betaFeedback

WorksMathematics

Uniform local smoothing for the derivative NLS and the soliton mass threshold

Abstract

Harrop-Griffiths, Killip, Ntekoume and Vişan proved that the derivative nonlinear Schrödinger equation (DNLS) is globally well-posed in L2(R)L^2(\mathbb{R}). A key ingredient of their proof is a local smoothing estimate. It holds on L2L^2-equicontinuous sets of initial data, but it cannot hold on L2L^2-balls, because rescaled stationary solitons, all of mass 2π2π, violate it. This raises a natural question: for which masses mm does local smoothing hold uniformly on the ball {M(q0)≤m}\{M(q_0)\le m\}? We prove three results. First, we give a quantitative version of the local smoothing estimate in which the data enter only through the mass and an explicit frequency scale κ(q0)κ(q_0). On time intervals of length 2T2T the bound grows like 1+κ(q0)2T1+κ(q_0)^2T, and the stationary soliton shows that the power of κκ cannot be lowered below one. Second, we prove uniform local smoothing on balls of small mass, together with a scale-invariant Morrey-type form and a localized Lt,x4L^4_{t,x} bound. Third, we analyse the soliton family from both sides. The uniform constant must blow up at least like (2π−m)−1(2π-m)^{-1} as m↑2πm\uparrow 2π, while every soliton of mass m<2πm<2π satisfies local smoothing with a constant O((2π−m)−1)O((2π-m)^{-1}). These facts single out the stationary-soliton mass 2π2π, which is also the Hayashi–Ozawa coercivity threshold, as the natural candidate for the sharp threshold. We formulate this as a conjecture, give a spectral reformulation in terms of the Kaup–Newell scattering data, and discuss the obstacles to a proof.

Introduction and main results

We consider the derivative nonlinear Schrödinger equation on the line,

i∂tq+∂x2q+i∂x(∣q∣2q)=0,q:Rt×Rx→C.(1)i\partial_t q+\partial_x^2q+i\partial_x\left(|q|^2q\right)=0,\qquad q:\mathbb{R}_t\times\mathbb{R}_x\to\mathbb{C}. \tag*{(1)}

It was derived as a model for circularly polarized Alfvén waves propagating along a magnetic field in a plasma [53, 54], and it is completely integrable through the Kaup–Newell Lax pair [34]. The equation conserves the mass M(q)=∫R∣q∣2 dxM(q)=\int_{\mathbb{R}}|q|^2\,dx and is invariant under the scaling

q(t,x)⟼qλ(t,x):=λ1/2q(λ2t,λx),λ>0,(2)q(t,x)\longmapsto q_\lambda(t,x):=\lambda^{1/2}q(\lambda^2t,\lambda x),\qquad\lambda>0, \tag*{(2)}

which leaves MM unchanged. Hence L2(R)L^2(\mathbb{R}) is the scaling-critical space.

In a recent landmark paper, Harrop-Griffiths, Killip, Ntekoume and Višan [20] proved that (1) is globally well-posed in L2(R)L^2(\mathbb{R}). More precisely, the data-to-solution map, defined on Schwartz data, extends to a jointly continuous map

Φ:R×L2(R)→L2(R),q(t)=Φ(t)q0.(3)\Phi:\mathbb{R}\times L^2(\mathbb{R})\to L^2(\mathbb{R}),\qquad q(t)=\Phi(t)q_0. \tag*{(3)}

This result is the end point of a long line of work, which we recall in § 1.2. Its proof relies on the second-generation method of commuting flows. The analytic heart of the method consists of two families of estimates. The first are local smoothing estimates, obtained from microscopic conservation laws for the logarithm of the Kaup–Newell perturbation determinant. The second express the tightness of orbits.

Let us recall the local smoothing statement of [20]. We use the localizing weights

ψ(x):=(sech⁡(x/99))1/2,W:=ψ12=sech⁡6(⋅/99),Wh:=W(⋅−h).(4)\psi(x):=\left(\operatorname{sech}(x/99)\right)^{1/2},\qquad W:=\psi^{12}=\operatorname{sech}^{6}(\cdot/99),\qquad W_h:=W(\cdot-h). \tag*{(4)}

By [20 Theorem 1.5], if Q⊂S(R)Q\subset\mathcal{S}(\mathbb{R}) is bounded and equicontinuous in L2L^2, then for every T>0T>0

sup⁡h∈R∫−TT∥Whq(t)∥H1/22 dt≲T,Q∥q(0)∥L22,q(0)∈Q.(5)\sup_{h\in\mathbb{R}}\int_{-T}^{T}\left\|W_hq(t)\right\|_{H^{1/2}}^2\,dt\lesssim_{T,Q}\left\|q(0)\right\|_{L^2}^2,\qquad q(0)\in Q. \tag*{(5)}

The authors stress that the estimate is claimed for equicontinuous sets and not for balls, and that this restriction is necessary. Indeed, [20 Proposition 1.11] exhibits solutions qnq_n, obtained by rescaling the stationary soliton, such that M(qn)≡2πM(q_n)\equiv2\pi while a localized local smoothing norm of qnq_n on [−1,1][-1,1] tends to infinity. In (5), the implicit constant comes from a rescaling that makes the relevant set of Schwartz functions “δ\delta-good” in the sense of [20 Definition 1.8]. As the authors note in the proof of [20 Theorem 1.5], this constant depends on QQ only through the rescaling parameter, but that dependence is not made explicit.

The counterexample of [20] lives exactly on the sphere M=2πM=2\pi, and 2π2\pi is a distinguished number for DNLS. It is the mass of the stationary soliton [20 (1.25), (1.28)]. It is also the threshold below which the energy is coercive, a fact used by Hayashi and Ozawa [24, 26] to obtain global H1H^1 solutions. Moreover, Bahouri and Perelman [2] point out that 2π2\pi and 4π4\pi are the masses of the static bright soliton and of the algebraic soliton, whose profiles are extremals of the relevant Gagliardo–Nirenberg inequalities. This suggests the following problem.

Problem. Determine the largest m∗∈(0,∞]m^*\in(0,\infty] such that local smoothing holds uniformly on every ball {q0∈L2:M(q0)≤m}\{q_0\in L^2:M(q_0)\le m\} with m<m∗m<m^*, with a constant depending only on mm. Describe the size of the optimal constant, and its dependence on the data when equicontinuity replaces smallness.

To make this precise, we introduce, for a solution qq of (1) and T>0T>0, the local smoothing functional

LS⁡T(q):=sup⁡h∈R∫−TT∥Whq(t)∥H1/22 dt.(6)\operatorname{LS}_T(q):=\sup_{h\in\mathbb{R}}\int_{-T}^{T}\left\|W_hq(t)\right\|_{H^{1/2}}^2\,dt. \tag*{(6)}

where ∥f∥Hs2=∫R(4+ξ2)s∣f^(ξ)∣2 dξ\|f\|_{H^s}^2=\int_{\mathbb{R}}(4+\xi^2)^s|\widehat{f}(\xi)|^2\,d\xi as in [20]. We also introduce the uniform local smoothing constant

C(m):=sup⁡{LS⁡1(Φ(⋅)q0)M(q0):q0∈L2(R), 0<M(q0)≤m}∈(0,∞].(7)C(m):=\sup\left\{\frac{\operatorname{LS}_1(\Phi(\cdot)q_0)}{M(q_0)}:q_0\in L^2(\mathbb{R}),\ 0<M(q_0)\le m\right\}\in(0,\infty]. \tag*{(7)}

For q0∈L2q_0\in L^2 the integrand in LS⁡1(Φ(⋅)q0)\operatorname{LS}_1(\Phi(\cdot)q_0) is a lower semicontinuous, hence measurable, function of tt, because t↦Φ(t)q0t\mapsto\Phi(t)q_0 is continuous in L2L^2 and the H1/2H^{1/2} norm is lower semicontinuous with respect to L2L^2 convergence (see the proof of Corollary 1.3). The function CC is nondecreasing, and we set m∗:=sup⁡{m>0:C(m)<∞}m^*:=\sup\{m>0:C(m)<\infty\}. By [20 Proposition 1.11], or by Theorem 1.4(i) below, m∗≤2πm^*\le2\pi.

Main results

Throughout the paper we fix 0<ε<1/20 < \varepsilon< 1/2 (for definiteness ε=1/4\varepsilon= 1/4; this parameter is denoted by σ\sigma in [20]). We let δ0>0\delta_0 > 0 be the smallness parameter for which [20 Proposition 4.1] holds; see Proposition 2.2 below. For κ>0\kappa> 0 we use the equicontinuity quantity of [20],

∥f∥Eε,κ2:=∫R∣ξ∣2ε(4κ2+ξ2)ε∣f^(ξ)∣2 dξ,(8)\lVert f\rVert_{E_{\varepsilon,\kappa}}^2 := \int_{\mathbb{R}} \frac{\lvert\xi\rvert^{2\varepsilon}}{(4\kappa^2+\xi^2)^{\varepsilon}} \lvert\widehat{f}(\xi)\rvert^2\,\mathrm{d}\xi, \tag*{(8)}

which is the norm ∥f∥Eσ,κ\lVert f\rVert_{E_{\sigma,\kappa}} of [20 (2.2)] with σ=ε\sigma= \varepsilon; for κ=1\kappa= 1 its square is the left-hand side of [20 (1.19)].

Definition 1.1 (Frequency scale of an orbit). For q0∈S(R)q_0 \in\mathcal{S}(\mathbb{R}) with solution q(t)=Φ(t)q0q(t) = \Phi(t)q_0 we set

κ(q0):=inf⁡{κ≥1:sup⁡t∈R∥q(t)∥Eε,κ≤δ0}.\kappa(q_0) := \inf\left\{\kappa\ge1 : \sup_{t\in\mathbb{R}} \lVert q(t)\rVert_{E_{\varepsilon,\kappa}} \le\delta_0\right\}.

Lemma 2.1 below shows that κ(q0)<∞\kappa(q_0) < \infty for every Schwartz datum, that the infimum is attained, and that κ(q0)=1\kappa(q_0) = 1 whenever M(q0)≤δ02M(q_0) \le\delta_0^2. It also shows that, apart from the normalization κ≥1\kappa\ge1, the frequency scale transforms linearly under the scaling (2). Our first result makes (5) quantitative.

Theorem 1.2 (Quantitative local smoothing). There is a constant C1C_1, depending only on the fixed parameter ε\varepsilon, such that for every q0∈S(R)q_0 \in\mathcal{S}(\mathbb{R}) and every T>0T > 0 the solution q=Φ(⋅)q0q = \Phi(\cdot)q_0 satisfies

LS⁡T(q)≤C1(1+κ(q0)2T)M(q0).(9)\operatorname{LS}_T(q) \le C_1(1+\kappa(q_0)^2T)M(q_0). \tag*{(9)}

Conversely, the power of κ\kappa in (9) cannot be lowered below one. More precisely, there are constants c∗>0c_* > 0, κs>0\kappa_s > 0 and λ0≥1\lambda_0 \ge1, and Schwartz data q0,λq_{0,\lambda}, λ≥1\lambda\ge1, such that M(q0,λ)=2πM(q_{0,\lambda}) = 2\pi, κ(q0,λ)=max⁡{1,κsλ}\kappa(q_{0,\lambda}) = \max\{1,\kappa_s\lambda\}, and

LS⁡T(Φ(⋅)q0,λ)≥c∗κ(q0,λ)TM(q0,λ)for all T>0 and all λ≥λ0.(10)\operatorname{LS}_T(\Phi(\cdot)q_{0,\lambda}) \ge c_*\kappa(q_{0,\lambda})T M(q_{0,\lambda}) \qquad\text{for all } T > 0 \text{ and all } \lambda\ge\lambda_0. \tag*{(10)}

Theorem 1.2 recovers [20 Theorem 1.5]. Indeed, if Q⊂S(R)Q \subset\mathcal{S}(\mathbb{R}) is bounded and equicontinuous, then [20 Corollary 1.9] gives sup⁡q0∈Qκ(q0)<∞\sup_{q_0\in Q}\kappa(q_0) < \infty; see Remark 2.3. The new information is twofold. All dependence on the data beyond the mass is carried by the single number κ(q0)\kappa(q_0), and this dependence is polynomial. When the mass is small, κ(q0)=1\kappa(q_0) = 1 and we obtain uniform estimates.

Corollary 1.3 (Uniform local smoothing for small mass). Let m0:=δ02m_0 := \delta_0^2. For every q0∈L2(R)q_0 \in L^2(\mathbb{R}) with M(q0)≤m0M(q_0) \le m_0 and every T>0T > 0, the solution q=Φ(⋅)q0q = \Phi(\cdot)q_0 of [20] satisfies

LS⁡T(q)≤C1(1+T)M(q0),(11)\operatorname{LS}_T(q) \le C_1(1+T)M(q_0), \tag*{(11)}
sup⁡h∈R∫−R2R2∥W(⋅−hR)q(t)∥H˙1/22 dt≤2C1RM(q0)for all R>0,(12)\sup_{h\in\mathbb{R}}\int_{-R^2}^{R^2}\left\lVert W\left(\frac{\cdot-h}{R}\right)q(t)\right\rVert_{\dot{H}^{1/2}}^2\,\mathrm{d}t \le2C_1RM(q_0) \qquad\text{for all } R > 0, \tag*{(12)}
sup⁡h∈R∥Whq∥L4([−T,T]×R)4≤C2(1+T)M(q0)2.(13)\sup_{h\in\mathbb{R}}\lVert W_hq\rVert_{L^4([-T,T]\times\mathbb{R})}^4 \le C_2(1+T)M(q_0)^2. \tag*{(13)}

In particular C(m0)≤2C1C(m_0) \le2C_1 and m∗≥m0>0m^* \ge m_0 > 0.

The estimate (12) is scale invariant. It is the nonlinear counterpart, on parabolic cylinders, of the Kato smoothing estimate for the free Schrödinger group. The point of Corollary 1.3 is not its difficulty, since it follows from the proof of [20 Proposition 4.1].

Its point is that it locates the obstruction of [20 Proposition 1.11] at positive mass. Our second result analyses this obstruction.

To state it, recall from [20 Section 1.3] that for each ϑ∈(0,π/2)\vartheta\in(0,\pi/2) DNLS has a soliton of mass 8ϑ8\vartheta travelling with velocity −2cot⁡(2ϑ)-2\cot(2\vartheta). The soliton with ϑ=π/4\vartheta= \pi/4 is stationary. We denote by Sϑ\mathcal{S}_\vartheta the family of all solutions obtained from it by translations in space and time, phase rotations and the scaling (1.2); see (2.12).

Theorem 1.4 (The stationary-soliton threshold). There are absolute constants c>0c > 0, C≥1C \ge1 and m1∈[π,2π)m_1 \in[\pi,2\pi) such that the following hold.

