Quasi-invariant Gaussian measures for the cubic fourth order nonlinear Schrödinger equation

We consider the cubic fourth order nonlinear Schrödinger equation on the circle. In particular, we prove that the mean-zero Gaussian measures on Sobolev spaces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^s({\mathbb {T}})$$\end{document}Hs(T), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s > \frac{3}{4}$$\end{document}s>34, are quasi-invariant under the flow.


Background
In this paper, we continue the program set up by the second author [70] and study the transport property of Gaussian measures on Sobolev spaces under the dynamics of a certain Hamiltonian partial differential equation (PDE).
In probability theory, there is an extensive literature on the transport property of Gaussian measures under linear and nonlinear transformations. See, for example, [3,6,19,24,25,45,62]. Classically, Cameron-Martin [19] studied the transport property of Gaussian measures under a shift and established a dichotomy between absolute continuity and singularity of the transported measure. In the context of nonlinear transformations, the work in [45,62] considers nonlinear transformations that are close to the identity, while the work in [24,25] considers the transport property under the flow generated by (non-smooth) vector fields. In particular, in [25], the existence of quasi-invariant measures under the dynamics was established under an exponential integrability assumption of the divergence of the corresponding vector field. We also note a recent work [53] establishing absolute continuity of the Gaussian measure associated to the complex Brownian bridge on the circle under certain gauge transformations.
In the field of Hamiltonian PDEs, Gaussian measures naturally appear in the construction of invariant measures associated to conservation laws such as Gibbs measures. These invariant measures associated to conservation laws are typically constructed as weighted Gaussian measures. There has been a significant progress over the recent years in this subject. See [8][9][10][11][12][14][15][16]18,27,28,30,31,[48][49][50]52,[55][56][57]61,63,65,[67][68][69]71,72,74,75]. On the one hand, in the presence of such an invariant weighted Gaussian measure, one can study the transport property of a specific Gaussian measure, relying on the mutual absolute continuity of the invariant measure and the Gaussian measure. On the other hand, the invariant measures constructed in the forementioned work are mostly supported on rough functions with the exception of completely integrable Hamiltonian PDEs such as the cubic nonlinear Schrödinger equation (NLS), the KdV equation, and the Benjamin-Ono equation [29,71,72,74,75].
These completely integrable equations admit conservation laws at high regularities, allowing us to construct weighted Gaussian measures supported on smooth functions. In general, however, it is rare to have a conservation law at a high regularity and thus one needs an alternative method to study the transport property of Gaussian measures supported on smooth functions under the dynamics of non-integrable PDEs.
In the following, we consider the cubic fourth order NLS as a model equation and study the transport property of Gaussian measures supported on smooth functions. In particular, we prove that the transported Gaussian measures and the original Gaussian measures are mutually absolutely continuous with respect to each other. Our approach combines PDE techniques such as an energy estimate and normal form reductions and probabilistic techniques in an intricate manner.

Cubic fourth order nonlinear Schrödinger equation
As a model dispersive equation, we consider the cubic fourth order nonlinear Schrödinger equation on T: where u is a complex-valued function on T × R with T = R/(2π Z). The Eq. (1.1) is also called the biharmonic NLS and it was studied in [40,66] in the context of stability of solitons in magnetic materials. The biharmonic NLS (1.1) is a special case of the following more general class of fourth order NLS: i∂ t u = λ∂ 2 x u + μ∂ 4 x u ± |u| 2 u. (1.2) The model (1.2) was introduced in [41,42] to include the effect of small fourth-order dispersion terms in the propagation of intense laser beams in a bulk medium with Kerr nonlinearity. See also [5,33,60] for the references therein. The Eq. (1.1) is a Hamiltonian PDE with the following Hamiltonian: See Appendix A for the proof. We point out that Proposition 1.1 is sharp in the sense that (1.1) is ill-posed below L 2 (T). See the discussion in Sect. A.2. See also [38,59].
Our main goal is to study the transport property of Gaussian measures on Sobolev spaces under the dynamics of (1.1).

Main result
We first introduce a family of mean-zero Gaussian measures on Sobolev spaces. Given s > 1 2 , let μ s be the mean-zero Gaussian measure on L 2 (T) with the covariance operator 2(Id − ) −s , written as While the expression dμ s = Z −1 s exp(− 1 2 u 2 H s )du may suggest that μ s is a Gaussian measure on H s (T), we need to enlarge a space in order to make sense of μ s .
The Gaussian measure μ s defined above is in fact the induced probability measure under the map 1 ω ∈ → u ω (x) = u(x; ω) = n∈Z g n (ω) n s e inx , (1.6) where · = (1 + | · | 2 ) 1 2 and {g n } n∈Z is a sequence of independent standard complexvalued Gaussian random variables, i.e. Var(g n ) = 2. Note that u ω in (1.6) lies in H σ (T) for σ < s − 1 2 but not in H s− 1 2 (T) almost surely. Moreover, for the same range of σ , μ s is a Gaussian probability measure on H σ (T) and the triplet (H s , H σ , μ s ) forms an abstract Wiener space. See [36,46].
Recall the following definition of quasi-invariant measures. Given a measure space (X, μ), we say that μ is quasi-invariant under a transformation T : X → X if the transported measure T * μ = μ • T −1 and μ are equivalent, i.e. mutually absolutely continuous with respect to each other. We now state our main result. 3 4 . Then, the Gaussian measure μ s is quasi-invariant under the flow of the cubic fourth order NLS (1.1).

