Semi-classical nonlinear quantum scattering

# On semi-classical limit of nonlinear quantum scattering

Rémi Carles CNRS & Univ. Montpellier
Mathématiques
CC 051
34095 Montpellier
France
###### Abstract.

We consider the nonlinear Schrödinger equation with a short-range external potential, in a semi-classical scaling. We show that for fixed Planck constant, a complete scattering theory is available, showing that both the potential and the nonlinearity are asymptotically negligible for large time. Then, for data under the form of coherent state, we show that a scattering theory is also available for the approximate envelope of the propagated coherent state, which is given by a nonlinear equation. In the semi-classical limit, these two scattering operators can be compared in terms of classical scattering theory, thanks to a uniform in time error estimate. Finally, we infer a large time decoupling phenomenon in the case of finitely many initial coherent states.

This work was supported by the French ANR projects SchEq (ANR-12-JS01-0005-01) and BECASIM (ANR-12-MONU-0007-04).

## 1. Introduction

We consider the equation

 (1.1) iε∂tψε+ε22Δψε=V(x)ψε+|ψε|2ψε,(t,x)∈R×R3,

and both semi-classical () and large time () limits. Of course these limits must not be expected to commute, and one of the goals of this paper is to analyze this lack of commutation on specific asymptotic data, under the form of coherent states as described below. Even though our main result (Theorem 1.6) is proven specifically for the above case of a cubic three-dimensional equation, two important intermediate results (Theorems 1.4 and 1.5) are established in a more general setting. Unless specified otherwise, we shall from now on consider , .

### 1.1. Propagation of initial coherent states

In this subsection, we consider the initial value problem, as opposed to the scattering problem treated throughout this paper. More precisely, we assume here that the wave function is, at time , given by the coherent state

 (1.2) ψε(0,x)=1εd/4a(x−q0√ε)eip0⋅(x−q0)/ε,

where denote the initial position and velocity, respectively. The function belongs to the Schwartz class, typically. In the case where is a (complex) Gaussian, many explicit computations are available in the linear case (see [33]). Note that the -norm of is independent of , .

Throughout this subsection, we assume that the external potential is smooth and real-valued, , and at most quadratic, in the sense that

 ∂αV∈L∞(Rd),∀|α|⩾2.

This assumption will be strengthened when large time behavior is analyzed.

#### 1.1.1. Linear case

Resume (1.1) in the absence of nonlinear term:

 (1.3) iε∂tψε+ε22Δψε=V(x)ψε,x∈Rd,

associated with the initial datum (1.2). To derive an approximate solution, and to describe the propagation of the initial wave packet, introduce the Hamiltonian flow

 (1.4) ˙q(t)=p(t),˙p(t)=−∇V(q(t)),

and prescribe the initial data , . Since the potential is smooth and at most quadratic, the solution is smooth, defined for all time, and grows at most exponentially. The classical action is given by

 (1.5) S(t)=∫t0(12|p(s)|2−V(q(s)))ds.

We observe that if we change the unknown function to by

 (1.6) ψε(t,x)=ε−d/4uε(t,x−q(t)√ε)ei(S(t)+p(t)⋅(x−q(t)))/ε,

then, in terms of , the Cauchy problem (1.3)–(1.2) is equivalent to

 (1.7) i∂tuε+12Δuε=Vε(t,y)uε;uε(0,y)=a(y),

where the external time-dependent potential is given by

 (1.8) Vε(t,y)=1ε(V(x(t)+√εy)−V(x(t))−√ε⟨∇V(x(t)),y⟩).

This potential corresponds to the first term of a Taylor expansion of about the point , and we naturally introduce solution to

 (1.9)

where

 Q(t):=∇2V(q(t)),so that 12⟨Q(t)y,y⟩=limε→0Vε(t,y).

The obvious candidate to approximate the initial wave function is then:

 (1.10) φε(t,x)=ε−d/4u(t,x−q(t)√ε)ei(S(t)+p(t)⋅(x−q(t)))/ε.

Indeed, it can be proven (see e.g. [2, 4, 17, 33, 35, 36]) that there exists independent of such that

Therefore, is a good approximation of at least up to time of order (Ehrenfest time).