(i) C(m)=∞C(m) = \infty for every m≥2πm \ge2\pi.

(ii) C(m)≥c(2π−m)−1C(m) \ge c(2\pi-m)^{-1} for every m∈[m1,2π)m \in[m_1,2\pi). In particular C(m)→∞C(m) \to\infty as m↑2πm \uparrow2\pi.

(iii) Every soliton u∈Sϑu \in\mathcal{S}_\vartheta with ϑ∈(0,π/4)\vartheta\in(0,\pi/4) satisfies

LS1(u)≤C(1+12π−M(u))M(u).(14)\mathrm{LS}_1(u) \le C\left(1 + \frac{1}{2\pi- M(u)}\right)M(u). \tag*{(14)}

Consequently m0≤m∗≤2πm_0 \le m^* \le2\pi. Moreover, whatever the value of m∗m^*, the constant C(m)C(m) is at least of order (2π−m)−1(2\pi-m)^{-1} near 2π2\pi, and the soliton family alone produces exactly this rate.

Part (i) is a reformulation of [20 Proposition 1.11]. We include a proof because the same lower bound, applied to slowly moving solitons, gives (ii). Part (iii) shows that solitons alone cannot push m∗m^* below 2π2\pi. Their contribution to local smoothing is controlled by the time they need to cross a window, which is of order (λcot⁡2ϑ)−1(\lambda\cot2\vartheta)^{-1}, against a local smoothing density of order λ\lambda; see Figure 2. These considerations lead to the following conjecture, which is the main open problem of this paper.

Figure 2. Trajectories x(t)=x0−2λbtx(t)=x_0-2\lambda b t, b=cot⁡(2ϑ)b=\cot(2\vartheta), of rescaled solitons in the space-time window [−1,1]×supp⁡Wh[-1,1]\times\operatorname{supp} W_h (schematic). Each soliton carries a local smoothing density of order λ\lambda, so its contribution to LS1\mathrm{LS}_1 is of order λ⋅min⁡{1,(λb)−1}≤b−1\lambda\cdot\min\{1,(\lambda b)^{-1}\}\le b^{-1}. This gives Proposition 4.1. The slow solitons with λb≃1\lambda b\simeq1 and λ≃b−1\lambda\simeq b^{-1} attain this order, which gives Theorem 1.4(ii).

Conjecture 1.5 (Mass threshold for uniform local smoothing). m∗=2πm^* = 2\pi. More precisely, C(m)<∞C(m) < \infty for every m<2πm < 2\pi, and

C(m)≃(2π−m)−1as m↑2π.C(m) \simeq(2\pi-m)^{-1} \qquad\text{as } m \uparrow2\pi.

In § 5 we explain why we believe this conjecture, using the trace formula for the Kaup–Newell scattering data. For Schwartz data of mass m<2πm < 2\pi, every eigenvalue ζj\zeta_j of the Kaup–Newell problem satisfies arg⁡ζj≤m/4<π/2\arg\zeta_j \le m/4 < \pi/2. Consequently, no component of the discrete spectrum can be stationary; see Remark 5.1. We also formulate a sharp form of Theorem 1.2 (Conjecture 5.3), in which κ2\kappa^2 is replaced by κ\kappa, and a global-in-time question for small mass (Question 5.4).

Figure 1 summarizes what is proved and what is conjectured about C(m)C(m), and Table 1 places our results among the known local smoothing statements.

Uniform local smoothing constant along the mass axis

Figure 1. The uniform local smoothing constant C(m)C(m) along the mass axis (not to scale). Green: proved finite. Orange: proved lower bound, finiteness conjectured. Red: proved infinite. The soliton bound of Theorem 1.4(iii) matches the lower bound of Theorem 1.4(ii).

SettingStatementConstantSource
Free Schrödinger groupsup⁡x∈R∫R∣Dx1/2eit∂x2u0(x)∣2 dt=c∥u0∥L22\sup_{x \in\mathbb{R}}\int_{\mathbb{R}} \lvert D_x^{1/2}e^{it\partial_x^2}u_0(x)\rvert^2\,dt=c\lVert u_0\rVert_{L^2}^2exact[33, 36]
DNLS, Q⊂S(R)Q \subset\mathcal{S}(\mathbb{R}) bounded and equicontinuousLST(q)≲T,QM(q0)\mathrm{LS}_T(q)\lesssim_{T,Q}M(q_0)implicit in QQ[20 Thm. 1.5]
DNLS, rescaled stationary solitons, M≡2πM\equiv2\pilocal smoothing fails on ballsno bound[20 Prop. 1.11]
DNLS, q0∈S(R)q_0\in\mathcal{S}(\mathbb{R})LST(q)≤C1(1+κ(q0)2T)M(q0)\mathrm{LS}_T(q)\le C_1(1+\kappa(q_0)^2T)M(q_0); power ≥1\ge1 necessaryexplicit in κ(q0)\kappa(q_0)Thm. 1.2 (new)
DNLS, M(q0)≤m0M(q_0)\le m_0uniform bound, Morrey form, localized Lt,x4L^4_{t,x}uniformCor. 1.3 (implicit in [20])
DNLS, m↑2πm\uparrow2\piC(m)≥c(2π−m)−1C(m)\ge c(2\pi-m)^{-1}blows upThm. 1.4(i),(ii) (new)
Solitons of mass <2π<2\piLS1(u)≤C(1+12π−M(u))M(u)\mathrm{LS}_1(u)\le C\left(1+\frac{1}{2\pi-M(u)}\right)M(u)sharp rateThm. 1.4(iii) (new)
DNLS, m<2πm<2\piC(m)<∞C(m)<\infty, C(m)≃(2π−m)−1C(m)\simeq(2\pi-m)^{-1}conjecturedConj. 1.5, 5.2

Table 1. Local smoothing statements on R\mathbb{R} and the position of the present results.

Related literature

Local and global well-posedness. Local well-posedness of (1.1) in Hs(R)H^s(\mathbb{R}) was proved for s>3/2s>3/2 by Tsutsumi and Fukuda [65], and in H1H^1 by Hayashi and Ozawa [23, 24, 26] by means of a gauge transformation; see also Ozawa [57]. Takaoka [61] reached HsH^s, s≥1/2s\ge1/2, by combining the gauge transformation with the Fourier restriction method. Below s=1/2s=1/2 the data-to-solution map fails to be C3C^3 [61, 62], and it is not even uniformly continuous on bounded sets [7]. Hence contraction arguments in HsH^s cannot go further. Local results in Fourier–Lebesgue and modulation spaces, closer to the critical scaling, are due to Grünrock [16] and to Guo, Ren and Wang [18].

Global results first required a mass restriction. Using the gauged energy and the sharp Gagliardo–Nirenberg inequality, Hayashi and Ozawa [26] and Ozawa [57] obtained global H1H^1 solutions for M<2πM < 2\pi. Colliander, Keel, Staffilani, Takaoka and Tao [9, 10] carried the 2π2\pi condition to HsH^s, s>1/2s > 1/2, by the II-method, and Miao, Wu and Xu [52] reached H1/2H^{1/2}. Wu exploited the momentum to go beyond 2π2\pi, first to 2π+ε∗2\pi+ \varepsilon_* [69] and then to 4π4\pi [70] in H1H^1. Guo and Wu [19] obtained the same threshold 4π4\pi in H1/2H^{1/2}; see also Fukaya, Hayashi and Inui [14]. The value 4π4\pi is the mass of the algebraic soliton [20 (1.26)], [2], for which all other polynomial conservation laws vanish; see the discussion following [20 (1.26)].

The inverse scattering approach, initiated by Lee [48], removed the mass restriction for data in weighted spaces [49, 30, 32, 58, 59]. It also produced long-time asymptotics for soliton-free data [50] and soliton resolution for generic weighted data [29]; see the survey [31]. Bahouri and Perelman [2] then proved global well-posedness in H1/2(R)H^{1/2}(\mathbb{R}) without restriction, with uniform-in-time bounds, by combining the integrable structure with profile decompositions. Bahouri, Leslie and Perelman [1] established uniform-in-time HsH^s bounds for s≥1/2s \ge1/2.

At the level of low-regularity conservation laws, Klaus and Schippa [45] obtained a priori estimates in Besov spaces of positive regularity for small mass, following the approach of Killip, Vișan and Zhang [43]. Tang and Xu [63] derived the corresponding microscopic conservation laws for small mass and pointed out that their coercive currents should yield local smoothing in HsH^s. They did not state an L2L^2-level estimate. Killip, Ntekoume and Vișan [41] proved equicontinuity of orbits for M<4πM < 4\pi and global well-posedness in H1/6H^{1/6} under the same restriction. Harrop-Griffiths, Killip and Vișan [21] then removed the mass restriction from the equicontinuity statement. The L2L^2 theorem [20] builds on all of these works, on the method of commuting flows of Killip and Vișan [42], and on its second-generation form developed for the cubic NLS and mKdV [22]. Table 2 summarizes this development.

SpaceMass conditionMain toolReferences
HsH^s, s>3/2s > 3/2 (local)noneenergy method[65]
H1H^1 (local)nonegauge transformation[23, 24, 26]
H1H^1 (global)M<2πM < 2\pigauged energy, sharp Gagliardo–Nirenberg[26, 57]
HsH^s, s≥1/2s \ge1/2 (local)nonegauge, Fourier restriction[61]
HsH^s, s>1/2s > 1/2 (global)M<2πM < 2\piII-method[9, 10]
H1/2H^{1/2} (global)M<2πM < 2\piII-method, resonant decomposition[52]
H1H^1 (global)M<2π+ε∗M < 2\pi+ \varepsilon_*, then M<4πM < 4\pimomentum, variational arguments[69, 70]
H1/2H^{1/2} (global)M<4πM < 4\piII-method, momentum[19]
weighted spaces (global)noneinverse scattering[49, 30, 32, 59, 58]
H1/2H^{1/2} (global)noneintegrability, profile decompositions[2]
H1/6H^{1/6} (global)M<4πM < 4\piconservation laws, equicontinuity[41]
L2L^2 (global)nonecommuting flows, local smoothing[20]

Table 2. Well-posedness of DNLS on the line (selection).

The periodic problem. On the circle, local well-posedness in H1/2H^{1/2} is due to Herr [27]. Grünrock and Herr [17] treated Fourier–Lebesgue data, and Deng, Nahmod and Yue [12] obtained the optimal Fourier–Lebesgue local theory. Global results under the restriction M<4πM < 4\pi were obtained by Mosincat and Oh [56] in H1(T)H^1(\mathbb{T}) and by Mosincat [55] in H1/2(T)H^{1/2}(\mathbb{T}). Bahouri and Perelman [3] first relaxed the restriction to M<8πM < 8\pi, and very recently they proved global well-posedness in H1(T)H^1(\mathbb{T}) with no mass restriction [4]. Well-posedness in L2(T)L^2(\mathbb{T}) remains open. Local smoothing, which is the tool studied here, is not available on the circle.

Solitons. The two-parameter soliton family and its orbital stability were studied by Colin and Ohta [8]. Kwon and Wu [46] treated the algebraic soliton, and Le Coz and Wu [47] multi-solitons. The coincidence of 2π2\pi and 4π4\pi with the masses of the static and algebraic solitons is noted in [2].

A parallel model. The continuum Calogero–Moser model, also called the Calogero–Moser derivative NLS, is another integrable, mass-critical derivative Schrödinger equation. There the stationary soliton also has mass 2π2\pi. Gérard and Lenzmann [15] proved global well-posedness of smooth solutions with mass at most 2π2\pi. Killip, Laurens and Vişan [40] proved well-posedness in L+2(R)L^2_+(\mathbb{R}) for focusing data of mass below 2π2\pi. Above 2π2\pi, Hogan and Kowalski [28] constructed solutions with unbounded Sobolev norms and mass arbitrarily close to 2π2\pi, and Kim, Kim and Kwon [44] constructed smooth finite-time blow-up solutions with mass arbitrarily close to that of the soliton. Small-mass dispersive decay is proved in [39]. In that model, therefore, 2π2\pi is a well-posedness threshold. For DNLS, well-posedness in L2L^2 holds for every mass [20], and Conjecture 1.5 proposes that 2π2\pi is instead the threshold for uniform smoothing. A threshold given by a ground-state mass is also familiar from the focusing mass-critical NLS [68, 13] and from the scattering theory below the ground state developed by Kenig and Merle [35]. For recent one-dimensional and inhomogeneous instances of this philosophy, we refer to [6, 51].

Local smoothing. For the free Schrödinger equation, Kato-type smoothing goes back to Kato [33], Constantin and Saut [11], Sjölin [60] and Vega [66]. In one dimension, Kenig, Ponce and Vega [36] proved the exact identity

sup⁡x∈R∫R∣Dx1/2eit∂x2u0(x)∣2 dt=c∥u0∥L22.\sup_{x\in\mathbb{R}}\int_{\mathbb{R}}\left|D_x^{1/2}e^{it\partial_x^2}u_0(x)\right|^2\,dt=c\|u_0\|_{L^2}^2.

The same authors used local smoothing to treat Schrödinger equations with derivative nonlinearities [37, 38]. The finite-scale behaviour of the smoothing quantity was studied by Vega and Visciglia [67]. An exact identity and sharp constants at finite scale were obtained by Bai and the author [5]. The monotonicity approach to local smoothing used in [20] is discussed in [64].

To the best of our knowledge, the literature contains no uniform local smoothing statement for DNLS on L2L^2-balls of positive mass beyond what is implicit in [20 Proposition 4.1]. Likewise, the only negative statement we are aware of is [20 Proposition 1.11], which shows failure at mass 2π2\pi; the remarks in [63] concern small mass in HsH^s, s>0s>0. The question whether 2π2\pi is sharp, and how the constant depends on the data, therefore appears to be open. We do not claim novelty for Corollary 1.3, which is implicit in the proof of [20 Proposition 4.1]. Our contributions are the explicit dependence in Theorem 1.2, the two-sided soliton analysis of Theorem 1.4, which pins down the expected rate (2π−m)−1(2\pi-m)^{-1}, and the formulation of Conjectures 1.5 and 5.3 together with a spectral reformulation.

The paper is organized as follows. § 2 collects notation, the results of [20] that we use, and two elementary tools, namely a window-summation lemma for H˙1/2\dot H^{1/2} and bounds on the soliton profiles. § 3 proves Theorem 1.2 and Corollary 1.3, and § 4 proves Theorem 1.4. § 5 discusses the method, its limitations and the evidence for the conjectures. § 6 contains concluding remarks and further directions.

