Abstract
We consider a damped/driven nonlinear Schrödinger equation in \(\mathbb {R}^n\), where n is arbitrary, \({\mathbb {E}}u_t-\nu \Delta u+i|u|^2u=\sqrt{\nu }\eta (t,x), \quad \nu >0,\) under odd periodic boundary conditions. Here \(\eta (t,x)\) is a random force which is white in time and smooth in space. It is known that the Sobolev norms of solutions satisfy \( \Vert u(t)\Vert _m^2 \le C\nu ^{-m}, \) uniformly in \(t\ge 0\) and \(\nu >0\). In this work we prove that for small \(\nu >0\) and any initial data, with large probability the Sobolev norms \(\Vert u(t,\cdot )\Vert _m\) with \(m>2\) become large at least to the order of \(\nu ^{-\kappa _{n,m}}\) with \(\kappa _{n,m}>0\), on time intervals of order \(\mathcal {O}(\frac{1}{\nu })\). It proves that solutions of the equation develop short space-scale of order \(\nu \) to a positive degree, and rigorously establishes the (direct) cascade of energy for the equation.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this work we study a damped/driven nonlinear Schrödinger equation
i.e. a CGL equation without linear dispersion, with cubic Hamiltonian nonlinearity and a random forcing. The dimension n is any, \(0<\nu \le 1\) is the viscosity constant and the random force \(\eta \) is white in time t and regular in x. The equation is considered under the odd periodic boundary conditions,
The latter implies that u vanishes on the boundary of the cube of half-periods \(K^{n}= [0, \pi ]^n\),
We denote by \(\{\varphi _d(\cdot ),\; d=(d_1,\dots ,d_n)\in {\mathbb {N}}^n\}\) the trigonometric basis in the space of odd periodic functions,
The basis is orthonormal with respect to the normalised scalar product \(\langle \!\langle \cdot ,\cdot \rangle \!\rangle \) in \(L_2(K^{n}, \pi ^{-n} dx)\),
where \(\langle \cdot , \cdot \rangle \) is the real scalar product in \({\mathbb {C}}\), \(\langle u, v \rangle =\mathfrak {R}u{\bar{v}}\). It is formed by eigenfunctions of the Laplacian:
The force \(\eta (t,x)\) is a random field of the form
Here \(\beta _d(t)=\beta ^R_d(t)+i\beta ^I_d(t)\), where \(\beta _d^R(t)\), \(\beta _d^I(t)\) are independent real-valued standard Brownian motions, defined on a complete probability space \((\Omega ,{\mathcal {F}},{\mathbb {P}})\) with a filtration \(\{{\mathcal {F}}_t;t\geqslant 0\}\). The set of real numbers \(\{b_d,\;d\in {\mathbb {N}}^n\}\) is assumed to form a non-zero sequence, satisfying
where for a real number k we set
For \(m\ge 0\) we denote by \(H^m\) the Sobolev space of order m, formed by complex odd periodic functions, equipped with the homogeneous norm,
where \(\Vert \cdot \Vert _0\) is the \(L^2\)-norm on \(K^{n}\), \( \Vert u\Vert ^2_0 = \langle \!\langle u,u\rangle \!\rangle \) (see (1.2)). If we write \(u\in H^m\) as Fourier series, \(u(x)=\sum _{d\in {\mathbb {N}}^n}u_d\varphi _d(x),\) then \( \Vert u\Vert _m^2=\sum _{d\in {\mathbb {N}}^n}|d|^{2m}| u_d|^2. \)
Equation (1.1) with small \(\nu \) belongs to a group of equations, describing turbulence in the CGL equations. These equations have got quite a lot of attention in physical literature as models for turbulence in various media, e.g. see [3, Chapter 5]. In particular – as a natural model for hydrodynamical turbulence since Eq. (1.1) is obtained from the Navier-Stokes system by replacing the Euler term \((u\cdot \nabla ) u \), which is a quadratic Hamiltonian nonlinearity, by \(i|u|^2u\), which is a cubic Hamiltonian nonlinearity, see [13].
The global solvability of Eq. (1.1) for any space dimension n is established in [8, 10]. It is proved there that if
where \(u_0\in H^m\cap C(K^n)\), \(m\in {\mathbb {N}}\), and if \(B_m<\infty \), then the problem (1.1), (1.5) has a unique strong solution u(t, x) in \(H^m\) which we write as \(u(t,x;u_0)\), or \(u(t;u_0)\), or \(u_\nu (t;u_0)\). Its norm satisfies
where \(C_m\) depends on \(\Vert u_0\Vert _m, |u_0|_\infty \) and \(B_m, B_{m_*}\). Furthermore, denoting by \(C_0(K^{n})\) the space of continuous complex functions on \(K^{n}\), vanishing at \(\partial K^{n}\), we have that the solutions u(t, x) define a Markov process in \(C_0(K^{n})\). Moreover, if the noise \(\eta (t,\cdot )\) is non-degenerate in the sense that in (1.3) all coefficients \(b_d\) are non-zero, then this process is mixing.Footnote 1
Our goal is to study the growth of higher Sobolev norms for solutions of Eq. (1.1) as \(\nu \rightarrow 0\) on time intervals of order \({\mathcal {O}}(\frac{1}{\nu })\). The main result of this work is the following.
