On the growth of high Sobolev norms for certain one-dimensional Hamiltonian PDEs

On the growth of high Sobolev norms for certain one-dimensional Hamiltonian PDEs

Joseph Thirouin
July 26, 2019

This paper is devoted to the study of large time bounds for the Sobolev norms of the solutions of the following fractional cubic Schrödinger equation on the torus :

where is a real parameter. We show that, apart from the case , which corresponds to a half-wave equation with no dispersive property at all, solutions of this equation grow at a polynomial rate at most. We also address the case of the cubic and quadratic half-wave equations.

MSC 2010 : 37K40, 35B45.

Keywords : Hamiltonian systems, fractional nonlinear Schrödinger equation, nonlinear wave equation, dispersive properties.



1 Introduction

In the study of Hamiltonian partial differential equations, understanding the large time dynamics of solutions is an important issue. In usual cases, the conservation of the Hamiltonian along trajectories enables to control one Sobolev norm of the solution (in the so-called energy space), but when solutions are globally defined and regular, higher norms could grow despite the conservation laws, reflecting an energy transfer to high frequencies. Even for notorious equations, such as the nonlinear Schrödinger equation on manifolds, it is an old problem to know whether such instability occurs [3], and often still an open question.

We start from the particular case of the defocusing Schrödinger equation on the torus of dimension one, with a cubic nonlinearity :


Here, is a function of time and of space variable , and . Equation (1) is Hamiltonian, and because of the energy conservation, its trajectories are bounded is . But it is also well-known that (1) is integrable (see [14], [26]), with conservation laws ensuring that if belongs to for some , then the solution remains bounded in (this is even true for any real [16]).

In order to track down large time instability for Hamiltonian systems, Majda, McLaughlin and Tabak [19] suggested to replace the Laplacian in (1) by a whole family of pseudo-differential operators : the operators (sometimes written as ) for real . Recall that if is a function on the torus, then

So we consider the following fractional Schrödinger equation :


where is any positive number. If , we recognize the classical Schrödinger equation (1). In the case , (2) is a non-dispersive equation, called the (defocusing) "half-wave" equation :


This half-wave equation has been studied by Gérard and Grellier in [9]. In particular, they showed that the dynamics of (3) is related to the behaviour of the solutions of a toy model equation, called the cubic Szegő equation :


where is the projection onto nonnegative Fourier modes. In a more precise way, equation (4) appears to be the completely resonant system of (3).

All the equations (2) derive from the Hamiltonian for the symplectic structure endowed by the form , where denotes the standard inner product on . The functional is therefore conserved along trajectories. Gauge invariance as well as translation invariance also imply the existence of two other conservation laws for equation (2) :

i.e. the mass and the momentum respectively. Starting from these observations, it has been proved that for , equation (3) admits a globally defined flow in with (see [9]). In the case of the half-wave equation, the Brezis-Gallouët inequality [4] also ensures that -norms of solutions grow at most like , for some constant depending on and on the initial data.

The question of the large time instability of global solutions of (2) thus naturally arises : is it possible to find smooth initial data whose corresponding orbits are not bounded in some space , or at least not polynomially bounded111We say that a solution is polynomially bounded in if there are positive constants and (not depending on time) such that for all , . ?

The cubic Szegő equation discloses this kind of instability, as recently shown in [10], [11] : for generic smooth initial data, the corresponding solution of the Szegő equation in is polynomially unbounded, for any . Therefore it is reasonable to think that the same statement should hold for the half-wave equation (3), though such a result seems far beyond our reach at this point. Nevertheless the theorem we prove in this paper gives an a priori bound for all solutions of the half-wave equation :

Theorem 1.

Let , and the solution of the half-wave equation (3) such that . Given any integer , there exist real constants , such that ,


Here, depends on and , whereas can be chosen equal to , with depending only on .

The bound appearing in (5) is an improvement the "double exponential bound" mentioned above. But as a matter of fact, finding any explicit non-trivial solution of (2) is still an open problem, and nothing is known about the optimality of (5). Solution with rapidly growing -norms could perfectly well exist. H. Xu [24] typically proved the existence of exponentially growing solutions for a perturbation of the Szegő equation. See also the result of Hani–Pausader–Tzvetkov–Visciglia [15], in the context of the Schrödinger equation, as well as its recent counterpart in [25].

Notice that in the case of the Szegő equation, the best bound quantifying the growth of Sobolev norms of solutions is : it is obtained by Gérard and Grellier in [8, section 3]. Hence (5) is likely to be improved, but recall that, as far as we know, the only way of proving the simple exponential bound for Szegő solutions makes use of the Lax pair structure associated with the equation. Elementary methods would only give an bound (see Appendix A). Unfortunately, such a Lax pair structure apparently does not exist as regards the half-wave equation.