Preliminaries and technical tools

Notation

We use the unitary Fourier transform f^(ξ)=(2π)−1/2∫Re−ixξf(x) dx\widehat{f}(\xi)=(2\pi)^{-1/2}\int_{\mathbb{R}}e^{-ix\xi}f(x)\,dx, so that ∥f^∥L2=∥f∥L2\|\widehat{f}\|_{L^2}=\|f\|_{L^2}. Following [20], we write

∥f∥Hs2=∫R(4+ξ2)s∣f^(ξ)∣2 dξ,∥f∥H˙1/22=∫R∣ξ∣∣f^(ξ)∣2 dξ.\|f\|_{H^s}^2=\int_{\mathbb{R}}(4+\xi^2)^s|\widehat{f}(\xi)|^2\,d\xi,\qquad\|f\|_{\dot H^{1/2}}^2=\int_{\mathbb{R}}|\xi||\widehat{f}(\xi)|^2\,d\xi.

Since ∣ξ∣≤(4+ξ2)1/2≤2+∣ξ∣|\xi|\le(4+\xi^2)^{1/2}\le2+|\xi|, we have

∥f∥H˙1/22≤∥f∥H1/22≤2∥f∥L22+∥f∥H˙1/22.(15)\|f\|_{\dot H^{1/2}}^2\le\|f\|_{H^{1/2}}^2\le2\|f\|_{L^2}^2+\|f\|_{\dot H^{1/2}}^2. \tag*{(15)}

We shall use the Gagliardo representation

∥f∥H˙1/22=12π∬R2∣f(x)−f(y)∣2∣x−y∣2 dx dy.(16)\|f\|_{\dot H^{1/2}}^2=\frac{1}{2\pi}\iint_{\mathbb{R}^2}\frac{|f(x)-f(y)|^2}{|x-y|^2}\,dx\,dy. \tag*{(16)}

It follows from Plancherel’s theorem, applied for fixed h=x−yh=x-y, and from the identity

∫R(2−2cos⁡hξ)h−2 dh=2π∣ξ∣.\int_{\mathbb{R}}(2-2\cos h\xi)h^{-2}\,dh=2\pi|\xi|.

We also use the elementary interpolation bound

∥f∥H1/22≤∥f∥L2(2∥f∥L2+∥f′∥L2)(17)\|f\|_{H^{1/2}}^{2}\leq\|f\|_{L^{2}}\left(2\|f\|_{L^{2}}+\|f'\|_{L^{2}}\right) \tag*{(17)}

which follows from the Cauchy–Schwarz inequality in ξ\xi and from ∥f∥H12=4∥f∥L22+∥f′∥L22≤(2∥f∥L2+∥f′∥L2)2\|f\|_{H^{1}}^{2}=4\|f\|_{L^{2}}^{2}+\|f'\|_{L^{2}}^{2}\leq\left(2\|f\|_{L^{2}}+\|f'\|_{L^{2}}\right)^{2}.

The weight W=sech⁡6(⋅/99)W=\operatorname{sech}^{6}(\cdot/99) from (4) satisfies 0<W≤10<W\leq1 and

∣W′∣≤699W≤116W,W(x)W(y)≤26e6∣x−y∣/99,∫RW2 dx=5127<74.(18)|W'|\leq\frac{6}{99}W\leq\frac{1}{16}W,\qquad\frac{W(x)}{W(y)}\leq2^{6}e^{6|x-y|/99},\qquad\int_{\mathbb{R}}W^{2}\,dx=\frac{512}{7}<74. \tag*{(18)}

The first inequality follows from W′=−699tanh⁡(⋅/99)WW'=-\frac{6}{99}\tanh(\cdot/99)W. The second follows from sech⁡a≤2e−∣a∣\operatorname{sech}a\leq2e^{-|a|} and sech⁡b≥e−∣b∣\operatorname{sech}b\geq e^{-|b|}. The value of ∫RW2=99∫Rsech⁡12\int_{\mathbb{R}}W^{2}=99\int_{\mathbb{R}}\operatorname{sech}^{12} is recorded in [20 (2.1)]. Moreover, W(s)≥sech⁡6(2/99)≥12W(s)\geq\operatorname{sech}^{6}(2/99)\geq\frac{1}{2} for ∣s∣≤2|s|\leq2.

Results from [20] used below

We use [20] as a black box through three statements. The first is the global well-posedness theorem [20 Theorem 1.1]. It provides (3) and the group property Φ(t+s)=Φ(t)Φ(s)\Phi(t+s)=\Phi(t)\Phi(s). Moreover, Schwartz data lead to global Schwartz solutions, which conserve the mass and the perturbation determinant a(k;q)a(k;q).

The second is the equicontinuity theorem of [21], quoted as [20 Theorem 1.7]. It states that if Q⊂S(R)Q\subset\mathcal{S}(\mathbb{R}) is bounded and equicontinuous in L2L^{2}, then so is the set

Q∗={q∈S(R):a(k;q)≡a(k;q~) for some q~∈Q}(19)Q_{*}=\left\{q\in\mathcal{S}(\mathbb{R}):a(k;q)\equiv a(k;\widetilde{q})\text{ for some }\widetilde{q}\in Q\right\} \tag*{(19)}

of [20 (1.11)]. We combine it with [20 Lemma 2.1], which states that sup⁡q∈Q∥q∥Eκε→0\sup_{q\in Q}\|q\|_{E_{\kappa}^{\varepsilon}}\to0 as κ→∞\kappa\to\infty for every bounded equicontinuous set QQ. (This last fact is elementary: splitting the integral (8) at a frequency KK, we get ∥q∥Eκε2≤(K/2κ)2ε∥q∥L22+∫∣ξ∣>K∣q^∣2 dξ\|q\|_{E_{\kappa}^{\varepsilon}}^{2}\leq(K/2\kappa)^{2\varepsilon}\|q\|_{L^{2}}^{2}+\int_{|\xi|>K}|\widehat{q}|^{2}\,d\xi, and equicontinuity makes the last term uniformly small.) For q0∈S(R)q_{0}\in\mathcal{S}(\mathbb{R}) we also set

κ♭(q0):=inf⁡{κ>0:sup⁡t∈R∥Φ(t)q0∥Eκε≤δ0}∈[0,∞).\kappa_{\flat}(q_{0}):=\inf\left\{\kappa>0:\sup_{t\in\mathbb{R}}\|\Phi(t)q_{0}\|_{E_{\kappa}^{\varepsilon}}\leq\delta_{0}\right\}\in[0,\infty).

Lemma 2.1. Let q0∈S(R)q_{0}\in\mathcal{S}(\mathbb{R}) and q(t)=Φ(t)q0q(t)=\Phi(t)q_{0}.

(i) κ(q0)<∞\kappa(q_{0})<\infty, and the infimum in Definition 1.1 is attained. More precisely, sup⁡t∥q(t)∥Eκε≤δ0\sup_{t}\|q(t)\|_{E_{\kappa}^{\varepsilon}}\leq\delta_{0} for every κ≥κ(q0)\kappa\geq\kappa(q_{0}).

(ii) If M(q0)≤δ02M(q_{0})\leq\delta_{0}^{2}, then κ(q0)=1\kappa(q_{0})=1.

(iii) For every λ>0\lambda>0 and κ>0\kappa>0 one has ∥(q0)λ∥Eκε=∥q0∥Eκ/λε\|(q_{0})_{\lambda}\|_{E_{\kappa}^{\varepsilon}}=\|q_{0}\|_{E_{\kappa/\lambda}^{\varepsilon}}, where (q0)λ=λ1/2q0(λ⋅)(q_{0})_{\lambda}=\lambda^{1/2}q_{0}(\lambda\cdot).

Consequently, κ(q0)=max⁡{1,κ♭(q0)}\kappa(q_{0})=\max\{1,\kappa_{\flat}(q_{0})\} and κ♭((q0)λ)=λκ♭(q0)\kappa_{\flat}((q_{0})_{\lambda})=\lambda\kappa_{\flat}(q_{0}).

Proof. For κ>0\kappa>0 set F(κ):=sup⁡t∈R∥q(t)∥EκεF(\kappa):=\sup_{t\in\mathbb{R}}\|q(t)\|_{E_{\kappa}^{\varepsilon}}. The weight in (8) is at most one and decreases in κ\kappa. Hence, by dominated convergence, each map κ↦∥q(t)∥Eκε\kappa\mapsto\|q(t)\|_{E_{\kappa}^{\varepsilon}} is continuous and nonincreasing. We infer that FF is nonincreasing and lower semicontinuous, being a supremum of continuous functions. Therefore the set S:={κ>0:F(κ)≤δ0}S:=\{\kappa>0:F(\kappa)\leq\delta_{0}\} is closed in (0,∞)(0,\infty) and is a half-line: either S=(0,∞)S=(0,\infty), in which case κ♭(q0)=0\kappa_{\flat}(q_{0})=0, or S=[κ♭(q0),∞)S=[\kappa_{\flat}(q_{0}),\infty).

(i) The DNLS flow conserves a(k;q)a(k;q). Hence the whole orbit {Φ(t)q0:t∈R}\{\Phi(t)q_{0}:t\in\mathbb{R}\} is contained in Q∗Q_{*} with Q={q0}Q=\{q_{0}\}. A single Schwartz function forms a bounded equicontinuous set, so Q∗Q_{*} is bounded and equicontinuous by [20 Theorem 1.7]. By [20 Lemma 2.1], F(κ)→0F(\kappa)\to0 as κ→∞\kappa\to\infty, so SS is nonempty and κ♭(q0)<∞\kappa_{\flat}(q_{0})<\infty. Since SS is a closed half-line, κ(q0)=inf⁡(S∩[1,∞))\kappa(q_{0})=\inf(S\cap[1,\infty)) belongs to SS, and S⊃[κ(q0),∞)S\supset[\kappa(q_{0}),\infty). (ii) If M(q0)≤δ02M(q_0) \le\delta_0^2, the weight in (8) being at most one, we get ∥q(t)∥E1ε≤∥q(t)∥L2=M(q0)1/2≤δ0\|q(t)\|_{E_1^\varepsilon} \le\|q(t)\|_{L^2} = M(q_0)^{1/2} \le\delta_0 for all tt by mass conservation. Thus 1∈S1 \in S and κ(q0)=1\kappa(q_0) = 1.

(iii) We have (q0)λ(ξ)=λ−1/2q^0(ξ/λ)(q_0)_\lambda(\xi) = \lambda^{-1/2}\widehat{q}_0(\xi/\lambda), and the substitution ξ=λη\xi= \lambda\eta gives

∫∣ξ∣2ελ−1∣q^0(ξ/λ)∣2(4κ2+ξ2)ε dξ=∫∣η∣2ε∣q^0(η)∣2(4(κ/λ)2+η2)ε dη.\int\frac{|\xi|^{2\varepsilon}\lambda^{-1}|\widehat{q}_0(\xi/\lambda)|^2}{(4\kappa^2+\xi^2)^\varepsilon}\,d\xi = \int\frac{|\eta|^{2\varepsilon}|\widehat{q}_0(\eta)|^2}{(4(\kappa/\lambda)^2+\eta^2)^\varepsilon}\,d\eta.

The identity κ(q0)=max⁡{1,κ♭(q0)}\kappa(q_0)=\max\{1,\kappa_\flat(q_0)\} follows from the half-line structure of SS. Finally, by the scaling invariance (2) and uniqueness of Schwartz solutions, Φ(t)(q0)λ=(Φ(λ2t)q0)λ\Phi(t)(q_0)_\lambda=(\Phi(\lambda^2t)q_0)_\lambda. Taking the supremum in tt and using the identity just proved, we obtain sup⁡t∥Φ(t)(q0)λ∥Eκε=F(κ/λ)\sup_t\|\Phi(t)(q_0)_\lambda\|_{E_\kappa^\varepsilon}=F(\kappa/\lambda), which yields κ♭((q0)λ)=λκ♭(q0)\kappa_\flat((q_0)_\lambda)=\lambda\kappa_\flat(q_0). □

The third statement is the local smoothing estimate [20 Proposition 4.1, (4.1)]. It reads as follows. Let Q⊂S(R)Q\subset\mathcal{S}(\mathbb{R}) be bounded and equicontinuous in L2L^2, and assume that the set of DNLS orbits

{Φ(t)q:∣t∣≤1, q∈Q}\{\Phi(t)q: |t|\le1,\ q\in Q\}

is δ\delta-good for a sufficiently small δ\delta. (In [20 Proposition 4.1] this set is written {etJ∇Hq:∣t∣≤1, q∈Q}\{e^{tJ\nabla H}q: |t|\le1,\ q\in Q\}, where HH is the DNLS Hamiltonian, and it is also denoted by Q∗Q_*; it should not be confused with the isospectral set (19) above.) Then ∥q∥X1/2≲∥q(0)∥L2\|q\|_{X_{1/2}}\lesssim\|q(0)\|_{L^2} uniformly for q(0)∈Qq(0)\in Q. Here δ\delta-good means bounded, equicontinuous and subject to the smallness condition [20 (1.19)], namely sup⁡∥q∥E1ε≤δ\sup\|q\|_{E_1^\varepsilon}\le\delta; see [20 Definition 1.8]. The local smoothing norm is defined in [20 (2.24), (2.25)] over the time interval [−1,1][-1,1], and by [20 (2.32)] with κ=1\kappa=1 it satisfies

∥q∥X1/22≃sup⁡h∈R∫−11∥Whq(t)∥H1/22 dt.\|q\|_{X_{1/2}}^2\simeq\sup_{h\in\mathbb{R}}\int_{-1}^{1}\|W_hq(t)\|_{H^{1/2}}^2\,dt.

with absolute implicit constants, because ψ1/2=W\psi^{1/2}=W and ∥(4−∂2)−1/2f∥H3/2=∥f∥H1/2\|(4-\partial^2)^{-1/2}f\|_{H^{3/2}}=\|f\|_{H^{1/2}}. Finally, a compact subset of L2L^2 is automatically bounded and equicontinuous, and the orbit {Φ(t)q0:∣t−s∣≤1}\{\Phi(t)q_0:|t-s|\le1\} of a Schwartz datum is compact. Applying [20 Proposition 4.1] with Q={Φ(s)q0}Q=\{\Phi(s)q_0\}, we obtain the result in the following form.

Proposition 2.2 ([20 Proposition 4.1]). There are constants δ0∈(0,1]\delta_0\in(0,1] and C0≥1C_0\ge1, depending only on ε\varepsilon, such that the following holds. Let q0∈S(R)q_0\in\mathcal{S}(\mathbb{R}) and s∈Rs\in\mathbb{R}, and suppose that ∥Φ(t)q0∥E1ε≤δ0\|\Phi(t)q_0\|_{E_1^\varepsilon}\le\delta_0 for all ∣t−s∣≤1|t-s|\le1. Then

sup⁡h∈R∫s−1s+1∥WhΦ(t)q0∥H1/22 dt≤C0M(q0).(20)\sup_{h\in\mathbb{R}}\int_{s-1}^{s+1}\|W_h\Phi(t)q_0\|_{H^{1/2}}^2\,dt\le C_0M(q_0). \tag*{(20)}

Decreasing δ0\delta_0 only strengthens the hypothesis, so the normalization δ0≤1\delta_0\le1 costs nothing. The time translation from [−1,1][-1,1] to [s−1,s+1][s-1,s+1] uses only the group property and mass conservation: one applies the estimate to the Schwartz datum Φ(s)q0\Phi(s)q_0.