#### 1.1.2. Nonlinear case

When adding a nonlinear term to (1.3), one has to be cautious about the size of the solution, which rules the importance of the nonlinear term. To simplify the discussions, we restrict our analysis to the case of a gauge invariant, defocusing, power nonlinearity, . We choose to measure the importance of nonlinear effects not directly through the size of the initial data, but through an -dependent coupling factor: we keep the initial datum (1.2) (with an -norm independent of ), and consider

 iε∂tψε+ε22Δψε=V(x)ψε+εα|ψε|2σψε.

Since the nonlinearity is homogeneous, this approach is equivalent to considering , up to multiplying the initial datum by . We assume , with if : for , defined by

we have, for fixed , , and the Cauchy problem is globally well-posed, (see e.g. [9]). It was established in [11] that the value

 αc=1+dσ2

is critical in terms of the effect of the nonlinearity in the semi-classical limit . If , then , given by (1.9)-(1.10), is still a good approximation of at least up to time of order . On the other hand, if , nonlinear effects alter the behavior of at leading order, through its envelope only. Replacing (1.9) by

 (1.11)

and keeping the relation (1.10), is now a good approximation of . In [11] though, the time of validity of the approximation is not always proven to be of order at least , sometimes shorter time scales (of the order ) have to be considered, most likely for technical reasons only. Some of these restrictions have been removed in [37], by considering decaying external potentials .

### 1.2. Linear scattering theory and coherent states

We now consider the aspect of large time, and instead of prescribing at (or more generally at some finite time), we impose its behavior at . In the linear case (1.3), there are several results addressing the question mentioned above, considering different forms of asymptotic states at . Before describing them, we recall important facts concerning quantum and classical scattering.

#### 1.2.1. Quantum scattering

Throughout this paper, we assume that the external potential is short-range, and satisfies the following properties:

###### Assumption 1.1.

We suppose that is smooth and real-valued, . In addition, it is short range in the following sense: there exists such that

 (1.12) |∂αV(x)|⩽Cα(1+|x|)μ+|α|,∀α∈Nd.

Our final result is established under the stronger condition (a condition which is needed in several steps of the proof), but some results are established under the mere assumption . Essentially, the analysis of the approximate solution is valid for (see Section 4), while the rest of the analysis requires .

Denote by

 Hε0=−ε22ΔandHε=−ε22Δ+V(x)

the underlying Hamiltonians. For fixed , the (linear) wave operators are given by

 Wε±=limt→±∞eitεHεe−itεHε0,

and the (quantum) scattering operator is defined by

 Sεlin=(Wε+)∗Wε−.

See for instance [20].

#### 1.2.2. Classical scattering

Let satisfying Assumption 1.1. For , we consider the classical trajectories defined by (1.4), along with the prescribed asymptotic behavior as :

 (1.13) limt→−∞∣∣q(t)−p−t−q−∣∣=limt→−∞|p(t)−p−|=0.

The existence and uniqueness of such a trajectory can be found in e.g. [20, 51], provided that . Moreover, there exists a closed set of Lebesgue measure zero in such that for all , there exists such that

 limt→+∞∣∣q(t)−p+t−q+∣∣=limt→+∞|p(t)−p+|=0.

The classical scattering operator is . Choosing implies that the following assumption is satisfied:

###### Assumption 1.2.

The asymptotic center in phase space, is such that the classical scattering operator is well-defined,

 Scl(q−,p−)=(q+,p+),p+≠0,

and the classical action has limits as :

 limt→−∞∣∣ ∣∣S(t)−t|p−|22∣∣ ∣∣=limt→+∞∣∣ ∣∣S(t)−t|p+|22−S+∣∣ ∣∣=0,

for some .

#### 1.2.3. Some previous results

It seems that the first mathematical result involving both the semi-classical and large time limits appears in [27], where the classical field limit of non-relativistic many-boson theories is studied in space dimension .