Theorem 1
For any real number \(m>2\), in addition to (1.4), assume that \(B_m<\infty \). Then there exists \(\kappa _{n,m}>0\) such that for every fixed quadruple \((\delta ,\kappa ,{\mathscr {K}}, T_0)\), where
there exists a \(\nu _0>0\) with the property that if \(0<\nu \le \nu _0\), then for every \(u_0\in H^m \cap C_0(K^{n})\), satisfying
the solution \(u(t,x;u_0)\) is such that
-
(1)
$$\begin{aligned} {\mathbb {P}}\big \{ \sup _{t \in [t_0, t_0+ T_0\nu ^{-1}]} \Vert u_\nu ^\omega (t)\Vert _m> \nu ^{-m\kappa }\big \} \ge 1-\delta , \quad \forall \, t_0\ge 0. \end{aligned}$$
-
(2)
If m is an integer, \(m\ge 3\), then a possible choice of \(\kappa _{n,m}\) is \(\kappa _{n,m}= \tfrac{1}{35}\), and there exists \(C\ge 1\), depending on \(\kappa < \tfrac{1}{35}\), \({\mathscr {K}}, m,B_{m_*}\) and \(B_m\), such that
$$\begin{aligned} C^{-1} \nu ^{-2m\kappa +1}\leqslant {\mathbb {E}}\left( \nu \int _{t_0}^{t_0+\nu ^{-1}}\Vert u_\nu (s)\Vert _m^2ds\right) \leqslant C \nu ^{-m}, \quad \forall \, t_0\ge 0. \end{aligned}$$(1.7)
A similar result holds for the classical \(C^k\)-norms of solutions:
Proposition 2
For any integer \(m \ge 2\) in addition to (1.4) assume that \(B_m<\infty \). Then for every fixed triplet \(K, {\mathcal {K}}, T_0>0\) and any \(0<\kappa <1/16\) we have
for each \(t_0\ge 0\), if \(u_0\) satisfies \(|u_0|_\infty \le {\mathcal {K}}\), \(|u_0|_{C^m} \le \nu ^{-\kappa m}\). The rate of convergence depends only on the triplet and \(\kappa \).
For a proof of this result see the extended version of our work [6]. Due to (1.8), for any \(m>2+n/2\) we have
for every \(K>0\) and \(0<\kappa <1/16\), where for \(a\in {\mathbb {R}}\) we denote \( \lfloor a \rfloor = \max \{ n\in {\mathbb {Z}}: n<a\}. \) This improves the first assertion of Theorem 1 for large m.
We have the following two corollaries from Theorem 1, valid if the Markov process defined by the Eq. (1.1) is mixing:
Corollary 3
Assume that \(B_m<\infty \) for all m and \(b_d\ne 0\) for all d. Then Eq. (1.1) is mixing and for any \(\kappa <1/35\) and \(0<\nu \le \nu _0\) its unique stationary measure \(\mu _\nu \) satisfies
Here C and \(\nu _0\) are as in Theorem 1.
Corollary 4
Under the assumptions of Corollary 3, for any \(u_0\in C^\infty \) we have
if \(s\ge T(\nu , u_0, \kappa , B_m, B_{m_*})\), where C is the same as in (1.9).
Theorem 1 rigorously establishes the energy cascade to high frequencies for solutions of Eq. (1.1) with small \(\nu \). Indeed, if \(u_0(x)\) and \(\eta (t,x)\) are smooth functions of x (or even trigonometric polynomials of x), then in view of (1.7) for \(0<\nu \ll 1\) and \(t > rsim \nu ^{-1}\) a substantial part of the energy \( \frac{1}{2} \sum |u_d(t)|^2 \) of a solution \(u(t,x;u_0)\) is carried by high modes \(u_d\) with \(|d|\gg 1\). Relation (1.7) (valid for all integer \(m\ge 3\)) also means that the averaged in time space-scale \(l_x\) of solutions for (1.1) satisfies \( l_x \in [\nu ^{1/2}, \nu ^{1/35}], \) and goes to zero with \(\nu \) (see [1, 9]). We recall that the energy cascade to high frequencies and formation of short space-scale is the driving force of the Kolmogorov theory of turbulence, see [5].
We mention that in the work [12] the stochastic CGL equation
with linear dispersion and white in time random force \(\eta \) as in (1.3) was considered under the odd periodic boundary conditions, and the inviscid limiting dynamics as \(\nu \rightarrow 0\) was examined. However, since the limiting Eq. (1.10)\({}_{\nu =0}\) is a regular PDE in difference with the Eq. (1.1)\({}_{\nu =0}\), the results on the inviscid limit in [12] differ in spirit from those in our work, and we do not discuss them now.
Deterministic versions of the result of Theorem 1 for Eq. (1.1) with \(\eta =0\), where \(\nu \) is a small non-zero complex number such that \(\mathfrak {R}\nu \ge 0\) and \(\mathfrak {I}\nu \le 0\) are known, see [9]. In particular, if \(\nu \) is a positive real number and \(u_0\) is a smooth function of order one, then for any integer \(m\ge 4\) a solution \(u_\nu (t,x;u_0)\) satisfies estimates (1.7) with the averaging \( \nu {\mathbb {E}}\int _t^{t+\nu ^{-1}}\! \!\dots ds \) replaced by \( \nu ^{1/3} \int _0^{\nu ^{-1/3}}\! \!\dots ds, \) with the same upper bound and with the lower bound \(C_m \nu ^{-\kappa _m m}\), where \(\kappa _m\rightarrow 1/3\) as \(m\rightarrow \infty \). Moreover, it was then shown in [2] that the lower bounds remain true with \(\kappa =1/3\), and that the estimates \( \sup _{t\in [0, |\nu |^{-1/3}]} \Vert u(t)\Vert _{C^m} \ge C_m |\nu |^{-m/3}, m\ge 2, \) hold for smooth solutions of Eq. (1.1) with \(\eta =0\) and any non-zero complex “viscosity” \(\nu \).
The better quality of the lower bounds for solutions of the deterministic equations is due to an extra difficulty which occurs in the stochastic case: when time grows, simultaneously with increasing of high Sobolev norms of a solution, its \(L_2\)-norm may decrease, which accordingly would weaken the mechanism, responding for the energy transfer to high modes. Significant part of the proof of Theorem 1 is devoted to demonstration that the \(L_2\)-norm of a solution cannot go down without sending up the second Sobolev norm.