Even so, the proof of (5) in theorem 1 suggests that we could get a simple exponential bound, instead of , if we could deal with a quadratic nonlinearity. Simply putting an -norm in the energy , instead of the one, would give rise to a nonlinearity of the form , and the singularity at the origin may lead to solutions less regular than their initial data. It is possible to avoid this phenomenon, by consider a system of two equations rather than a single scalar equation :


with . System (6) only involves (analytic) quadratic nonlinearities. It happens that Schrödinger systems of that kind frequently appear in physics : they are closely linked with the SHG (Second-Harmonic Generation) theory in optics, and the study of propagation of solitons in so-called (or quadratic) media or materials (for a review, see e.g. [18, section 4]). Quadratic systems are also relevant in fluid mechanics, to describe the interaction between long nonlinear waves in fluid flows [13]. From a mathematical point of view, interest on quadratic systems is more recent [20].

For the case of system (6), we prove the following theorem :

Theorem 2.

Let , and the solution of (6) such that . Given any integer , there exist real constants , such that ,


Here, depends on and on the sum , whereas can be chosen equal to , with depending only on .

Let us now return to equation (2). When , (2) has dispersive properties. Using them for and proving some Strichartz estimate for the operator , Demirbas, Erdoğan and Tzirakis show in [7] that (2) is globally well-posed in the energy space (and even below). Their method rely on Bourgain’s high-low frequency decomposition, but it does not say anything about the possible growth of solutions.

Still for , a naive calculation leads to an exponential bound for -norms of solutions, i.e. a bound of the form , but results such as Bourgain’s [2] or Staffilani’s [22] suggest that because of dispersion, solutions should be polynomially bounded. On the other hand, the polynomial growth of solutions of (2) for is announced to be true in the work by Demirbas–Erdoğan–Tzirakis. Indeed we establish a theorem also involving (part of) the case :

Theorem 3.

Let , and . There exist a unique solution of (2) with . Furthermore, given any integer , there exist real constants , , and a constant depending only on , such that ,


When , we can choose

where is a constant which only depends on and .

When , a possible choice is

Here again, it is not known whether (8) is optimal or not. When , Demirbas [6] found, by a probabilistic way inspired by the works of Bourgain [1], solutions growing at most like a power of . But in any case, proving that some solutions of (2) do blow up for large time, even at a very low rate, would be a big step forward.

Combining theorem 1 and 3 thus indicate that is an isolate point in the family of equations (2). Notice that, when , theorem 3 includes the existence of a flow, which had not been proved so far. As for the condition , it appears to be convenient in the proof for technical reasons ; but since the heart of our work is to prove that the case is more likely to disclose weak turbulence phenomena than other cases, we postpone discussions and comments concerning the relevance of the value until Appendix B.

The proof of theorems 1, 2 and 3 is based on an idea developped by Ozawa and Visciglia in [21] : the authors introduce a modified energy method, in order to sharpen -estimates and thus prove well-posedness for the half-wave equation with quartic nonlinearity. To put it shortly, their idea is to introduce a nonlinear energy which is in fact a perturbation of the norm they wish to bound. The perturbation does not modify the size of the norm, but induces simplifications while differentiating, so that time-differentiation behaves like a self-adjoint operator.

In the sequel of this paper, we begin by proving theorems 1, 2, and the first part of theorem 3, with elementary tools. Then we address the case of , using Bourgain spaces as in [5], which we fully develop for the convenience of the reader.

We were about to finish this paper when we were informed of a work by Planchon and Visciglia also applying the modified energy method to solutions of nonlinear Schrödinger equations on certain Riemaniann manifolds, and for every power nonlinearity.

The author would like to express his gratitude towards P. Gérard for his deep insight and generous advice. He also thanks J.-C. Saut for the references concerning quadratic medias.

2 The case

Throughout this section, we suppose that . We fix , and we study , the associated solution of (2) (or (3), depending on the value of ).

2.1 The modified energy method

Fix . The following lemma gathers some standard inequalities of which we will make an extensive use.

Lemma 2.1.

There exist an absolute constant , a constant depending on and a constant depending on such that, for every ,

  1. ,

  2. if , ,

  3. if , .

We justify briefly these inequalities : 1 derives from the conservation of together with the Sobolev embeddings in dimension one, and 2 is a consequence of 1 and the injection . As for 3, it follows from the classical Brezis-Gallouët inequality : for , and ,


(see e.g. [8]). Here, to infer 3, we begin by squaring the ratio of the two norms, and we then take into account the fact that the function is increasing.