Theorem 1.2 Let s >
When s = 2, one may obtain Theorem 1.2 by establishing invariance of the Gibbs measure "dρ = Z −1 exp(−H (u))du" and appealing to the mutual absolute continuity of the Gibbs measure ρ and the Gaussian measure μ 2 , at least in the defocusing case. Such invariance, however, is a very rigid statement and is not applicable to other values of s > 3 4 . Instead, we follow the approach introduced by the second author in the context of the (generalized) BBM equation [70]. In particular, we combine both PDE techniques and probabilistic techniques in an intricate manner. Moreover, we perform both local and global analysis on the phase space. An example of local analysis is an energy estimate (see Proposition 6.1 below), where we study a property of a particular trajectory, while examples of global analysis include the transport property of Gaussian measures under global transformations discussed in Sect. 4 and a change-of-variable formula (Proposition 6.6).
As in [70], it is essential to exhibit a smoothing on the nonlinear part of the dynamics of (1.1). Furthermore, we crucially exploit the invariance property of the Gaussian measure μ s under some nonlinear (gauge) transformation. See Sect. 4. In the context of the generalized BBM considered in [70], there was an obvious smoothing coming from the smoothing operator applied to the nonlinearity. There is, however, no apparent smoothing for our Eq. (1.1). In fact, a major novelty compared to [70] is that in this work we exploit the dispersive nature of the equation in a fundamental manner. Our main tool in this context is normal form reductions analogous to the approach employed in [4,37,47]. In [4], Babin-Ilyin-Titi introduced a normal form approach for constructing solutions to dispersive PDEs. It turned out that this approach has various applications such as establishing unconditional uniqueness [37,47] and exhibiting nonlinear smoothing [32]. The normal form approach is also effective in establishing a good energy estimate, though such an application of the normal form reduction in energy estimates is more classical and precedes the work of [4]. See Sect. 6.1.
In [62], Ramer proved a criterion on quasi-invariance of a Gaussian measure on an abstract Wiener space under a nonlinear transformation. In the context of our problem, this result basically states that μ s is quasi-invariant if the nonlinear part is (1 + ε)smoother than the linear part. See [45] for a related previous result. In Sect. 5, we perform a normal form reduction on the renormalized Eq. (3.6) and exhibit (1 + ε)smoothing on the nonlinear part if s > 1. This argument provides the first proof of Theorem 1.2 when s > 1. It seems that the regularity restriction s > 1 is optimal for the application of Ramer's result. See Remark 5.4.
When s ≤ 1, we need to go beyond Ramer's argument. In this case, we follow the basic methodology in [70], combining an energy estimate and global analysis of truncated measures. Due to a lack of apparent smoothing, our energy estimate is more intricate. Indeed, we need to perform a normal form reduction and introduce a modified energy for this purpose. This introduces a further modification to the argument from [70]. See Sect. 6. Lastly, let us point out the following. While the regularity restriction s > 3 4 in Theorem 1.2 comes from the energy estimate (Proposition 6.1), we expect that, by introducing some new ideas related to more refined normal form reductions developed in [37], the result may be extended to the (optimal) regularity range s > 1 2 . We plan to address this question in a future work. Remark 1.3 (i) In the higher regularity setting s > 1, we can reduce the proof of Theorem 1.2 to Ramer's result [62]. See Sect. 5. While there is an explicit representation for the Radon-Nikodym derivative in [62], we do not know how to gain useful information from it at this point (ii) In the low regularity case 3 4 < s ≤ 1, we employ the argument introduced in [70]. See Sect. 6. This argument is more quantitative and in particular, it allows us to obtain a polynomial upper bound on the growth of the Sobolev norm. However, such a polynomial growth bound may also be obtained by purely deterministic methods. See Remark 7.4 in [70]. A quasi-invariance result with better quantitative bounds may lead to an improvement of the known deterministic bounds. At the present moment, however, we do not know how to make such an idea work. (iii) We point out that the existence of a quasi-invariant measure is a qualitative statement, showing a delicate persistency property of the dynamics. In particular, this persistence property due to the quasi-invariance is stronger than the (usual) persistence of regularity. In a future work, we plan to construct Hamiltonian dynamics possessing the persistence of regularity such that the Gaussian measure μ s and the transported measure under the dynamics are mutually singular. Remark 1.4 Let us briefly discuss the situation for the related cubic (second order) NLS: It is known to be completely integrable and possesses an infinite sequence of conservation laws H k , k ∈ N ∪ {0}, controlling the H k -norm [1,2,35]. Associated to the conservation laws H k , k ≥ 1, there exists an infinite sequence of invariant weighted Gaussian measures ρ k supported on H k− 1 2 −ε (T), ε > 0 [8,74]. As mentioned above, one may combine this invariance and the mutual absolute continuity of ρ k and the Gaussian measure μ k to deduce quasi-invariance of μ k under the dynamics of (1.7), k ≥ 1. It may be of interest to investigate quasi-invariance of μ s for non-integer values of s.

Organization of the paper
In Sect. 2, we introduce some notations. In Sect. 3, we apply several transformations to (1.1) and derive a new renormalized equation. We also prove a key factorization lemma (Lemma 3.1) which play a crucial role in the subsequent nonlinear analysis. We then investigate invariance properties of Gaussian measures under several transformations in Sect. 4. In Sect. 5, we prove Theorem 1.2 for s > 1 as a consequence of Ramer's result [62]. By establishing a crucial energy estimate and performing global analysis of truncated measures, we finally present the proof of Theorem 1.2 for the full range s > 3 4 in Sect. 6. In Appendix A, we discuss the well-posedness issue of the Cauchy problem (1.1). Then, we use it to study the approximation property of truncated dynamics in Appendix B, which is used in the proof of Theorem 1.2 in Sect. 6.