Remark 2.3. The essential point in Proposition 2.2 is that δ0\delta_0 and C0C_0 depend on nothing but ε\varepsilon. We have checked this against the proof of [20 Proposition 4.1] on pp. 895–896 of the published version. Combining the identity [20 (4.4)] with [20 Lemmas 4.2 and 4.9], [20 (2.4)] and [20 (2.13)], the authors arrive at [20 (4.33)], which for κ=1\kappa=1 reads

∥q∥X1/22≲∥q∥Lt∞Lx∞E1ε∥q∥X1/22+∥q∥Lt∞Lx∞E1ε2[1+∥q∥Lt∞Lx22]+∥q∥Lt∞Lx22.\|q\|_{X_{1/2}}^2\lesssim\|q\|_{L^\infty_tL^\infty_xE_1^\varepsilon}\|q\|_{X_{1/2}}^2+\|q\|_{L^\infty_tL^\infty_xE_1^\varepsilon}^2\left[1+\|q\|_{L^\infty_tL^2_x}^2\right]+\|q\|_{L^\infty_tL^2_x}^2.

If ∥q∥Lt∞E1ε≤δ\|q\|_{L^\infty_tE_1^\varepsilon}\le\delta with δ\delta small compared with the implicit constant, the first term is absorbed into the left-hand side; this is legitimate since ∥q∥X1/2<∞\|q\|_{X_{1/2}}<\infty for Schwartz solutions. Using ∥q∥Eε1≤min⁡{δ,∥q∥L2}\|q\|_{E^1_\varepsilon} \le\min\{\delta,\|q\|_{L^2}\} and mass conservation, we deduce that ∥q∥X1/22≲(2+δ2)M(q0)\|q\|_{X^{1/2}}^2 \lesssim(2+\delta^2)M(q_0).

In the proof of (4.1), the set QQ enters only through the smallness [20 (1.19)] along the trajectory and through [20 Lemmas 4.2 and 4.9], whose estimates are stated uniformly for qq in any set satisfying the hypotheses of [20 Proposition 4.1]. Equicontinuity is used quantitatively in that section only through [20 (2.3)], and only for the refined estimate [20, (4.2)}, which we do not need. This is consistent with the proof of [20 Theorem 1.5] on p. 921, where the authors note that the resulting constant depends on QQ solely through the rescaling parameter. The paper [20] does not fix a normalization of the Fourier transform. A different normalization multiplies every norm in (8), in [20 (1.19)] and in H1/2H^{1/2} by the same absolute factor, so only the values of δ0\delta_0 and C0C_0 change.

For a bounded equicontinuous set Q⊂S(R)Q \subset\mathcal{S}(\mathbb{R}), [20 Corollary 1.9] provides λ=λ(Q,δ0)\lambda=\lambda(Q,\delta_0) such that the rescaled isospectral set Qλ∗={λq(λ⋅):q∈Q∗}Q^*_\lambda=\{\sqrt{\lambda}q(\lambda\cdot):q\in Q^*\} is δ0\delta_0-good. Since Q∗Q^* contains every orbit issued from QQ, Lemma 2.1(iii) gives κ(q0)≤max⁡{1,1/λ}\kappa(q_0)\le\max\{1,1/\lambda\} for all q0∈Qq_0\in Q, which is the bound used after Theorem 1.2. In this sense Theorem 1.2 is a quantitative form of the proof of [20 Theorem 1.5], in which a uniform rescaling makes all orbits δ\delta-good and the estimate is then summed over unit time intervals.

Summing small windows

The following lemma converts local smoothing on windows of size r≤1r\le1 into local smoothing on unit windows. For r>0r>0 and η∈R\eta\in\mathbb{R} we set

φr,η(x):=W((x−η)/r).(21)\varphi_{r,\eta}(x):=W\left((x-\eta)/r\right). \tag*{(21)}

Lemma 2.4 (Window summation). There is an absolute constant C3C_3 such that for every r∈(0,1]r\in(0,1], every h∈Rh\in\mathbb{R} and every f∈H1/2(R)f\in H^{1/2}(\mathbb{R}),

∥Whf∥H˙1/22+1r∥Whf∥L22≤C3r∫RW(η−h)2(∥φr,ηf∥H˙1/22+1r∥φr,ηf∥L22) dη.(22)\|W_hf\|_{\dot H^{1/2}}^2+\frac{1}{r}\|W_hf\|_{L^2}^2\le\frac{C_3}{r}\int_{\mathbb{R}}W(\eta-h)^2\left(\|\varphi_{r,\eta}f\|_{\dot H^{1/2}}^2+\frac{1}{r}\|\varphi_{r,\eta}f\|_{L^2}^2\right)\,d\eta. \tag*{(22)}

Proof. By translation invariance we may take h=0h=0. We first treat the L2L^2 term. Fix xx and let ∣η−x∣≤r|\eta-x|\le r. Then φr,η(x)≥12\varphi_{r,\eta}(x)\ge\frac{1}{2}, and (18) gives W(x)≤26e6/99W(η)≤27W(η)W(x)\le2^6e^{6/99}W(\eta)\le2^7W(\eta). We average the identity ∣Wf(x)∣2=12r∫∣η−x∣≤r∣W(x)f(x)∣2 dη|Wf(x)|^2=\frac{1}{2r}\int_{|\eta-x|\le r}|W(x)f(x)|^2\,d\eta over these η\eta, insert the two bounds, and integrate in xx. We obtain

∥Wf∥L22≤216r∫RW(η)2∥φr,ηf∥L22 dη,(23)\|Wf\|_{L^2}^2\le\frac{2^{16}}{r}\int_{\mathbb{R}}W(\eta)^2\|\varphi_{r,\eta}f\|_{L^2}^2\,d\eta, \tag*{(23)}

which controls the second term on the left of (22).

For the homogeneous term, we split the double integral in (16) into the far region ∣x−y∣≥r|x-y|\ge r and the near region ∣x−y∣<r|x-y|<r. In the far region we use ∣Wf(x)−Wf(y)∣2≤2∣Wf(x)∣2+2∣Wf(y)∣2|Wf(x)-Wf(y)|^2\le2|Wf(x)|^2+2|Wf(y)|^2 and ∫∣y−x∣≥r∣x−y∣−2 dy=2/r\int_{|y-x|\ge r}|x-y|^{-2}\,dy=2/r. It follows that the far region contributes at most 8r−1∥Wf∥L228r^{-1}\|Wf\|_{L^2}^2, which is controlled by (23).

In the near region we again average over the η\eta with ∣η−x∣≤r|\eta-x|\le r. For such η\eta, both xx and yy lie in Jη=[η−2r,η+2r]J_\eta=[\eta-2r,\eta+2r], where φr,η≥12\varphi_{r,\eta}\ge\frac12. On JηJ_\eta we write Wf=mηgηWf=m_\eta g_\eta with gη=φr,ηfg_\eta=\varphi_{r,\eta}f and mη=W/φr,ηm_\eta=W/\varphi_{r,\eta}. From (18) and r≤1r\le1 we get ∣mη∣≤28W(η)|m_\eta|\le2^8W(\eta) on JηJ_\eta. Moreover, ∣φr,η′∣≤φr,η/(16r)|\varphi'_{r,\eta}|\le\varphi_{r,\eta}/(16r) and ∣W′∣≤W/16|W'|\le W/16, so that

∣mη′∣≤∣W′∣φr,η+W∣φr,η′∣φr,η2≤W4r≤25W(η)ron Jη.|m'_\eta|\le\frac{|W'|}{\varphi_{r,\eta}}+\frac{W|\varphi'_{r,\eta}|}{\varphi_{r,\eta}^2}\le\frac{W}{4r}\le\frac{2^5W(\eta)}{r}\quad\text{on }J_\eta.

Writing Wf(x)−Wf(y)=mη(x)(gη(x)−gη(y))+(mη(x)−mη(y))gη(y)Wf(x)-Wf(y)=m_{\eta}(x)\bigl(g_{\eta}(x)-g_{\eta}(y)\bigr)+\bigl(m_{\eta}(x)-m_{\eta}(y)\bigr)g_{\eta}(y) and using the mean value theorem on the interval JηJ_{\eta}, we deduce that

∣Wf(x)−Wf(y)∣≤28W(η)∣gη(x)−gη(y)∣+25W(η)r−1∣x−y∣∣gη(y)∣.\left|Wf(x)-Wf(y)\right|\le2^{8}W(\eta)\left|g_{\eta}(x)-g_{\eta}(y)\right|+2^{5}W(\eta)r^{-1}|x-y|\left|g_{\eta}(y)\right|.

We square this inequality, divide by ∣x−y∣2|x-y|^{2} and integrate over the near region, using ∫∣x−y∣<rdx=2r\int_{|x-y|<r}dx=2r for the second term. We conclude that the near region contributes at most

12r∫RW(η)2(217⋅2π∥gη∥H˙1/22+212r∥gη∥L22) dη\frac{1}{2r}\int_{\mathbb{R}}W(\eta)^{2}\left(2^{17}\cdot2\pi\left\|g_{\eta}\right\|_{\dot H^{1/2}}^{2}+\frac{2^{12}}{r}\left\|g_{\eta}\right\|_{L^{2}}^{2}\right)\,d\eta

to the double integral. Dividing by 2π2\pi and collecting the three contributions gives (22). □

The soliton family

We now record what we use about the solitons. Following [20 Section 1.3], for ϑ∈(0,π/2)\vartheta\in(0,\pi/2) we put

q0(x;ϑ)=A(x;ϑ)e−ixcot⁡(2ϑ),A(x;ϑ)=2sin⁡(2ϑ)Z3∣Z∣4,Z=cosh⁡(x−iϑ).(24)q_{0}(x;\vartheta)=A(x;\vartheta)e^{-ix\cot(2\vartheta)},\qquad A(x;\vartheta)=\frac{\sqrt{2\sin(2\vartheta)}Z^{3}}{|Z|^{4}},\qquad Z=\cosh(x-i\vartheta). \tag*{(24)}

Note that Z=cos⁡ϑcosh⁡x−isin⁡ϑsinh⁡xZ=\cos\vartheta\cosh x-i\sin\vartheta\sinh x, so that (24) is exactly the formula of [20]. (We have also checked numerically that (25) below solves (1) with this sign of the phase, and not with the opposite one, when ϑ≠π/4\vartheta\ne\pi/4. Only ∣∂xe−ixcot⁡2ϑ∣=∣cot⁡2ϑ∣|\partial_xe^{-ix\cot2\vartheta}|=|\cot2\vartheta| enters our estimates.) By [20 (1.24)],

q(t,x;ϑ)=q0(x+2cot⁡(2ϑ)t;ϑ)eitcsc⁡2(2ϑ)(25)q(t,x;\vartheta)=q_{0}\bigl(x+2\cot(2\vartheta)t;\vartheta\bigr)e^{it\csc^{2}(2\vartheta)} \tag*{(25)}

is a Schwartz solution of (1) of mass ∥q0(⋅;ϑ)∥L22=8ϑ\|q_{0}(\cdot;\vartheta)\|_{L^{2}}^{2}=8\vartheta [20 (1.25)]. For ϑ=π/4\vartheta=\pi/4 it reduces to the stationary soliton [20 (1.28)]. We let Sϑ\mathcal{S}_{\vartheta} be the set of all solutions

u(t,x)=eiγλ1/2q(λ2(t−t0),λ(x−x0);ϑ),λ>0, γ,t0,x0∈R.(26)u(t,x)=e^{i\gamma}\lambda^{1/2}q\bigl(\lambda^{2}(t-t_{0}),\lambda(x-x_{0});\vartheta\bigr),\qquad\lambda>0,\ \gamma,t_{0},x_{0}\in\mathbb{R}. \tag*{(26)}

Every u∈Sϑu\in\mathcal{S}_{\vartheta} has mass 8ϑ8\vartheta, and by (25)

∣u(t,x)∣=λ1/2∣A(λ(x−x(t));ϑ)∣,∣∂xu(t,x)∣=λ3/2∣∂xq0(λ(x−x(t));ϑ)∣,(27)|u(t,x)|=\lambda^{1/2}|A(\lambda(x-x(t));\vartheta)|,\qquad|\partial_xu(t,x)|=\lambda^{3/2}|\partial_xq_{0}(\lambda(x-x(t));\vartheta)|, \tag*{(27)}

with centre x(t)=x0−2λcot⁡(2ϑ)(t−t0)x(t)=x_{0}-2\lambda\cot(2\vartheta)(t-t_{0}). The second identity holds because the phase factor in (25) depends on tt only.

Lemma 2.5 (Profile bounds). Let ϑ∈(0,π/4]\vartheta\in(0,\pi/4] and b:=cot⁡(2ϑ)≥0b:=\cot(2\vartheta)\ge0. Then for all x∈Rx\in\mathbb{R}

∣A(x;ϑ)∣2=4sin⁡(2ϑ)cosh⁡(2x)+cos⁡(2ϑ)≤2(8ϑ)e−2∣x∣,∣∂xq0(x;ϑ)∣≤(9+b)∣A(x;ϑ)∣.(28)|A(x;\vartheta)|^{2}=\frac{4\sin(2\vartheta)}{\cosh(2x)+\cos(2\vartheta)}\le2(8\vartheta)e^{-2|x|},\qquad|\partial_xq_{0}(x;\vartheta)|\le(9+b)|A(x;\vartheta)|. \tag*{(28)}

Moreover, ϑ↦q0(⋅;ϑ)\vartheta\mapsto q_{0}(\cdot;\vartheta) is continuous from [π/8,π/4][\pi/8,\pi/4] into H1(R)H^{1}(\mathbb{R}).

Proof. Since ∣cosh⁡(a+ib)∣2=sinh⁡2a+cos⁡2b|\cosh(a+ib)|^{2}=\sinh^{2}a+\cos^{2}b and ∣sinh⁡(a+ib)∣2=sinh⁡2a+sin⁡2b|\sinh(a+ib)|^{2}=\sinh^{2}a+\sin^{2}b, we have

∣Z∣2=12(cosh⁡2x+cos⁡2ϑ),∣Z′∣2=sinh⁡2x+sin⁡2ϑ≤cosh⁡2x.|Z|^{2}=\frac{1}{2}\bigl(\cosh2x+\cos2\vartheta\bigr),\qquad|Z'|^{2}=\sinh^{2}x+\sin^{2}\vartheta\le\cosh^{2}x.

