exist polynomials P m and Q m , m ∈ N , such that forany u ∈ H m , we have || d Ψ( u ) || m,m P m ( || u || m ) , || d Ψ − ( v ) || m,m Q m ( || v || m ) , where v = Ψ( u ) . Prove similar polynomial bounds for the norms of higher differentialsof Ψ and Ψ − . It seems that to solve the problem a new proof of Theorem 3.2 has to be found (notethat the existing proof is rather bulky and occupies half of the book [31]).The difficulty in resolving the problem above streams from the fact that Ψ( u ) isconstructed in terms of spectral characteristics of the Strum-Liouville operator L u ,and their dependence on large potentials u ( x ) is poorly understood. Accordingly, thefollowing question seems to be very complicated: Problem 3.8
Let u ( t, x ) = u ( t, x ; λ ) be a solution of (1.1) such that u (0 , x ) = λu ( x ) ,where u is a given smooth function with zero mean-value, and λ > is a largeparameter. The task is to study behaviour of u ( t, x ; λ ) when λ → ∞ . Let u ( t, x ) be as above. Then || u (0 , x ) || m = λC ( m, u ). By Theorem 3.6, | v (0) | m P m ( λC ( m, u )). Since | v ( t ) | m is an integral of motion, then using again the theoremwe get that || u ( t ) || m Q m (cid:16) P m ( λC ( m, u )) (cid:17) . In particular, by (3.9) and (3.10) we have || u ( t ) || C (1 + λ ) / . A lower bound forthe Sobolev norms comes from the fact that || u ( t ) || is an integral of motion. So || u ( t ) || m > || u (0) || = λC (0 , u ) . We have demonstrated:
Proposition 3.9
Let u ( t, x ) be a solution of (1.1) such that u (0 , x ) = λu ( x ) , where u ∈ C ∞ ∩ H and λ > . Then for m ≥ we have λC ( m, u )) A m > lim sup t →∞ || u ( t ) || m > lim inf t →∞ || u ( t ) || m > c ( u ) λ, (3.11)for a suitable A m >
1. E.g. A = 21 / t →∞ || u ( t ) || m λC ( m, u ) , since the curve u ( t ) is almost periodic. The first estimate with the exponent A m which follows from Theorem 3.6 certainly is not optimal. But the assertion thatlim sup t →∞ || u ( t ) || m grows with λ as λ A m , where A m goes to infinity with m , is correct.It follows from our next result: Theorem 3.10
Let k > . Then there exists α > and, for any λ > there exists t ∗ = t ∗ ( u , λ ) such that || u ( t ∗ ) || k > c ′ u λ αk . (3.12) ONTENTS u = λw and pass to fast time τ = λt . Then the function w ( τ, x ; λ ) satisfies w τ + λ − w xxx − ww x = 0 , w (0 , x ) = u ( x ) , (3.13)and we are interested in w ( λt, x ; λ ) when λ → ∞ . For λ = ∞ the equation abovebecomes the Hopf equation. Since u ( x ) is a periodic non-constant function, then thesolution of (3.13) λ = ∞ developes a shock at time τ ∗ , 0 < τ ∗ < ∞ . Accordingly, theelementary perturbation theory allows to study solutions of (3.13) when λ → ∞ for τ < τ ∗ , but not for τ > τ ∗ . The problem to study this limit for τ > τ ∗ is addressedby the Lax-Levermore theory (mostly for the case when x ∈ R and u ( x ) vanishes atinfinity). There is vast literature on this subject, e.g. see [54, 18] and references in[18]. The existing Lax-Levermore theory does not allow to study solutions w ( τ, x ) for τ ∼ λ − , as is required by Problem 3.8. Let us denote W ( I ) = ( W ( I ) , W ( I ) , . . . ) , W i ( I ) = ∂H K ∂I i , i ∈ N . (3.14)This is the frequency map for KdV. By Theorem 3.2 each its component is an analyticfunction, defined in the vicinity of h I + in h I . Lemma 3.11 a) For i, j ≥ we have ∂ W (0) /∂I i ∂I j = − δ i,j .b) For any n ∈ N , if I n +1 = I n +2 = . . . = 0 , then det (cid:16)(cid:0) ∂W i ∂I j (cid:1) i,j n (cid:17) = 0 . For a) see [9, 31, 44]. For a proof of b) and references to the original works ofKrichever and Bikbaev-Kuksin see Section 3.3 of [44].Let l ∞ i , i ∈ Z , be the Banach spaces of all real sequences l = ( l , l , . . . ) withnorms | l | ∞ i = sup n > n i | l n | < ∞ . Denote κ = ( κ n ) n ∈ N , where κ n = (2 πn ) . For the following result see [31],Theorem 15.4. Lemma 3.12
The normalized frequency map I W ( I ) − κ is real analytic as amapping from h to l ∞− . From these two lemmata we known that the Hamiltonian H K ( I ) of KdV is non-degenerated in the sense of Kolmogorov and its nonlinear part is more regular thanits linear part. These properties are very important to study perturbations of KdV. ONTENTS H K ( I )By Theorem 3.2, the dynamics of KdV is determined by the Hamiltonian H K ( I ).To understand the properties of the latter is an important step toward the study ofperturbations of KdV.Denote by P j the moments of the actions, given by P j = X i > (2 πn ) j I i , j ∈ Z . Due to Theorem 3.2, the linear part of H K ( I ) at the origin dH K (0)( I ) equals to R T ( ∂∂x (Ψ − v )) dx = P . So we can write H K ( I ) as H K ( I ) = P ( I ) − V ( I ) , V ( I ) = O ( | I | ) . (The minus-sign here is convenient since, as we will see, V ( I ) is non-negative.) Forany N >
1, denote ˜ l N ⊂ l the N -dimensional subspace˜ l N = { l = ( l , . . . ) ∈ R ∞ : l n = 0 , ∀ n > N } , and set ˜ l ∞ = ∪ N ∈ N ˜ l N . Clearly the function V is analytic on each octant ˜ l N + .So it is Gato-analytic on the octant ˜ l ∞ + . That is, it is analytic on every interval { ( ta + (1 − t ) c ) ∈ ˜ l ∞ + : 0 ≤ t ≤ } , where a, c ∈ ˜ l ∞ + .By Lemma 3.11 a), d V (0)( I ) = 6 || I || . This suggests that the Hilbert space l rather than the Banach space h I (which is contained in l ) is a distinguished phasespace for the Hamiltonian H K ( I ). This guess is justified by the following result: Theorem 3.13 (see [37]). (i) The function V : ˜ l ∞ + → R extends to a non-negativecontinuous function on the l -octant l , such that V ( I ) = 0 for some I ∈ l iff I = 0 .Moreover V ( I ) P P − . (ii) For any I ∈ l , the following estimates hold true: π || I || P / − V ( I ) (8 (1 + P / − ) / P − + 6 πe P / − / || I || ) || I || . (iii) The function V ( I ) is convex on l . Note that the assertion (iii) follows from (i) and Lemma 3.11. Indeed, since V ( I )is analytic on ˜ l N + , then Lemma 3.11 assures that the Hessian { ∂ V∂I i ∂I j } i,j N is positivedefinite on ˜ l N + . Thus V is convex on ˜ l N + , for each N ∈ N . Then the assertion (iii) isdeduced from the fact the ˜ l ∞ + = ∪ N ∈ N ˜ l N + is dense in l , where V ( I ) is continuous.Assertion ii) of the theorem shows that l is the biggest Banach space on which V ( I ) is continuous + . Jointly with the convexity of V ( I ) on l , this hints that l is thenatural space to study the long-time dynamics of actions I ( u ( t )) for solutions u ( t ) ofa perturbed KdV.This theorem and the analyticity of the KdV Hamiltonian (1.6) do not leavemuch doubts that V ( I ) is analytic on l . If so, then by Lemma 3.11, this function isstrictly convex in a neighbourhood of the origin in l . Most likely, it is strictly convexeverywhere on l .In difference with V ( I ), the total Hamiltonian H K ( I ) is not continuous on l sinceits linear part P ( I ) is there an unbounded linear functional. But P ( I ) contributes + That is, if V ( I ) continuously extends to a Banach space B of the sequences ( I , I , . . . ), then B may be continuously embedded in l . ONTENTS v k = J (2 πn ) v k , k ∈ N . So the properties ofequation (3.7) essentially are determined by the component V ( I ) of the Hamiltonan.Note that since P ( I ) is a bounded linear functional on the space h ⊂ l , then theHamiltonian H K ( I ) is concave in h . Problem 3.14
Is it true that the function V ( I ) − | I | l extends analytically (orcontinuously) to a space, bigger than l ?