Notations
Given N ∈ N, we use P ≤N to denote the Dirichlet projection onto the frequencies {|n| ≤ N } and set P >N := Id − P ≤N . Define E N and E ⊥ N by Given s > 1 2 , let μ s be the Gaussian measure on L 2 (T) defined in (1.5). Then, we can write μ s as where μ s,N and μ ⊥ s,N are the marginal distributions of μ s restricted onto E N and E ⊥ N , respectively. In other words, μ s,N and μ ⊥ s,N are induced probability measures under the following maps: respectively. Formally, we can write μ s,N and μ ⊥ s,N as Given r > 0, we also define a probability measure μ s,r by The defocusing/focusing nature of the Eq. (1.1) does not play any role, and thus we assume that it is defocusing, i.e. with the + sign in (1.1). Moreover, in view of the time reversibility of the equation, we only consider positive times in the following.

Reformulation of the cubic fourth order NLS
In this section, we apply several transformations to (1.1) and reduce it to a convenient form on which we perform our analysis. Given t ∈ R, we define a gauge transformation G t on L 2 (T) by setting where Note that G is invertible and its inverse is given by Let u ∈ C(R; L 2 (T)) be a solution to (1.1). Define u by Then, it follows from the the mass conservation that u is a solution to the following renormalized fourth order NLS: x . For simplicity of notations, we use v n to denote the Fourier coefficient of v in the following, when there is no confusion. By writing (3.4) on the Fourier side, we have v n (t) = e itn 4 u n (t).
In the remaining part of the paper, we present the proof of Theorem 1.2 by performing analysis on (3.6). In view of Lemma 3.1, we refer to the first term N (v) and the second term R(v) on the right-hand side of (3.6) as the non-resonant and resonant terms, respectively. While we do not have any smoothing on R(v) under a time integration, Lemma 3.1 shows that there is a smoothing on the non-resonant term N (v). We will exploit this fact in Sect. 5. In Sect. 6, we will exploit a similar non-resonant behavior in establishing a crucial energy estimate (Proposition 6.1).

Gaussian measures under transformations
In this section, we discuss invariance properties of Gaussian measures under various transformations. Proof Note that μ s can be written as an infinite product of Gaussian measures: where ρ n is the probability distribution for u n . In particular, ρ n is a mean-zero Gaussian probability measure on C with variance 2 n −2s . Then, noting that the action of S(t) on u n is a rotation by e −itn 4 , the lemma follows from the rotation invariance of each ρ n . Lemma 4.2 Given a complex-valued mean-zero Gaussian random variable g with variance σ , i.e. g ∈ N C (0, σ ), let T g = e it|g| 2 g for some t ∈ R. Then, T g ∈ N C (0, σ ).
Proof By viewing C R 2 , let x = (x, y) = (Re g, Im g) and u = (u, v) = (Re T g, Im T g). Noting that |T g| = |g|, we have T −1 g = e −it|g| 2 g. In terms of x and u, we have Then, with C t = cos t|u| 2 and S t = sin t|u| 2 , a direct computation yields Let μ and μ be the probability distributions for g and T g. Then, for a measurable set A ⊂ C R 2 , we have This proves the lemma.
Next, we extend Lemma 4.2 to the higher dimensional setting. While we could adapt the proof of Lemma 4.2 to the higher dimensional setting, this would involve computing determinants of larger and larger matrices. Hence, we present an alternative proof in the following.
for t ∈ R. Then, noting that ∂ t M(v(t)) = 0, where M(v(t)) = |n|≤N |v n (t)| 2 , we see that v n satisfies the following system of ODEs: With a n = Re v n and b n = Im v n , we can rewrite (4.1) as da n = −2M(v)b n dt db n = 2M(v)a n dt, |n| ≤ N . (4.2) Let L N be the infinitesimal generator for (4.2). Then, μ s,N is invariant under G t for any t ∈ R if and only if (L N ) * μ s,N = 0. See [44]. Note that the last condition is equivalent toˆ( for all test functions F ∈ C ∞ (R 2N +2 ; R). From (4.2), we have Then, by integration by parts, we havê This proves (4.3).
In the following, we assume that s > 1 2 such that μ s is a well-defined probability measure on L 2 (T) and G t defined in (3.1) makes sense on supp(μ s ) = L 2 (T).
Then, for any t ∈ R, the Gaussian measure μ s defined in Note that, when s = 1, Lemma 4.4 basically follows from Theorem 3.1 in [53] which exploits the properties of the Brownian loop under conformal mappings. For general s > 1 2 , such approach does not seem to be appropriate. In the following, we present the proof, using Lemma 4.3.
be a test function depending only on the frequencies {|n| ≤ N }. Then, we claim that With a slight abuse of notations, we write By the independence of {g n } |n|≤N and {g n } |n|>N , we can write = 0 × 1 such that where I N (ω 1 ) is given by < ∞ almost surely. For fixed ω 1 ∈ 1 , define { g ω 1 n } |n|≤N by setting g ω 1 n = e 2itμ(ω 1 ) g n , |n| ≤ N . Then, by the rotational invariance of the standard complex-valued Gaussian random variables and independence of {g n } |n|≤N and {g n } |n|>N , we see that, for almost every ω 1 ∈ 1 , { g ω 1 n } |n|≤N is a sequence of independent standard complex-valued Gaussian random variables (in ω 0 ∈ 0 ). In particular, the law of { g ω 1 n } |n|≤N is the same as that of {g n } |n|≤N , almost surely in ω 1 ∈ 1 . Then, from the definitions of μ s,N and G t , we can rewrite (4.7) as for almost every ω 1 ∈ 1 , where u N = P ≤N u is as in (2.2). Then, it follows from Lemma 4.3 with (4.5) and (2.1) that for almost every ω 1 ∈ 1 . Note that the right-hand side of (4.8) is independent of ω 1 ∈ 1 . Therefore, from (4.6) and (4.8), we havê This proves (4.4). Next, given almost surely with respect to μ s . Then, from the dominated convergence theorem and (4.4), we havê for all F ∈ C b (L 2 (T); R). Hence, the lemma follows (see, for example, [26, Proposition 1.5]).
Lastly, we conclude this section by stating the invariance property of quasiinvariance under a composition of two maps.