As ∣A∣2=2sin⁡(2ϑ)∣Z∣−2|A|^{2}=2\sin(2\vartheta)|Z|^{-2}, this gives the identity in (28). For ϑ≤π/4\vartheta\le\pi/4 we have cos⁡2ϑ≥0\cos2\vartheta\ge0. Hence ∣A∣2≤4sin⁡(2ϑ)/cosh⁡2x≤8sin⁡(2ϑ)e−2∣x∣|A|^{2}\le4\sin(2\vartheta)/\cosh2x\le8\sin(2\vartheta)e^{-2|x|}, and sin⁡2ϑ≤2ϑ\sin2\vartheta\le2\vartheta gives the first bound.

Next, A′/A=3Z′/Z−2(∣Z∣2)′/∣Z∣2A'/A=3Z'/Z-2(|Z|^{2})'/|Z|^{2}. Because cosh⁡2x≥cosh⁡2x\cosh2x\ge\cosh^{2}x and cos⁡2ϑ≥0\cos2\vartheta\ge0, we have ∣Z∣2≥12cosh⁡2x|Z|^{2}\ge\frac{1}{2}\cosh^{2}x, and therefore ∣Z′/Z∣≤2|Z'/Z|\le\sqrt{2}. Furthermore, (∣Z∣2)′/∣Z∣2=2sinh⁡2x/(cosh⁡2x+cos⁡2ϑ)(|Z|^{2})'/|Z|^{2}=2\sinh2x/(\cosh2x+\cos2\vartheta) is bounded by 22 in absolute value. We infer that ∣A′∣≤(32+4)∣A∣≤9∣A∣|A'|\le(3\sqrt{2}+4)|A|\le9|A|, and the product rule applied to (24) gives the second bound in (28). For the continuity statement, note that ZZ does not vanish on R×[π/8,π/4]\mathbb{R} \times[\pi/8,\pi/4]. Hence q0(x;ϑ)q_0(x;\vartheta) and ∂xq0(x;ϑ)\partial_xq_0(x;\vartheta) are continuous in (x,ϑ)(x,\vartheta) on this set. By (28) (with b≤1b \le1), they are dominated by Ce−∣x∣Ce^{-|x|} uniformly in ϑ\vartheta. Dominated convergence yields continuity in H1H^1. □

Quantitative local smoothing

Local smoothing at the natural scale

The first step is to rescale Proposition 2.2 to the scale on which the orbit is δ0\delta_0-good. In the original variables, this gives local smoothing on windows of size κ−1\kappa^{-1} over times of length κ−2\kappa^{-2}.

Proposition 3.1 (Scale-adapted local smoothing). Let q0∈S(R)q_0 \in\mathcal{S}(\mathbb{R}) and κ≥κ(q0)\kappa\ge\kappa(q_0), κ≥1\kappa\ge1. Then for every s,η∈Rs,\eta\in\mathbb{R},

∫s−κ−2s+κ−2∥φ1/κ,ηq(t)∥H˙1/22 dt≤C0M(q0)κ,(29)\int_{s-\kappa^{-2}}^{s+\kappa^{-2}} \|\varphi_{1/\kappa,\eta}q(t)\|_{\dot H^{1/2}}^2\,dt \le\frac{C_0M(q_0)}{\kappa}, \tag*{(29)}

where q=Φ(⋅)q0q=\Phi(\cdot)q_0 and φ1/κ,η=W(κ(⋅−η))\varphi_{1/\kappa,\eta}=W(\kappa(\cdot-\eta)).

Proof. Let v(τ,y):=κ−1/2q(κ−2τ,κ−1y)v(\tau,y):=\kappa^{-1/2}q(\kappa^{-2}\tau,\kappa^{-1}y). This is the rescaling (2) with λ=κ−1\lambda=\kappa^{-1}. Hence vv is the Schwartz solution with data (q0)1/κ(q_0)_{1/\kappa}, and M(v)=M(q0)M(v)=M(q_0). By Lemma 2.1(iii) with λ=1/κ\lambda=1/\kappa, and then by Lemma 2.1(i), we have ∥v(τ)∥E1ε=∥q(κ−2τ)∥Eκε≤δ0\|v(\tau)\|_{E^ \varepsilon_1}=\|q(\kappa^{-2}\tau)\|_{E^\varepsilon_\kappa}\le\delta_0 for all τ\tau. Proposition 2.2 applied at time κ2s\kappa^2s, together with (15), gives

∫κ2s−1κ2s+1∥Wκηv(τ)∥H˙1/22 dτ≤C0M(q0).\int_{\kappa^2s-1}^{\kappa^2s+1}\|W_{\kappa\eta}v(\tau)\|_{\dot H^{1/2}}^2\,d\tau\le C_0M(q_0).

Now Wκη(y)v(τ,y)=κ−1/2Fτ(y/κ)W_{\kappa\eta}(y)v(\tau,y)=\kappa^{-1/2}F_\tau(y/\kappa) with Fτ(x)=W(κ(x−η))q(κ−2τ,x)F_\tau(x)=W(\kappa(x-\eta))q(\kappa^{-2}\tau,x). A change of variables in the Fourier integral shows that ∥κ−1/2F(⋅/κ)∥H˙1/22=κ−1∥F∥H˙1/22\|\kappa^{-1/2}F(\cdot/\kappa)\|_{\dot H^{1/2}}^2=\kappa^{-1}\|F\|_{\dot H^{1/2}}^2. Substituting τ=κ2t\tau=\kappa^2t, we obtain

κ∫s−κ−2s+κ−2∥φ1/κ,ηq(t)∥H˙1/22 dt≤C0M(q0),\kappa\int_{s-\kappa^{-2}}^{s+\kappa^{-2}}\|\varphi_{1/\kappa,\eta}q(t)\|_{\dot H^{1/2}}^2\,dt\le C_0M(q_0),

which is (29). □

Proof of Theorem 1.2

Let q0∈S(R)q_0\in\mathcal{S}(\mathbb{R}), write κ=κ(q0)\kappa=\kappa(q_0) and M=M(q0)M=M(q_0), and fix h∈Rh\in\mathbb{R}. Consider a time interval B=[s−κ−2,s+κ−2]B=[s-\kappa^{-2},s+\kappa^{-2}]. We apply Lemma 2.4 with r=κ−1≤1r=\kappa^{-1}\le1 to f=q(t)f=q(t) for each t∈Bt\in B, and we integrate over BB. On the right-hand side, Proposition 3.1 bounds the time integral of ∥φ1/κ,ηq(t)∥H˙1/22\|\varphi_{1/\kappa,\eta}q(t)\|_{\dot H^{1/2}}^2 by C0M/κC_0M/\kappa for every η\eta. Since φ1/κ,η≤1\varphi_{1/\kappa,\eta}\le1, mass conservation gives ∫Bκ∥φ1/κ,ηq(t)∥L22 dt≤κ⋅2κ−2M\int_B\kappa\|\varphi_{1/\kappa,\eta}q(t)\|_{L^2}^2\,dt\le\kappa\cdot2\kappa^{-2}M. Hence, by Fubini’s theorem and (18),

∫B∥Whq(t)∥H˙1/22 dt≤C3κ∫RW(η−h)2(C0Mκ+2Mκ) dη≤74C3(C0+2)M.\int_B\|W_hq(t)\|_{\dot H^{1/2}}^2\,dt\le C_3\kappa\int_{\mathbb{R}}W(\eta-h)^2\left(\frac{C_0M}{\kappa}+\frac{2M}{\kappa}\right)\,d\eta\le74C_3(C_0+2)M.

Moreover, ∫B∥Whq(t)∥L22 dt≤2κ−2M≤2M\int_B\|W_hq(t)\|_{L^2}^2\,dt\le2\kappa^{-2}M\le2M. Combining these two bounds with (15), we get ∫B∥Whq(t)∥H1/22 dt≤C4M\int_B\|W_hq(t)\|_{H^{1/2}}^2\,dt\le C_4M with C4=74C3(C0+2)+4C_4=74C_3(C_0+2)+4. The interval [−T,T][-T,T] is covered by at most 1+κ2T1+\kappa^2T intervals of the form BB. Summing their contributions and taking the supremum over hh, we obtain (9) with C1=C4C_1=C_4.

For the lower bound, let qsq_s be the stationary soliton, that is (25) with ϑ=π/4\vartheta=\pi/4, and set q0,λ:=(qs(0))λq_{0,\lambda}:=(q_s(0))_\lambda. By (25), Φ(t)q0,λ=eiλ2t(qs(0))λ\Phi(t)q_{0,\lambda}=e^{i\lambda^2t}(q_s(0))_\lambda, so that sup⁡t∥Φ(t)q0,λ∥Eκε=∥qs(0)∥Eκ/λε\sup_t\|\Phi(t)q_{0,\lambda}\|_{E^\varepsilon_\kappa}=\|q_s(0)\|_{E^\varepsilon_{\kappa/\lambda}} by Lemma 2.1(iii). Set κs:=κ♭(qs(0))\kappa_s:=\kappa_\flat(q_s(0)). It is finite by Lemma 2.1(i). It is positive, because ∥qs(0)∥Eκε→∥qs(0)∥L2=2π>δ0\|q_s(0)\|_{E^\varepsilon_\kappa}\to\|q_s(0)\|_{L^2}=\sqrt{2\pi}>\delta_0 as κ↓0\kappa\downarrow0, by monotone convergence. Lemma 2.1(iii) then gives κ(q0,λ)=max⁡{1,κsλ}≤K∗λ\kappa(q_{0,\lambda})=\max\{1,\kappa_s\lambda\}\le K_*\lambda for λ≥1\lambda\ge1, where K∗:=max⁡{1,κs}K_*:=\max\{1,\kappa_s\}. In particular κ(q0,λ)\kappa(q_{0,\lambda}) grows linearly in λ\lambda. On the other hand, Φ(⋅)q0,λ∈Sπ/4\Phi(\cdot)q_{0,\lambda} \in\mathcal{S}_{\pi/4} with x0=t0=0x_0=t_0=0, and its centre x(t)≡0x(t)\equiv0 does not move. Proposition 4.2 below therefore gives ∥WΦ(t)q0,λ∥H1/22≥c5λ\|W\Phi(t)q_{0,\lambda}\|_{H^{1/2}}^2 \ge c_5\lambda for all tt and all λ≥λ0\lambda\ge\lambda_0. Integrating over [−T,T][-T,T], we find

LS⁡T(Φ(⋅)q0,λ)≥2c5λT≥2c5K∗κ(q0,λ)T.\operatorname{LS}_T(\Phi(\cdot)q_{0,\lambda}) \ge2c_5\lambda T \ge\frac{2c_5}{K_*}\kappa(q_{0,\lambda})T.

Since M(q0,λ)=2πM(q_{0,\lambda})=2\pi, this is (10) with c∗=c5/(πK∗)c_* = c_5/(\pi K_*). Finally, if (9) held with κ2\kappa^2 replaced by κα\kappa^\alpha for some α<1\alpha<1, then 2c5λT≤C(1+K∗αλαT)2π2c_5\lambda T\le C(1+K_*^\alpha\lambda^\alpha T)2\pi for all λ≥λ0\lambda\ge\lambda_0, which is impossible as λ→∞\lambda\to\infty. □

Remark 3.2. The exponent 2 in (9) comes from parabolic scaling. Proposition 3.1 controls windows of size κ−1\kappa^{-1} only during times of order κ−2\kappa^{-2}. Moreover, the smallness that drives the bootstrap of [20] is available only at frequencies above κ\kappa, so one cannot simply enlarge the windows. The stationary soliton shows that the true growth is at least linear in κ\kappa; see Conjecture 5.3.

Proof of Corollary 1.3

Let first q0∈S(R)q_0\in\mathcal{S}(\mathbb{R}) with M(q0)≤m0=δ02M(q_0)\le m_0=\delta_0^2. By Lemma 2.1(ii), κ(q0)=1\kappa(q_0)=1, and (11) is (9).

For (12), let R>0R>0 and let q(R)(t,x)=R1/2q(R2t,Rx)q^{(R)}(t,x)=R^{1/2}q(R^2t,Rx) be the solution with data (q0)R(q_0)_R, which has the same mass. Applying (11) with T=1T=1 to q(R)q^{(R)} and using (15), we obtain

∫−11∥Whq(R)(t)∥H˙1/22 dt≤2C1M(q0)\int_{-1}^{1}\|W_hq^{(R)}(t)\|_{\dot H^{1/2}}^2\,dt\le2C_1M(q_0)

for every hh. Now Wh(x)q(R)(t,x)=R1/2G(Rx)W_h(x)q^{(R)}(t,x)=R^{1/2}G(Rx) with G(y)=W((y−Rh)/R)q(R2t,y)G(y)=W((y-Rh)/R)q(R^2t,y), and ∥R1/2G(R⋅)∥H˙1/22=R∥G∥H˙1/22\|R^{1/2}G(R\cdot)\|_{\dot H^{1/2}}^2=R\|G\|_{\dot H^{1/2}}^2. Hence the substitution t′=R2tt'=R^2t turns the last bound into

∫−R2R2∥W(⋅−RhR)q(t′)∥H˙1/22 dt′≤2C1RM(q0).\int_{-R^2}^{R^2}\left\|W\left(\frac{\cdot-Rh}{R}\right)q(t')\right\|_{\dot H^{1/2}}^2\,dt'\le2C_1RM(q_0).

Since RhRh ranges over R\mathbb{R}, this is (12).

For (13), we use the one-dimensional inequality ∥f∥L4≤C∥f∥H˙1/4≤C∥f∥L21/2∥f∥H˙1/21/2\|f\|_{L^4}\le C\|f\|_{\dot H^{1/4}}\le C\|f\|_{L^2}^{1/2}\|f\|_{\dot H^{1/2}}^{1/2}, which follows from Sobolev embedding and interpolation. We apply it to f=Whq(t)f=W_hq(t). As ∥Whq(t)∥L2≤M(q0)\|W_hq(t)\|_{L^2}\le M(q_0), integration in time and (11) give (13) with C2=C4C1C_2=C^4C_1.

Now let q0∈L2q_0\in L^2 with 0<M(q0)≤m00<M(q_0)\le m_0. Choose Schwartz functions pn→q0p_n\to q_0 in L2L^2, with pn≠0p_n\ne0, and set q0,n=∥q0∥L2pn/∥pn∥L2q_{0,n}=\|q_0\|_{L^2}p_n/\|p_n\|_{L^2}. Then q0,n∈S(R)q_{0,n}\in\mathcal{S}(\mathbb{R}), M(q0,n)=M(q0)M(q_{0,n})=M(q_0), and q0,n→q0q_{0,n}\to q_0 in L2L^2. By the joint continuity of Φ\Phi [20 Theorem 1.1] and the compactness of [−T,T][-T,T], we get qn(t)=Φ(t)q0,n→q(t)q_n(t)=\Phi(t)q_{0,n}\to q(t) in L2L^2 uniformly for t∈[−T,T]t\in[-T,T]. For fixed hh and tt, it follows that Whqn(t)→Whq(t)W_hq_n(t)\to W_hq(t) in L2L^2.