If \(\eta =0\) and \(\nu =i\delta \in i{\mathbb {R}}\), then (1.1) is a Hamiltonian PDE (the defocusing Schrödinger equation), and the \(L_2\)-norm is its integral of motion. If this integral is of order one, then the results of [9] (see there Appendix 3) imply that at some point of each time-interval of order \(\delta ^{-1/3}\) the \(C^m\)-norm of a corresponding solution will become \(\ > rsim \delta ^{-m\kappa }\) if \(m\ge 2\), for any \(\kappa <1/3\). Furthermore, if \(n=2\) and \(\delta =1\), then due to [4] for \(m>1\) and any \(M>1\) there exists a \(T=T(m,M)\) and a smooth \(u_0(x)\) such that \(\Vert u_0\Vert _m< M^{-1}\) and \(\Vert u(T;u_0)\Vert _m>M\).
The paper is organized as follows. In Sect. 2, we recall the results from [8, 10] on solutions of the Eq. (1.1). Next we show in Sect. 3 that if the noise \(\eta \) is non-degenerate, the \(L^2\)-norm of a solution of Eq. (1.1) cannot stay too small on time intervals of order \({\mathcal {O}}(\frac{1}{\nu })\) with high probability, unless its \(H^2\)-norm gets very large
(see Lemma 12). Then in Sect. 4 we derive from this fact the assertion (1) of Theorem 1. We prove assertion (2) and both corollaries in Sect. 5.
Constants in estimates never depend on \(\nu \), unless otherwise stated. For a metric space M we denote by \({\mathcal {B}}(M)\) the Borel \(\sigma \)-algebra on M, and by \({{\mathcal {P}}}(M)\) – the space of probability Borel measures on M. By \({\mathcal {D}}(\xi )\) we denote the law of a r.v. \(\xi \), and by \(|\cdot |_p\) – the norm in \(L_p(K^n)\).
2 Solutions and estimates
Strong solutions for the Eq. (1.1) are defined in the usual way:
Definition 5
Let \((\Omega , {\mathcal {F}},\{{\mathcal {F}}_t\}_{t\geqslant 0}, {\mathbb {P}})\) be the filtered probability space as in the introduction. Let \(u_0\) in (1.5) be a r.v., measurable in \({\mathcal {F}}_0\) and independent from the Wiener process \(\xi \) (e.g., \(u_0(x)\) may be a non-random function). Then a random process \(u(t)=u(t;u_0)\in C_0(K^{n}) \), \(t\in [0,T]\), adapted to the filtration, is called a strong solution of (1.1), (1.5), if
-
(1)
a.s. its trajectories u(t) belong to the space
$$\begin{aligned} {\mathcal {H}}([0,T]):=C([0,T],C_0(K^{n}))\cap L^2([0,T], H^1); \end{aligned}$$ -
(2)
we have
$$\begin{aligned} u(t)=u_0+\int _0^t(\nu \Delta u-i|u|^2u)ds+\sqrt{\nu }\,\xi (t),\quad \forall t\in [0,T], \quad a.s., \end{aligned}$$where both sides are regarded as elements of \(H^{-1}\).
If (1)-(2) hold for every \(T<\infty \), then u(t) is a strong solution for \(t\in [0,\infty )\). In this case a.s. \(u\in C([0,\infty ),C_0(K^{n}))\cap L_{loc}^2([0,\infty ), H^1). \)
Everywhere below when we talk about solutions for the problem (1.1), (1.5) we assume that the r.v. \(u_0\) is as in the definition above.
The global well-posedness of Eq. (1.1) was established in [8, 10]:
Theorem 6
For any \(u_0\in C_0(K^{n})\) the problem (1.1), (1.5) has a unique strong solution \(u^\omega (t,x;u_0)\), \(t\ge 0\). The family of solutions \(\{ u^\omega (t;u_0)\}\) defines in the space \(C_0(K^{n})\) a Fellerian Markov process.
In [8, 10] the theorem above is proved when (1.4) is replaced by the weaker assumption \(B_*<\infty \), where \(\ B_* =\sum |b_d| \) (note that \(B_*\le C_n B_{m_*}^{1/2}\)).
The transition probability for the obtained Markov process in \(C_0(K^{n})\) is
and the corresponding Markov semigroup in the space \({\mathscr {P}}(C_0(K^{n}))\) of Borel measures on \(C_0(K^{n})\) is formed by the operators \(\{ {\mathcal {B}}_t^*, t\ge 0\}\),
Then \({\mathcal {B}}_t^*\mu = {\mathcal {D}}u(t;u_0)\) if \(u_0\) is a r.v., independent from \(\xi \) and such that \({\mathcal {D}}(u_0)=\mu \).
Introducing the slow time \(\tau =\nu t\) and denoting \(v(\tau ,x)=u(\frac{\tau }{\nu },x)\), we rewrite Eq. (1.1) in the following form, more convenient for some calculations:
where
and \({\tilde{\beta }}_d(\tau ):=\nu ^{1/2}\beta _d(\tau \nu ^{-1})\), \(d\in {\mathbb {N}}^d\), is another set of independent standard complex Brownian motions.
Let \(\Upsilon \in C^\infty ({\mathbb {R}})\) be any smooth function such
Writing \(v\in {\mathbb {C}}\) in the polar form \(v=re^{i\Phi }\), where \(r=|v|\), and recalling that \(\langle \cdot ,\cdot \rangle \) stands for the real scalar product in \({\mathbb {C}}\), we apply Itô’s formula to \(\Upsilon (|v|)\) and obtain that the process \(\Upsilon (\tau ):=\Upsilon (|v(\tau )|)\) satisfies
where \(\Upsilon _0=\Upsilon (|v(0)|)\) and \({\mathbb {W}}(\tau )\) is the stochastic integral
In [10] Eq. (2.1) is considered with \(\nu =1\) and, following [8], the norm \(|v(t)|_\infty \) of a solution v is estimated via \(\Upsilon (t)\) (since \(|v| \le \Upsilon +1/2\)). But the nonlinear term \(i\nu ^{-1}|v|^2v\) does not contribute to Eq. (2.2), which is the same as the \(\Upsilon \)-equation (2.3) in [10] (and as the corresponding equation in [8, Section 3.1]). So the estimates on \(|\Upsilon (t)|_\infty \) and the resulting estimates on \(|v(t)|_\infty \), obtained in [10], remain true for solutions of (2.1) with any \(\nu \). Thus we get the following upper bound for quadratic exponential moments of the \(L_\infty \)-norms of solutions:Footnote 2
Theorem 7
For any \(T>0\) there are constants \(c_*>0\) and \(C>0\), depending only on \(B_*\) and T, such that for any r.v. \(v_0^\omega \in C_0(K^{n})\) as in Definition 5, any \(\tau \geqslant 0\) and any \(c\in (0,c_*]\), a solution \(v(\tau ; v_0)\) of Eq. (2.1) satisfies
In [10] the result above is proved for a deterministic initial data \(v_0\). The theorem’s assertion follows by averaging the result of [10] in \(v_0^\omega \).