Lemma 4.5 Let (X, μ) be a measure space. Suppose that T 1 and T 2 are maps on X into itself such that μ is quasi-invariant under T j for each
Conversely, if T * μ(A) = 0, then we have μ(T −1 1 A) = 0, which in turn implies μ(A) = 0. Hence, μ and T * μ are mutually absolutely continuous.
As it is written, the Eq. (1.1) or (3.6) does not manifest a smoothing in an explicit manner. In the following, we perform a normal form reduction and establish a nonlinear smoothing by exploiting the dispersion of the equation.

Normal form reduction
By writing (3.6) (5.1) Lemma 3.1 states that we have a non-trivial (in fact, fast) oscillation caused by the phase function φ(n) in the non-resonant part N(v). The main idea of a normal form reduction is to transform the non-resonant part N(v) into smoother terms of higher degrees, exploiting this rapid oscillation. More concretely, integrating by parts, we formally have In view of Lemma 3.1, the phase function φ(n) appearing in the denominators allows us to exhibit a smoothing in N(v). See Lemma 5.1 below. At this point, the computation in (5.2) is rather formal and thus requires justification in several steps. In the first step, we switched the order of the time integration and the summation: is absolutely convergent with a bound uniform in time t , provided that v ∈ C(R; H s (T)) with s ≥ 1 6 . This justifies (5.3). If v ∈ C(R; H s (T)) with s ≥ 1 6 , it follows from (3.6) and a computation similar to (5.4) that v n ∈ C 1 (R). This allows us to apply integration by parts and the product rule. Lastly, we need to justify the switching of the time integration and the summation in the last equality of (5.2). By crudely estimating with (3.6), (5.4) and Lemma 3.1 (note that |φ(n)| ≥ 1 on (n)), we have Hence, the series on the left-hand side of (5.5) is absolutely convergent with a bound uniform in time t , provided that v ∈ C(R; H s (T)) with s ≥ 1 6 . This justifies the last equality in (5.2).
The following lemma shows a nonlinear smoothing for (3.6). Note that the amount of smoothing for R(v) depends on the regularity s > 1 2 .
Proof By Lemma 3.1 and the algebra property of H s (T), s > 1 2 , we have The second term II in (5.2) can be estimated in an analogous manner. Similarly, by Lemma 3.1, (3.6), and the algebra property of H s (T), s > 1 2 , we have The fourth term IV in (5.2) can be estimated in an analogous manner. From (5.1) and 2 n ⊂ 6 n , we have This proves the second estimate (5.7).