In [56], the case of a short range potential (Assumption 1.1) is considered, with asymptotic states under the form of semi-classically concentrated functions,

 e−iεt2Δψε(t)∣t=−∞=1εd/2ˆf(x−q−ε),f∈L2(Rd),

where denotes the standard Fourier transform (whose definition is independent of ). The main result from [56] shows that the semi-classical limit for can be expressed in terms of the classical scattering operator, of the classical action, and of the Maslov index associated to each classical trajectory. We refer to [56] for a precise statement, and to [57] for the case of long range potentials, requiring modifications of the dynamics.

In [34, 35], coherent states are considered,

 (1.14)

More precisely, in [34, 35], the asymptotic state is assumed to be a complex Gaussian function. Introduce the notation

 δ(t)=S(t)−q(t)⋅p(t)−q−⋅p−2.

Then Assumption 1.2 implies that there exists such that

 δ(t)⟶t→−∞0andδ(t)⟶t→+∞δ+.

In [17, 35], we find the following general result (an asymptotic expansion in powers of is actually given, but we stick to the first term to ease the presentation):

###### Theorem 1.3.

Let Assumptions 1.1 and 1.2 be satisfied, and let

 u−(y)=a−exp(i2⟨Γ−y,y⟩),

where and is a complex symmetric matrix whose imaginary part is positive and non-degenerate. Consider solution to (1.3), with (1.14). Then the following asymptotic expansion holds in :

 Sεlinψε−=1εd/4eiδ+/εeip+⋅(x−q+)/ε+iq+⋅p+/(2ε)^R(G+)u−(x−q+√ε)+O(√ε),

where is the metaplectic transformation associated to .

As a corollary, our main result yields another interpretation of the above statement. It turns out that a complete scattering theory is available for (1.9). As a particular case of Theorem 1.5 (which addresses the nonlinear case), given , there exist a unique solution to (1.9) and a unique such that

 ∥e−it2Δu(t)−u±∥Σ⟶t→±∞0.

Then in the above theorem (where is restricted to be a Gaussian), we have

 u+=^R(G+)u−.

Finally, we mention in passing the paper [48], where similar issues and results are obtained for

 iε∂tψε+ε22Δψε=V(xε)ψε+U(x)ψε,

for a short-range potential, and is bounded as well as its derivatives. The special scaling in implies that initially concentrated waves (at scaled ) first undergo the effects of , then exit a time layer of order , through which the main action of corresponds to the above quantum scattering operator (but with due to the new scaling in the equation). Then, the action of becomes negligible, and the propagation of the wave is dictated by the classical dynamics associated to .

### 1.3. Main results

We now consider the nonlinear equation

 (1.15) iε∂tψε+ε22Δψε=V(x)ψε+εα|ψε|2σψε,

along with asymptotic data (1.14). We first prove that for fixed , a scattering theory is available for (1.15): at this stage, the value of is naturally irrelevant, as well as the form (1.14). To establish a large data scattering theory for (3.1), we assume that the attractive part of the potential,

 (∂rV(x))+=(x|x|⋅∇V(x))+

is not too large, where for any real number .

###### Theorem 1.4.

Let , , and satisfying Assumption 1.1 for some . There exists such that if the attractive part of the potential satisfies

 (∂rV(x))+⩽M(1+|x|)μ+1,∀x∈Rd,

one can define a scattering operator for (3.1) in : for all , there exist a unique solution to (3.1) and a unique such that

 ∥ψε(t)−eiεt2Δψε±∥H1(Rd)⟶t→±∞0.

The (quantum) scattering operator is the map .

We emphasize the fact that several recent results address the same issue, under various assumptions on the external potential : [58] treats the case where is an inverse square (a framework which is ruled out in our contribution), while in [12], the potential is more general than merely inverse square. In [12], a magnetic field is also included, and the Laplacian is perturbed with variable coefficients. We make more comparisons with [12] in Section 3.

The second result of this paper concerns the scattering theory for the envelope equation:

###### Theorem 1.5.

Let , , and satisfying Assumption 1.1 for some . One can define a scattering operator for (1.11) in : for all , there exist a unique solution to (1.11) and a unique such that

 ∥e−it2Δu(t)−u±∥Σ⟶t→±∞0.