The estimate (2.3) is crucial for derivation of further properties of solutions, including the given below upper bounds for their Sobolev norms, obtained in the work [8]. Since the scaling of the equation in [8] differs from that in (2.1) and the result there is a bit less general than in the theorem below, a sketch of the proof is given in Appendix B.
Theorem 8
Assume that \(B_m<\infty \) for some \( m\in {\mathbb {N}}\), and \(v_0=v_0^\nu \in H^m\cap C_0(K^n)\) satisfies
Then
where \(C_{M,m}\) also depends on M, \(M_m\) and \(B_m\), \(B_{m_*}\).
Neglecting the dependence on \(\nu \), we have that if \(B_m<\infty \), \(m\in {\mathbb {N}}\), and a r.v. \(v_0^\omega \in H^m\cap C_0(K^n)\) satisfies \({\mathbb {E}}\Vert v_0\Vert _m^2<\infty \) and \({\mathbb {E}}\exp (c\ |v_0|^2_\infty )<\infty \) for some \(c>0\), then Eq. (2.1) has a solution, equal \(v_0\) at \(t=0\), such that
where \( C>0\) depend on \(c, \nu , {\mathbb {E}}\exp (c\ |v_0|^2_\infty ), B_{m_*}\) and \(B_m\), while \(C'\) also depends on \({\mathbb {E}}\Vert v_0\Vert _m^2<\infty \) and T. See Appendix B.
As it is shown in [10], the estimate (2.3) jointly with an abstract theorem from [11], imply that under a mild nondegeneracy assumption on the random force the Markov process in the space \(C_0(K^{n})\), constructed in Theorem 6, is mixing:
Theorem 9
For each \(\nu >0\), there is an integer \(N=N(B_*,\nu )> 0\) such that if \(b_d\ne 0\) for \(|d|\leqslant N\), then the Eq. (1.1) is mixing. I.e. it has a unique stationary measure \(\mu _\nu \in {\mathscr {P}}(C_0(K^{n}))\), and for any probability measure \(\lambda \in {\mathscr {P}}(C_0(K^{n}))\) we have \({\mathcal {B}}^*_t\lambda \rightharpoonup \mu _\nu \) as \(t\rightarrow \infty \).
Under the assumption of Theorem 8, for any \(u_0\in H^m\) the law \({\mathcal {D}}u(t;u_0)\) of a solution \(u(t;u_0)\) is a measure in \(H^m\). The mixing property in Theorem 9 and (2.4) easily imply
Corollary 10
If under the assumptions of Theorem 9\(B_m<\infty \) for some \(m\in {\mathbb {N}}\) and \(u_0\in H^m\), then \({\mathcal {D}}(u(t;u_0)) \rightharpoonup \mu _\nu \) in \({\mathcal {P}}(H^m)\).
In view of Theorems 7, 8 with \(v_0=0\) and the established mixing, we have:
Corollary 11
Under the assumptions of Theorem 9, if \(v^{st}(\tau )\) is the stationary solution of the equation, then
where the constant \({\mathcal {C}}>0\) depends only on T and \(B_*\). If in addition \(B_m<\infty \) for some \(m\in {\mathbb {N}}\cup \{0\}\), then \( {\mathbb {E}}\Vert v^{st}(\tau )\Vert _m^2\leqslant C_{m}\nu ^{-m}, \) where \(C_m\) depends on \(B_*\) and \(B_m\).
Finally we note that applying Itô’s formula to \(\Vert v^{st}(\tau )\Vert _0^2\), where \(v^{st}\) is a stationary solution of (2.1), and taking the expectation we get the balance relation
We cannot prove that \({\mathbb {E}}\Vert v^{st}(\tau )\Vert _0^2 \ge B'>0\) for some \(\nu \)-independent constant \(B'\), and cannot bound from below the energy \(\tfrac{1}{2} {\mathbb {E}}\Vert v(\tau ; v_0)\Vert ^2_0\) of a solution v by a positive \(\nu \)-independent quantity. Instead in next section we get a weaker conditional lower bound on the energies of solutions.
3 Conditional lower bound for the \(L^2\)-norm of solutions
In this section we prove the following result:
Lemma 12
Let \(B_2<\infty \) and \(u(\tau ;u_0)\), where \(u_0\in H^2\cap C_0(K^n)\) is non-random, be a solution of Eq. (2.1). Take any constants \( \chi >0, \Gamma \ge 1, \tau _0\ge 0, \) and define the stopping time
(as usual, \(\tau _\Gamma =\infty \) if the set under the \(\inf \)-sign is empty). Then
for any \(\tau >\tau _0\).
Proof
We establish the result by adapting the proof from [16] (also see [11, Theorem 5.2.12]) to non-stationary solutions. The argument relies on the concept of local time for semi-martingales (see e.g. [14, Chapter VI.1] for details of the concept). By \([\cdot ]_b\) we denote the quasinorm \( [u]_b^2 = \sum _d |u_d|^2 b_d^2. \)
Without loss of generality we assume \(\tau _0=0\). Otherwise we just need to replace \(u( \tau ,x)\) by the process \({\tilde{u}}(\tau ,x):=u(\tau +\tau _0,x)\), apply the lemma with \(\tau _0=0\) and with \(u_0\) replaced by the initial data \({\tilde{u}}_0^\omega =u^\omega (\tau _0; u_0)\), and then average the estimate in the random \({\tilde{u}}_0^\omega \).