4. Perturbations of KdV
In the theory of integrable systems in finite dimension, there are two types ofperturbative results concerning long-time stability of solutions. The first type is theKAM theory. Roughly, it says that among a family of invariant tori of the unperturbedsystem, given by the Liouville-Arnold-Jost theorem, there exists (under genericassumptions) a large set of tori which survive under sufficiently small hamiltonianperturbations, deforming only slightly. In particular, the perturbed system admitsplenty of quasi-periodic solutions [1, 2]. Results of the second type are obtainedby the techniques of averaging which applies to a larger class of dynamical systems,characterized by the existence of fast and slow variables. This method has a muchlonger history which dates back to the epoch of Lagrange and Laplace, who appliedit to the problems of celestial mechanics, without proper justifications. Only in thelast fifty years rigorous mathematical justification of the principle has been obtained,see in [64, 1, 56]. If the unperturbed system is hamiltonian integrable, then for theslow-fast variables one can choose the action-angle variables. The averaging theoremssay that under appropriate assumptions, the action variables, calculated for solutionsof the perturbed system, can be well approximated by solutions of a suitable averagedvector field, over an extended time interval. If the perturbation is hamiltonian, thenthe averaged vector filed vanishes. The strongest result in this direction is due toNekhoroshev [66, 55], who proved that in the hamiltonian case the action variablesvary just a bit over exponentially long time intervals.Concerning instability of solutions, also two types of phenomena are known. Oneis called the Arnold diffusion. In [3], Arnold observed that despite most of the phasespace of near integrable hamiltonian systems with more than two degrees of freedomis foliated by invariant KAM tori, still there can exist solutions such that their actionsadmit increments of order one during sufficiently long time. Arnold conjectured thatthis phenomenon is generic. Though there are many developments in this direction inthe last ten years, the mechanism of the Arnold diffusion still is far from being wellunderstood. Another instability mechanism is known as the capture in resonance.The essence of this phenomenon is that a solution of a perturbed system reaches aresonant zone and begins drifting along it in such a way that the resonance conditionapproximately holds. Therefore, solutions of the original perturbed system and theaveraged one diverge by a quantity of order one on a time interval of order ǫ − (seee.g. [65]).Now return to PDEs. Two types of perturbations of the KdV equation (1.1) havebeen considered: when the boundary condition is perturbed but the equation is not,and other way round. Problems of the first type lie outside the scope of our work,and very few results (if any) are proved there rigorously, see [39] for discussion andsome related statements. Problems of the second type are much closer to the finite-dimensional situation and they are discussed below. Several attempts were made to ONTENTS
Consider the hamiltonian perturbation of KdV, corresponding to a Hamiltonian H ǫ = H ( u ) + ǫF ( u ):˙ u + u xxx − uu x − ǫ ∂∂x ∇ F ( u ) = 0 , u (0) ∈ H . (4.1)Here F ( u ) is an analytic functional on H and ∇ F is its L -gradient in u . Let n ∈ N ,and Γ ⊂ R n + be a compact set of positive Lebesgue measure. Consider a family of the n -gap tori: T Γ = ∪ I ∈ Γ T nI ⊂ J n , T nI = T ( I, ,... ) , where I = ( I , . . . , I n ), see (3.2) and (3.3). It turns out that most of them persist asinvariant tori of the perturbed equation (4.1): Theorem 4.1
For some M > , assume that the Hamiltonian F analytically extendsto a complex neighbourhood U c of T Γ in H M ⊕ C and satisfies there the regularitycondition ∇ F : U c → H M ⊗ C , sup u ∈ U c ( | F ( u ) | + ||∇ F ( u ) || M ) . Then, there exists an ǫ > and for ǫ < ǫ there exist(i) a nonempty Cantor set Γ ǫ ⊂ Γ with mes (Γ \ Γ ǫ ) → as ǫ → ;(ii) a Lipshcitiz mapping Ξ : T n × Γ ǫ → U ∩ H M , such that its restriction toeach torus T n × I , I ∈ Γ ǫ , is an analytical embedding;(iii) a Lipschitz map χ : Γ ǫ → R n , | χ − Id | Const · ǫ ,such that for every ( ϕ, I ) ∈ T n × Γ ǫ , the curve u ( t ) = Ξ( ϕ + χ ( I ) t, I ) is an quasi-periodic solution of (4.1) winding around the invariant torus Ξ( T n × { I } ) . Moreover,these solutions are linearly stable.Proof: In the coordinates v as in Theorem 3.2, the Hamiltonian H ǫ becomes H K ( I ) + ǫF ( v ). For any I ∈ h MI , using Taylor’s formula, we write H K ( I + I ) = H K ( I ) + X i > ∂H K ∂I i ( I ) I i + Z (1 − t ) X ∂ H K ∂I i ∂I j I i I j dt := const + X i > W i ( I ) I i + Q ( I , I ) , (4.2)where W is the frequency map, see (3.14). Now we introduce the symplectic polarcoordinates around the tori in the family T Γ . Namely, for each ξ ∈ Γ, we set ( v i = p ξ + y i cos ϕ, v − i = p ξ + y i sin ϕ, i n,b i = v i , b − i = v − i , i > n + 1 . ONTENTS b = ( b n +1 , b − n − , . . . ). The transformation above is real analytic andsymplectic on D ( s, r ) = {| Im ϕ | < s } × {| y | < r } × {| b | M < r } , for all s > r > I = ( ξ , H K = N + Q = N ( y, ξ , b ) + Q ( y, ξ , b ) , where N = P i n W i ( ξ ) y i + P i > W n + i ( b n + i + b − n − i ) , with I i = y i for 1 i n ,and 2 I i = b i + b − i for i > n . Then the whole Hamiltonian of the perturbed equation(4.1) can be written as H ǫ = N + Q + ǫF. (4.3)We consider the new perturbation term P = Q + ǫF . By Lemma 3.12 and the regularityassumption, if r = √ ǫ , thensup D ( s,r ) || X P || M − c √ ǫ. Using the non-degeneracy Lemma 3.11, we can take the vector { W i ( ξ ) , ≤ i ≤ n } fora free n -dimensional parameter of the problem and apply an abstract KAM theorem(see Theorem 8.3 in [44] and Theorem 18.1 in [31]) to obtain the statements of thetheorem. For a complete proof, see [44, 31]. (cid:3) Example.