As mentioned above, the proof includes the construction of a linear scattering operator, comparing the dynamics associated to (1.9) to the free dynamics . In the above formula, we have incorporated the information that is unitary on , but not on (see e.g. [13]).

We can now state the nonlinear analogue to Theorem 1.3. Since Theorem 1.4 requires , we naturally have to make this assumption. On the other hand, we will need the approximate envelope to be rather smooth, which requires a smooth nonlinearity, . Intersecting this property with the assumptions of Theorem 1.4 leaves only one case: and , that is (1.1), up to the scaling. We will see in Section 5 that considering is also crucial, since the argument uses dispersive estimates which are known only in the three-dimensional case for satisfying Assumption 1.1 with (larger values for could be considered in higher dimensions, though). Introduce the notation

 Σk={f∈Hk(Rd),x↦|x|kf(x)∈L2(Rd)}.
###### Theorem 1.6.

Let Assumptions 1.1 and 1.2 be satisfied, with and as in Theorem 1.4. Consider solution to

 iε∂tψε+ε22Δψε=V(x)ψε+ε5/2|ψε|2ψε,(t,x)∈R×R3,

and such that (1.14) holds, with . Then the following asymptotic expansion holds in :

 (1.16) Sεψε−=1ε3/4eiδ+/εeip+⋅(x−q+)/ε+iq+⋅p+/(2ε)u+(x−q+√ε)+O(√ε),

where is given by Theorem 1.4 and stems from Theorem 1.5.

###### Remark 1.7.

In the subcritical case, that is if we consider

 iε∂tψε+ε22Δψε=V(x)ψε+εα|ψε|2ψε,(t,x)∈R×R3,

along with (1.14), for some , the argument of the proof shows that (1.16) remains true, but with given by the scattering operator associated to (1.9) (as opposed to (1.11)), that is, the same conclusion as in Theorem 1.3 when is a Gaussian.

As a corollary of the proof of the above result, and of the analysis from [11], we infer:

###### Corollary 1.8 (Asymptotic decoupling).

Let Assumption 1.1 be satisfied, with and as in Theorem 1.4. Consider solution to

 iε∂tψε+ε22Δψε=V(x)ψε+ε5/2|ψε|2ψε,(t,x)∈R×R3,

with initial datum

 ψε(0,x)=N∑j=11ε3/4aj(x−q0j√ε)eip0j⋅(x−q0j)/ε=:ψε0(x),

where , , so that scattering is available as for , in the sense of Assumption 1.2, and . We suppose for . Then we have the uniform estimate:

 supt∈R∥∥ ∥∥ψε(t)−N∑j=1φεj(t)∥∥ ∥∥L2(R3)⟶ε→00,

where is the approximate solution with the -th wave packet as an initial datum. As a consequence, the asymptotic expansion holds in , as :

 (Wε±)−1ψε0=N∑j=11ε3/4eiδ±j/εeip±j⋅(x−q±j)/ε+iq±j⋅p±j/(2ε)uj±(x−q±j√ε)+o(1),

where the inverse wave operators stem from Theorem 1.4, the ’s are the asymptotic states emanating from , and

 δ±j=limt→±∞(Sj(t)−qj(t)⋅pj(t)−q0j⋅p0j2)∈R.
###### Remark 1.9.

In the case , the approximation by wave packets is actually exact, since then , hence . For one wave packet, Theorem 1.6 becomes empty, since it is merely a rescaling. On the other hand, for two initial wave packets, even in the case , Corollary 1.8 brings some information, reminiscent of profile decomposition. More precisely, define by (1.6), and choose (arbitrarily) to privilege the trajectory . The Cauchy problem is then equivalent to

 ⎧⎪ ⎪ ⎪⎨⎪ ⎪ ⎪⎩i∂tuε+12Δuε=|uε|2uε,uε(0,y)=a1(y)+a2(y+q01−q02√ε)eip02⋅δq0/ε−iδp0⋅y/√ε,

where we have set and . Note however that the initial datum is uniformly bounded in , but in no for (if ), while the equation is -critical, Therefore, even in the case , Corollary 1.8 does not seem to be a consequence of profile decompositions like in e.g. [21, 42, 45]. In view of (1.4), the approximation provided by Corollary 1.8 reads, in that case:

 uε(t,y)=u1(t,y)+u2(t,y+tδp0+δq0√ε)eiϕε2(t,y)+o(1)% in L∞(R;L2(R3)),

where the phase shift is given by

 ϕε2(t,y) =1εp02⋅(tδp0+δq0)−1√εδp0⋅y+t2ε(|p02|2−|p01|2) =1εp02⋅δq0−1√εδp0⋅y−t2ε|δp0|2.