Let us write the solution \(u(\tau ;u_0)\) as \(u(\tau )=\) \(\sum _{d\in {\mathbb {N}}^n}u_d(\tau )\varphi _d\). For any fixed function \(g\in C^2({\mathbb {R}})\), consider the process
Since
then by Itô’s formula we have
Step 1: We firstly show that for any bounded measurable set \({G}\subset {\mathbb {R}}\), denoting by \({\mathbb {I}}_G\) its indicator function, we have the following equality
Let \(L(\tau ,a)\), \((\tau ,a)\in [0,\infty )\times {\mathbb {R}}\), be the local time for the semi martingale \(f(\tau )\) (see e.g. [14, Chapter VI.1]). Since in view of (3.2) the quadratic variation of the process \(f(\tau )\) is
then for any bounded measurable set \(G\subset {\mathbb {R}}\), we have the following equality (known as the occupation time formula, see [14, Corollary VI.1.6]),
For the local time \(L(\tau ,a)\), due to Tanaka’s formula (see [14, Theorem VI.1.2]) we have
Taking expectation of both sides of (3.5) and (3.6) we obtain the required equality (3.4).
Step 2: Let us choose \({G}=[\rho _0,\rho _1]\) with \(\rho _1>\rho _0>0\), and \(g(x) =g_{\rho _0} (x) \in C^2({\mathbb {R}})\) such that \(g'(x)\geqslant 0\), \(g(x)=\sqrt{x}\) for \(x\geqslant \rho _0\) and \(g(x)=0\) for \(x\leqslant 0\). Then due to the factors \({\mathbb {I}}_G(f)\) and \({\mathbb {I}}_G(a)\) in (3.4), we may there replace g(x) by \(\sqrt{x}\), and accordingly replace \( g(\Vert u\Vert ^2_0), g'(\Vert u\Vert ^2_0)\) and \( g''(\Vert u\Vert ^2_0) \) by \(\Vert u\Vert _0\), \( \tfrac{1}{2}\Vert u\Vert _0^{-1}\) and \( -\tfrac{1}{4}\Vert u\Vert _0^{-3}\). So the relation (3.4) takes the form
Since the l.h.s. of the above equality is non-negative, we have
Noting that
that by the definition of the stopping time \(\tau _\Gamma \)
and that by interpolation,
we derive from (3.7) the relation
When \(\rho _0 \rightarrow 0\), we have \(g(x) \rightarrow \sqrt{x}\) and \(f(\tau ) \rightarrow \Vert u(\tau \wedge \tau _\Gamma )\Vert _0\). So sending \(\rho _0\) to 0 and using Fatou’s lemma we get from the last estimate that
As the l.h.s. above is not smaller than
then
By the monotone convergence theorem
so we get from (3.8) that
Step 3: We continue to verify that
To do this let us fix any index \(d\in {\mathbb {N}}^n\) such that \(b_d\ne 0\). The process \(u_d(\tau )\) is a semimartingale, \(\ du_d = v_d ds +b_d d\beta _d, \, \) where \(v_d(s)\) is the d-th Fourier coefficient of \(\, \Delta u+\frac{1}{\nu }i|u|^2u\,\) for the solution \(u(\tau )=\sum _d u_d(\tau ) \varphi _d\) which we discuss. Consider the stopping time
Due to (2.3) and (2.6), \( {\mathbb {P}} (\tau _R= \tau \wedge \tau _\Gamma ) \rightarrow 1\) as \(R\rightarrow \infty \). Let us denote \( u_d^R(\tau ) = u_d(\tau \wedge \tau _R). \) To prove (3.10) it suffices to verify that
If we replace above \(u_d\) by \(u_d^R\), then the obtained new quantity \( \pi ^R(\delta )\) differs from \( \pi (\delta )\) at most by \( {\mathbb {P}} (\tau _R< \tau \wedge \tau _\Gamma )\). The process \(u_d^R\) is an Ito process with a bounded drift. So by [7, Theorem 2.2.2, p. 52], \( \pi ^R(\delta )\) goes to zero with \(\delta \). Thus, given any \(\varepsilon >0\), we firstly choose R sufficiently big and then \(\delta \) sufficiently small to achieve \(\pi (\delta ) <\varepsilon \), for a suitable \(\delta (\varepsilon )>0\). So (3.10) is verified. Jointly with (3.9) this proves (3.1). \(\square \)
4 Lower bounds for Sobolev norms of solutions
In this section we work with Eq. (1.1) in the original time scale t and provide lower bounds for the \(H^m\)-norms of its solutions with \(m>2\). This will prove the assertion (1) of Theorem 1. As always, the constants do not depend on \(\nu \), unless otherwise stated.
Theorem 13
For any integer \(m\geqslant 3\), if \(B_m<\infty \) and
then for any r.v. \(u_0(x)\in H^m\cap C_0(K^n)\), satisfying
for some \(c, C >0\), we have
for every \(K>0\).
Proof
Consider the complement to the event in (4.2):
We will prove the assertion (4.2) by contradiction. Namely, we assume that there exists a \(\gamma >0\) and a sequence \(\nu _j\rightarrow 0\) such that
and will derive a contradiction. Below we write \( Q^{\nu _j}\) as Q and always suppose that
The constants in the proof may depend on \({\mathcal {K}}, K, \gamma \), \(B_{\,m\vee m_*}\), but not on \(\nu \).
Without lost of generality we assume that \(T_1=1\). For any \(T_0>0\), due to (2.5) and (2.3) the r.v. \( {\tilde{u}}_0 := u(T_1) \) satisfies (4.1) with c replaced by c/5. So considering \({\tilde{u}}(t,x) = u(t+T_0,x)\) we may assume that \(T_0=0\).