The theorem applies if (4.1) is the hamiltonian PDE with the localHamiltonian H ( u ) + ǫ Z T g ( u ( x ) , x ) dx = Z T (cid:0) u x u + ǫg ( u ( x ) , x ) (cid:1) dx, where g ( u, x ) is a smooth function, periodic in x and analytic in u . In this case in(4.1) we have ∂∂x ∇ F ( u ( · ))( x ) = ∂∂x g u ( u ( x ) , x ). (cid:3) It is easy to see that for the proof of Theorem 4.1, explained above, the normalform (3.4) (weaker and much simpler than that in Theorem 3.2) is sufficient, see[42, 44]. Normal forms, similar to (3.4), exist for integrable PDEs for which a normalform as in Theorem 3.2 does not hold, e.g, for the Sine-Gordon equations, see [44].
Remark . Recently Berti and Baldi announced a KAM-theorem, similar toTheorem 4.1, which applies to perturbations of KdV, given by operators of orderhigher than one.Theorem 4.1 allows to perturb a set of KdV-solutions which form a null-set forany reasonable measure in the function space (in difference with the finite-dimensionalKAM theory which insures the persistence of a set of almost-periodic solutions whichoccupy the phase space up to a set of measure . ǫ γ , γ > Problem 4.2
Do typical tori T I as in (3.3), where I = ( I , I , . . . ) ∈ h pI , p < ∞ ,persist in the perturbed equation (4.1)? If they do not, what happens to them? Though there are KAM-theorems for perturbations of infinite-dimensionalinvariant tori, e.g. see [68, 13], they are not applicable to the problem above since,firstly, those works do not apply to KdV due to the strong nonlinear effects andlong-range coupling between the modes and, secondly, they only treat invariant toricorresponding to the actions I , I , . . . , decaying very fast (faster than exponentially). ONTENTS Compare to KAM, averaging type theorems for perturbed KdV are more recentand less developed. Their stochastic versions, which we discuss in Section 4.3, aresignificantly stronger than corresponding deterministic statements in Section 4.4. Wewill explain reasons for that a bit later. Let us start with the ‘easiest’ case, whereKdV is stabilized by small dissipation:˙ u + u xxx − uu x = ǫu xx , u (0) ∈ H . (4.4)A simple calculation shows that a solution u ( t ) satisfies || u ( t ) || e − ǫt || u (0) || . So u ( t ) becomes negligible for t ≫ ǫ − . But what happens during time intervals oforder ǫ − ? Let us pass to the slow time τ = ǫt and apply to equation (4.4) the nonlinearFourier transform Ψ, denoting for k = 1 , , . . . , Ψ k ( u ) = v k if Ψ( u ) = v = ( v , . . . ): ddτ v k = ǫ − J W k ( I ) v k + d Ψ k (Ψ − ( v ))( △ Ψ − ( v )) =: ǫ − J W k ( I ) v k + P k ( v ) , k ∈ N . (4.5)Since I k = | v k | is an integral of motion for KdV, then ddτ I k = (cid:0) P k ( v ) , v k (cid:1) := F k ( v ) , k ∈ N , (4.6)where ( · , · ) stands for the Euclidian scalar product in R . Using (4.5) we get ddτ ϕ k = ǫ − W k ( I ) + h term of order 1 i , if v k = 0 , k ∈ N . (4.7)We have written equation (4.4) in the action-angle variables ( I, ϕ ). Consider the averaged equation for actions: ddτ J k = h F k i ( J ) , h F k i ( J ) = Z T ∞ F k ( J, ϕ ) dϕ, k ∈ N , J (0) = I ( u (0)) , (4.8)where F k ( I, ϕ ) = F k ( v ( I, ϕ )), k ∈ N , and dϕ is the Haar measure on the infinitedimensional torus T ∞ . The main problem of the averaging theory is to see if thefollowing holds true: Averaging principle : Fix any
T >
0. Let ( I ( τ ) , ϕ ( τ )) be a solution of (4.6), (4.7),and J ( τ ) be a solution of (4.8). Then (either for all or, for ’typical’ initial data u (0))we have || I ( τ ) − J ( τ ) || ρ ( ǫ ) , ∀ τ T, where || · || is a suitable norm, and ρ ( ǫ ) → ǫ → a (see (3.5)).To handle the third difficulty observe that eq. (4.7) with a specific k is singularwhen I k is small. But then v k is small, the k -th mode does not affect much thedynamics, so the equation for ϕ k may be excluded from the considereation. Concerningthe second difficulty, in [46] the second author of this work established that theaveraged equation (4.8) may be lifted to a regular system in the space h p , which ONTENTS ddτ J k = h F k i = Z T ∞ ( ¯Φ θ k v k , P k (Φ θ v )) dθ = ( v k , R k ( v )) , (4.9) R k = Z T ∞ ¯Φ − θ k P k (Φ θ v ) dθ, where the maps Φ θ and ¯Φ θ k are defined in (3.8). Now consider equation ddτ v = R ( v ) . (4.10)Then relation (4.9) implies: Lemma 4.3 If v ( τ ) satisfies (4.10), then I ( v ( τ )) satisfies (4.8). Equation (4.10) is called the effective equation for the perturbed KdV equation (4.4).It is rotation-invariant: if v ( τ ) is a solution of (4.10), then for each θ ∈ T ∞ , Φ θ ( v ( τ ))also is a solution. Since the map Ψ is quasilinear by Theorem 3.4, we may write R ( v ) more explicitly. Namely, denote by ˆ △ the Fourier-image of the Laplacian,ˆ △ = diag {− k , k ∈ N } , and set L = d Ψ(0) , Ψ = Ψ − L, G = Ψ − = L − + G . Then G : h s → H s +1 is analytic for any s >