The norms ∥⋅∥H1/2\|\cdot\|_{H^{1/2}} and ∥⋅∥H˙1/2\|\cdot\|_{\dot H^{1/2}} (with value +∞+\infty allowed) are lower semicontinuous with respect to L2L^2 convergence. Indeed, if fn→ff_n\to f in L2L^2, choose a subsequence along which ∥fn∥H1/2\|f_n\|_{H^{1/2}} tends to lim inf⁡n∥fn∥H1/2\liminf_n\|f_n\|_{H^{1/2}}, and a further subsequence along which fn^→f^\widehat{f_n}\to\widehat f almost everywhere. Fatou’s lemma in ξ\xi then gives ∥f∥H1/2≤lim inf⁡n∥fn∥H1/2\|f\|_{H^{1/2}}\le\liminf_n\|f_n\|_{H^{1/2}}, and the same argument applies to H˙1/2\dot H^{1/2}. Fatou’s lemma in tt therefore transfers (11) and (12) from qnq_n to qq. Finally, Whqn→WhqW_hq_n\to W_hq in Lt,x2([−T,T]×R)L^2_{t,x}([-T,T]\times\mathbb{R}). Hence a subsequence converges almost everywhere, and Fatou’s lemma also transfers (13). The bound C(m0)≤2C1C(m_0)\le2C_1 is (11) with T=1T=1. □

Solitons and the threshold 2π2\pi

The two propositions of this section quantify a simple geometric mechanism, illustrated in Figure 2. A soliton of scale λ\lambda carries a local smoothing density of order λ\lambda, and it contributes only while it stays in the window. A soliton of mass 8ϑ<2π8\vartheta< 2\pi moves with speed 2λcot⁡(2ϑ)2\lambda\cot(2\vartheta), so it crosses a unit window in a time of order (λcot⁡2ϑ)−1(\lambda\cot2\vartheta)^{-1}. Its total contribution is therefore of order (cot⁡2ϑ)−1(\cot2\vartheta)^{-1}, uniformly in λ\lambda. By contrast, the stationary soliton never leaves the window.

An upper bound for moving solitons.

Proposition 4.1. Let ϑ∈(0,π/4)\vartheta\in(0,\pi/4), b=cot⁡(2ϑ)>0b=\cot(2\vartheta)>0 and u∈Sϑu\in\mathcal{S}_{\vartheta}. Then

LS1(u)≤(86+666b)M(u).(30)\mathrm{LS}_1(u)\le\left(86+\frac{666}{b}\right)M(u). \tag*{(30)}

Proof. Write M=M(u)=8ϑM=M(u)=8\vartheta, fix h∈Rh\in\mathbb{R}, and let λ\lambda and x(t)x(t) be as in (27). Define

I(t):=∫RWh(x)2λe−2λ∣x−x(t)∣ dx.I(t):=\int_{\mathbb{R}}W_h(x)^2\lambda e^{-2\lambda|x-x(t)|}\,dx.

Since Wh≤1W_h\le1, we have I(t)≤1I(t)\le1. By (27) and (28),

∥Whu(t)∥L22=∫Wh2λ∣A(λ(x−x(t)))∣2 dx≤2MI(t).\|W_hu(t)\|_{L^2}^2=\int W_h^2\lambda\left|A(\lambda(x-x(t)))\right|^2\,dx\le2M I(t).

Using ∣Wh′∣≤Wh/16|W_h'|\le W_h/16 and the second bound in (28), we also get

∥(Whu(t))′∥L2≤116∥Whu(t)∥L2+λ(9+b)(∫Wh2λ∣A(λ(x−x(t)))∣2 dx)1/2≤(116+λ(9+b))2MI(t).\begin{aligned} \|(W_hu(t))'\|_{L^2} &\le\frac{1}{16}\|W_hu(t)\|_{L^2} +\lambda(9+b)\left(\int W_h^2\lambda\left|A(\lambda(x-x(t)))\right|^2\,dx\right)^{1/2}\\ &\le\left(\frac{1}{16}+\lambda(9+b)\right)\sqrt{2M I(t)}. \end{aligned}

By (17), it follows that ∥Whu(t)∥H1/22≤2MI(t)(3+λ(9+b))\|W_hu(t)\|_{H^{1/2}}^2\le2M I(t)(3+\lambda(9+b)).

We now integrate in time. The centre x(t)x(t) moves with speed 2λb2\lambda b. Denote by txt_x the time at which x(t)=xx(t)=x, so that ∣x−x(t)∣=2λb∣t−tx∣|x-x(t)|=2\lambda b|t-t_x|. Fubini’s theorem gives

∫−11I(t) dt≤∫RWh(x)2(∫Rλe−4λb∣t−tx∣ dt) dx=∥Wh∥L222λb≤37λb.\int_{-1}^{1}I(t)\,dt\le\int_{\mathbb{R}}W_h(x)^2\left(\int_{\mathbb{R}}\lambda e^{-4\lambda b|t-t_x|}\,dt\right)\,dx=\frac{\|W_h\|_{L^2}^2}{2\lambda b}\le\frac{37}{\lambda b}.

Combining this with I≤1I \le1, we obtain

∫−11∥Whu(t)∥H1/22 dt≤2M(3+9λ)min⁡{2,37λb}+λb⋅37λb≤2M(6+333b+37),(31)\int_{-1}^{1}\|W^{h}u(t)\|_{H^{1/2}}^{2}\,dt \le2M(3+9\lambda)\min\left\{2,\frac{37}{\lambda b}\right\}+\lambda b\cdot\frac{37}{\lambda b}\le2M\left(6+\frac{333}{b}+37\right), \tag*{(31)}

where we used the bound 3⋅23\cdot2 for the first summand and 9λ⋅37/(λb)9\lambda\cdot37/(\lambda b) for the second. Taking the supremum over hh gives (4.1).

The bound (4.1) is uniform in the scaling parameter λ\lambda and in the translation and phase parameters. It degenerates only as b↓0b\downarrow0, that is, as ϑ↑π/4\vartheta\uparrow\pi/4. This is the regime of slowly moving solitons with mass close to 2π2\pi. For small mass the bound is uniform as well, because small solitons travel fast.

A lower bound for slow solitons.

Proposition 4.2. There are absolute constants c5>0c_{5}>0 and λ0≥1\lambda_{0}\ge1 such that the following holds. Let ϑ∈[π/8,π/4]\vartheta\in[\pi/8,\pi/4] and λ≥λ0\lambda\ge\lambda_{0}, and let u∈Sϑu\in\mathcal{S}_{\vartheta} have parameters x0=t0=0x_{0}=t_{0}=0 in (26). Then for every tt with ∣x(t)∣≤2|x(t)|\le2,

∥Wu(t)∥H1/22≥c5λ.(32)\|W u(t)\|_{H^{1/2}}^{2}\ge c_{5}\lambda. \tag*{(32)}

Proof. By (26) and (25), W(x)u(t,x)=ωλ1/2G(λ(x−x(t)))W(x)u(t,x)=\omega\lambda^{1/2}G(\lambda(x-x(t))) for a unimodular constant ω\omega, where G(y)=W(x(t)+y/λ)q0(y;ϑ)G(y)=W(x(t)+y/\lambda)q_{0}(y;\vartheta). By (15), translation invariance and the scaling of H˙1/2\dot{H}^{1/2},

∥Wu(t)∥H1/22≥∥Wu(t)∥H˙1/22=λ∥G∥H˙1/22.\|Wu(t)\|_{H^{1/2}}^{2}\ge\|Wu(t)\|_{\dot{H}^{1/2}}^{2}=\lambda\|G\|_{\dot{H}^{1/2}}^{2}.

Set μ:=min⁡ϑ∈[π/8,π/4]∥q0(⋅;ϑ)∥H˙1/2\mu:=\min_{\vartheta\in[\pi/8,\pi/4]}\|q_{0}(\cdot;\vartheta)\|_{\dot{H}^{1/2}}. This minimum is attained, by Lemma 2.5 and the continuity of the H˙1/2\dot{H}^{1/2} norm on H1H^{1}. It is positive, because a function with vanishing H˙1/2\dot{H}^{1/2} norm vanishes, while ∥q0(⋅;ϑ)∥L22=8ϑ>0\|q_{0}(\cdot;\vartheta)\|_{L^{2}}^{2}=8\vartheta>0. Since ∣x(t)∣≤2|x(t)|\le2, we have W(x(t))≥12W(x(t))\ge\frac{1}{2}, and hence ∥W(x(t))q0∥H˙1/2≥μ/2\|W(x(t))q_{0}\|_{\dot{H}^{1/2}}\ge\mu/2.

It remains to estimate the difference D(y)=(W(x(t)+y/λ)−W(x(t)))q0(y;ϑ)D(y)=(W(x(t)+y/\lambda)-W(x(t)))q_{0}(y;\vartheta). By (18), ∣W(x(t)+y/λ)−W(x(t))∣≤∣y∣/(16λ)|W(x(t)+y/\lambda)-W(x(t))|\le|y|/(16\lambda). By Lemma 2.5 with b≤1b\le1 and 16ϑ≤4π16\vartheta\le4\pi, we also have ∣q0∣+∣∂yq0∣≤124πe−∣y∣|q_{0}|+|\partial_{y}q_{0}|\le12\sqrt{4\pi}e^{-|y|}. Moreover, ∣D′(y)∣≤λ−1∣W′′∣∣q0∣+(16λ)−1∣y∣∣∂yq0∣|D'(y)|\le\lambda^{-1}|W''||q_{0}|+(16\lambda)^{-1}|y||\partial_{y}q_{0}|. We deduce that

∥D∥H˙1/22≤∥D∥L22+∥D′∥L22≤C62λ2.\|D\|_{\dot{H}^{1/2}}^{2}\le\|D\|_{L^{2}}^{2}+\|D'\|_{L^{2}}^{2}\le\frac{C_{6}^{2}}{\lambda^{2}}.

with an absolute constant C6C_{6}. We choose λ0:=max⁡{1,4C6/μ}\lambda_{0}:=\max\{1,4C_{6}/\mu\}. Then ∥G∥H˙1/2≥μ/2−C6/λ≥μ/4\|G\|_{\dot{H}^{1/2}}\ge\mu/2-C_{6}/\lambda\ge\mu/4, and (32) holds with c5=μ2/16c_{5}=\mu^{2}/16.

Proof of Theorem 1.4.

(i) Let m≥2πm\ge2\pi. For λ≥λ0\lambda\ge\lambda_{0}, let uλ∈Sπ/4u_{\lambda}\in\mathcal{S}_{\pi/4} be the rescaled stationary soliton with x0=t0=0x_{0}=t_{0}=0. It has mass 2π≤m2\pi\le m and centre x(t)≡0x(t)\equiv0. By Proposition 4.2, LS⁡1(uλ)≥2c5λ\operatorname{LS}_{1}(u_{\lambda})\ge2c_{5}\lambda. Since uλ=Φ(⋅)uλ(0)u_{\lambda}=\Phi(\cdot)u_{\lambda}(0), we get C(m)≥c5λ/πC(m)\ge c_{5}\lambda/\pi for every λ≥λ0\lambda\ge\lambda_{0}, and therefore C(m)=∞C(m)=\infty.

(ii) Let λ0\lambda_{0} be as in Proposition 4.2 and set m1:=max⁡{π,2π−4arctan⁡(1/λ0)}m_{1}:=\max\{\pi,2\pi-4\arctan(1/\lambda_{0})\}. Given m∈[m1,2π)m\in[m_{1},2\pi), let ϑ=m/8∈[π/8,π/4)\vartheta=m/8\in[\pi/8,\pi/4), so that b=cot⁡(2ϑ)=tan⁡((2π−m)/4)∈(0,1/λ0]b=\cot(2\vartheta)=\tan((2\pi-m)/4)\in(0,1/\lambda_{0}]. Take u∈Sϑu\in\mathcal{S}_{\vartheta} with x0=t0=0x_{0}=t_{0}=0 and λ=1/b≥λ0\lambda=1/b\ge\lambda_{0}. Then ∣x(t)∣=2λb∣t∣≤2|x(t)|=2\lambda b|t|\le2 for ∣t∣≤1|t|\le1, and Proposition 4.2 gives LS⁡1(u)≥2c5/b\operatorname{LS}_{1}(u)\ge2c_{5}/b. Since (2π−m)/4≤π/4(2\pi-m)/4\le\pi/4 and tan⁡s≤4s/π\tan s\le4s/\pi on [0,π/4][0,\pi/4] by convexity, we have b≤(2π−m)/πb\le(2\pi-m)/\pi. As M(u)=mM(u)=m, we conclude that

C(m)≥LS⁡1(u)M(u)≥2c5mb≥2c5π2π(2π−m)=c52π−m.C(m)\ge\frac{\operatorname{LS}_{1}(u)}{M(u)}\ge\frac{2c_{5}}{mb}\ge\frac{2c_{5}\pi}{2\pi(2\pi-m)}=\frac{c_{5}}{2\pi-m}.

(iii) For u∈Sϑu \in\mathcal{S}_{\vartheta} with ϑ∈(0,π/4)\vartheta\in(0,\pi/4) we have M(u)=8ϑM(u)=8\vartheta and b=tan⁡((2π−M(u))/4)≥(2π−M(u))/4b=\tan((2\pi-M(u))/4)\ge(2\pi-M(u))/4. Hence 666/b≤2664/(2π−M(u))666/b\le2664/(2\pi-M(u)), and Proposition 4.1 gives (14) with C=2664C=2664.

The final assertions follow from Corollary 1.3 and parts (i)–(iii), with c=c5c=c_5. □

Remark 4.3. Proposition 4.1 also covers solitons of very small mass. As ϑ↓0\vartheta\downarrow0, the amplitude of q0(⋅;ϑ)q_0(\cdot;\vartheta) is of order ϑ1/2\vartheta^{1/2}, while its frequency cot⁡(2ϑ)\cot(2\vartheta) is of order ϑ−1\vartheta^{-1}. These solitons are thus high-frequency wave packets of small mass, travelling at a speed comparable to their frequency. They are invisible to the threshold question. They do show, however, that small-mass solutions need not scatter. This is why the global-in-time question below is phrased in terms of smoothing and not of scattering; compare with the weighted-space asymptotics of [25, 50].