Notation. We write whenever there exists independent of and such that .

## 2. Spectral properties and consequences

In this section, we derive some useful properties for the Hamiltonian

 H=−12Δ+V.

Since the dependence upon is not addressed in this section, we assume .

First, it follows for instance from [46] that Assumption 1.1 implies that has no singular spectrum. Based on Morawetz estimates, we show that has no eigenvalue, provided that the attractive part of is sufficiently small. Therefore, the spectrum of is purely absolutely continuous. Finally, again if the attractive part of is sufficiently small, zero is not a resonance of , so Strichartz estimates are available for .

### 2.1. Morawetz estimates and a first consequence

In this section, we want to treat both linear and nonlinear equations, so we consider

 (2.1) i∂tψ+12Δψ=Vψ+λ|ψ|2σψ,λ∈R.

Morawetz estimate in the linear case will show the absence of eigenvalues. In the nonlinear case , these estimates will be a crucial tool for prove scattering in the quantum case. The following lemma and its proof are essentially a rewriting of the presentation from [3].

###### Proposition 2.1 (Morawetz inequality).

Let , and satisfying Assumption 1.1 for some . There exists such that if the attractive part of the potential satisfies

 (∂rV(x))+⩽M(1+|x|)μ+1,∀x∈Rd,

then any solution to (2.1) satisfies

 (2.2) λ∬R×Rd|ψ(t,x)|2σ+2|x|dtdx+∬R×Rd|ψ(t,x)|2(1+|x|)μ+1dtdx≲∥ψ∥2L∞(R;H1).

In other words, the main obstruction to global dispersion for comes from , which is the attractive contribution of in classical trajectories, while is the repulsive part, which does not ruin the dispersion associated to (it may reinforce it, see e.g. [8], but repulsive potentials do not necessarily improve the dispersion, see [32]).

###### Proof.

The proof follows standard arguments, based on virial identities with a suitable weight. We resume the main steps of the computations, and give more details on the choice of the weight in our context. For a real-valued function , we compute, for solution to (3.1),

 ddt∫h(x)|ψ(t,x)|2dx=Im∫¯ψ(t,x)∇h(x)⋅∇ψ(t,x)dx,
 (2.3) ddtIm∫¯ψ(t,x)∇h(x)⋅∇ψ(t,x)dx =∫∇¯ψ(t,x)⋅∇2h(x)∇ψ(t,x)dx −14∫|ψ(t,x)|2 Δ2h(x)dx−∫|ψ(t,x)|2∇V⋅∇h(x)dx +λσσ+1∫|ψ(t,x)|2σ+2Δh(x)dx.

In the case , the standard choice is , for which

 ∇h=x|x|,∇2jkh=1|x|(δjk−xjxk|x|2),Δh⩾d−1h,and Δ2h⩽0 for d⩾3.

This readily yields Proposition 2.1 in the repulsive case , since .

In the same spirit as in [3], we proceed by perturbation to construct a suitable weight when the attractive part of the potential is not too large. We seek a priori a radial weight, , so we have

 Δh=h′′+d−1rh′, Δ2h=h(4)+2d−1rh(3)+(d−1)(d−3)r2h′′−(d−1)(d−3)r3h′, ∇2jkh=1r(δjk−xjxkr2)h′+xjxkr2h′′.

We construct a function such that , so the condition will remain. The goal is then to construct a radial function such that the second line in (2.3) is non-negative, along with for some .

Case . In this case, the expression for is simpler, and the above conditions read

 14h(4)+1rh(3)+∇V(x)⋅∇h⩽0, h′′+2rh′⩾ηr,h′,h′′⩾0.