Consequence of Ramer's result
In this subsection, we present the proof of Theorem 1.2 for s > 1. The main ingredient is Ramer's result [62] along with the nonlinear smoothing discussed in the previous subsection. We first recall the precise statement of the main result in [62] for readers' convenience.
Then, μ and μ • T are mutually absolutely continuous measures on U .
Here, H S(H ) denotes the space of Hilbert-Schmidt operators on H and G L(H ) denotes invertible linear operators on H with a bounded inverse.
Given t, τ ∈ R, let (t): L 2 → L 2 be the solution map for (1.1) and (t, τ ): L 2 → L 2 be the solution map for (3.6), 2 sending initial data at time τ to solutions at time t. When τ = 0, we may denote (t, 0) by (t) for simplicity.
By inverting the transformations (3.2) and (3.4) with (5.1), we have where and v is the solution to (3.6) with v| t=0 = u 0 . In view of Lemmas 4.1, 4.4, and 4.5, it suffices to show that μ s is quasi-invariant under (t). Fix s > 1 and σ 1 > 1 2 sufficiently close to 1 2 . First, note that μ s is a probability measure on H s−σ 1 (T). Given R > 0, let B R be the open ball of radius R centered at the origin in H s−σ 1 (T). The following proposition shows that the hypotheses of Ramer's result in [62] are indeed satisfied.
, the following statements hold: We first present the proof of Theorem 1.2 for s > 1, assuming Proposition 5.3. Thanks to Ramer's result (Proposition 5.2 above), Proposition 5.3 implies that μ s and the pullback measure (t) * μ s := μ s • (t) are mutually absolutely continuous as measures restricted to the ball B R for any t ∈ (0, τ (R)]. 1 be the open ball of radius R centered at the origin as above. Fix T > 0. It follows from the growth Given R > 0, let R * be as in (5.9). Then, from (5.9), we have Now, letting R → ∞, it follows from the continuity from below of a measure that μ s ( (T )(A)) = 0. Note that the choice of T was arbitrary. In view of the time reversibility of the Eq. (3.6), we conclude that μ s is quasi-invariant under the flow of (3.6). Therefore, Theorem 1.2 follows from (5.8) with Lemmas 4.1, 4.4, and 4.5. The remaining part of this section is devoted to the proof of Proposition 5.3. The claim (i) follows from the well-posedness of (3.6) in H s−σ 1 . In particular, the continuity of (t) on H s−σ 1 with the time reversibility implies (i). As before, from (A.12), we have the uniform growth bound: for all solutions v to (3.6) with v| t=0 = u 0 ∈ B R . Then, the claim (ii) follows from Lemma 5.1 and the continuity of (t) on H s−σ 1 .
We postpone the proof of the claim (iii) and first prove the claim (iv). For fixed Then, by computing a derivative of F at the origin, we have 3 This is clearly a linear map. Moreover, the boundedness of D F| 0 on H s follows from the claim (iii). Note that F is invertible with the inverse F −1 given by Hence, it follows from the chain rule that Hence, we proved the claim (iv) except for the boundedness of (D F| 0 ) −1 . We will prove the boundedness of (D F| 0 ) −1 at the end of this section. Next, we prove the claim (iii). In the following, we will prove that DK (t)| u 0 is Hilbert-Schmidt on H s for u 0 ∈ B R ⊂ H s−σ 1 as long as t = t (R) 1. Given u 0 ∈ B R ⊂ H s−σ 1 , let v be the global solution to (3.6) with v| t=0 = u 0 .
We first introduce some notations. Given a multilinear 4    Let w(t) be a solution to the following linear equation: (5.14) Given (m 1 , m 2 , m 3 ) ∈ Z 3 and n ∈ Z, we use the following shorthand notation: Then, by a direct computation with (3.6), (5.1), and (5.2), we have where φ(n) and (n) are as in (3.7) and (3.8), respectively. Fix σ 2 > 1 2 (to be chosen later) and write where A t (w(0)) is given by Applying Young's inequality and 2 n ⊂ 6 n to (5.16) with Lemma 3.1 and (5.17), we have In particular, by choosing τ = τ (R) > 0 sufficiently small, we obtain Therefore, A t is bounded on H s and hence DK (t)| u 0 is a Hilbert-Schmidt operator on H s for all t ∈ [0, τ ]. The second claim in (iii) basically follows from the continuous dependence of (3.6) and (5.14) (in v) and thus we omit details. It remains to prove the boundedness of (D F| 0 ) −1 = (Id H s + DK (t)| u 0 ) −1 By the time reversibility of the equation and (5.13), the argument above shows that (D F| 0 ) −1 − Id H s is Hilbert-Schmidt on H s by choosing τ = τ (R) sufficiently small. In particular, (D F| 0 ) −1 is bounded on H s . This completes the proof of Proposition 5.3.

Remark 5.4
The condition s > 1 is necessary for this argument. In estimating the resonant term, i.e. the first term in (5.16) by the H s−σ 1 -norms of its arguments, we need to use the second condition s+σ 2 3 ≤ s − σ 1 in (5.17). Thus, we must have In this section, we present the proof of Theorem 1.2 for s > 3 4 . The basic structure of our argument follows the argument introduced in [70] by the second author in the context of the (generalized) BBM equation, with one importance difference. While the energy estimate in [70] was carried out on the H s -norm of solutions (to the truncated equations), we carry out our energy estimate on a modified energy. This introduction of a modified energy is necessary to exhibit a hidden nonlinear smoothing, exploiting the dispersion of the equation. See Proposition 6.1 below. This, in turn, forces us to work with the weighted Gaussian measure ρ s,N ,r,t and ρ s,r,t adapted to this modified energy, instead of the Gaussian measure μ s,r with an L 2 -cutoff. See (6.21) and (6.22) below for the definitions of ρ s,N ,r,t and ρ s,r,t . Lastly, we point out that this usage of the modified energy is close to the spirit of higher order modified energies in the I -method introduced by Colliander-Keel-Staffilani-Takaoka-Tao [22,23].
As in Sect. 5, we carry out our analysis on (3.6). Let us first introduce the following truncated approximation to (3.6): where N (n) is defined by A major part of this section is devoted to the study of the dynamical properties of (6.1). Note that (6.1) is an infinite dimensional system ODEs for the Fourier coefficients {v n } n∈Z , where the flow is constant on the high frequencies {|n| > N }. We also consider the following finite dimensional system of ODEs: Given t, τ ∈ R, denote by N (t, τ ) and N (t, τ ) the solution maps of (6.1) and (6.3), sending initial data at time τ to solutions at time t, respectively. For simplicity, we set when τ = 0. Then, we have the following relations: N (t, τ ) = N (t, τ )P ≤N + P >N and P ≤N N (t, τ ) = N (t, τ )P ≤N . (6.5)

Energy estimate
In this subsection, we establish a key energy estimate. Before stating the main proposition, let us first perform a preliminary computation. Given a smooth solution u to (1.1), let v be as in (3.4). Then, from (3.6), we have Then, differentiating by parts, i.e. integrating by parts without an integral sign, 5 we obtain This motivates us to define the following quantity. Given s > 1 2 , define the modified Then, we have the following energy estimate. 3 4 . Then, for any sufficiently small ε > 0, there exist small θ > 0 and C > 0 such that