0, and direct calculation shows that R ( v ) = ˆ △ v + R ( v ) , where R ( v ) = Z T ∞ [Φ − θ L △ ( G Φ θ v ) + Φ − θ d Ψ ( G Φ θ v ) △ ( G Φ θ v )] dθ. Hence R ( v ) is an operator of order one and the effective equation (4.10) is a Fouriertransform of a quasi-linear heat equation with a non-local nonlinearity of first order.Such equations are locally well posed. Due to the direct relation between the effectiveequation and the averaged equation, the former can be used to study the latter.The first difficulty is serious. Sometime it cannot be overcome, and then theaveraging fails. A way to handle it is discussed in the next section. A way to handle the first obstacle – the resonant tori – is to add to the perturbedequation (4.4) a random force which would shake solutions u ( t ) off a resonant torus(as well as off any other invariant torus T I ). So let us consider a randomly perturbedKdV: ˙ u + u xxx − uu x = ǫu xx + √ ǫ η ( t, x ) , u (0) = u ∈ C ∞ ∩ H , (4.11) η ( t, x ) = ∂∂t X j ∈ Z b j β j ( t ) e j ( x ) . Here Z is the set of all non-zero integers, { e j ( x ) } j ∈ Z is the basis (1.2) and • all b j > | j | → ∞ . • { β j ( t ) } j ∈ Z are independent standard Wiener processes (so β j ( t ) = β ωj ( t ), where ω is a point in a probability space (Ω , F , P )). ONTENTS √ ǫ in the r.h.s is natural since only with this scaling do solutionsof equation (4.11) remain of order one when t → ∞ and ǫ →
0. To simplify formulaswe assume that b j = b − j for all j .In [48, 46], Kuksin and Piatnitski justified the averaging principle for thestochastic equation (4.11). To explain their result we pass to the slow time τ = ǫt anduse Itˆo’s formula (e.g. see in [35]) to write the corresponding equation for the vectorof actions I ( u ( ǫ − τ )) = I ω ( τ ): dI k dτ = F k ( I, ϕ ) + K k ( I, ϕ ) + X j G jk ( I, ϕ ) ∂∂τ β j ( τ ) , k ≥ . (4.12)Here F is defined as in (4.6), K is the Ito term K k = 12 X j ∈ Z b j (cid:0) ( d Ψ k ( u )[ e j , e j ] , v k ) + | d Ψ k ( u ) e j | (cid:1) , and G is the dispersion matrix, G jk = b j d Ψ k ( u ) e j .Let us average the equation above: dJdτ = h F i ( J ) + h K i ( J ) + X j h G j i ( J ) ∂∂τ β j ( τ ) . (4.13)Here h F i are the same as in (4.8), h K i is the average of K and h G j i ( J ), j ∈ Z ,are column-vectors, forming an infinite matrix h G i ( J ). The latter is defined as asquare root of the averaged diffusion matrix R T ∞ G ( J, ϕ ) G T ( J, ϕ ) dϕ, where G ( J, ϕ ) isformed by the columns G j ( J, ϕ ). Similar to Section 4.2, equation (4.11) also admitsan effective equation of the form d v k dτ = − k v k + R ′ k ( v ) + X j ( R ′′ ) jk ( v ) ∂∂t β j ( τ ) , k ≥ , where R ′ ( v ) is an operator of first order, R ′′ ( v ) is a Hilbert-Schmidt matrix, whichis an analytic function of v , and { β j ( · ) , j > } are standard independent Wienerprocesses. This is a quasilinear stochastic heat equation with a non-local nonlinearity,written in the Fourier coordinates. It is well posed in the spaces h p , p ≥ v ( τ ) is a solution of (4.9), then I ( v ( τ )) is a weak solution of(4.13). See [48, 46] for details.We recall (e.g. see [35]) that a random process J = J ω ( τ ) is a weak solution (inthe sense of stochastic analysis) of equation (4.13), if for almost every ω it satisfies theintegrated version of equation (4.13), where the processes β j ’s are replaced by someother independent standard Brownian motions ˆ β j ’s: J ω ( τ ) = Z τ ( h F i + h K i )( J ω ( s )) ds + Z τ X j h G j i ( J ω ( s )) d ˆ β ωj ( s ) , ∀ τ ∈ [0 , T ] . Fix any
T >
0. Let u ǫ ( t ), 0 t ǫ − T , be a solution of (4.11). Introduce slowtime τ = ǫt and denote I ǫ ( τ ) = I ( u ǫ ( ǫ − τ )). Consider the distribution of this randomprocess. This is a measure in the space C ([0 , T ] , h pI ). We assume p > Theorem 4.4 (i) The limiting measure lim ǫ → D ( I ǫ ( · )) exists. It is the law of a weaksolution I ( τ ) of (4.13) with the initial data I (0) = I ( u ) .ONTENTS (ii)The law D ( I ( τ )) equals to that of I ( v ( τ )) , where v ( τ ) , τ T , is a regularsolution of the corresponding effective system (4.9) with the initial data v = Ψ( u ) .(iii) Let f > be a continuous function such that R T f ( s ) ds = 1 . Thenthe measure R T f ( τ ) D ( I ǫ ( τ ) , ϕ ǫ ( τ )) dτ on the space h pI × T ∞ weakly converges, as ǫ → , to the measure (cid:16) R T f ( τ ) D ( I ( τ )) ds (cid:17) × dϕ . In particular, the measure R T f ( τ ) D ( ϕ ǫ ( τ )) ds weakly converges to dϕ , where dϕ is the Haar measure on theinfinite dimensional torus T ∞ .(iv) Every sequence ǫ ′ j → contains a subsequence ǫ j → such that the doublelimit lim ǫ j → lim t →∞ D (cid:0) Ψ( u ǫ ( t )) (cid:1) exists for any solution u ǫ ( t ) and is a stationarymeasure ∗ for the effective equation (4.9). For the proof see [48, 46]. For the last assertion also see the argument Section 4in [51]. The proof of the theorem applies to other stochastic perturbations of KdV. Inparticular, assertions (i)-(iii) hold for equations˙ u + u xxx − uu x = ǫg ( u ( x ) , x ) + √ ǫ η, where η is the same as in (4.11) and g is a smooth function, periodic in x , which hasat most a linear growth in u , and is such that g ( u ( · ) , x ) ∈ H p if u ∈ H p (this holdse.g. if g ( u, x ) is even in u and odd in x ).The key to the proof of Theorem 4.4 is the following result (see Lemma 5.2 in[48]), where for any m ∈ N , K > δ > δ, m, K ) := (cid:8) I : | W ( I ) k + . . . + W m ( I ) k m | < δ, for some k ∈ Z m such that 1 | k | K (cid:9) . (4.14) Lemma 4.5
For any m ∈ N , K > , T > and δ > we have Z T P { I ( τ ) ∈ Ω( δ, m, K ) } dτ κ ( δ, K, m, T ) , where κ ( δ, K, m, T ) goes to zero with δ , for any fixed K , m and T . This lemma assures that in average, solutions of (4.11) do not spend much time inthe vicinity of resonant tori. The stochastic nature of the equation is crucial for thisresult.