Since we do not suppose a priori that is a radial potential, the first condition is not rigorous. We actually use the fact that for , Assumption 1.1 implies

 ∇V(x)⋅∇h⩽(∂rV(x))+h′(r)⩽M(1+r)μ+1h′(r).

To achieve our goal, it is therefore sufficient to require:

 (2.4) 14h(4)+1rh(3)+M(1+r)μ+1h′⩽0, (2.5) h′′+2rh′⩾ηr,h′∈L∞(R+), h′,h′′⩾0.

In view of (2.5), we seek

 h′(r)=η+∫r0h′′(ρ)dρ.

Therefore, if with , (2.5) will be automatically fulfilled. We now turn to (2.4). Since we want , we may even replace by a constant in (2.4), and solve, for , the ODE

 14h(4)+1rh(3)+C(1+r)μ+1=0.

 h(3)(r)=−4Cr4∫r0ρ4(1+ρ)μ+1dρ,

along with the properties ,

 h(3)(r)∼r→∞−krmin(μ,4),for some k>0.

It is now natural to set

 h′′(r)=−∫∞rh(3)(ρ)dρ,

so we have and

 h′′(r)∼r→∞κrmin(μ−1,3),for some κ>0.

This function is indeed in if and only if . We define by ,

 (2.6) h(3)(r)=−Kr4∫r0ρ4(1+ρ)μ+1dρ,

for some , and being given by the above relations: (2.5) is satisfied for any value of , and (2.4) boils down to an inequality of the form

 (2.7) −K4+M(η+C(μ)K)⩽0,

where is proportional to

 1K∥h′∥L∞=∫∞0∫∞r1ρ4∫ρ0s4(1+s)μ+1dsdρdr.

We infer that (2.6) is satisfied for , provided that . Note then that by construction, we may also require

 14Δ2h+∇V⋅∇h⩽−c0(1+|x|)μ+1,

for morally very small.

Case . Resume the above reductions, pretending that the last two terms in are not present: (2.6) just becomes

 h(3)(r)=−Kr2d−2∫r0ρ2d−2(1+ρ)μ+1dρ,

and we see that with and defined like before, we have

 rh′′−h′=−η−∫r0h′′+rh′′.

Since this term is negative at and has a non-positive derivative, we have , so finally . ∎

We infer that has no eigenvalue. Indeed, if there were an solution to , , then , and would be an solution to (2.1) for . This is contradiction with the global integrability in time from (2.2), so .

### 2.2. Strichartz estimates

In [3, Proposition 3.1], it is proved that zero is not a resonance of , but with a definition of resonance which is not quite the definition in [52], which contains a result that we want to use. So we shall resume the argument.

By definition (as in [52]), zero is a resonance of , if there is a distributional solution , such that for all , to .

###### Corollary 2.2.

Under the assumptions of Proposition 2.1, zero is not a resonance of .

###### Proof.

Suppose that zero is a resonance of . Then by definition, we obtain a stationary distributional solution of (2.1) (case ), , and we may assume that it is real-valued. Since , Assumption 1.1 implies

 ⟨x⟩μ−sΔψ∈L2(Rd),∀s>12.

This implies that , by taking for instance in

 ∫|∇ψ|2=−∫⟨x⟩−sψ⟨x⟩sΔψ.

By definition, for all test function ,

 (2.8) 12∫Rd∇φ(x)⋅∇ψ(x)dx+∫RdV(x)φ(x)ψ(x)dx=0.

Let be the weight constructed in the proof of Proposition 2.1, and consider

 φ=ψΔh+2∇ψ⋅∇h.

Since , , and , we see that , and that this choice is allowed in (2.8). Integration by parts then yields (2.3) (where the left hand side is now zero):

 0=∫∇ψ⋅∇2h∇ψ−14∫ψ2Δ2h−∫ψ2∇V⋅∇h.

By construction of , this implies

 ∫Rdψ(x)2(1+|x|)μ+1dx⩽0,

hence . ∎

Therefore, [52, Theorem 1.4] implies non-endpoint global in time Strichartz estimates. In the case , we know from [31] that (in view of the above spectral properties)

 ∥e−itH∥