Discussion: methods, obstacles and conjectures

What the method gives and where it stops

The proof of Theorem 1.2 uses [20] in an essentially optimal way. Smallness of the EεE^\varepsilon norm at frequency scale κ\kappa is the only source of smallness in their bootstrap, and rescaling converts it into local smoothing at the natural parabolic scale. All further losses come from geometry, namely from summing κ\kappa small windows into a unit window and κ2T\kappa^2T short time intervals into [−T,T][-T,T]. Lemma 2.4 shows that the first step costs nothing. The second step is where the factor κ2\kappa^2 appears.

For data of mass below m0m_0 the frequency scale is irrelevant, and Corollary 1.3 follows. Between m0m_0 and 2π2\pi the method gives no uniform information. The reason is that the E1εE^ \varepsilon_1 norm of a datum of moderate mass need not be small, and making it small requires a rescaling that depends on the frequency distribution of the datum. This is not a technical defect. By Theorem 1.4(i), the same mechanism must fail at mass 2π2\pi. Any proof of Conjecture 1.5 must therefore use a property of the region M<2πM<2\pi that the stationary soliton violates.

Heuristic evidence for Conjecture 1.5

A unit window can accumulate a large amount of local smoothing only if a solution concentrated at a high frequency λ\lambda remains near the window for a time much longer than λ−1\lambda^{-1}. Linear wave packets at frequency λ\lambda move with speed of order λ\lambda and do not do this, but coherent structures could. Proposition 4.1 shows that single solitons of mass 8ϑ<2π8\vartheta<2\pi cannot. Indeed, their velocity 2λcot⁡(2ϑ)2\lambda\cot(2\vartheta) is proportional to their frequency, with a coefficient bounded below in terms of 2π−8ϑ2\pi-8\vartheta. Theorem 1.4(ii) shows that this coefficient is the only source of growth among solitons. Figure 3 illustrates the resulting two-sided picture near 2π2\pi.

Schematic behaviour near $m=2\pi$: lower bound $c(2\pi-m)^{-1}$ for $C(m)$, soliton upper bound $C(1+(2\pi-m)^{-1})$, and conjectured behaviour of $C(m)$

Figure 3. Schematic behaviour near m=2πm=2\pi (constants not to scale). Both proved bounds blow up at the same rate (2π−m)−1(2\pi-m)^{-1}. Conjecture 1.5 asserts that no other mechanism produces a faster rate below 2π2\pi.

The scattering theory of the Kaup–Newell problem gives a more structural version of this observation. For u∈S(R)u\in\mathcal{S}(\mathbb{R}), let a~u(ζ)\widetilde{a}_u(\zeta) be the transmission coefficient in the variable ζ=k2\zeta=k^2, where kk is the Kaup–Newell spectral parameter, and let ζj\zeta_j be its zeros in the upper half-plane, with arg⁡ζj∈(0,π)\arg\zeta_j\in(0,\pi). Bahouri and Perelman record the trace formula [4 Lemma 2.3(ii)]; see also [2 (2.15)]:

M(u)=4∑jarg⁡ζj+2π∫Rdμu(ξ),(33)M(u)=4\sum_j\arg\zeta_j+\frac{2}{\pi}\int_{\mathbb{R}}d\mu_u(\xi), \tag*{(33)}

where μu\mu_u is a nonnegative measure accounting for the radiation.

Remark 5.1. [Spectral picture below 2π2\pi] By (33) and the nonnegativity of μu\mu_u, every eigenvalue of a Schwartz function uu with M(u)<2πM(u)<2\pi satisfies 0<arg⁡ζj≤M(u)/4<π/20<\arg\zeta_j\le M(u)/4<\pi/2. For a single soliton the picture is explicit. In the normalization of [2], the bright soliton with parameters (E,c)(E,c), E>0E>0, travels with velocity cc. Its transmission coefficient has the single zero ζE,c=−c/4+iE/2\zeta_{E,c}=-c/4+i\sqrt{E}/2, and its mass is given in [2 (1.12)]. A direct computation shows that cos⁡(arg⁡ζE,c)=−c/c2+4E=cos⁡(M(uE,c)/4)\cos(\arg\zeta_{E,c})=-c/\sqrt{c^2+4E}=\cos\left(M(u_{E,c})/4\right), hence M(uE,c)=4arg⁡ζE,cM(u_{E,c})=4\arg\zeta_{E,c}. Consequently, there is no radiation in (33), the velocity c=−4Re⁡ζE,cc=-4\operatorname{Re}\zeta_{E,c} is proportional to Re⁡ζE,c\operatorname{Re}\zeta_{E,c}, and the soliton is stationary if and only if its eigenvalue lies on the imaginary axis, that is, if and only if its mass equals 2π2\pi. In the parametrization (24) this corresponds to arg⁡ζ=2ϑ\arg\zeta=2\vartheta. The equation of [2] is related to (1) by the reflection x↦−xx\mapsto-x, which reverses the direction of motion but does not affect these statements. For several solitons, breathers and general data, the analogous statement, namely that each coherent component attached to ζj\zeta_j moves with a velocity proportional to Re⁡ζj\operatorname{Re}\zeta_j, is suggested by the Lax-pair time dependence and by soliton resolution [29].

Under soliton resolution [29], a solution decomposes for large times into solitons, attached to the eigenvalues, and dispersive radiation. If all eigenvalues lie in the sector arg⁡ζ≤π/2−η\arg\zeta\le\pi/2-\eta, every soliton moves with a speed bounded below by a multiple of its frequency, and radiation disperses. This motivates the following spectral strengthening of Conjecture 1.5.

Conjecture 5.2 (Spectral form). For every η∈(0,π/2)\eta\in(0,\pi/2) and m>0m>0 there is C(η,m)C(\eta,m) with the following property. Every q0∈S(R)q_0 \in\mathcal{S}(\mathbb{R}) with M(q0)≤mM(q_0)\le m, whose transmission coefficient has no zeros in the sector {π/2−η<arg⁡ζ<π/2+η}\{\pi/2-\eta<\arg\zeta<\pi/2+\eta\}, satisfies LS1(Φ(⋅)q0)≤C(η,m)M(q0)\mathrm{LS}_1(\Phi(\cdot)q_0)\le C(\eta,m)M(q_0). Moreover, one can take C(η,m)≲mη−1C(\eta,m)\lesssim_m\eta^{-1}.

By Remark 5.1, Schwartz data of mass m<2πm<2\pi satisfy the hypothesis with η=(2π−m)/4\eta=(2\pi-m)/4. Therefore Conjecture 5.2, combined with the density argument in the proof of Corollary 1.3, implies Conjecture 1.5, including the rate (2π−m)−1(2\pi-m)^{-1}, which Theorem 1.4 shows to be optimal. The spectral formulation also explains why breathers or multi-soliton bound states should not produce stationary structures below 2π2\pi. Indeed, a stationary bound state would need all its eigenvalues on the imaginary axis, and each of them contributes 2π2\pi to (33). This last assertion relies on the heuristic velocity–eigenvalue relation of Remark 5.1 for multi-soliton states.

The sharp dependence on the frequency scale

The lower bound (10) and the soliton heuristics suggest that the factor κ\kappa in (9) is an artefact of the method.

Conjecture 5.3 (Sharp quantitative local smoothing). There is a nondecreasing function C:(0,∞)→(0,∞)C:(0,\infty)\to(0,\infty) such that for all q0∈S(R)q_0\in\mathcal{S}(\mathbb{R}) and T>0T>0,

LST(Φ(⋅)q0)≤C(M(q0))(1+κ(q0)T)M(q0).\mathrm{LS}_T(\Phi(\cdot)q_0)\le C(M(q_0))(1+\kappa(q_0)T)M(q_0).

The stationary soliton shows that the exponent one cannot be improved. A proof would require local smoothing on windows of size much larger than κ−1\kappa^{-1} during times of order one. This seems to need a substitute for the parabolic rescaling that preserves the smallness used in [20].

Global-in-time smoothing for small mass

The estimate (12) holds on parabolic cylinders of every size, which suggests asking whether smoothing for small mass holds globally in time. Some care is needed in the formulation. For the free flow with low-frequency data, ∣u(t,0)∣2|u(t,0)|^2 decays like t−1t^{-1}. Hence ∫R∥Wu(t)∥H˙1/22 dt\int_{\mathbb{R}}\|Wu(t)\|_{\dot{H}^{1/2}}^2\,dt can diverge logarithmically, because ∥Wu(t)∥H˙1/2\|Wu(t)\|_{\dot{H}^{1/2}} is comparable to ∥Wu(t)∥L2\|Wu(t)\|_{L^2} for such slowly varying functions, which reflects the commutator between WW and ∣D∣1/2|D|^{1/2}. The natural global quantity is the Kato–Kenig–Ponce–Vega one.

Question 5.4. Is there m2>0m_2>0 such that every q0∈L2q_0\in L^2 with M(q0)≤m2M(q_0)\le m_2 satisfies

sup⁡h∈R∬R×RWh(x)2∣Dx1/2q(t,x)∣2 dx dt≤CM(q0) ?\sup_{h\in\mathbb{R}}\iint_{\mathbb{R}\times\mathbb{R}} W_h(x)^2\left|D_x^{1/2}q(t,x)\right|^2\,dx\,dt\le C M(q_0)\,?

For the free Schrödinger group the answer is yes, by the identity of [36]. Small-mass solitons, which by Remark 4.3 do not scatter, appear to satisfy this bound, since their frequency and speed are both of order λcot⁡(2ϑ)\lambda\cot(2\vartheta). A positive answer would be a nonlinear Kato smoothing theorem at the critical regularity, in a setting without scattering. A natural tool is the family of microscopic conservation laws of [63, 20] with the spectral parameter coupled to time, in the spirit of an interaction Morawetz estimate. We have not been able to close such an argument. The finite-scale analysis of [5] may help to understand the linear part of the problem.

Strichartz-type quantities

By [20 Proposition 1.11], the rescaled stationary solitons also satisfy ∫−11∫R∣qn∣6 dx dt→∞\int_{-1}^{1}\int_{\mathbb{R}} |q_n|^6\,dx\,dt \to\infty. (This is also immediate from scaling, since ∫R∣uλ(t)∣6 dx=λ2∫R∣qs(0)∣6 dx\int_{\mathbb{R}} |u_\lambda(t)|^6\,dx = \lambda^2 \int_{\mathbb{R}} |q_s(0)|^6\,dx for the rescaled stationary soliton uλu_\lambda.) Our methods yield the localized Lt,x4L^4_{t,x} bound (13) for small mass, but not an Lt,x6L^6_{t,x} bound. The latter would require control of ∥Whq∥H˙1/2\|W_hq\|_{\dot H^{1/2}} in Lt4L^4_t rather than in Lt2L^2_t, since ∥f∥L66≲∥f∥L22∥f∥H˙1/24\|f\|_{L^6}^6 \lesssim\|f\|_{L^2}^2\|f\|_{\dot H^{1/2}}^4. We do not know whether ∫−11∫R∣Whq∣6 dx dt≤CM3\int_{-1}^{1}\int_{\mathbb{R}} |W_hq|^6\,dx\,dt \le CM^3 holds even for small mass. Its threshold, if it exists, is presumably also 2π2\pi.

Possible routes to Conjecture 1.5 and their obstacles

A natural approach is by contradiction and concentration compactness, in the spirit of Kenig and Merle [35]. Suppose that C(m)=∞C(m)=\infty for some m<2πm<2\pi, and take a minimal such mm. One would extract a sequence qnq_n with LS1(qn)/M(qn)→∞LS_1(q_n)/M(q_n)\to\infty. One would then rescale it so that the bad behaviour occurs at unit frequency on time intervals of length Tn→∞T_n\to\infty, and try to pass to a limit which is almost periodic modulo symmetries and has mass at most mm. Three obstacles stand in the way. First, a profile decomposition adapted to the DNLS flow in L2L^2 is not available. Bahouri and Perelman [2] use profile decompositions at the H1/2H^{1/2} and H1H^1 levels, and the absence of uniform continuity of the flow on bounded sets of HsH^s, s<1/2s<1/2 [7], prevents the usual nonlinear profile approximation. Second, minimality of mm and the Lipschitz stability theory that usually comes with it have no counterpart here. Third, the rigidity step would have to exclude non-dispersing almost periodic solutions of mass below 2π2\pi. Remark 5.1 suggests doing so through the spectral data, which are conserved and somehow well behaved for L2L^2 data in the sense of [4 Corollary 2.1].

A second approach is to strengthen the bootstrap of [20 Section 4] directly. There, the quartic and higher terms of the current are controlled by two copies of the local smoothing norm times a small factor ∥q∥Eε\|q\|_{E^\varepsilon}. Below 2π2\pi, one would need to replace this smallness by coercivity. The Hayashi–Ozawa coercivity [24, 26] is an H1H^1-level statement for the gauged energy, and it does not obviously localize to the frequency-dependent quantities A(κ;q)A(\kappa;q). Finding a coercive L2L^2-level substitute would be a significant step, for instance a lower bound on Re⁡A(κ;q)\operatorname{Re} A(\kappa;q) for M(q)<2πM(q)<2\pi that is uniform across frequency scales.

Concluding remarks and further directions

We have isolated a quantitative question left open by the L2L^2 well-posedness theory of [20]: on which balls of L2(R)L^2(\mathbb{R}) does local smoothing for DNLS hold uniformly? We proved that it holds for small mass, with a scale-invariant Morrey-type form, and that it fails from mass 2π2\pi on. We also proved that the uniform constant must blow up at least like (2π−m)−1(2\pi-m)^{-1}, which is exactly the rate exhibited by the soliton family. Finally, we made explicit the dependence on the data in the local smoothing estimate of [20]. It is carried by a single frequency scale κ(q0)\kappa(q_0), and it is polynomial with an exponent between one and two.

The results proved here rest on [20] as a black box together with elementary arguments, and their value lies mainly in the questions they make precise. Conjecture 1.5 asserts that 2π2\pi, the mass of the stationary soliton and the Hayashi–Ozawa coercivity threshold, is sharp. Conjecture 5.2 gives a spectral form, and Conjecture 5.3 the sharp dependence on the frequency scale. Question 5.4 asks for global-in-time smoothing at small mass. A positive answer would be a nonlinear Kato smoothing theorem at critical regularity, in a setting where small solitons prevent scattering. Several further directions are natural. The commuting-flow method of [20] applies, at least formally, to the whole Kaup–Newell hierarchy, and the threshold question can be asked for each flow. For odd flows the relevant soliton velocities change, and the threshold may differ from 2π2\pi. The models (1.33) and (1.36) of [20] are obtained from (1) by a Galilean-type modulation and by a gauge transformation, respectively, and they are well-posed in L2L^2 by [20 Corollary 1.13]. The threshold can be studied for them as well, keeping in mind that gauge transformations do not commute with spatial localization. For the continuum Calogero–Moser model, where 2π2\pi is already a well-posedness threshold [40, 44], it would be interesting to know whether uniform local smoothing holds up to 2π2\pi. Finally, the periodic problem in L2(T)L^2(\mathbb{T}) remains open. Since local smoothing is unavailable on the circle, a different mechanism will be needed there.