Let us denote \(J_1=[0,\frac{1}{\nu }]\). Due to Theorem 7,
uniformly in \(\nu \), for a suitable \(C_1(\gamma )\). Then, by the definition of Q and Sobolev’s interpolation,
for \(l\in [0,m]\) (and any \(\nu \in \{\nu _1,\nu _2,\dots \}\)).
Denote \(J_2=[0,\frac{1}{2\nu }]\) and consider the stopping time
Then \(\tau _1=\frac{1}{2\nu }\) for \(\omega \in Q\cap Q_1\). So due to (3.1) with \(\Gamma =C_{2,\gamma }\nu ^{-2\kappa }\), for any \(\chi >0\), we have
Consider the event
Due to the above, we have,
So \({\mathbb {P}}(\Lambda )\leqslant \gamma \) if we choose
Let us set
and for \(\chi \) as in (4.5), consider the stopping time
Then \({{\tilde{\tau }}}_1\leqslant \frac{1}{2\nu }\) for all \(\omega \in Q_2\). Consider the function
It solves Eq. (1.1) with modified Wiener processes and with initial data \(v_0(x) = u^\omega ({{\tilde{\tau }}}_1,x)\), satisfying
Now we introduce another stopping time, in terms of v(t, x):
For \(\omega \in Q_2\), \(\tau _2=\tfrac{1}{2\nu }\) and in view of (4.4)
Step 1: Let us estimate from above the increment \({\mathscr {E}}(t,x)=|v(t\wedge \tau _2,x)|^2-|v_0(x)|^2\). Due to Itô’s formula, we have that
We treat M as a martingale M(t) in the space \(H^1\). Since in view of (A.3) for \(0\le s<\tau _2\) we have
where \(\zeta = \sup _{0\leqslant s\leqslant \frac{1}{\nu }}|u(s)|_\infty \) (the assertion is empty if \(\tau _2=0\)), then for any \(0<T_*\le \tfrac{1}{2\nu }\)
where we used that \(B_1<\infty \). So by Doob’s inequality
Let us choose
where b will be specified later. Then \( 1\le T_*\le \tfrac{1}{2\nu } \) if \(\nu \) is sufficiently small, so due to (4.10)
for a suitable \(C_4(\gamma )\) (and for \(\nu \ll 1\)); thus \( {\mathbb {P}} (Q_2\cap Q_3) \ge 2\gamma \). Since \( \Vert \langle v , \Delta v\rangle \Vert _1 \le C|v|_\infty \Vert v\Vert _3\) by (A.2) and \( \Vert \sum _d b_d \varphi _d\Vert _1 \le C, \) then in view of (4.8) and the definition of \(Q_3\),
Step 2: For any \(x\in K^n\), denoting \( R(t) = |v(t,x)|^2, \, a(t) = \Delta v(t,x) \) and \(\xi (t) = \xi (t,x)\), we write the equation for \(v(t) := v(t,x)\) as an Itô process:
Setting \(w(t) = e^{i\int _0^t R(s)ds} v(t)\), we observe that w also is an Itô process, \( w(0) =v_0\) and \( dv = e^{-i\int _0^t R(s)ds} dw -i Rv\, dt. \) From here and (4.12),
So \(v(t\wedge \tau _2)= v(t\wedge \tau _2,x)\) can be written as
where
Our next goal is to obtain a lower bound for \( \Vert v(T_*)\Vert _1\) when \(\omega \in Q_2\cap Q_3\), using the above decomposition (4.13).
Step 3: We first deal with the stochastic term \(I_3(t)\). For \(0 \leqslant s\leqslant s_1\leqslant T_*\wedge \tau _2\) we set
then \( W(s,s_1, x) = \exp \big (i F(s,s_1, x)\big )\). The functions F and W are periodic in x, but not odd. Speaking about them we understand \(\Vert \cdot \Vert _m\) as the non-homogeneous Sobolev norm, so \( \Vert F\Vert _m^2 = \Vert F\Vert _0^2 + \Vert ( -\Delta )^{m/2} F\Vert _0^2, \) etc. We write \(I_3\) as
In view of (A.1),
For any \(s\in J =[ 0, T_*\wedge \tau _2)\), by (A.3) and the definition of \( \tau _2\), we have that \(v:=v(s)\) satisfies
(this assertion is empty if \( \tau _2=0\) since then \(J=\emptyset \)). So for \(s,s_1 \in J\),
for \(k\le m\). Then, due to (4.16),
Consider the stochastic integral in (4.15),
The process \( t \mapsto W(0,t,x) \) is adapted to the filtration \(\{ {\mathcal {F}}_t\}\), and
So integrating by parts (see, e.g., [14, Proposition IV.3.1]) we re-write N as
and we see from (4.15) that
Due to (1.4) and since \(B_m<\infty \), the Wiener process \(\xi (t,x)\) satisfies
and
(we recall that \(B_*=\sum _{d\in {\mathbb {N}}^n}|b_d|<\infty \)). Therefore,
with a suitable \(C=C(\gamma )\). Let
then \({\mathbb {P}}({\tilde{Q}})\geqslant \gamma \). As \(\tau _2 = T_*\) for \(\omega \in {\tilde{Q}}\), then due to (4.17), (4.18), (4.19) and (A.3), for \(\omega \in {\tilde{Q}}\) we have
Setp 4: We then consider the term \( I_{2}= \nu \int _0^{t\wedge \tau _2} {\bar{W}}(s, t\wedge \tau _2,x) \Delta v(s,x)ds \). To bound its \(H^1\)-norm we need to estimate \(\Vert W\Delta v\Vert _1\). Since
(see [17, Proposition 3.6]), we have
Then in view of (4.16) and (4.8), for \(\omega \in {\tilde{Q}}\)
and accordingly
Step 5: Now we estimate from below the \(H^1\)-norm of the term \(I^\omega _{1}(T_*,x)\), \(\omega \in {\tilde{Q}}\). Writing it as \( I^\omega _1(T_*,x) = e^{-i T_*|v_0(x)|^2} e^{ -i \int _0^{T_*} {\mathscr {E}} (s,x) ds} v_0(x) \) wee see that
This first term on the r.h.s is
where we have used the fact that \(u|_{\partial K^{n}}=0\), Poincaré’s inequality and (4.7).