To prove the estimates we have in mind, we are going to establish an inequality between the -norm of the solution and its derivative, and apply a Gronwall lemma. As announced, we define for this purpose a well-chosen nonlinear functional222From now on, the time-dependence of the terms will always be implicit. In addition, we will always restrict ourselves to nonnegative times , since it is possible to reverse the evolution of (2) via the transformation . :


Roughly speaking, is a perturbation of the square of the -norm of (the first two terms) by the means of two corrective quantities. First of all, let us show that the latter do not substantially modify the size of . To turn this into a rigorous statement, we begin by restricting ourselves to intervals of time on which is larger than a certain constant depending on , and we show that on such intervals,


for a suitable choice of which we precise later.

Set . We can write333The symbol is understood as refering to constants depending only on , or absolute constants, whose explicit form is not particularly meaningful.

where we used the tame estimates for products in , for . Then interpolate between and (or just bound the -norm by a constant if ), and using lemma 2.1, get

On the other side, introducing , we similarly obtain :

In sight of 2 and 3, all these estimates show that and are of lower order than . More precisely, if we now set, for real ,

we see that it suffices to request for instance that , which holds true whenever is greater than a certain . As a conclusion, (11) is proved on intervals of the form which satisfy

Now we study the evolution of on . As the -norm of is conserved, we denote by , and compute at once :

where the dot refers to the time-derivative, and commutes with for all . According to equation (2), , so

Because of the imaginary part, the first term of this sum is zero. As for the second one, it combines with the time-derivative of , and we thus have

Applying Leibniz formula, we get three terms (or only two when ) :

Each of these terms has to be estimated. The first one and the third one are more tricky, since all the time- and space-derivative are concentrated on the same function.

First term : A simplification fortunately occurs. Rewrite

and observe that . The first term then equals . Let be this new quantity. Assuming that , we can bound

Indeed, because of the equation, we have for any , so that (using again the property of the interval ). Hence with the same as above (notice that it is true even if ).

Second term : As announced, we suppose here that , and fix a . We must estimate . Using the Sobolev embedding as well as tame estimates again, write

Interpolate the -norms between et , and get finally

where .

In the same way, can be proven to be controlled by the same quantity (with the same exponents).

Third term : This term is the most delicate. We have

where the sign means that the equality is true up to terms of order (which we control in the same way as for the second term). Thus we would like to estimate , but as is real, it appears, expanding the real part, that , which combines with the time-derivative of , and finally leads to the following expression :


Now we have a very simple Leibniz lemma on the operator , which we will prove in section 2.2 :

Lemma 2.2.

Let . For any integer , there is a constant depending only on , such that for all function ,

Such a result is better than the crude - estimate, because of the exponent of the -norm (which is strictly less than as soon as ).

Consequently, expression (12) is controlled by as well. To sum up, only the second and the third term really matter, whence

First, assume . In this situation, we can incorporate the -norm into the one (see 2 of lemma 2.1), so for ,

Furthermore, remembering our estimates (11) on ,

Let . The above calculation ensures that for some depending on and ,

Now is positive. A "Gronwall’s lemma" argument (which is also known as "Osgood’s lemma") thus proves that for . Notice that the value of , i.e. of , only depends on the value of . Moreover, this inequality remains true even for outside any interval of type , so it globally holds and the first part of theorem 3 is proved. At last, the constant can be set to , i.e. , which implies the statement.

It remains to consider the case . This time, for any , and in addition, the -norm of is not bounded by a constant anymore. Using part 3 of lemma 2.1, and going on as in the previous case with an auxiliary function , we find,

for all . Osgood’s lemma then yields , and the proof of theorem 1 is complete.

2.2 A Leibniz lemma

We eventually turn to the

Proof of lemma 2.2.

Let . We intend to control the -norm of . A straightforward computation yields

The key idea is to replace by a more symmetric coefficient, and then to recognize a convolution product. More precisely, define a continuous function of the real variable with

and . For every , with , we then have , which is true even if , once we have assigned , by convention.

Now, to show that is bounded on , it suffices to check that it is bounded near and , and then to invoke the symmetry of around . And actually, since , and . Studying the variations of even show that , and hence is independent of .

As a consequence,

because of the inequality , satisfied for any , any and .

When , the result is rather immediate, since

where . Using as always the embedding , and interpolating between and , we get the result.

From here on, we suppose . We shall deal with the first part of the above sum (the last one follows identically). We consider two sequences and . With these notations,

By Schur’s lemma, . But . As for , write