Proposition 6.1 Let s >
for all N ∈ N and any solution v to (6.1), uniformly in t ∈ R.
Recall that the probability measures μ s and μ s,r defined in (1.5) and (2.5) are supported on H s− 1 2 −ε (T) for any ε > 0, while we have v L 2 ≤ r in the support of μ s,r .
Before proceeding to the proof of this proposition, recall the following arithmetic fact [39]. Given n ∈ N, the number d(n) of the divisors of n satisfies d(n) ≤ C δ n δ (6.10) for any δ > 0.
Proof Let v be a solution to (6.1). Then, from (6.7) and (6.8) with (6.1), we have 11) where N j (v) and R j (v), j = 1, 2, 3, are defined by  .15). For simplicity of the presentation, we drop the restriction on the summations in (6.12) with the understanding that v n = 0 for |n| > N . Moreover, we can assume that all the Fourier coefficients are non-negative. In the following, we establish uniform (in t) estimates for these multilinear terms N j and R j , j = 1, 2, 3. For simplicity, we suppress the t-dependence with the understanding that all the estimates hold with implicit constants independent of t ∈ R. Given n, μ ∈ Z, define (n, μ) by Then, given δ > 0, it follows from the divisor counting estimate (6.10) that In the following, we use (6.13) to estimate N j (v) and R j (v), j = 1, 2, 3. For simplicity of the presentation, we drop multiplicative constants depending on δ > 0. We now estimate N 1 (v). We first consider the case s < 1. By Sobolev's inequality and interpolation, we have for small γ > 0 and some θ = θ(γ ) > 0. Then, by Lemma 3.1 and Cauchy-Schwarz inequality (in n and then in n 2 , n 3 ) with (6.14), we have for small γ, δ > 0 such that 5 − 4s − 2δ > 1, where n max := max(|n|, |n 1 |, |n 2 |, |n 3 |). From the divisor counting argument (6.13), we have for sufficiently small δ, ε, γ > 0, provided that s > 3 4 .

Weighted Gaussian measures
Our main goal in this subsection is to define weighted Gaussian measures adapted to the modified energy E t (P ≤N v) and E t (v) defined in the previous section. Given N ∈ N and r > 0, define F N ,r,t (v) and F r,t (v) by where R t is defined in (6.8). Then, we would like to construct probability measures ρ s,N ,r,t and ρ s,r,t of the form: 6 and dρ s, The following proposition shows that they are indeed well defined probability measures on H s− 1 2 −ε (T), ε > 0.
Proposition 6.2 Let s > 1 2 and r > 0. Then, F N ,r,t (v) ∈ L p (μ s ) for any p ≥ 1 with a uniform bound in N ∈ N and t ∈ R, depending only on p ≥ 1 and r > 0. Moreover, for any finite p ≥ 1, F N ,r,t (v) converges to F r,t (v) in L p (μ s ), uniformly in t ∈ R, as N → ∞.
In the following, we restrict our attention to s > 1 2 . Hence, we view ρ s,N ,r,t and ρ s,r,t as probability measures on L 2 (T).
Let μ s,r be as in (2.5). Then, it follows from Proposition 6.2 that ρ s,r,t and μ s,r are mutually absolutely continuous. Moreover, we have the following 'uniform convergence' property of ρ s,N ,r,t to ρ s,r,t . The proof of Proposition 6.2 follows closely Bourgain's argument in constructing Gibbs measures [8]. We first recall the following basic tail estimate. See [ Proof of Proposition 6.2 Fix r > 0. We first prove F N ,r,t L p (μ s ) , F r,t L p (μ s ) ≤ C p,r < ∞ (6.23) for all N ∈ N and t ∈ R. From the distributional characterization of the L p -norm and (6.20), we have In the following, we estimate μ s |R t (v)| ≥ K , v L 2 ≤ r for K ≥ 1, using the dyadic pigeon hole principle and Lemma 6.4. Let us divide the argument into two cases: s > 1 and 1 2 < s ≤ 1. Note that, while R t depends on t ∈ R, all the estimates below hold uniformly in t ∈ R.
First, suppose that s > 1. Then, from Lemma 3.1 and the divisor counting argument as in the proof of Proposition 6.1 (see (6.19)), we have For j ∈ N, let M j = 2 j M 0 and σ j = C ε 2 −εj for some small ε > 0 such that j∈N σ j = 1 2 . Then, from (6.24) and (6.25), we have 1 2 j . Here, we used that r −2 K ∼ M s−1 0 r 2 1 in view of (6.26). Then, applying Lemma 6.4 with (6.26), we obtain This proves (6.23) for F r,t when s > 1. A similar argument holds for F N ,r,t with a uniform bound in N ∈ N. Next, suppose that 1 2 < s ≤ 1. Proceeding with Lemma 3.1 as before, we have under v L 2 ≤ r . Hence, (6.23) trivially follows in this case. It remains to show that F N ,r,t converges to F r,t in L p (μ s ). It follows from a small modification of (6.24) and (6.27) that R t (P ≤N v) converges to R t (v) almost surely with respect to μ s , uniformly in t ∈ R. Indeed, when s > 1, we have Hence, F N ,r,t converges to F r,t almost surely with respect to μ s . As a consequence of Egoroff's theorem, we see that F N ,r,t converges to F r,t almost uniformly and hence in measure (uniformly in t ∈ R). Namely, given ε > 0, if we let Then, by Cauchy-Schwarz inequality and (6.23), we have for all sufficiently large N ∈ N, uniformly in t ∈ R. Therefore, F N ,r,t converges to F r,t in L p (μ s ) for any p ≥ 1.
We conclude this subsection by stating a large deviation estimate on the quantity appearing in the energy estimate (Proposition 6.1). Here, the second to the last inequality follows from the hypercontractvity estimate due to Nelson [54,Theorem 2]. See also [17, Lemma 3.1].