For \(\omega \in {\tilde{Q}}\) and \(0\le s\le T_*\), in view of (4.11), the second term is bounded by
Therefore, using (4.11), we get for the term \(I^\omega _1(T_*)\) the following lower bound:
Recalling \(T_*=\nu ^{-b}\) we see that if we assume that
then for \(\omega \in {\tilde{Q}}\),
provided that \(\nu \) is sufficiently small.
Step 6: Finally, remembering that \(\tau _2=T_*\) for \(\omega \in {\tilde{Q}}\) and combining the relations (4.20), (4.21) and (4.23) to estimate the terms of (4.13), we see that for \(\omega \in {\tilde{Q}}\) we have
if we assume in addition to (4.22) that
and \(\nu \) is small. Note that this relation implies the last two in (4.22).
Combining (4.8) and (4.24) we get that
for all sufficiently small \(\nu \). Thus we have obtained a contradiction with the existence of the sets \(Q^{\nu _j}\) as at the beginning of the proof if (for a chosen \(\kappa \)) we can find a \(b\in (0,1)\) which meets (4.22), (4.25) and
Noting that this is nothing but the first relation in (4.22), we see that we have obtained a contradiction if
for some \(b\in (0,1)\). We see immediately that such a b exists if and only if \(\kappa < \tfrac{1}{35} \).
\(\square \)
Amplification If we replace the condition \(m\geqslant 3\) with the weaker assumption
then the statement (4.2) remains true for \(0<\kappa <\kappa (n,m)\) with a suitable (less explicit) constant \(\kappa (n,m)>0\). In this case we obtain a contradiction with the assumption (4.3) by deriving a lower bound for \(\Vert v(T_*)\Vert _\alpha \), where \(\alpha =\min \{1,m-2\}\in (0,1]\), using the decomposition (4.13). The proof remains almost identical except that now, firstly, we bound \(\Vert I_{2}\Vert _\alpha \) (\(\alpha <1\)) from above using the following estimate from [15, Theorem 5, p. 206] (also see there p. 14):
and, secondly, estimate \(\Vert I_{1}^\omega (T_*)\Vert _\alpha \) (\(\alpha <1\)) from below as
which directly follows from Sobolev’s interpolation. See [6] for more details.
5 Lower bounds for time-averaged Sobolev norms
In this section we prove the assertion (2) of Theorem 1. We provide each space \(H^r\), \(r\ge 0\), with the scalar product
corresponding to the norm \(\Vert u\Vert _r\). Let \( u(t) =\sum u_d(t) \varphi _d \) be a solution of Eq. (1.1). Applying Itô’s formula to the functional \(\Vert u\Vert _m^2\), we have for any \(0\leqslant t<t'<\infty \) the relation
where M stands for the real scalar productis the stochastic integral
Let us fix a \(\gamma \in (0,\frac{1}{8})\). Due to Theorems 7 and 13, for small enough \(\nu \) there exists an event \(\Omega _1\subset \Omega \), \( {\mathbb {P}}(\Omega _1)\geqslant 1-\gamma /2, \) such that for all \(\omega \in \Omega _1\) we have:
-
a)
\(\sup _{0\leqslant t\leqslant \frac{1}{\nu }}|u^\omega (t)|_\infty \leqslant C(\gamma )\), for a suitable \(C(\gamma )>0\);
-
b)
there exist \(t_\omega \in [0,\frac{1}{3\nu }]\) and \(t_\omega '\in [\frac{2}{3\nu },\frac{1}{\nu }]\) satisfying
$$\begin{aligned} \Vert u^\omega (t_\omega )\Vert _m,\;\Vert u^\omega (t_\omega ')\Vert _m\geqslant \nu ^{-m\kappa }. \end{aligned}$$(5.2)
Since for the martingale M(0, t) we have that
then by Doob’s inequality
Now let us set \(\ {\hat{\Omega }}=\Omega _1\cap \Omega _2.\) Then \({\mathbb {P}}({\hat{\Omega }})\geqslant 1-\gamma \) for small enough \(\nu \), and for any \(\omega \in {\hat{\Omega }}\) there are two alternatives:
(i) there exists a \(t_\omega ^0\in [0,\frac{1}{3\nu }]\) such that \(\Vert u^\omega (t^0_\omega )\Vert _m=\frac{1}{3}\nu ^{-\kappa m}\). Then from (5.1) and (5.2) in view of (A.4) we get
(ii) There exists no \(t\in [0,\frac{1}{3\nu }]\) with \(\Vert u^\omega (t)\Vert _m=\frac{1}{3}\nu ^{-\kappa m}\). In this case, since \(\Vert u^\omega (t)\Vert _m\) is continuous with respect to t, then due to (5.2) \( \Vert u^\omega (t)\Vert _m > \frac{1}{3}\nu ^{-m\kappa } \) for all \( t\in [0,\frac{1}{3\nu }]\). This leads to the relation
In both cases for \(\omega \in {\hat{\Omega }}\) we have:
It implies that
(for small enough \(\nu \)), and gives the lower bound in (1.7).
The upper bound follows directly from Theorem 8.
Proof of Corollaries 3 and 4
Since \(B_k<\infty \) for each k and all coefficients \(b_d\) are non-zero, then Eq. (1.1) is mixing in the spaces \(H^m\), \(m\in {\mathbb {N}}\), see Corollary 10. As the stationary solution \(v^{st} \) satisfies Corollary 11 with any m, then for each \(\mu \in {\mathbb {N}}\) and \(M>0\), interpolating the norm \(\Vert u\Vert _\mu \) via \(\Vert u\Vert _0\) and \(\Vert u\Vert _m\) with m sufficiently large we get that the stationary measure \(\mu _\nu \) satisfies
Similar, in view of (2.5) and Theorem 7,
for each \(u_0\in C^\infty \) and every \(\mu \) and M as in (5.3). Now let us consider the integral in (1.7) and write it as
Replacing the integrand in \(J_t\) with \({\mathbb {E}}( \Vert u_\nu (s)\Vert _m \wedge N)^2\), \( N\ge 1\), using the convergence
which follows from Corollary 10, and the estimates (5.3), (5.4) we get that
This convergence and (1.7) imply the assertion of Corollary 3.