It is plausible that the averaging principle also holds for equation (4.4). But withoutrandomness, it is unclear how to assure that solutions of (4.4) ‘pass the resonant zonequickly’ (in analogy with Lemma 4.5). This naturally leads to the question: for whichdeterministic perturbations of KdV it is possible to prove the property of fast crossingthe resonant zones and verify the averaging principle? Some results in this directionare obtained by the first author in [26, 27]. Now we discuss them.Consider a deterministically perturbed KdV equation:˙ u + u xxx − uu x = ǫf ( u ) , x ∈ T , u ∈ H p , (4.15) ∗ For this notion and its discussion see [50]. We are certain that eq. (4.11) has a unique stationarymeasure. When this is proven, it would imply that the convergence in (iv) holds as ǫ → ONTENTS p > f ( u ) = f ( u ( · )) may be non-local. I.e., f ( u )( x ) maydepend on values of u ( y ), where | y − x | ≥ κ >
0. We are going to discuss solutions of(4.15) on time-intervals of order ǫ − . Accordingly we fixe some ζ ≤ p > T >
Assumption A . (i) There exists p ′ = p ′ ( p ) < p , such that for any q ∈ [ p ′ , p ] theperturbation in (4.15) defines an analytic mapping of order ζ : H q → H q − ζ , u ( · ) f ( u ( · )) . (ii) For any u ∈ H p , there exists a unique solution u ( t ) ∈ H p of (4.15) with u (0) = u . For ≤ t T ǫ − its norm satisfies || u ( t ) || p C ( T, p, || u || p ) . It will be convenient for us to discuss equation (4.15) in the v -variables. We willdenote B p ( M ) = { v ∈ h p : | v | p M } . With some abuse of notation we will denote by S t , 0 t T ǫ − , the flow-maps ofequation (4.15), both in the u - and in the v -variables. Definition 4.6
1) A Borelian measure µ on h p is called regular if for any analyticfunction g on h p , we have µ ( { v ∈ h p : g ( v ) = 0 } ) = 0 .2) A measure µ on h p is said to be ǫ -quasi-invariant for equation (4.15) if it isregular and for any M > there exists a constant C ( T, M ) such that for every Borelset A ⊂ B p ( M ) we have ♯e − ǫtC ( T,M ) µ ( A ) µ ( S t ( A )) e ǫtC ( T,M ) µ ( A ) , ∀ ≤ t ǫ − T. (4.16)Similarly, these definitions can be carried to measures on the space H p and the flowmaps of equation (4.15) on H p .For an ǫ -perturbed finite-dimensional hamiltonian system the Lebesgue measure is ǫ -quasi-invariant by the Liouville theorem. This fact is crucial for the Anosov approachto justify the classical averaging principle (see in [1, 56]). In infinite dimension thereis no Lebesgue measure, and existence of an ǫ -quasi-invariant measure is a seriousrestriction.If equation (4.15) has an ǫ -quasi-invariant measure µ , then the argument, inventedby Anosov for the finite dimensional averaging, insures the required analogy ofLemma 4.5 for equation (4.15). Indeed, let us define the resonant subset B of h p × R as B := (cid:8) ( v, t ) : v ∈ B p ( M ) , t ∈ [0 , ǫ − T ] and S t v ∈ Ω( δ, m, K ) (cid:9) (see (4.14)), and consider the measure µ on h p × R , where d µ = dµdt . Then by (4.16)we have µ ( B ) = Z ǫ − T µ (cid:16) B p ( M ) ∩ S − t (cid:0) Ω( δ, m, K ) (cid:1)(cid:17) dt ǫ − T e C ( T,M ) µ (Ω( δ, m, K )) . For any v ∈ B p ( M ), define Res ( v ) as the set of resonant instants of time for atrajectory, which starts from v : Res ( v ) = { τ ∈ [0 , ǫ − T ] : S t ( v ) ∈ Ω( δ, m, K ) } . ♯ This specifies the usual definition of a quasi-invariant measure. We recall that if a flow { S t } ofsome equation exists for all t ≥ t ∈ R ), then a measure m is called quasi-invariant for thisequation if the measures S t ◦ m are absolutely continuous with respect to m for all t ≥ t ∈ R ). ONTENTS µ ( B ) = Z B p ( M ) mes (cid:0) Res ( v ) (cid:1) dµ ( v )by the Fubini theorem, where mes ( · ) stands for the Lebesgue measure on [0 , ǫ − T ]. Iffor ρ > V Res( ρ ) := { v ∈ B p ( M ) : mes ( Res ( v )) > ǫ − ρ } , then in view of the Chebyshev inequality µ ( V Res( ρ )) ǫρ µ ( B ) T e C ( M,T ) ρ µ (Ω( δ, m, K )) . By Theorem 3.2 the functions v W ( I ( v )) k + . . . + W m ( I ( v )) k m , k ∈ Z m \ { } , areanalytic on h p . Since they do not vanish identically by Lemma 3.11 and the measure µ is regular, then µ (Ω(0 , m, K )) = 0. Accordingly µ (Ω( δ, m, K )) goes to zero with δ ,and µ ( V Res( ρ )) → δ → , (4.17)for any ρ . This gives us a sought for analogy of Lemma 4.5 for deterministicperturbations of KdV which have ǫ -quasi-invariant measures.The averaged equation for actions, corresponding to (4.15), reads dJ k dτ = h F k i ( J ) , k = 1 , , . . . , (4.18)where F k = (cid:0) d Ψ k (Ψ − ( v ))( f (Ψ − ( v )( · )) , v k (cid:1) (cf. (4.6)). Due to item (i) ofAssumption A, the r.h.s. of (4.18) defines a Lipschitz vector filed on h pI , so theaveraged equation is well posed locally on h pI . We denote by J I ( τ ) a solution of(4.18) with an initial data J I (0) = I ∈ h pI . It is shown in [26, 27] that relation (4.17)and the well-posedness of the averaged equation jointly allow to establish an averagingtheorem for equation (4.15), provided that it has an ǫ -quasi-invariant mesure.In the statement below u ǫ ( t ) stand for solutions of equation (4.15) and v ǫ ( τ ) – forthese solutions, written using the v -variables and slow time τ = ǫt . By Assumption A,for τ ∈ [0 , T ] we have | I ( v ǫ ( τ )) | ∼ p C (cid:0) T, | I ( v ǫ (0)) | ∼ p (cid:1) . Denote˜ T ( I ) := min { τ ∈ R + : | J I ( τ ) | ∼ p > C ( T, | I | ∼ p ) + 1 } . Theorem 4.7
Fix some
M > . Suppose that Assumption A holds and equation (4.15)has an ǫ -quasi-invariant measure µ on h p such that µ ( B p ( M )) > . Then(i) For any ρ > and any q < p − max { ζ , − } , there exist ǫ ρ,q > and aBorel subset Γ ǫρ,q ⊂ B p ( M ) , satisfying lim ǫ → µ ( B p ( M ) \ Γ ǫρ,q ) = 0 , with the followingproperty:if ǫ ǫ ρ,q and v ǫ (0) ∈ Γ ǫρ,q , then | I ( v ǫ ( τ )) − J I ǫ ( τ ) | ∼ q ρ for τ min { T, ˜ T ( I ǫ ) } , (4.19) where I ǫ = I ( v ǫ (0)) .(ii) Let λ v ǫ be a probability measure on T ∞ , defined by the relation Z T ∞ f ( ϕ ) dλ v ǫ ( dϕ ) = 1 T Z T f ( ϕ ( v ǫ ( τ )) dτ, ∀ f ∈ C ( T ∞ ) , ONTENTS where v := v ǫ (0) ∈ B p ( M ) . Then the measure µ (cid:0) B p ( M ) (cid:1) − R B p ( M ) λ v ǫ dµ ( v ) converges weakly, as ǫ → , to the Haar measure dϕ on T ∞ †† Proposition 4.8
If Assumption A holds with ζ < , then for ρ < , p q and ǫ ǫ ρ,q , we have ˜ T := ˜ T (cid:0) I ( v ǫ (0)) (cid:1) > T for v ǫ (0) ∈ Γ ǫρ,q . So (4.19) holds for ≤ τ ≤ T .Proof: Assume that ˜ T T . By (4.19) for 0 ≤ τ ≤ ˜ T we have | I ǫ ( τ ) − J I ǫ ( τ ) | ∼ p ρ .Therefore | J I ǫ ( τ ) | ∼ p C ( T, | I ǫ | ∼ p ) + ρ < C ( T, | I ǫ | ∼ p ) + 1. This contradicts thedefinition of ˜ T , so ˜ T > T . (cid:3) Remark . Assume that ǫ -quasi-invariant measure µ depends on ǫ , i.e., µ = µ ǫ , anda) µ ǫ (Ω( δ, m, K )) goes to zero with δ uniformly in ǫ ,b) the constants C ( T, M ) in (4.16) are bounded uniformly in ǫ .Then assertion (i) holds with µ replaced by µ ǫ . For assertion (ii) to hold, morerestrictions should be imposed, see [27].Theorem 4.7 gives rise to the questions: Problem 4.9
Does a version of the averaging theorem above holds without assumingthe existence of an ǫ -quasi-invariant measure? Problem 4.10
Which equations (4.15) have ǫ -quasi-invariant measures? See the next subsection for some results in this direction.