A change-of-variable formula
In this subsection, we establish an important change-of-variable formula (Proposition 6.6). It is strongly motivated by the work [71,72]. We closely follow the argument presented in [70].
Given N ∈ N, let d L N = |n|≤N d u n denote the Lebesgue measure on C 2N +1 . Then, from (6.20) and (6.21) with (2.4), we have where Z s,N ,r is a normalizing constant defined by 7 Then, we have the following change-of-variable formula: Proposition 6.6 Let s > 1 2 , N ∈ N, and r > 0. Then, we have for any t, τ ∈ R and any measurable set A ⊂ L 2 .
We first state the basic invariance property of L N . Proof The finite dimensional system (6.3) basically corresponds to the finite dimensional Hamiltonian approximation to (1.1) under two transformations (3.2) and (3.4). Therefore, morally speaking, the lemma should follow from the inherited Hamiltonian structure and Liouville's theorem. In the following, however, we provide a direct proof. Write (6.3) as ∂ t v n = X n , |n| ≤ N . Then, by Liouville's theorem, it suffices to show |n|≤N ∂Re X n ∂Re v n + ∂Im X n ∂Im v n = 0, or equivalently, |n|≤N ∂ X n ∂v n + ∂ X n ∂v n = 0. (6.29) Note that the first sum in (6.3) does not have any contribution to (6.29) due to the frequency restriction n 1 , n 3 = n. Hence, we have ∂ X n ∂v n + ∂ X n ∂v n = 2i|v n | 2 − 2i|v n | 2 = 0 for each |n| ≤ N . Therefore, (6.29) holds.
We now present the proof of Proposition 6.6.

On the evolution of the truncated measures
In this subsection, we establish a growth estimate on the truncated measure ρ s,N ,r,t .
The key ingredients are the energy estimate (Proposition 6.1), the large deviation estimate (Lemma 6.5), and the change-of-variable formula (Proposition 6.6) from the previous subsections.
As in [69,71,72], the main idea of the proof of Lemma 6.8 is to reduce the analysis to that at t = 0.
As a corollary to Lemma 6.8, we obtain the following control on the truncated measures ρ s,N ,r,t . Lemma 6.9 Let s > 3 4 . Then, given t ∈ R, r > 0, and δ > 0, there exists C = C(t, r, δ) > 0 such that

r,t (A)
1−δ for any N ∈ N and any measurable set A ⊂ L 2 (T).
Proof Given R > 0, let B R denote the ball of radius R centered at the origin in L 2 (T). We first consider the case when A is compact in L 2 and A ⊂ B R for some R > 0. It follows from Proposition 6.21 and Corollary 6.3 that, given ε, γ > 0, there exists for any N ≥ N 0 . Then, by Lemma 6.9 and Corollary 6.3, we have Hence, by taking a limit of (6.37) as ε, γ → 0 (with the continuity from above of a probability measure), we obtain (6.36) in this case. Next, let A be any measurable set in L 2 . Then, by the inner regularity of ρ s,r,t , there exists a sequence {K j } j∈N of compact sets such that K j ⊂ (t)(A) and ρ s,r,t ( (t)(A)) = lim j→∞ ρ s,r,t (K j ). (6.38) By the bijectivity of (t, τ ), we have Note that (0, t)(K j ) is compact since it is the image of a compact set K j under the continuous map (0, t). Moreover, we have (0, t)( Then, by (6.36) applied to (0, t)(K j ), we have By taking a limit as j → ∞, we obtain (6.36) from (6.38) and (6.39).
Finally, we present the proof of Theorem 1.2.
Proof of Theorem 1.2 As in Sect. 5, it follows from Lemmas 4.1, 4.4, and 4.5 that it suffices to prove that μ s is quasi-invariant under (t), i.e. the dynamics of (3.6). Fix t ∈ R. Let A ⊂ L 2 (T) be a measurable set such that μ s (A) = 0. Then, for any r > 0, we have In this appendix, we discuss the well-posedness issue for the Cauchy problem (1.1). In particular, we prove global well-posedness of (1.1) in L 2 (T) and ill-posedness below L 2 (T).

Well-posedness in L 2 (T)
We say u is a solution to (1.1) if u satisfies the following Duhamel formulation x . The main result of this section is the following local wellposedness of (1.1). See Remark 6.12 below for the precise uniqueness statement. Once we prove Proposition 6.11, global well-posedness (Proposition 1.1) follows from the conservation of mass (1.4). We prove Proposition 6.11 via the Fourier restriction norm method [7]. While the argument is standard, we present the details of the proof for the sake of completeness. Given s, b ∈ R, define X s,b as the completion of S(T × R) under the following norm: Given a time interval I ⊂ R, we define the local-in-time version X s,b I restricted to the time interval I by setting Before presenting the proof of Proposition 6.11, we first go over preliminary lemmas. Let η ∈ C ∞ c (R) be a smooth cut off function such that η(t) ≡ 1 for |t| ≤ 1 and η(t) ≡ 0 for |t| ≥ 2. Given T > 0, set η T (t) = η(T −1 t). Then, we have the following basic linear estimates. See [7,34,43,64] Lemma 6.13 Let s ∈ R.
(i) For any b ∈ R, we have Next, we state the L 4 -Strichartz estimate. Now, we are ready to prove Proposition 6.11.
Proof of Proposition 6.11 Let u 0 ∈ L 2 (T). Given 0 < T ≤ 1, let Let b > 1 2 and small δ > 0. Then, from Lemma 6.13, a duality argument, Hölder's inequality, and Lemma 6.14, we have as long as b ≤ 11 16 − δ. Similarly, we have Hence, it follows from (A.9) and (A.10) that is a contraction on some ball in X 0,b as long as T = T ( u 0 L 2 ) > 0 is sufficiently small. Now, suppose that u 0 ∈ H s (T) for some s > 0. Then, proceeding as in (A.9) with T = T ( u 0 L 2 ) > 0 as above, we have yielding (A.1). A similar argument yields local Lipschitz dependence of the solution map on H s (T). This completes the proof of Proposition 6.11.
Remark 6.15 When s = 0, the conservation of mass yields u(t) L 2 = u 0 L 2 for all t ∈ R. Now, suppose that s > 0. Then, by iterating (A.1) along with the mass conservation, we conclude that there exists θ > 0 such that the following growth estimate on the H s -norm holds: for any τ > 0 and all u 0 ∈ H s (T) with u 0 L 2 ≤ K .