Now the convergence (5.5) jointly with estimates (5.3), (5.4) and (1.8) imply Corollary 4. \(\square \)
Notes
We note that solutions of eqs. (1.1) with complex \(\nu \) behave differently, and solubility of those equations with large n is unknown.
References
Bartuccelli, M., Gibbon, J., Oliver, M.: Length scales in solutions of the complex Ginzburg–Landau equation. Physica D 89, 267–286 (1996)
Biryuk, A.: Lower bounds for derivatives of solutions for nonlinear Schrödinger equations. Proc. A R. Soc. Edinburgh 139, 237–251 (2009)
Bohr, T., Jensen, M., Paladin, G., Vulpiani, A.: Dynamical Systems Approach to Turbulence. Cambridge University Press, Cambridge (1998)
Colliander, J., Keel, M., Staffilani, G., Takaoka, H., Tao, T.: Transfer of energy to high frequencies in the cubic defocusing nonlinear Schrödinger equation. Invent. Math. 181, 39–113 (2010)
Frisch, U.: Turbulence: the Legacy of A. N. Kolmogorov. Cambridge University Press, Cambridge (1995)
Huang, G., Kuksin, S.: On the energy transfer to high frequencies in damped/driven nonlinear Schrödinger equation (extended version). arXiv:2006.11518, (2020)
Krylov, N.V.: Controlled Diffusion Processes. Springer, Berlin (1980)
Kuksin, S.: A stochastic nonlinear Schrödinger equation I. A priori estimates. Tr. Mat. Inst. Stekl. 225, 232–256 (1999)
Kuksin, S.: Spectral properties of solutions for nonlinear PDEs in the turbulent regime. GAFA 9, 141–184 (1999)
Kuksin, S., Nersesyan, V.: Stochastic CGL equations without linear dispersion in any space dimension. Stoch PDE Anal. Comp. 1(3), 389–423 (2013)
Kuksin, S., Shirikyan, A.: Mathematics of Two-Dimensional Turbulence. Cambridge University Press, Cambridge (2012)
Kuksin, S.B., Shirikyan, A.: Randomly forced CGL equation: stationary measures and the inviscid limit. J. Phys. A Math. Gen. 37, 1–18 (2004)
Levermore, C.D., Oliver, M.: The complex Ginzburg–Landau equation as a model problem. Lect. Appl. Math. 31, 141–189 (1996)
Revuz, D., Yor, M.: Continuous Martingales and Brownian Motion. Springer, Berlin (2005)
Runst, T., Sickel, W.: Sobolev Spaces of Fractional Order, Nemytskij Operators, and Nonlinear Partial Differential Equations. Walter de Gruyter, New York (2011)
Shirikyan, A.: Local times for solutions of the complex Ginzburg-Landau equation and the inviscid limit. J. Math. Anal. Appl. 384, 130–137 (2011)
Taylor, M.E.: Partial Differential Equations III. Applied Mathematical Sciences, Springer (2011)
Acknowledgements
GH is supported by NSFC (Significant project No.11790273) in China and SK thanks the Russian Science Foundation for support through the grant 18-11-00032.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A. Some estimates
For any integer \(l\in {\mathbb {N}}\) and \(F\in H^l\) we have that
Indeed, to verify (A.1) it suffices to check that for any non-zero multi-indices \(\beta _1,\dots ,\beta _{l'}\), where \(1\le l'\leqslant l\) and \( |\beta _1| +\cdots + |\beta _{l'}| =l, \) we have
But this is the assertion of Lemma 3.10 in [17]. Similarly,
see [17, Proposition 3.7] (this relation is known as Moser’s estimate). Finally, since for \(| \beta | \le m\) we have \( | \partial _x^\beta v|_{2m/ \beta |} \le C |v|_\infty ^{1- |\beta | /m} \Vert v\Vert _m^{|\beta | /m} \) (see relation (3.17) in [17]), then
Appendix B. Proof of Theorem 8
Applying Ito’s formula to a solution \(v(\tau )\) of Eq. (2.1) we get a slow time version of the relation (5.1):
where \(M(\tau )= \int _0^{\tau }\sum _{d} b_d|d|^{2m} \langle v_d(s), d\beta _d(s)\rangle .\) Since in view of (A.4)
then denoting \( {\mathbb {E}}\Vert v(\tau )\Vert _r^2 =: g_r(\tau ), \ r \in {\mathbb {N}}\cup \{0\}, \) taking the expectation of (B.1), differentiating the result and using (2.3), we get that
since \( g_m \le g_0^{1/(m+1)} g_{m+1}^{m/(m+1)} \le C_m g_{m+1}^{m/(m+1)} . \) We see that if \(g_m \ge (2\nu ^{-1} C'_m)^m\), then the r.h.s. of (B.2) is
which is negative if \(\nu \ll 1\). So if
at \(\tau =0\), then (B.4) holds for all \(\tau \ge 0\) and (2.4) follows. If \(g_m(0)\) violates (B.4), then in view of (B.2) and (B.3), for \(\tau \ge 0\), while (B.4) is false, we have that
which again implies (2.4). Besides, in view of (B.2),
This relation immediately implies (2.5).
Now let us return to Eq. (B.1). Using Doob’s inequality and (2.4) we find that
Next, applying (A.4) and Young’s inequality we get
Finally, using in (B.1) the last two displayed formulas jointly with (2.3) we obtain (2.6).
Rights and permissions
About this article
Cite this article
Huang, G., Kuksin, S. On the energy transfer to high frequencies in the damped/driven nonlinear Schrödinger equation. Stoch PDE: Anal Comp 9, 867–891 (2021). https://doi.org/10.1007/s40072-020-00187-2
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40072-020-00187-2