Problem 4.11
Find an averaging theorem for equations (4.15), where thenonlinearity defines an unbounded operator, i.e. in Assumption A(i) we have ζ > (note that in the equation from Example in Section 4.1 we have ζ = 1 , and in equation(4.4) ζ = 2 ). It is unlikely that the assertion of Theorem 4.7 holds for all initial data, and we believethat the phenomenon of capture in resonance happens for some solutions of (4.15):
Problem 4.12
Prove that (4.19) does not hold for some solutions of (4.15).4.5. Existence of ǫ -quasi-invariant measures Clearly every regular measure, invariant for equation (4.15), is ǫ -quasi-invariant.Gibbs measures for some equations of the KdV type are regular and invariant,they were studied by a number of people (e.g., see [12, 74]). However, for generichamiltonian perturbations of KdV it is difficult, probably impossible, to constructinvariant measures in higher order Sobolev spaces due to the lack of high orderconservation laws. Below we give some examples of ǫ -quasi-invariant measures forsmoothing perturbations of KdV, which are Gibbs measures of KdV (so they areKdV-invariant). Note that some of these perturbed equations do not have non-trivialinvariant measures. For example, our argument applies to equations which in the v -variables read as ˙ v j = J W j ( I ) v j − ǫj − ρ v j , j ∈ N , where ρ >
1. But all trajectories of this equation converge to zero, so the onlyits invariant measure is the δ -measure at the origin. That is, for averaging in theperturbed KdV (4.15) (various) Gibbs measures of KdV play the same role as theLebesgue measure plays for the classical averaging. †† In [26] a stronger assertion was claimed. Namely, that the measure λ v ǫ converges to dϕ for µ -a.a. v in B p ( M ). Unfortunately, the proof in [26] contains a gap which we still cannot fix. ONTENTS Definition 4.13
For any ζ ′ < − , a Gaussian measure µ on the Hilbert space h p is called ζ ′ -admissible if it has zero mean value and a diagonal correlation operator ( v , . . . ) ( σ v , . . . ) , where < j ζ ′ /σ j Const for each j . For any ζ ′ < − ζ ′ -admissible measure µ is a well-defined probability measureon h p , which can be formally written as µ = ∞ Y j =1 (2 πj ) p πσ j exp {− (2 πj ) p | v j | σ j } d v j , (4.20)where d v j , j >
1, is the Lebesgue measure on R v j . It is known that (4.20) is awell-defined measure on h p if and only if P σ j < ∞ (see [10]). It is regular andnon-degenerate in the sense that its support equals h p (see [10, 11]). Writing KdVin the v -variables we see that µ is invariant under the KdV flow. Note that µ is aGibbs measure for KdV, written in the form (3.7), since formally it may be writtenas µ = Z − exp {−h Qv, v i} dv , where h Qv, v i = P c j | v j | is an integrals of motion forKdV (the statistical sum Z = ∞ , so indeed this is a formal expression).For a perturbed KdV (4.15) we define P ( v ) = d Ψ( u )( f ( u )), where u = Ψ − ( v ).A non-complicated calculation (see in [26]) shows that: Theorem 4.14
If Assumption A holds and ( i ) ′ the operator P analytically maps the space h p to h p − ζ ′ with some ζ ′ < − ,then every ζ ′ -admissible Gaussian measure on h p is ǫ -quasi-invariant for equa-tion (4.15) on the space h p . However, due to the complexity of the nonlinear Fourier transform Ψ, it is not easyto verify the condition ( i ) ′ of Theorem 4.14 for specific equation (4.15). Now we willgive other examples of ǫ -quasi-invariant measures on the space H p , by strengtheningthe restrictions in Assumption A. Suppose that there p ∈ N . Let µ p be the centeredGaussian measure on H p with the correlation operator △ − . Since △ − is an operatorof the trace type, then µ p is a well-defined probability measure on H p .We recall (see Remark 3.3) that KdV has infinitely many conservation laws J n ( u ), n >
0, of the form J n = || u || n + J n − ( u ), where J − ( u ) = 0 and for n > J n − ( u ) = Z T (cid:8) c n u ( ∂ n − x u ) + Q n ( u, . . . , ∂ n − x u ) (cid:9) dx . (4.21)Here c n are real constants and Q n are polynomial in their arguments. From (4.21),we know that the functional J p is bounded on bounded sets in H p . We consider aGibbs measure η p for KdV, defined by its density against µ p , η p ( du ) = e − J p ( u ) µ p ( du ) . It is regular and non-degenerated in the sense that its support contains the wholespace H p (see [10]). Moreover, it is invariant for KdV [74]. The following theoremwas shown in [27]: Theorem 4.15
If Assumption A holds with ζ − , then the measure η p is ǫ -quasi-invariant for perturbed KdV (4.15) on the space H p . Corollary 4.16 If ζ − , then the assertions of Theorem 4.7 hold with µ = η p ,and we have ˜ T > T .ONTENTS u + u xxx − uu x = ǫf ( x ) , x ∈ T , u ∈ H p , where f ( x ) is a smooth function with zero mean-value. This equation may be viewedas a model for shallow water wave propagation under small external force. Note thatthe KAM-Theorem 4.1 also applies to it. Problem 4.17
Besides the class of ζ ′ -admissible Gaussian measures and Gibbsmeasure η p , there are many other KdV-invariant measures. How to check if a measurelike that is ǫ -quasi-invariant for a given ǫ -perturbation of KdV?4.6. Nekhoroshev type results (long-time stability)? In the finite dimensional case, the strict convexity of the unperturbed integrableHamiltonian assures the long-time stability of solutions for perturbed hamiltonianequations ([66, 55, 69]). Theorem 3.13 tells us the the KdV Hamiltonian H K ( I ) isconvex in l and hints that it is strictly convex (at least) in a neighborhood of theorigin in l . This suggests that an Nekhoroshev type stability may hold for perturbedKdV under hamiltonian perturbations (see equation (4.1)), at least for initial datain a neighborhood of the origin, where the strict convexity should hold. But at themoment of writing no exact statement is available.There are several ad hoc quasi-Nekhoroshev theorems for hamiltonian PDEs, see[5, 14] and references therein. However, these results only apply in a small (of thesize of the perturbation) neighborhood of the origin. Nonetheless, we believe thatthe corresponding technique and the results in Theorem 3.13 will lead to results onlong-time stability (at least in time interval of order ǫ − p , p ≥ p = 1 and for typical initial data, if the perturbation f is smoothing. Stability ontime-intervals of order ǫ − seems to be a much harder question. Problem 4.18
Is it possible to prove a long-time stability result for perturbed KdVunder hamiltonian perturbations, e.g. for equation (4.1), that holds for all ‘smooth’initial data?