Declarations

Funding. This research received no funding.

Conflict of interest. The author declares no conflict of interest.

Data availability. Data sharing is not applicable to this article, as no datasets were generated or analysed.

References

Department of Mathematics, College of Science, Imam Abdulrahman Bin Faisal University, P.O. Box 1982, Dammam, Saudi Arabia

Basic and Applied Scientific Research Center, Imam Abdulrahman Bin Faisal University, P.O. Box 1982, 31441, Dammam, Saudi Arabia

Email address: [email protected]

Email address: [email protected]

Email address: [email protected]

URL: https://orcid.org/0000-0001-6038-1069

References

  1. [1]H. Bahouri, T. M. Leslie and G. Perelman, Hˢ bounds for the derivative nonlinear Schrödinger equation, Math. Res. Lett. 30 (2023), 1299–1333.
  2. [2]H. Bahouri and G. Perelman, Global well-posedness for the derivative nonlinear Schrödinger equation, Invent. Math. 229 (2022), 639–688.arxiv.org/abs/2012.01923
  3. [3]H. Bahouri and G. Perelman, Global well-posedness for the derivative nonlinear Schrödinger equation with periodic boundary condition, Int. Math. Res. Not. IMRN 2024, no. 24, 14479–14518.DOI
  4. [4]H. Bahouri and G. Perelman, Global well-posedness for the derivative nonlinear Schrödinger equation on the circle, preprint, arXiv:2609.10005 (2026).arxiv.org/abs/2609.10005
  5. [5]R. Bai and M. Majdoub, Local smoothing at finite scale: an exact identity and sharp constants, in preparation.
  6. [6]R. Bai, M. Majdoub and R. Jia, Scattering below the ground state for the inhomogeneous nonlinear Schrödinger equation in one dimension, in preparation.
  7. [7]H. A. Biagioni and F. Linares, Ill-posedness for the derivative Schrödinger and generalized Benjamin–Ono equations, Trans. Amer. Math. Soc. 353 (2001), 3649–3659.DOI
  8. [8]M. Colin and M. Ohta, Stability of solitary waves for derivative nonlinear Schrödinger equation, Ann. Inst. H. Poincaré Anal. Non Linéaire 23 (2006), 753–764.DOI
  9. [9]J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, Global well-posedness for Schrödinger equations with derivative, SIAM J. Math. Anal. 33 (2001), 649–669.DOI
  10. [10]J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, A refined global well-posedness result for Schrödinger equations with derivative, SIAM J. Math. Anal. 34 (2002), 64–86.
  11. [11]P. Constantin and J.-C. Saut, Local smoothing properties of dispersive equations, J. Amer. Math. Soc. 1 (1988), 413–439.DOI
  12. [12]Y. Deng, A. R. Nahmod and H. Yue, Optimal local well-posedness for the periodic derivative nonlinear Schrödinger equation, Comm. Math. Phys. 384 (2021), 1061–1107.DOI
  13. [13]B. Dodson, Global well-posedness and scattering for the mass critical nonlinear Schrödinger equation with mass below the mass of the ground state, Adv. Math. 285 (2015), 1589–1618.arxiv.org/abs/1104.1114
  14. [14]N. Fukaya, M. Hayashi and T. Inui, A sufficient condition for global existence of solutions to a generalized derivative nonlinear Schrödinger equation, Anal. PDE 10 (2017), 1149–1167.DOI
  15. [15]P. Gérard and E. Lenzmann, The Calogero–Moser derivative nonlinear Schrödinger equation, Comm. Pure Appl. Math. 77 (2024), 4008–4062.arxiv.org/abs/2208.04105
  16. [16]A. Grünrock, Bi- and trilinear Schrödinger estimates in one space dimension with applications to cubic NLS and DNLS, Math. Res. Not. 2005, 2525–2558.
  17. [17]A. Grünrock and S. Herr, Low regularity local well-posedness of the derivative nonlinear Schrödinger equation with periodic initial data, SIAM J. Math. Anal. 39 (2008), 1890–1920.DOI
  18. [18]S. Guo, X. Ren and B. Wang, Local well-posedness for the derivative nonlinear Schrödinger equation with L²-subcritical data, Discrete Contin. Dyn. Syst. 41 (2021), 4207–4253.DOI
  19. [19]Z. Guo and Y. Wu, Global well-posedness for the derivative nonlinear Schrödinger equation in H¹/²(ℝ), Discrete Contin. Dyn. Syst. 37 (2017), 257–264.
  20. [20]B. Harrop-Griffiths, R. Killip, M. Ntekoume and M. Vişan, Global well-posedness for the derivative nonlinear Schrödinger equation in L²(ℝ), J. Eur. Math. Soc. 28 (2026), 843–924.
  21. [21]B. Harrop-Griffiths, R. Killip and M. Vişan, Large-data equicontinuity for the derivative NLS, Int. Math. Res. Not. IMRN 2023, 4601–4642.
  22. [22]B. Harrop-Griffiths, R. Killip and M. Vişan, Sharp well-posedness for the cubic NLS and mKdV in H³(ℝ), Forum Math. Pi 12 (2024), e6.
  23. [23]N. Hayashi, The initial value problem for the derivative nonlinear Schrödinger equation in the energy space, Nonlinear Anal. 20 (1993), 823–833.DOI
  24. [24]N. Hayashi and T. Ozawa, On the derivative nonlinear Schrödinger equation, Phys. D 55 (1992), 14–36.DOI
  25. [25]N. Hayashi and T. Ozawa, Modified wave operators for the derivative nonlinear Schrödinger equation, Math. Ann. 298 (1994), 557–576.DOI
  26. [26]N. Hayashi and T. Ozawa, Finite energy solutions of nonlinear Schrödinger equations of derivative type, SIAM J. Math. Anal. 25 (1994), 1488–1503.DOI
  27. [27]S. Herr, On the Cauchy problem for the derivative nonlinear Schrödinger equation with periodic boundary condition, Int. Math. Res. Not. 2006, Art. ID 96763.DOI
  28. [28]J. Hogan and M. Kowalski, Turbulent threshold for continuum Calogero–Moser models, Pure Appl. Anal. 6 (2024), 941–954.DOI
  29. [29]R. Jenkins, J. Liu, P. Perry and C. Sulem, Soliton resolution for the derivative nonlinear Schrödinger equation, Comm. Math. Phys. 363 (2018), 1003–1049.DOI
  30. [30]R. Jenkins, J. Liu, P. A. Perry and C. Sulem, Global well-posedness for the derivative nonlinear Schrödinger equation, Comm. Partial Differential Equations 43 (2018), 1151–1195.
  31. [31]R. Jenkins, J. Liu, P. Perry and C. Sulem, The derivative nonlinear Schrödinger equation: global well-posedness and soliton resolution, Quart. Appl. Math. 78 (2020), 33–73.arxiv.org/abs/1905.02866
  32. [32]R. Jenkins, J. Liu, P. Perry and C. Sulem, Global existence for the derivative nonlinear Schrödinger equation with arbitrary spectral singularities, Anal. PDE 13 (2020), 1539–1578.DOI
  33. [33]T. Kato, On the Cauchy problem for the (generalized) Korteweg–de Vries equation, in: Studies in Applied Mathematics, Adv. Math. Suppl. Stud. 8, Academic Press, New York, 1983, 93–128.
  34. [34]D. J. Kaup and A. C. Newell, An exact solution for a derivative nonlinear Schrödinger equation, J. Math. Phys. 19 (1978), 798–801.DOI
  35. [35]C. E. Kenig and F. Merle, Global well-posedness, scattering and blow-up for the energy-critical, focusing, non-linear Schrödinger equation in the radial case, Invent. Math. 166 (2006), 645–675.
  36. [36]C. E. Kenig, G. Ponce and L. Vega, Oscillatory integrals and regularity of dispersive equations, Indiana Univ. Math. J. 40 (1991), 33–69.
  37. [37]C. E. Kenig, G. Ponce and L. Vega, Small solutions to nonlinear Schrödinger equations, Ann. Inst. H. Poincaré Anal. Non Linéaire 10 (1993), 255–288.DOI
  38. [38]C. E. Kenig, G. Ponce and L. Vega, Smoothing effects and local existence theory for the generalized nonlinear Schrödinger equations, Invent. Math. 134 (1998), 489–545.DOI
  39. [39]R. Killip, T. Laurens, J. Murphy and M. Vişan, Dispersive decay and scattering for continuum Calogero–Moser models, preprint, arXiv:2609.12985 (2026).arxiv.org/abs/2609.12985
  40. [40]R. Killip, T. Laurens and M. Vişan, Scaling-critical well-posedness for continuum Calogero–Moser models on the line, Comm. Amer. Math. Soc. 5 (2025), 284–320.DOI
  41. [41]R. Killip, M. Ntekoume and M. Vişan, On the well-posedness problem for the derivative nonlinear Schrödinger equation, Anal. PDE 16 (2023), 1245–1270.DOI
  42. [42]R. Killip and M. Vişan, KdV is well-posed in H⁻¹, Ann. of Math. (2) 190 (2019), 249–305.DOI
  43. [43]R. Killip, M. Vişan and X. Zhang, Low regularity conservation laws for integrable PDE, Geom. Funct. Anal. 28 (2018), 1062–1090.arxiv.org/abs/1708.05362
  44. [44]K. Kim, T. Kim and S. Kwon, Construction of smooth chiral finite-time blow-up solutions to Calogero–Moser derivative nonlinear Schrödinger equation, preprint, arXiv:2404.09603 (2024).arxiv.org/abs/2404.09603
  45. [45]F. Klaus and R. Schippa, A priori estimates for the derivative nonlinear Schrödinger equation, Funkcial. Ekvac. 65 (2022), 329–346.
  46. [46]S. Kwon and Y. Wu, Orbital stability of solitary waves for derivative nonlinear Schrödinger equation, J. Anal. Math. 135 (2018), 473–486.arxiv.org/abs/1603.03745
  47. [47]S. Le Coz and Y. Wu, Stability of multisolitons for the derivative nonlinear Schrödinger equation, Int. Math. Res. Not. IMRN 2018, 4120–4170.DOI
  48. [48]J.-H. Lee, Global solvability of the derivative nonlinear Schrödinger equation, Trans. Amer. Math. Soc. 314 (1989), 107–118.
  49. [49]J. Liu, P. A. Perry and C. Sulem, Global existence for the derivative nonlinear Schrödinger equation by the method of inverse scattering, Comm. Partial Differential Equations 41 (2016), 1692–1760.arxiv.org/abs/1511.01173
  50. [50]J. Liu, P. A. Perry and C. Sulem, Long-time behavior of solutions to the derivative nonlinear Schrödinger equation for soliton-free initial data, Ann. Inst. H. Poincaré Anal. Non Linéaire 35 (2018), 217–265.DOI
  51. [51]M. Majdoub and T. Saanouni, Normalized standing waves for the focusing inhomogeneous Schrödinger equation with spatially growing nonlinearity, preprint, arXiv:2602.07505 (2026).arxiv.org/abs/2602.07505
  52. [52]C. Miao, Y. Wu and G. Xu, Global well-posedness for Schrödinger equation with derivative in H¹/²(ℝ), J. Differential Equations 251 (2011), 2164–2195.DOI
  53. [53]K. Mio, T. Ogino, K. Minami and S. Takeda, Modified nonlinear Schrödinger equation for Alfvén waves propagating along the magnetic field in cold plasmas, J. Phys. Soc. Japan 41 (1976), 265–271.
  54. [54]E. Mjølhus, On the modulational instability of hydromagnetic waves parallel to the magnetic field, J. Plasma Phys. 16 (1976), 321–334.DOI
  55. [55]R. Mosincat, Global well-posedness of the derivative nonlinear Schrödinger equation with periodic boundary condition in H¹/², J. Differential Equations 263 (2017), 4658–4722.DOI
  56. [56]R. Mosincat and T. Oh, A remark on global well-posedness of the derivative nonlinear Schrödinger equation on the circle, C. R. Math. Acad. Sci. Paris 353 (2015), 837–841.arxiv.org/abs/1502.02261
  57. [57]T. Ozawa, On the nonlinear Schrödinger equations of derivative type, Indiana Univ. Math. J. 45 (1996), 137–163.DOI
  58. [58]D. E. Pelinovsky and Y. Shimabukuro, Existence of global solutions to the derivative NLS equation with the inverse scattering transform method, Int. Math. Res. Not. IMRN 2018, 5663–5728.arxiv.org/abs/1602.02118
  59. [59]D. E. Pelinovsky, A. Saalmann and Y. Shimabukuro, The derivative NLS equation: global existence with solitons, Dyn. Partial Differ. Equ. 14 (2017), 271–294.arxiv.org/abs/1703.05277
  60. [60]P. Sjölin, Regularity of solutions to the Schrödinger equation, Duke Math. J. 55 (1987), 699–715.DOI
  61. [61]H. Takaoka, Well-posedness for the one-dimensional nonlinear Schrödinger equation with the derivative nonlinearity, Adv. Differential Equations 4 (1999), 561–580.
  62. [62]H. Takaoka, Global well-posedness for Schrödinger equations with derivative in a nonlinear term and data in low-order Sobolev spaces, Electron. J. Differential Equations 2001, No. 42.
  63. [63]X. Tang and G. Xu, Microscopic conservation laws for the derivative nonlinear Schrödinger equation, Lett. Math. Phys. 111 (2021), Paper No. 138.
  64. [64]T. Tao, Nonlinear Dispersive Equations: Local and Global Analysis, CBMS Regional Conference Series in Mathematics 106, Amer. Math. Soc., Providence, RI, 2006.
  65. [65]M. Tsutsumi and I. Fukuda, On solutions of the derivative nonlinear Schrödinger equation. Existence and uniqueness theorem, Funkcial. Ekvac. 23 (1980), 259–277.DOI
  66. [66]L. Vega, Schrödinger equations: pointwise convergence to the initial data, Proc. Amer. Math. Soc. 102 (1988), 874–878.
  67. [67]L. Vega and N. Visciglia, On the local smoothing for the Schrödinger equation, Proc. Amer. Math. Soc. 135 (2007), 119–128.DOI
  68. [68]M. I. Weinstein, Nonlinear Schrödinger equations and sharp interpolation estimates, Comm. Math. Phys. 87 (1983), 567–576.
  69. [69]Y. Wu, Global well-posedness for the nonlinear Schrödinger equation with derivative in energy space, Anal. PDE 6 (2013), 1989–2002.arxiv.org/abs/1310.7166
  70. [70]Y. Wu, Global well-posedness on the derivative nonlinear Schrödinger equation, Anal. PDE 8 (2015), 1101–1112.DOI

Paper details

vlnc:W_93zmyj2e
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents