betaFeedback

WorksMathematics

Tunneling effect in some singular wells

Abstract

The Schrödinger operator in one dimension is considered. When the electric potential has a linear/V-shaped well, an approximation of the ground-state is established at fixed distance from the well. In this situation, the classical WKB analysis and the localization techniques à la Agmon near a non-degenerate minimum do not work. We propose an alternative to the WKB analysis better suited to the shape of the well and we prove that our Ansatz is an exponentially good approximation by means of revisited (and simplified) Agmon estimates.

Introduction and motivation

About semiclassical tunneling in one dimension

This article is devoted to the spectral analysis of Schrödinger operators in one dimension, i.e., operators of the form

L=−h2∂x2+V(x),\mathcal{L}=-h^2\partial_x^2+V(x),

acting on L2(I)L^2(I), for some interval II and some real-valued, confining, continuous, electric potential VV. By confining, we mean that, if II is unbounded, we have inf⁡V<lim inf⁡∣x∣→+∞V(x)\inf V < \liminf_{|x|\to+\infty} V(x). We are especially interested in the approximation of the groundstate eigenfunction in the semiclassical limit h→0h \to0, which is equivalent to the large electric field limit.

Such a study is motivated by the tunneling effect, and probably began with the celebrated article by Harrell [8]. This work was devoted to symmetric double-well potentials, and he established formulas describing the first spectral gap in terms of the eigenfunctions, see [8 Theorem 2.1]. The exponentially small size of this gap is directly related to the exponential smallness of one-well eigenfunctions in the so-called classically forbidden zone, where the electric potential is larger than the eigenvalue, see [8 Theorems 2.8 & 2.22]. These considerations are then applied to the case of an explicit smooth non-degenerate double-well, where the one-well eigenfunctions are estimated at a fixed distance from the well, see [8 Section III]. However, for more singular wells, such estimates are not as easy to obtain.

This relation between exponential decay of the one-well eigenfunctions and the spectral gaps was then used by Simon [20, 21] and Helffer-Sjöstrand [10, 11] in order to consider quantum systems of higher dimensions. The reader might also want to consider the monograph [5 Chapters 3 & 6], where WKB constructions are performed (in any dimension) in smooth non-degenerate wells and proved to be exponentially good approximations in Appendix A, see also [9] and [19].

In this article we analyze these questions for one-dimensional singular wells. In fact this analysis is also the occasion to present an alternative to the standard WKB constructions, together with a revisited (and simplified) approach to Agmon estimates. In particular, the techniques we present below could be adapted to non-degenerate wells as well, providing slightly better decay estimates.

The problem under consideration

The purpose of the present article is to approximate the groundstate eigenfunction of the Dirichlet realization of L\mathcal{L} when I=R+I = \mathbb{R}_{+}, with a smooth confining VV attaining a unique minimum at x=0x = 0 and such that V′(0)>0V'(0) > 0. This simple model appears in the study of the so-called Stark effect, see, for instance, [3]. It is also motivated by the analysis of broken waveguides (or asymptotically flat triangles), see [4, 17] where effective operators involve the Dirichlet realization of L\mathcal{L} on a half-line. Hopefully, our results will help establish accurate tunneling estimates by means of dimensional reductions in these contexts, see, for instance, [17 Section 1.6].

Let us gather our main assumptions.

Assumption 1.1. We consider a non-negative smooth function VV on [0,+∞)[0,+\infty) such that V(0)=0V(0) = 0 and V′(0)>0V'(0) > 0. We assume that it has a unique minimum at 00 and that it is not attained at infinity. For λ>0\lambda> 0 small enough, we denote by xλ>0x_\lambda> 0 the unique solution to V(x)=λV(x) = \lambda.

Under Assumption 1.1, we know that L\mathcal{L} has discrete eigenvalues below the essential spectrum, which we denote by

0<λ1<λ2<⋯<λj<⋯0 < \lambda_1 < \lambda_2 < \cdots< \lambda_j < \cdots

As h→0h \to0, there are increasingly many such eigenvalues, even if for fixed hh there may be only a finite number of them. We are especially interested in the asymptotic behaviour of the normalized groundstate ff, in the semiclassical limit h→0h \to0. Our main result is the following.

Theorem 1.2. Under Assumption 1.1, for all j≥1j \ge1, we have

λj=∣zj∣V′(0)23h23+O(h43),(1)\lambda_j = |z_j|V'(0)^{\frac{2}{3}}h^{\frac{2}{3}} + O(h^{\frac{4}{3}}), \tag*{(1)}

where zjz_j is the jj-th zero of the Airy function. For all x>0x > 0, the L2L^2-normalized positive groundstate satisfies, in the limit h→0h \to0,

f(x)=V′(0)132π∣Ai′(z1)∣e−1hS(x)V(x)14h16(1+o(1)),S(x)=∫0x(V(s)−λ1)+ ds.f(x) = \frac{V'(0)^{\frac{1}{3}}}{2\sqrt{\pi}|Ai'(z_1)|}\frac{e^{-\frac{1}{h}S(x)}}{V(x)^{\frac{1}{4}}h^{\frac{1}{6}}}(1+o(1)), \qquad S(x) = \int_0^x \sqrt{(V(s)-\lambda_1)_+}\,\mathrm{d}s.

Remark 1.3. Our proof actually shows that the λj\lambda_j have asymptotic expansions in powers of h23h^{\frac{2}{3}}. The main novelty in Theorem 1.2 is the asymptotic estimate of the groundstate eigenfunction at a fixed distance from its linear minimum. It relies on two independent steps, which are based on classical ideas when considering tunneling effect, namely:

  • construction of exponentially good quasimodes,

  • sharp exponential decay estimates of the true eigenfunctions and their approximations.

The combination of these two ideas allows one to replace ff by the quasimode, for which we have the pointwise asymptotic formula. This strategy is now standard, and was used in particular in the setting of non-degenerate wells, as described in [5], but also in the case of degenerate wells (V(x)∼xαV(x) \sim x^{\alpha} with α≥2\alpha\ge2, see [14]). In these situations, the exponential decay is given by the Agmon estimates [1], and quasimodes are in the WKB form.

We stress that in our situation, with linear wells, both the standard WKB analysis and the standard Agmon strategy fail due to the linear behavior. The strategy of [14 Théorème 3.1] consisting in gluing an Ansatz living at the bottom of the well, i.e., in {V(x)<λ}\{V(x)<\lambda\}, and a WKB Ansatz away from the well, does not work in our context due to a more complicated transition of scales. Indeed, the localisation near the bottom of the well is much stronger in the case of linear wells: the characteristic scale near the minimum is h23≪h12h^{\frac{2}{3}}\ll h^{\frac{1}{2}} making the connection between the classically allowed and forbidden regions, {V(x)<λ}\{V(x)<\lambda\} and {V(x)>λ}\{V(x)>\lambda\} respectively, more abrupt than for non-degenerate wells. Therefore, we have to provide substitutes to the WKB quasimodes to deal with this transition, and to revisit the usual Agmon estimates.

This yields a robust approach, that could be used in a more general setting. In fact, our contribution gives an alternative to the WKB constructions in the case of non-degenerate wells, and it simplifies the usual Agmon estimates in this case also. For our linear wells, one will use an Airy profile to connect the classically allowed and forbidden regions. Our quasimode construction is reminiscent of Liouville’s technique [13 Section III], see also [7 Section 2] and [15 Section 4]. More precisely, our first Ansatz (see (3.8) with N=0N=0) is Langer’s uniform approximation near a turning point [12], see also [16 Chapter 11]. Our contribution lies in its iterative improvement and in the Agmon estimates which allow us to compare it with the true groundstate.

Organization of the article

This article is organized as follows. § 2 is devoted to the exposition of the main insights behind the proof of Theorem 1.2. We state the optimal Agmon estimate and introduce the quasimode. In § 3, we provide the reader with the detailed proofs of our statements. This section contains the proof of the Agmon estimates, a linear approximation inside the well, the detailed iterative quasimode construction, and the conclusion of the proof of Theorem 1.2. In § 4, an application to a tunneling result is discussed.

Main insights

Agmon estimates

A first key ingredient to prove Theorem 1.2 is given by the celebrated Agmon estimates (see [1]), which provide exponential decay estimates on the eigenfunctions of Schrödinger operators, of the form

∥eΦ/hf∥≤Ch−κ∥f∥,h∈(0,1].(2)\lVert e^{\Phi/h}f\rVert\le Ch^{-\kappa}\lVert f\rVert,\qquad h\in(0,1]. \tag*{(2)}

Two natural questions arise: for which weights Φ\Phi can we obtain such estimates, and with which power κ\kappa? This second question, which may seem of lower interest, will actually have implications if one wants to consider potentials of lower regularity.

In our specific situation, we obtain the following Agmon-type elliptic estimate, with optimal weight Φ=S\Phi=S, and a relatively small power κ\kappa. The proof is given in § 3.1 (where a slightly stronger version – which controls 2 derivatives – is established).

Proposition 2.1. Let KK, E>0E>0. There exist h0h_0 and C>0C>0 such that for all h∈(0,h0)h\in(0,h_0), all 0≤λ≤Eh230\le\lambda\le Eh^{\frac{2}{3}} and all u∈H01(R+)∩H2(R+)u\in H^1_0(\mathbb{R}_+)\cap H^2(\mathbb{R}_+),

∥eS/hu∥L2(0,K)≤Ch−43∥eS/h(L−λ)u∥+Ch−13∥u∥,\lVert e^{S/h}u\rVert_{L^2(0,K)}\le Ch^{-\frac{4}{3}}\lVert e^{S/h}(\mathcal{L}-\lambda)u\rVert+Ch^{-\frac{1}{3}}\lVert u\rVert,

where S(x)=∫0x(V(t)−λ)+ dtS(x)=\int_0^x\sqrt{(V(t)-\lambda)_+}\,\mathrm{d}t.

Remark 2.2. When uu is an eigenfunction, we obtain an estimate of the form (2) on compact sets, with κ=13\kappa=\frac{1}{3}. The weight Φ=S\Phi=S is in fact optimal in the sense that it corresponds to the decay announced in Theorem 1.2.

The idea behind the Agmon estimates is to use the ellipticity of the conjugate operator LΦ−λ=eΦ/h(L−λ)e−Φ/h\mathcal{L}_\Phi-\lambda=e^{\Phi/h}(\mathcal{L}-\lambda)e^{-\Phi/h} away from a small neighborhood of x=0x=0, for a suitable weight Φ\Phi. A straightforward computation gives that, in the sense of quadratic forms,

Re⁡LΦ−λ=h2Dx2+V(x)−Φ′(x)2−λ,(3)\operatorname{Re}\mathcal{L}_\Phi-\lambda=h^2D_x^2+V(x)-\Phi'(x)^2-\lambda, \tag*{(3)}

which naturally leads us to consider the weight

Φ(x)=1−ε∣∫0xV(s) ds∣,ε∈(0,1),(4)\Phi(x)=\sqrt{1-\varepsilon}\left|\int_0^x\sqrt{V(s)}\,\mathrm{d}s\right|,\qquad\varepsilon\in(0,1), \tag*{(4)}

and gives

Re⁡LΦ−λ≥εV(x)−λ.\operatorname{Re}\mathcal{L}_\Phi-\lambda\ge\varepsilon V(x)-\lambda.

This allows one to establish an estimate in the form

∥eΦ/hu∥≤Ch−κ1∥eΦ/h(L−λ)u∥+Ch−κ2∥u∥,(5)\lVert e^{\Phi/h}u\rVert\le Ch^{-\kappa_1}\lVert e^{\Phi/h}(\mathcal{L}-\lambda)u\rVert+Ch^{-\kappa_2}\lVert u\rVert, \tag*{(5)}

for some powers κ1,κ2>0\kappa_1,\kappa_2>0. The rather naive choice (4) allows one to show in our specific situation that the groundstate eigenfunction is exponentially localized near 0 at a scale of order h23h^{\frac{2}{3}} since, for xx small, we have Φ(x)∼21−ε3V′(0)x32\Phi(x)\sim\frac{2\sqrt{1-\varepsilon}}{3}\sqrt{V'(0)}x^{\frac{3}{2}}.

The choice (4) was made in the case of non-degenerate wells, see [10 Lemma 2.3] but it prevents one from taking the limiting weight Φ=∣∫0xV(s) ds∣\Phi=\left|\int_0^x\sqrt{V(s)}\,\mathrm{d}s\right|, which is needed to get an exponentially good approximation of the eigenfunctions far from the well. To fix this issue in [10 Prop. 5.5], the authors introduce a slightly better weight, which is a logarithmic correction of the previous one satisfying

Φ≃∣∫0xV(s) ds∣−Nh∣log⁡h∣,\Phi\simeq\left|\int_0^x \sqrt{V(s)}\,\mathrm{d}s\right|-Nh|\log h|,

with a large constant NN. With their more involved choice, one can actually replace Φ\Phi by ∫0xV(s) ds\int_0^x\sqrt{V(s)}\,\mathrm{d}s at the price of increasing the powers κ1\kappa_1 and κ2\kappa_2 by NN. Note that smooth versions of such weights have recently been considered in a pseudodifferential framework in [6 Section 4.2].

In our situation with linear wells, none of these weights works since the eigenvalues are much larger (they are of order λ≃h23\lambda\simeq h^{\frac{2}{3}}) than in the non-degenerate case (when they are of order hh). For this reason, we rather consider the weight:

Φh(x)=1−h∫0x(V(s)−λ)+ ds,(6)\Phi_h(x)=\sqrt{1-h}\int_0^x\sqrt{(V(s)-\lambda)_+}\,\mathrm{d}s, \tag*{(6)}

with which we can directly prove (5) with κ1=43\kappa_1=\frac{4}{3}, κ2=13\kappa_2=\frac{1}{3}, and no fancy logarithmic correction is needed. Then, we can easily replace Φh\Phi_h by SS since ∣Φh−S∣=O(h)|\Phi_h-S|=O(h) on any compact set.

Remark 2.3. Our strategy also provides a simpler proof for non-degenerate wells, for which one can still use the weight (6), without any logarithmic correction. This choice allows one to prove simultaneously the optimal exponential decay of the eigenfunctions and their exponential approximation far from the well.

Quasimodes and approximation

Thanks to the Agmon estimate (5) with the naive weight (4), we can see that most of the mass of the groundstate ff lives on an interval of size h23h^{\frac{2}{3}} around the origin. Thus, a Taylor expansion near x=0x=0 suggests considering the Airy operator

A=−h2∂x2+V′(0)x.\mathcal{A}=-h^{2}\partial_x^2+V'(0)x.

A first approximation of ff is then given by the groundstate of A\mathcal{A}, i.e., uniformly on [0,1][0,1],

f(x)=c0h−13Ai⁡(V′(0)13h−23x+z1)+O(h13),(7)f(x)=c_0h^{-\frac{1}{3}}\operatorname{Ai}\left(V'(0)^{\frac{1}{3}}h^{-\frac{2}{3}}x+z_1\right)+O\left(h^{\frac{1}{3}}\right), \tag*{(7)}

where Ai⁡\operatorname{Ai} denotes the Airy function, z1z_1 its first negative zero, and

c0=V′(0)16(∫0+∞Ai⁡2(x+z1) dx)12=V′(0)16∣Ai⁡′(z1)∣.c_0=\frac{V'(0)^{\frac{1}{6}}}{\left(\int_0^{+\infty}\operatorname{Ai}^2(x+z_1)\,\mathrm{d}x\right)^{\frac{1}{2}}}=\frac{V'(0)^{\frac{1}{6}}}{|\operatorname{Ai}'(z_1)|}.

is a normalizing constant. In L2L^2 norm, the error in (7) is even O(h23)O(h^{\frac{2}{3}}). It also suggests that the first eigenvalues behave like in (1). These approximations are made rigorous in Proposition 3.2.

However (7) is only good for xx of order h23h^{\frac{2}{3}}. Indeed, Ai decays exponentially fast at infinity, thus making the error term O(h23)O(h^{\frac{2}{3}}) too large to describe the announced behavior of ff, when x≫h23x \gg h^{\frac{2}{3}}, see Theorem 1.2.

For larger values of xx, one may want to consider a WKB Ansatz, i.e., a trial function of the form

fWKB=e−Sh(a0+ha1+h2a2+⋯ ).f^{\mathrm{WKB}}=e^{-\frac{S}{h}}(a_0+ha_1+h^2a_2+\cdots).

It is well-known that such an Ansatz can be built, and that it is a good approximation of the solution for large values of xx (i.e. far inside the classically forbidden region {V(x)>λ}={x>xλ}\{V(x)>\lambda\}=\{x>x_\lambda\}). However, this construction degenerates when we reach the classically allowed region, when xx is of order h23h^{\frac{2}{3}}. To overcome this problem, we give an alternative to the classical WKB method by changing the exponential profile of the WKB Ansatz, especially in the region x≳h23x\gtrsim h^{\frac{2}{3}}, see § 3.3. This Ansatz already appeared (for N=0N=0) a long time ago in [15 Section 5 (iii)].

Proposition 2.4. Let N≥0N\geq0, K≥1K\geq1; let λ=λ1\lambda=\lambda_1. There exist h0,c>0h_0,c>0 and, for h∈(0,h0)h\in(0,h_0), a smooth function φ\varphi on [0,K][0,K] such that φ′≥c\varphi'\geq c and φ′(xλ)=V′(0)13+O(h23)\varphi'(x_\lambda)=V'(0)^{\frac{1}{3}}+O(h^{\frac{2}{3}}) and such that the function

fNapp(x)=αhφ′(x)Ai⁡(h−23φ(x)),withαh=f(xλ)φ′(xλ)Ai⁡(h−23φ(xλ)),f_N^{\mathrm{app}}(x)=\frac{\alpha_h}{\sqrt{\varphi'(x)}}\operatorname{Ai}\left(h^{-\frac{2}{3}}\varphi(x)\right),\qquad\text{with}\qquad\alpha_h=f(x_\lambda)\frac{\sqrt{\varphi'(x_\lambda)}}{\operatorname{Ai}\left(h^{-\frac{2}{3}}\varphi(x_\lambda)\right)},

satisfies fNapp(xλ)=f(xλ)f_N^{\mathrm{app}}(x_\lambda)=f(x_\lambda) and

(L−λ)fNapp=O(h2N+2)fNapp,(8)(\mathcal{L}-\lambda)f_N^{\mathrm{app}}=O(h^{2N+2})f_N^{\mathrm{app}}, \tag*{(8)}

uniformly on [xλ,K][x_\lambda,K]. Moreover,

φ(x)=(32∫0x(V(s)−λ1)+ ds)23+O(h2),\varphi(x)=\left(\frac{3}{2}\int_0^x\sqrt{(V(s)-\lambda_1)_+}\,\mathrm{d}s\right)^{\frac{2}{3}}+O(h^2),

for all x∈[xλ,K]x\in[x_\lambda,K], with equality when N=0N=0.

Compared to the WKB Ansatz, our construction is better suited for the shape of the well, through the use of the Airy function instead of the exponential. This idea could be generalized to other kinds of wells by using other special functions.

Remark 2.5. In fact, in order to prove Theorem 1.2, we only need the construction with N=0N=0. This is due to the low powers of h−1h^{-1} appearing in our Agmon estimates. For larger NN, we get higher order approximations of ff. See § 3.6.

Remark 2.6. – The function φ\varphi that we construct satisfies φ(0)=z1h23+O(h43)\varphi(0)=z_1h^{\frac{2}{3}}+O(h^{\frac{4}{3}}) and thus fNappf_N^{\mathrm{app}} almost satisfies the Dirichlet boundary condition. However, we will mainly use our Ansatz in the classically forbidden region.

  • In Lemma 3.6, it is shown that (f~Napp)′(xλ)(\tilde{f}^{\mathrm{app}}_N)'(x_\lambda) almost coincides with f′(xλ)f'(x_\lambda).

  • We recall that

Ai⁡(x)=x→+∞12πx1/4e−23x3/2(1+o(1)).\operatorname{Ai}(x) \underset{x\to+\infty}{=} \frac{1}{2\sqrt{\pi}x^{1/4}}e^{-\frac{2}{3}x^{3/2}}(1+o(1)).

Thus, for all fixed x>0x>0, we have, when h→0h\to0,

fNapp(x)=ch16V(x)14e−1hS(x)(1+o(1)),(9)f_N^{\mathrm{app}}(x)=\frac{c}{h^{\frac{1}{6}}V(x)^{\frac{1}{4}}}e^{-\frac{1}{h}S(x)}(1+o(1)), \tag*{(9)}

with c=c0V′(0)162πc=\frac{c_0V'(0)^{\frac{1}{6}}}{2\sqrt{\pi}} and where (7) is used to estimate αh\alpha_h. This matches the asymptotic behavior in Theorem 1.2.

In § 3.4, the function fNappf_N^{\mathrm{app}} is used to construct an exponentially good quasimode defined as follows. We consider the function defined by

ψ(x)={f(x)if x∈[0,xλ],χ(x)f~Napp(x)if xλ≤x,\psi(x)= \begin{cases} f(x) & \text{if }x\in[0,x_\lambda],\\ \chi(x)\tilde{f}_N^{\mathrm{app}}(x) & \text{if }x_\lambda\le x, \end{cases}

where χ\chi is a smooth cutoff function and f~Napp\tilde{f}_N^{\mathrm{app}} is a very small correction of fNappf_N^{\mathrm{app}} locally near xλx_\lambda to make it C1C^1 at xλx_\lambda and belong to the domain of L\mathcal{L}, see (19). In Proposition 3.7, we prove that ψ\psi is a good L2L^2-approximation of ff.

In § 3.5, we prove that ψ\psi is an exponentially good quasimode, see Lemma 3.8. The crucial ingredient is (8), combined with the asymptotic behavior of the Airy function.

Finally in § 3.6, we use the Agmon estimates (apply Proposition 3.1 to u=f−ψu=f-\psi) and we get, for all fixed x>0x>0,

eΦ(x)/h(f(x)−ψ(x))=O(h2N),e^{\Phi(x)/h}(f(x)-\psi(x))=O(h^{2N}),

with a weight Φ\Phi such that Φ(x)=S(x)+O(h)\Phi(x)=S(x)+O(h). Then, we have

f(x)=ψ(x)+O(h2N)e−S(x)/h.f(x)=\psi(x)+O(h^{2N})e^{-S(x)/h}.

For fixed xx, and hh small enough, ψ(x)=fNapp(x)\psi(x)=f_N^{\mathrm{app}}(x), which is estimated with (9), and thus Theorem 1.2 follows by taking N=0N=0.

Analysis of the single singular well

Proof of the optimal Agmon estimates

As explained in the introduction, the main idea to prove the Agmon estimates with optimal weight SS (as stated in Proposition 2.1), is to first consider the quasi-optimal weight Φ\Phi.

Proposition 3.1. Let E>0E>0. There exist h0h_0 and C>0C>0 such that for all h∈(0,h0)h\in(0,h_0), all 0≤λ≤Eh230\le\lambda\le Eh^{\frac{2}{3}}, the following holds. Let M≥1M\ge1. Let us consider the bounded and Lipschitz function

Φ(x)=1−hmin⁡(S(x),S(M2)),\Phi(x)=\sqrt{1-h}\min\left(S(x),S\left(\frac{M}{2}\right)\right),

with S(x)=∫0x(V−λ)12S(x)=\int_0^x(V-\lambda)^{\frac{1}{2}}. For u∈H2(R+)∩H01(R+)u\in H^2(\mathbb{R}_+)\cap H^1_0(\mathbb{R}_+),

∥eΦ/hu∥≤Ch−43∥eΦ/h(L−λ)u∥+Ch−13∥u∥,\left\|e^{\Phi/h}u\right\|\leq Ch^{-\frac{4}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+Ch^{-\frac{1}{3}}\|u\|,
∥h∂xeΦ/hu∥≤Ch−23∥eΦ/h(L−λ)u∥+Ch13∥u∥.\left\|h\partial_xe^{\Phi/h}u\right\|\leq Ch^{-\frac{2}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+Ch^{\frac{1}{3}}\|u\|.

Additionally, there exists a constant CM>0C_M>0 so that

∥eΦ/hh2∂x2u∥L2(0,M/2)≤CMh−43∥eΦ/h(L−λ)u∥+CMh−13∥u∥.\left\|e^{\Phi/h}h^2\partial_x^2u\right\|_{L^2(0,M/2)}\leq C_Mh^{-\frac{4}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+C_Mh^{-\frac{1}{3}}\|u\|.

Proof. By the Agmon formula (see [5 Proposition 6.1]) or by using (3),

Re⁡⟨(L−λ)u,e2Φ/hu⟩=∥h∂x(eΦ/hu)∥2+∫R+(V−λ−Φ′2)∣eΦ/hu∣2dx.\operatorname{Re}\langle(\mathcal{L}-\lambda)u,e^{2\Phi/h}u\rangle=\left\|h\partial_x(e^{\Phi/h}u)\right\|^2+\int_{\mathbb{R}_+}(V-\lambda-\Phi'^2)\left|e^{\Phi/h}u\right|^2dx.

The weight Φ\Phi is chosen so that Φ′2≤(1−h)(V−λ)+\Phi'^2\leq(1-h)(V-\lambda)_+. Thus,

Re⁡⟨(L−λ)u,e2Φ/hu⟩≥∥h∂x(eΦ/hu)∥2+∫R+(h(V−λ)+−(V−λ)−)∣eΦ/hu∣2dx\operatorname{Re}\langle(\mathcal{L}-\lambda)u,e^{2\Phi/h}u\rangle\geq\left\|h\partial_x(e^{\Phi/h}u)\right\|^2+\int_{\mathbb{R}_+}\left(h(V-\lambda)_+-(V-\lambda)_-\right)\left|e^{\Phi/h}u\right|^2dx

Note that, since Φ=0\Phi=0 on {V≤λ}\{V\leq\lambda\} and λ=O(h23)\lambda=O(h^{\frac{2}{3}}),

∫R+(V−λ)−∣eΦ/hu∣2dx≤Ch23∥u∥2.\int_{\mathbb{R}_+}(V-\lambda)_-\left|e^{\Phi/h}u\right|^2dx\leq Ch^{\frac{2}{3}}\|u\|^2.

Therefore,

∥h∂x(eΦ/hu)∥2+h∫R+(V−λ)∣eΦ/hu∣2dx≤Re⁡⟨(L−λ)u,e2Φ/hu⟩+Ch23∥u∥2.\left\|h\partial_x(e^{\Phi/h}u)\right\|^2+h\int_{\mathbb{R}_+}(V-\lambda)\left|e^{\Phi/h}u\right|^2dx\leq\operatorname{Re}\langle(\mathcal{L}-\lambda)u,e^{2\Phi/h}u\rangle+Ch^{\frac{2}{3}}\|u\|^2.

The operator −h22∂x2+hV−hλ-\frac{h^2}{2}\partial_x^2+hV-h\lambda is bounded below by ch43ch^{\frac{4}{3}}, with some positive constant cc. Thus, after keeping half of the kinetic energy we find

12∥h∂x(eΦ/hu)∥2+ch43∥eΦ/hu∥2≤Re⁡⟨(L−λ)u,e2Φ/hu⟩+Ch23∥u∥2.(10)\frac{1}{2}\left\|h\partial_x(e^{\Phi/h}u)\right\|^2+ch^{\frac{4}{3}}\left\|e^{\Phi/h}u\right\|^2\leq\operatorname{Re}\langle(\mathcal{L}-\lambda)u,e^{2\Phi/h}u\rangle+Ch^{\frac{2}{3}}\|u\|^2. \tag*{(10)}

Using the Young inequality we have

Re⁡⟨(L−λ)u,e2Φ/hu⟩≤12ch−43∥eΦ/h(L−λ)u∥2+c2h43∥eΦ/hu∥2,\operatorname{Re}\langle(\mathcal{L}-\lambda)u,e^{2\Phi/h}u\rangle\leq\frac{1}{2c}h^{-\frac{4}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|^2+\frac{c}{2}h^{\frac{4}{3}}\left\|e^{\Phi/h}u\right\|^2,

and we deduce from (10) the first two estimates:

∥eΦ/hu∥≤Ch−43∥eΦ/h(L−λ)u∥+Ch−13∥u∥,\left\|e^{\Phi/h}u\right\|\leq Ch^{-\frac{4}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+Ch^{-\frac{1}{3}}\|u\|,
∥h∂x(eΦ/hu)∥≤Ch−23∥eΦ/h(L−λ)u∥+Ch13∥u∥.\left\|h\partial_x(e^{\Phi/h}u)\right\|\leq Ch^{-\frac{2}{3}}\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+Ch^{\frac{1}{3}}\|u\|.

Finally, to get the second order estimates we use the first estimate and the triangle inequality,

∥eΦ/hh2∂x2u∥L2(0,M/2)≤C∥eΦ/h(L−λ)u∥+CM∥eΦ/hu∥L2(0,M/2).\left\|e^{\Phi/h}h^2\partial_x^2u\right\|_{L^2(0,M/2)}\leq C\left\|e^{\Phi/h}(\mathcal{L}-\lambda)u\right\|+C_M\left\|e^{\Phi/h}u\right\|_{L^2(0,M/2)}.

To get the result as stated in Proposition 2.1, we simply use Proposition 3.1 with M=max⁡(1,2K)M=\max(1,2K), so that S≤(1+h)ΦS\le(1+h)\Phi on (0,K)(0,K), and therefore

∥eS/hu∥L2(0,K)≤C∥eΦ/hu∥≤Ch−4/3∥eS/h(L−λ)u∥+Ch−1/3∥u∥,\lVert e^{S/h}u\rVert_{L^2(0,K)}\le C\lVert e^{\Phi/h}u\rVert\le Ch^{-4/3}\lVert e^{S/h}(\mathcal{L}-\lambda)u\rVert+Ch^{-1/3}\lVert u\rVert,

where in the last inequality we also used Φ≤S\Phi\le S.

Linear approximation inside the well

The Agmon estimates show that the groundstate ff of L\mathcal{L} is localized on an h2/3h^{2/3} neighborhood of 00. For this reason we can approximate ff by the groundstate gh0g_h^0 of the Airy operator

A=−h2∂x2+V′(0)x.\mathcal{A}=-h^{2}\partial_x^2+V'(0)x.

gh0g_h^0 can be expressed in terms of the Airy function, solution to

−∂x2Ai⁡+xAi⁡=0,lim⁡x→+∞Ai⁡(x)=0.-\partial_x^2\operatorname{Ai}+x\operatorname{Ai}=0,\qquad\lim_{x\to+\infty}\operatorname{Ai}(x)=0.

The eigenvalue associated to gh0g_h^0 is ∣z1∣V′(0)2/3h2/3|z_1|V'(0)^{2/3}h^{2/3}, where z1z_1 is the first negative zero of Ai⁡\operatorname{Ai}. Higher order Taylor expansions of VV yield good approximations of ff in L2L^2 and even C1\mathcal{C}^1 norm.

Proposition 3.2. For all k≥1k\ge1, we have

λk(h)=∣zk∣V′(0)2/3h2/3+O(h4/3),\lambda_k(h)=|z_k|V'(0)^{2/3}h^{2/3}+O(h^{4/3}),

where zkz_k is the kk-th negative zero of the Airy function. Moreover, there exists a family of functions in the Schwartz class (gj)j≥0(g_j)_{j\ge0}, all vanishing at 00, the first one being given by

g0(x)=c0Ai⁡(V′(0)1/3x+z1),g_0(x)=c_0\operatorname{Ai}(V'(0)^{1/3}x+z_1),

with c0=V′(0)1/6∣Ai⁡′(z1)∣−1c_0=V'(0)^{1/6}|\operatorname{Ai}'(z_1)|^{-1}, such that, for N≥0N\ge0, the function

ghN(x)=h−1/3∑j=0Nh2j/3gj(h−2/3x),g_h^N(x)=h^{-1/3}\sum_{j=0}^{N}h^{2j/3}g_j(h^{-2/3}x),

satisfies

∥f−χghN∥L2(R+)=O(h(2N+2)/3)and∥f−χghN∥C1(R+)=O(h2N/3−1/2).\lVert f-\chi g_h^N\rVert_{L^2(\mathbb{R}_+)}=O(h^{(2N+2)/3})\quad\text{and}\quad\lVert f-\chi g_h^N\rVert_{\mathcal{C}^1(\mathbb{R}_+)}=O(h^{2N/3-1/2}).

for any cutoff function χ\chi supported and equal to 11 near 00. For some such cutoff, set gh=χghNg_h=\chi g_h^N.

Proof. The proof follows classical arguments. Let us recall the main steps. We start by constructing a rather naive quasimode. By rescaling, the operator L\mathcal{L} is unitarily equivalent to Lresc=−h2/3∂y2+V(h2/3y)\mathcal{L}^{\mathrm{resc}}=-h^{2/3}\partial_y^2+V(h^{2/3}y) which we can formally Taylor expand as

Lresc∼h2/3∑j≥0h2j/3Lj,L0=−∂y2+V′(0)y,Lj=V(j+1)(0)(j+1)!yj+1 (j≥1),\mathcal{L}^{\mathrm{resc}}\sim h^{2/3}\sum_{j\ge0}h^{2j/3}\mathcal{L}_j,\qquad\mathcal{L}_0=-\partial_y^2+V'(0)y,\qquad\mathcal{L}_j=\frac{V^{(j+1)}(0)}{(j+1)!}y^{j+1}\ (j\ge1),

This suggests looking for a quasi-eigenvalue μ=h23∑j=0Nh2j3μj\mu= h^{\frac{2}{3}}\sum_{j=0}^{N}h^{\frac{2j}{3}}\mu_j and for a quasimode in the form

g(y)=∑j=0Nh2j3gj(y).g(y)=\sum_{j=0}^{N}h^{\frac{2j}{3}}g_j(y).

Solving formally

(Lresc−μ)g=0(\mathcal{L}^{\mathrm{resc}}-\mu)g=0

gives a sequence of equations. The first one is

(L0−μ0)g0=0,(\mathcal{L}_0-\mu_0)g_0=0,

which leads to the choice μ0=∣z1∣V′(0)23\mu_0=|z_1|V'(0)^{\frac{2}{3}} (the groundstate eigenvalue of the Airy operator L0\mathcal{L}_0) and to

g0(y)=c0Ai⁡(V′(0)13y+z1),g_0(y)=c_0\operatorname{Ai}\left(V'(0)^{\frac{1}{3}}y+z_1\right),

the corresponding normalized eigenfunction.

Then, the next (μj,gj)(\mu_j,g_j) are determined by an induction procedure requiring us to solve an equation in the form

(L0−μ0)gj=μjg0+Rj(11)(\mathcal{L}_0-\mu_0)g_j=\mu_jg_0+R_j \tag*{(11)}

where RjR_j is known and belongs to the Schwartz class. Since the left-hand side must be orthogonal to g0g_0, we need to choose μj\mu_j such that the right-hand side also is, i.e. μj=−⟨Rj,g0⟩\mu_j=-\langle R_j,g_0\rangle. Then, for such right-hand side, there is a unique solution gjg_j to (11) that

∥∑ℓ=0jh2ℓ3gℓ∥2=1+O(h2(j+1)3).\left\|\sum_{\ell=0}^{j}h^{\frac{2\ell}{3}}g_\ell\right\|^2=1+O\left(h^{\frac{2(j+1)}{3}}\right).

This determines the sequence μj\mu_j and gjg_j. The gjg_j are Schwartz class, this is shown by variation of parameters. For any C>0C>0, using the Taylor formula with remainder to order N+2N+2 for VV, and the fact that the gjg_j's are Schwartz class, we get

(Lresc−μ)∑j=0Nh2j3gj=OL2([0,Ch−2/3])(h2N+43).(\mathcal{L}^{\mathrm{resc}}-\mu)\sum_{j=0}^{N}h^{\frac{2j}{3}}g_j=O_{L^2([0,Ch^{-2/3}])}\left(h^{\frac{2N+4}{3}}\right).

(We have to restrict to C<∞C<\infty because we have not made any assumption on the size of the derivatives of VV at infinity).

Rescaling back, and using the fact [L,χ][\mathcal{L},\chi] is supported where ghNg_h^N is O(h∞)O(h^\infty), we deduce that

∥(L−μ)(χghN)∥L2(R+)=O(h2N+43).(12)\|(\mathcal{L}-\mu)(\chi g_h^N)\|_{L^2(\mathbb{R}_+)}=O\left(h^{\frac{2N+4}{3}}\right). \tag*{(12)}

Thanks to the spectral theorem, this shows that dist⁡(sp⁡(L),μ)≤Ch2N+43\operatorname{dist}(\operatorname{sp}(\mathcal{L}),\mu)\leq Ch^{\frac{2N+4}{3}}. Playing the same game by choosing other eigenfunctions of the Airy operator associated with the next zeroes, we build a sequence of quasi-eigenvalues, all of order O(h23)O(h^{\frac{2}{3}}).

By means of Proposition 3.1, it can be established that there exists c>0c>0 such that, for all F∈E:=ker⁡(L−λ1(h))⊕ker⁡(L−λ2(h))F\in\mathcal{E}:=\ker(\mathcal{L}-\lambda_1(h))\oplus\ker(\mathcal{L}-\lambda_2(h)),

∫R+ecxh2/3∣F(x)∣2dx≤C∥F∥2.(13)\int_{\mathbb{R}_+}e^{\frac{cx}{h^{2/3}}}|F(x)|^2\mathrm{d}x\le C\|F\|^2. \tag*{(13)}

To prove this, one can proceed directly with the usual strategy for Agmon estimate. Considering Q\mathcal{Q} the quadratic form associated with L\mathcal{L}, we have

Q(F)=∫R+(h2∣F′(x)∣2+V(x)∣F(x)∣2)dx≤λ2(h)∥F∥2.\mathcal{Q}(F)=\int_{\mathbb{R}_+}\left(h^2|F'(x)|^2+V(x)|F(x)|^2\right)\mathrm{d}x\le\lambda_2(h)\|F\|^2.

and then, by using the Taylor expansion of VV and estimating the remainders thanks to (13),

Q(F)≥∫R+(h2∣F′(x)∣2+V′(0)x∣F(x)∣2)dx−Ch43∥F∥2.\mathcal{Q}(F)\ge\int_{\mathbb{R}_+}\left(h^2|F'(x)|^2+V'(0)x|F(x)|^2\right)\mathrm{d}x-Ch^{\frac{4}{3}}\|F\|^2.

By the min-max theorem, this shows that

λ2(h)≥∣z2∣∣V′(0)∣23h23−Ch43,\lambda_2(h)\ge|z_2||V'(0)|^{\frac{2}{3}}h^{\frac{2}{3}}-Ch^{\frac{4}{3}},

and therefore ∣λ1(h)−μ∣=dist⁡(sp⁡(L),μ)=O(h2N+43)|\lambda_1(h)-\mu|=\operatorname{dist}(\operatorname{sp}(\mathcal{L}),\mu)=O\left(h^{\frac{2N+4}{3}}\right). (The same argument gives similar estimation for the next eigenvalues that are thus all simple). With the spectral theorem and (12), we deduce that

∥f−gh∥L2(R+)=O(h2N+23).\|f-g_h\|_{L^2(\mathbb{R}_+)}=O\left(h^{\frac{2N+2}{3}}\right).

To obtain a uniform estimate, we can use that for u∈H2∩H01u\in H^2\cap H_0^1 supported in [0,K][0,K],

∥(L−μ)u∥2=∥h2u′′∥2+∥(V−μ)u∥2+2h2Re⁡⟨(V−μ)u′,u′⟩−h2⟨V′′u,u⟩.\|(\mathcal{L}-\mu)u\|^2=\|h^2u''\|^2+\|(V-\mu)u\|^2+2h^2\operatorname{Re}\langle(V-\mu)u',u'\rangle-h^2\langle V''u,u\rangle.
h2∥u′′∥2≤⟨(L−μ)u,u⟩+μ∥u∥2.h^2\|u''\|^2\le\langle(\mathcal{L}-\mu)u,u\rangle+\mu\|u\|^2.

It follows that

∥h2∂x2u∥L22≤Ch−2/3∥(L−μ)u∥L22+Ch2/3∥u∥L22.\|h^2\partial_x^2u\|_{L^2}^2\le Ch^{-2/3}\|(\mathcal{L}-\mu)u\|_{L^2}^2+Ch^{2/3}\|u\|_{L^2}^2.

with constants depending on KK (actually ∥V′′∥L∞(0,K)\|V''\|_{L^\infty(0,K)}). Inserting a cutoff against f−ghf-g_h, and using that ff is O(h∞)O(h^\infty) for x>1x>1, we deduce that

∥h2∂x2(f−gh)∥L2(0,K)≤Ch2N+33.\|h^2\partial_x^2(f-g_h)\|_{L^2(0,K)}\le Ch^{\frac{2N+3}{3}}.

For large values of x>Kx>K, we use the fact that V>λV>\lambda and f>0f>0, so that

f′′=h−2(V−λ)f>0.f''=h^{-2}(V-\lambda)f>0.

This implies that f′f' is non-decreasing, but it goes to 00, so f′(K)≤f′(x)≤0f'(K)\le f'(x)\le0, and 0≤f(x)≤f(K)0\le f(x)\le f(K). □

Since Ai is exponentially decaying at infinity, the approximation of ff given by Proposition 3.2 is only good on an h2/3h^{2/3} neighborhood of 00. Still, this allows to compute the asymptotic value of ff at the boundary of the well, i.e. the point xλx_\lambda such that V(xλ)=λ1V(x_\lambda)=\lambda_1.

Corollary 3.3. We have

f(xλ)∼h→0c0h−13Ai⁡(0).f(x_\lambda) \sim_{h\to0} c_0h^{-\frac{1}{3}}\operatorname{Ai}(0).

Proof. First note that xλx_\lambda must be of order h23h^{\frac{2}{3}}. By a Taylor approximation we deduce that

λ1=V(xλ)=V′(0)xλ+O(h43),\lambda_1=V(x_\lambda)=V'(0)x_\lambda+O(h^{\frac{4}{3}}),

and since λ1=∣z1∣V′(0)23h23+O(h43)\lambda_1=|z_1|V'(0)^{\frac{2}{3}}h^{\frac{2}{3}}+O(h^{\frac{4}{3}}) we obtain

xλ=∣z1∣V′(0)−13h23+O(h43).(14)x_\lambda=|z_1|V'(0)^{-\frac{1}{3}}h^{\frac{2}{3}}+O(h^{\frac{4}{3}}). \tag*{(14)}

Using the uniform estimates in Proposition 3.2, we find that

f(xλ)=h−13g0(h−23xλ)+O(h13)f(x_\lambda)=h^{-\frac{1}{3}}g_0(h^{-\frac{2}{3}}x_\lambda)+O(h^{\frac{1}{3}})
=c0h−13Ai⁡(V′(0)13h−23xλ+z1)+O(h13),=c_0h^{-\frac{1}{3}}\operatorname{Ai}(V'(0)^{\frac{1}{3}}h^{-\frac{2}{3}}x_\lambda+z_1)+O(h^{\frac{1}{3}}),

and the result follows using (14). □

A construction of quasimodes

We now construct a quasimode that will be exponentially good. It will be shown later to be a correct approximation of the true groundstate ff at large and intermediate distance from the well.

In the following, we denote by λ=λ1\lambda=\lambda_1 the first eigenvalue of L\mathcal{L}. We recall that xλx_\lambda denotes the boundary of the well, i.e. the solution to V(xλ)=λV(x_\lambda)=\lambda. It will be convenient to extend VV to a smooth function on (−ε,∞)(-\varepsilon,\infty), for some small ε>0\varepsilon>0.

We introduce the weight

φ0(x)={(32∫xλxV−λ)23if x≥xλ,−(32∫xxλλ−V)23if x≤xλ.(15)\varphi_0(x)= \begin{cases} \left(\frac{3}{2}\int_{x_\lambda}^{x}\sqrt{V-\lambda}\right)^{\frac{2}{3}} & \text{if }x\geq x_\lambda,\\ -\left(\frac{3}{2}\int_{x}^{x_\lambda}\sqrt{\lambda-V}\right)^{\frac{2}{3}} & \text{if }x\leq x_\lambda. \end{cases} \tag*{(15)}

In fact, φ0\varphi_0 defines a smooth function.

Lemma 3.4. Let M≥1M\geq1. There exists ε>0\varepsilon>0 such that the function φ0\varphi_0 is smooth and increasing on (−ε,M)(-\varepsilon,M). Moreover, there exist c>0c>0 and h0>0h_0>0 such that, for all h∈(0,h0)h\in(0,h_0) and all x∈(−ε,M)x\in(-\varepsilon,M), we have φ0′(x)≥c>0\varphi_0'(x)\geq c>0. In addition, we have

V(x)−λ=φ0(x)φ0′(x)2.(16)V(x)-\lambda=\varphi_0(x)\varphi_0'(x)^2. \tag*{(16)}

Proof. The regularity problem appears near x=0x=0. Thanks to the Taylor formula, we get

V(x)−V(xλ)=(x−xλ)∫01V′(xλ+t(x−xλ)) dt.V(x)-V(x_\lambda)=(x-x_\lambda)\int_0^1 V'(x_\lambda+t(x-x_\lambda))\,\mathrm{d}t.

For hh small enough and xx in a neighborhood of 00, we can write

∫01V′(xλ+t(x−xλ)) dt=R(x−xλ)2,(17)\int_0^1 V'(x_\lambda+t(x-x_\lambda))\,\mathrm{d}t=R(x-x_\lambda)^2, \tag*{(17)}

for some smooth function RR such that R≥c0>0R \ge c_0 > 0. Then, for x≥xλx \ge x_\lambda, we have

∫xλxV(s)−λ ds=∫0x−xλθR(θ) dθ=23(x−xλ)32∫01R(ν23(x−xλ)) dν\int_{x_\lambda}^{x} \sqrt{V(s)-\lambda}\,\mathrm{d}s = \int_{0}^{x-x_\lambda} \sqrt{\theta}R(\theta)\,\mathrm{d}\theta= \frac{2}{3}(x-x_\lambda)^{\frac{3}{2}}\int_{0}^{1}R\bigl(\nu^{\frac{2}{3}}(x-x_\lambda)\bigr)\,\mathrm{d}\nu

so that

φ0(x)=(x−xλ)(∫01R(ν23(x−xλ)) dν)23.\varphi_0(x)=(x-x_\lambda)\left(\int_{0}^{1}R\bigl(\nu^{\frac{2}{3}}(x-x_\lambda)\bigr)\,\mathrm{d}\nu\right)^{\frac{2}{3}}.

We can check that this formula also holds for x≤xλx \le x_\lambda. We see that φ0\varphi_0 is a smooth function of xx, since R≥c0>0R \ge c_0 > 0 everywhere. Finally we can compute φ0′\varphi_0' using (15) and we immediately see that φ0\varphi_0 is increasing and satisfies (16). □

The next proposition gives the construction of our quasimode, which is a combination of the Airy function (well-suited for the shape of the well) and an iterative perturbation φ\varphi of φ0\varphi_0.

Proposition 3.5. Let K≥1K \ge1. There exist ε0,h0>0\varepsilon_0,h_0>0 and, for h∈(0,h0)h\in(0,h_0), smooth increasing functions φ1,…,φN,…\varphi_1,\ldots,\varphi_N,\ldots defined on φ0((−2ε0,K+ε0))\varphi_0((-2\varepsilon_0,K+\varepsilon_0)) such that the following holds. Uniformly and for all j≥1j\ge1 we have

φj(x)=x+OC∞(h2j).\varphi_j(x)=x+O_{C^\infty}(h^{2j}).

Thus, for N>0N>0 and h>0h>0 small enough, φ=φN∘⋯∘φ1∘φ0\varphi=\varphi_N\circ\cdots\circ\varphi_1\circ\varphi_0 is well defined and we let

fN(x)=1φ′(x)Ai⁡(h−23φ(x)).f_N(x)=\frac{1}{\sqrt{\varphi'(x)}}\operatorname{Ai}\bigl(h^{-\frac{2}{3}}\varphi(x)\bigr).

Then, uniformly on (−ε0,K)(-\varepsilon_0,K) we have

(L−λ)fN=O(h2N+2)fN.(\mathcal{L}-\lambda)f_N=O(h^{2N+2})f_N.

Proof. To construct the announced quasimode, we first look for it in the form

u(x)=α(x)v(φ(x)).u(x)=\alpha(x)v(\varphi(x)).

The second derivative of such a function is given by

u′′(x)=α′′v+2α′φ′v′+αφ′′v′+αφ′2v′′u''(x)=\alpha''v+2\alpha'\varphi'v'+\alpha\varphi''v'+\alpha\varphi'^{2}v''

and we choose

α(x)=1φ′(x),\alpha(x)=\frac{1}{\sqrt{\varphi'(x)}},

in order to kill first order derivatives in vv. Letting y=φ(x)y=\varphi(x), a computation gives

(L−λ)u=α(x)φ′(x)2(−h2v′′(y)+V(φ−1(y))−λφ′(φ−1(y))2v(y)−h2α′′α[φ′(φ−1(y))]2v(y)).(\mathcal{L}-\lambda)u=\alpha(x)\varphi'(x)^2\left(-h^2v''(y)+\frac{V(\varphi^{-1}(y))-\lambda}{\varphi'(\varphi^{-1}(y))^2}v(y)-h^2\frac{\alpha''}{\alpha[\varphi'(\varphi^{-1}(y))]^2}v(y)\right).