Acknowledgments.
This work was supported by l’Agence Nacionale de la Recherche,grant ANR-10-BLAN 0102.
References [1] V. Arnold, V. V. Kozlov, and A. I. Neistadt.
Mathematical Aspects of Classical and CelestialMechanics . Springer, Berlin, third edition, 2006.[2] V. I. Arnold. Proof of a theorem of A. N. Kolmogorov on the conservation of quasiperiodicmotions under a small change of the Hamiltonian function.
Russ. Math. Surv. , 18(5):9–36,1963.[3] V. I. Arnold. Instability of dynamical systems with several degrees of freedom.
SovietMathematics , 5:581–585, 1964.[4] V. I. Arnold.
Mathematical methods in classical mechanics . Springer, Berlin, 3rd edition, 1989.[5] D. Bambusi. Nekhoroshev theorem for small amplitude solutions in nonlinear Schr¨odingerequations.
Mathematische Zeitschrift , 230(2):345–387, 1999.[6] D. Bambusi and B. Gr´ebert. Birkhoff normal form for partial differential equations with tamemodulus.
Duke Mathematical Journal , 135(3):507–567, 2006.[7] E. D. Belokolos, A. I. Bobenko, V. Z. Enolskii, A. R. Its, and V. B. Matveev.
Algebro-geometricapproach to nonlinear integrable equations . Springer, 1994.[8] A. Biryuk. On multidimensional Burgers type equations with small viscosity. In
Contributionsto Current Challenges in Mathematical Fluid Mechanics , pages 1–30. Springer, 2004.
ONTENTS [9] A. Bobenko and S. Kuksin. Finite-gap periodic solutions of the KdV equation are non-degenerate. Physics Letters A , 161(3):274–276, 1991.[10] V. Bogachev.
Differentiable measures and the Malliavin calculus . American MathematicalSociety, 2010.[11] V. I. Bogachev and I. I. Malofeev. On the absolute continuity of the distributions of smoothfunctions on infinite-dimensional spaces with measures. preprint , 2013.[12] J. Bourgain. Periodic nonlinear Schr¨odinger equation and invariant measures.
Comm. Math.Phys. , 166:1–26, 1994.[13] J. Bourgain. Construction of approximative and almost periodic solutions of perturbed linearSchr¨odinger and wave equations.
Geometric & Functional Analysis GAFA , 6(2):201–230,1996.[14] J. Bourgain. On diffusion in high-dimensional Hamiltonian systems and PDEs.
Journald’Analyse Math´ematique , 80(1):1–35, 2000.[15] J. Boussinesq. Th´eorie de l’intumescence liquid appel´ee onde solitaire ou de translation, sepropageant dans un canal rectangulaire.
Comptes Rend. Acad. Sci (Paris) , 72:755–759, 1871.[16] J. Colliander, M.Keel, G.Staffiliani, and T. Tao. Symplectic nonsqueezing of the Korteweg-deVries flow.
Acta Math. , 195:197–252, 2005.[17] W. Craig.
Probl`emes de Petitts Diviseurs dans les ´Equations aux D´eriv´ees Partielles . Number 9in Panoramas et Synth´eses. Soci´et´e Math´ematique de France, 2000.[18] P. Deift, S. Venakides, and X. Zhou. New results in small dispersion KdV by an extension of thesteepest descent method for Riemann-Hilbert problems.
International Mathematics ResearchNotices , 1997(6):285–299, 1997.[19] B. Dubrovin. Theta functions and nonlinear equations.
Russ. Math. Surv , 36:11–92, 1981.[20] B. Dubrovin, V. Matveev, and S. Novikov. Nonlinear equations of Korteweg-de Vries type, finitezone linear operators, and Abelian varieties.
Russ. Math. Surv , 31(1):55–136, 1976.[21] H. Eliasson. Normal forms for Hamiltonian systems with Poisson commuting integrals — ellipticcase.
Commentarii Mathematici Helvetici , 65(1):4–35, 1990.[22] H. Flaschka and D. McLaughlin. Canonically conjugate variables for the Koreweg-de Vriesequation and the Toda lattice with periodic boundary conditions.
Progress of TheoreticalPhysics , 55(2):438–457, 1976.[23] C. Gardner. Korteweg-de Vries equation and generalizations. i.v. the Korteweg-de Vries equationas a Hamiltonian system.
J. Math. Phys , 12:1548–1551, 1971.[24] J. Garnett and E. Trubowitz. Gaps and bands of one dimensional periodic Schr¨odinger operators.
Comment. Math. Helv. , 59:258–312, 1984.[25] B. Gr´ebert and T. Kappeler.
The defocusing NLS equation and its normal form . EuropeanPublishing House, 2013.[26] G. Huang. An averaging theorem for a perturbed KdV equation.
Nonlinearity , 26:1599–1621,2013.[27] G Huang. An averaging theorem for perturbed KdV equations. preprint , 2013.[28] H. Ito. Convergence of Birkhoff normal forms for integrable systems.
Commentarii MathematiciHelvetici , 64(1):412–461, 1989.[29] A. Its and V. Matveev. Schr¨odinger operators with the finite-band spectrum and the N -solitonsolutions of the Korteweg-de Vries equation. Teoret. Mat. Fiz. , 23:51–68, 1975.[30] T. Kappeler, C. M¨ohr, and P. Topalov. Birkhoff coordinates for KdV on phase spaces ofdistributions.
Selecta Mathematica, New Series , 11(1):37–98, 2005.[31] T. Kappeler and J. P¨oschel.
KdV & KAM . Springer, 2003.[32] T. Kappeler, B. Schaad, and P. Topalov. Asymptotics of spectral quantities of Schr¨odingeroperators. In
Spectral geometry , Proc. Sympos. Pure Math. 84, pages 243–284. AmericanMathematical Society, 2012.[33] T. Kappeler, B. Schaad, and P. Topalov. Qualitative features of periodic solutions of KdV.
Communications in Partial Differential Equations , 2013. arXiv:1110.0455.[34] T. Kappeler and P. Topalov. Global wellposedness of KdV in H − ( T, R ). Duke Math. J. ,135:327–360, 2006.[35] I. Karatzas and S. E. Shreve.
Brownian motion and stochastic calculus . Springer, 1991.[36] E. Korotyaev. Estimates for the Hill operator, II.
Journal of Differential Equations , 223(2):229–260, 2006.[37] E. Korotyaev and S. Kuksin. KdV Hamiltonian as the function of actions. preprint , 2011.[38] D. Korteweg and G. de Vries. On the change of form of long waves advancing in a rectangularcanal, and on a new type of long stationary waves.
Philos. Mag. , 539:422–443, 1895.[39] I. Krichever. Perturbation theory in periodic problems for two-dimensional integrable systems.
Sov. Sci. Rev. C. Math. Phys. , 9:1–101, 1991.
ONTENTS [40] M. Kruskal and N. Zabusky. Interaction of “solitons” in a collision-less plasma and the recurrenceof initial states. Phys. Rev. Lett. , 15:240–243, 1965.[41] S. Kuksin. Hamiltonian perturbations of infinite-dimensional linear systems with an imaginaryspectrum.
Functional Analysis and Its Applications , 21(3):192–205, 1987.[42] S. Kuksin. Perturbation theory for quasi-periodic solutions of infinite dimensional Hamiltoniansystems, and its application to the Koreteweg-de Vries equation.
Math. USSR Subornik ,67:397–413, 1989.[43] S. Kuksin. Spectral properties of solutions for nonlinear PDEs in the turbulent regime.
Geometric & Functional Analysis GAFA , 9(1):141–184, 1999.[44] S. Kuksin.
Analysis of Hamiltonian PDEs . Oxford University Press, Oxford, 2000.[45] S. Kuksin. Hamiltonian PDEs. In
Handbook of dynamical systems , volume 1, pages 1087–1133.Elsevier, 2006.[46] S. Kuksin. Damped driven KdV and effective equations for long-time behaviour of its solutions.
GAFA , 20:1431–1463, 2010.[47] S. Kuksin and G. Perelman. Vey theorem in infinite dimensions and its application to KdV.
DCDS-A , 27:2010, 1-24.[48] S. Kuksin and A. Piatnitski. Khasminskii-Whitham averaging for randomly perturbed KdVequation.
J. Math. Pures Appl. , 89:400–428, 2008.[49] S. Kuksin and J. Poschel. Invariant Cantor manifolds of quasi-periodic oscillations for a nonlinearSchr¨odinger equation.
Annals of Mathematics , pages 149–179, 1996.[50] S. Kuksin and A. Shirikyan.
Mathematics of Two-Dimensional Turbulence . CambridgeUniversity Press, Cambridge, 2012.[51] S. B. Kuksin. Weakly nonlinear stochastic CGL equations.
Ann. Inst. H. Poincar´e - PR ,49:1033–1056, 2013.[52] P. Lax. Integrals of nonlinear equations of evolution and solitary waves.
Comm. Pure Appl.Math. , 21:467–490, 1968.[53] P. Lax. Outline of a theory of the KdV equation. In
Recent mathematical methods in nonlinearwave propagation , Lecture Notes in Mathematics 1640, pages 70–102. Springer, 1996.[54] P. Lax and C. Levermore. The zero dispersion limit for KdV equation.
Proc. Nat. Acad. Sci.USA , pages 3602–3606, 1979.[55] P. Lochak. Canonical perturbation theory via simultaneous approximation.
Uspekhi Mat. Nauk ,47(6):59–140, 1992.[56] P. Lochak and C. Meunier.
Multiphase averaging for classical systems: with applications toadiabatic theorems . Springer, 1988.[57] V. Marcˆenko.
Sturm-Liouville operators and applications . Birkh¨auser,Basel, 1986.[58] H. McKean and E. Trubowitz. Hill’s operator and hyper-elliptic function theory in the presenceof infinitely many branch points.
Comm. Pure Appl. Math , 29:143–226, 1976.[59] H. McKean and E. Trubowitz. Hill’s surfaces and their theta functions.
Bull. Am. Math. Soc ,84:1042–1085, 1978.[60] R. Miura. Korteweg-de Vries equation and generalizations. I. A remarkable explicit nonlineartransformation.
J. Math. Phys. , 9:1202–1204, 1968.[61] J. Moser. Lectures on Hamiltonian systems.
Mem. Amer. Math. Soc. , 81:1–60, 1968.[62] J. Moser and C. L. Siegel.
Lectures on celestial mechanics . Springer, Berlin, 1971.[63] R. Muira, C. Gardner, and M. Kruskal. Korteweg-de Vries equation and generalizations. II.Existence of conservation laws and constants of motion.
J. Math. Phys , 9:1204–1209, 1968.[64] A. I. Neishtadt. Averaging in multi-frequency systems, I and II.
Soviet Phys, Doklady , 20(7)and 21(2):492–494 and 80–82, 1975.[65] A. I. Neishtadt. Capture into resonance and scattering on resonances in two frequency systems.
Proceedings of the Steklov Institute of Mathematics , 250:183–203, 2005.[66] N. Nekhoroshev. An exponential estimate of the time of stability of nearly-integrableHamiltonian systems I.
Uspekhi Mat. Nauk , 32(6):5–66, 1972.[67] S. Novikov. A periodic problem for the Korteweg-de Vries equation.
I. Funkt.Anal. Prilozh ,8:54–66, 1974.[68] J. P¨oschel. Small divisors with spatial structure in infinite dimensional Hamiltonian systems.
Comm. Math. Phys , 127(2):351–393, 1990.[69] J. P¨oschel. Nekhoroshev estimates for quasi-convex Hamiltonian systems.
Math. Z. , 213:187–216, 1993.[70] J. Saut and R. Teman. Remarks on the Kortewed-de Vries equation.
Israel J. Math , 24:1976,78-87.[71] J. Vey. Sur certain syst`emus dynamiques s´eparables.
Am. J. Math , 100:591–614, 1978.[72] C. Wayne. Periodic and quasi-periodic solutions of nonlinear wave equations via KAM theory.
ONTENTS Comm. Math. Phys , 127:479–528, 1990.[73] V. E. Zakharov, S. V. Manakov, S. P. Novikov, and L. P. Pitaevskij.
Theory of Solitons: TheInverse Scattering Method . Plenum Press, 1984.[74] P. E. Zhidkov.