Ill-posedness below L 2 (T)
In the following, we briefly discuss the ill-posedness of (1.1) below L 2 (T). We first present the following failure of uniform continuity of the solution map on bounded sets below L 2 (T). Then, it is easy to see that u (N ,a) is a smooth global solution to (1.1).
Given n ∈ N, let u 0,n = u (N n ,1) (0) and u 0,n = u (N n ,1+ 1 n ) (0), where N n ∈ N is to be chosen later. Then, we have u 0,n H s , u 0,n H s 1 (A.13) uniformly in n ∈ N. Moreover, we have (A.14) Note that (A.13) and (A.14) hold independently of a choice of N n ∈ N. Let u n and u n be the solutions to (1.1) with initial data u n | t=0 = u 0,n and u n | t=0 = u 0,n , respectively. Namely, u n = u (N n ,1) and u n = u (N n ,1+ 1 n ) . Given n ∈ N, define t n > 0 by Since s < 0, we can choose N n ∈ N sufficiently large such that t n ≤ 1 n . Then, we have Noting that t n → 0 as n → ∞, Lemma 6.16 follows from (A.13), (A.14), and (A.15).

Remark 6.17
The cubic NLS (1.7) enjoys the Galilean symmetry, which preserves the L 2 -norm. Namely, L 2 is critical with respect to the Galilean symmetry. Indeed, the cubic NLS is known to be ill-posed below L 2 (T). See [13,20,21,38,51]. As for the fourth order NLS (1.1), there seems to be no Galilean symmetry 10 and it is not clear why the regularity s = 0 plays a role as a critical value. H s (T)) as n → ∞.
Note that this is one of the strongest forms of ill-posedness.
In the following, we present a sketch of the argument. See [38,59] for details. Namely, first use the short time Fourier restriction norm method and establish an a priori bound in H s , s < 0, to the renormalized Eq. (3.3). Here, the main observation is that Lemma 3.1 guarantees that the a priori bound for the renormalized cubic NLS also holds for the renormalized fourth order NLS (3.3). 11

Appendix B. On the approximation property of the truncated dynamics
In this appendix, we perform further analysis on the Eq. (3.6) and its truncated approximation (6.1) and establish a certain approximation property. See Proposition 6.21 below.
Given N ∈ N, we first consider the following approximation to (1.1): (A. 16) We first study the approximation property of (A.16) to (1.1). By a slight modification of the proof of Proposition 6.11, it is easy to see that (A.16) is globally well-posed in L 2 (T). Let (t) and N (t) be the solution maps to (1.1) and (A.16), respectively. Given R > 0, let B R be the ball of radius R centered at the origin in L 2 (T). Let τ > 0. By iterating the local-in-time argument (see (A.11)), we have the following uniform estimate: for all u 0 ∈ A and N ≥ N 0 .
Proof By the continuity of the map: u 0 ∈ L 2 → (t)u 0 ∈ L 4 ([0, τ ]; L 4 x ) and the compactness of K , we see that (t)K is compact in L 4 ([0, τ ]; L 4 x ). Hence, there exists a finite index set J and {u 0, j } j∈J ⊂ K such that, given u 0 ∈ K , we have We first establish the following approximation property of (A.16) to (1.1).

Lemma 6.20
Given R > 0, let A ⊂ B R be a compact set in L 2 (T). Then, for any τ > 0 and ε > 0, there exists N 0 ∈ N such that for all u 0 ∈ A and N ≥ N 1 . It remains to control P >N N (t)(u 0 ). Recall that the solution map N (t) to (A.16) is locally uniformly continuous. Moreover, it follows from a slight modification of the proof of Proposition 6.11 that the modulus of continuity is uniform in N ∈ N. Hence, for any ε > 0, there exists δ > 0 such that if u 0 , u 1 ∈ B R satisfies u 0 − u 1 L 2 < δ, then we have for all N ∈ N. By the compactness of A, we can cover A by finitely many ball of radius δ centered at u 0, j , j = 1, . . . , J for some J < ∞ such that, given u 0 ∈ A, there exists j ∈ {1, . . . , J } such that We conclude this appendix by establishing the following approximation property of (6.1) to (3.6). Lemma 6.20 and (A.29) play an important role. Proposition 6.21 Given R > 0, let A ⊂ B R be a compact set in L 2 (T). Fix t ∈ R. Then, for any ε > 0, there exists N 0 = N 0 (t, R, ε) ∈ N such that we have By the mean value theorem with (3.2) and (A.28) followed by Lemma 6.20 and the unitarity of (t), we have for all sufficiently large N 1, uniformly in u 0 ∈ A ⊂ B R . This proves (A.30).