By choosing φ(x)=φ0(x)\varphi(x)=\varphi_0(x) for all x∈(−ε,M)x\in(-\varepsilon,M), we get

V(φ−1(y))−λφ′(φ−1(y))2=y,\frac{V(\varphi^{-1}(y))-\lambda}{\varphi'(\varphi^{-1}(y))^2}=y,

for all y∈(φ0(−ε),φ0(M))y \in(\varphi_0(-\varepsilon),\varphi_0(M)). We let the remainder term be

r0=[(φ0′)−3/2((φ0′)−1/2)′′]∘φ0−1,r_0=\left[(\varphi_0')^{-3/2}\left((\varphi_0')^{-1/2}\right)''\right]\circ\varphi_0^{-1},

so that

(L−λ)u=(φ0′)32(−h2∂y2v+yv−h2r0v).(\mathcal{L}-\lambda)u=(\varphi_0')^{\frac{3}{2}}\left(-h^2\partial_y^2v+yv-h^2r_0v\right).

If r0r_0 were 0 we could choose v=Ai⁡(h−2/3y)v=\operatorname{Ai}(h^{-2/3}y). Instead we will recursively reduce the size of the remainder term. We look for vv in the form

v(y)=1φ1′(y)w(φ1(y)).v(y)=\frac{1}{\sqrt{\varphi_1'(y)}}w(\varphi_1(y)).

The map y↦y−h2r0(y)=W1(y)y\mapsto y-h^2r_0(y)=W_1(y) is increasing when hh is small enough. There exists a unique yλy_\lambda such that yλ−h2r0(yλ)=0y_\lambda-h^2r_0(y_\lambda)=0. We let, for all y∈(φ0(−ε),φ0(M))y\in(\varphi_0(-\varepsilon),\varphi_0(M)),

φ1(y)={(32∫yλyW1(s) ds)23if y≥yλ,−(32∫yyλ−W1(s) ds)23if y≤yλ.\varphi_1(y)= \begin{cases} \left(\dfrac{3}{2}\int_{y_\lambda}^{y}\sqrt{W_1(s)}\,\mathrm{d}s\right)^{\frac{2}{3}} & \text{if } y\ge y_\lambda,\\ -\left(\dfrac{3}{2}\int_{y}^{y_\lambda}\sqrt{-W_1(s)}\,\mathrm{d}s\right)^{\frac{2}{3}} & \text{if } y\le y_\lambda. \end{cases}

The function φ1\varphi_1 is smooth, increasing, and we have, uniformly in y∈(φ0(−ε),φ0(M))y\in(\varphi_0(-\varepsilon),\varphi_0(M)),

φ1′(y)=1+O(h2).\varphi_1'(y)=1+O(h^2).

In the same way as previously, we get

(−h2∂y2+y−h2r0(y))v=φ1′(y)32(−h2∂z2+z−h2r1(z))w,(-h^2\partial_y^2+y-h^2r_0(y))v=\varphi_1'(y)^{\frac{3}{2}}(-h^2\partial_z^2+z-h^2r_1(z))w,

with z=φ1(y)z=\varphi_1(y) and

r1=[(φ1′)−3/2((φ1′)−1/2)′′]∘φ1−1=O(h2).r_1=\left[(\varphi_1')^{-3/2}\left((\varphi_1')^{-1/2}\right)''\right]\circ\varphi_1^{-1}=O(h^2).

This procedure can be continued by induction: at step NN we obtain φ=φN∘…∘φ1∘φ0\varphi=\varphi_N\circ\ldots\circ\varphi_1\circ\varphi_0, and

(L−λ)u=(φ′)32(−h2∂z2+z+O(h2N+2))w∘φ.(\mathcal{L}-\lambda)u=(\varphi')^{\frac{3}{2}}\left(-h^2\partial_z^2+z+O(h^{2N+2})\right)w\circ\varphi.

We end the induction by taking w=Ai⁡(h−2/3z)w=\operatorname{Ai}(h^{-2/3}z), which gives

(L−λ)u=O(h2N+2)u.(\mathcal{L}-\lambda)u=O(h^{2N+2})u.

Approximation of the groundstate in L2L^2 norm

Proposition 3.5 suggests considering the function given, for all K>x≥xλK>x\ge x_\lambda, by

fNapp(x)=αhφ′(x)Ai⁡(h−23φ(x)),(18)f_N^{\mathrm{app}}(x)=\frac{\alpha_h}{\sqrt{\varphi'(x)}}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x)), \tag*{(18)}

where φ=φN∘…∘φ1∘φ0\varphi=\varphi_N\circ\ldots\circ\varphi_1\circ\varphi_0, and with normalization constant

αh=f(xλ)φ′(xλ)Ai⁡(h−23φ(xλ)).\alpha_h=f(x_\lambda)\frac{\sqrt{\varphi'(x_\lambda)}}{\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x_\lambda))}.

The constant αh\alpha_h is chosen so that fNapp(xλ)=f(xλ)f_N^{\mathrm{app}}(x_\lambda)=f(x_\lambda), and we find that

αh∼c0h−13V′(0)16.\alpha_h\sim c_0h^{-\frac{1}{3}}V'(0)^{\frac{1}{6}}.

However, the derivatives do not match at x=xλx=x_\lambda. In fact, (fNapp)′(xλ)(f_N^{\mathrm{app}})'(x_\lambda) almost equals f′(xλ)f'(x_\lambda).

Lemma 3.6. We have

(f−fNapp)′(xλ)=O(h2N+13).\left(f-f_N^{\mathrm{app}}\right)'(x_\lambda)=O\left(h^{2N+\frac{1}{3}}\right).

Proof. We consider the Wronskian

W=f(fNapp)′−f′fNapp\mathcal{W}=f\left(f_N^{\mathrm{app}}\right)'-f'f_N^{\mathrm{app}}

which satisfies

−h2W′(x)=f(x)(L−λ)fNapp.-h^2\mathcal{W}'(x)=f(x)(\mathcal{L}-\lambda)f_N^{\mathrm{app}}.

We get

f(xλ)(fNapp−f′)(xλ)=W(xλ)=W(K)+∫xλKh−2f(L−λ)fNapp.f(x_\lambda)\left(f_N^{\mathrm{app}}-f'\right)(x_\lambda)=\mathcal{W}(x_\lambda)=\mathcal{W}(K)+\int_{x_\lambda}^{K}h^{-2}f(\mathcal{L}-\lambda)f_N^{\mathrm{app}}.

The conclusion follows by using that W(K)=O(h∞)\mathcal{W}(K)=O(h^\infty) and by recalling Corollary 3.3.

We make a small perturbation of fNappf_N^{\mathrm{app}} in order to correct its derivative at x=xλx=x_\lambda. In what follows, we consider a smooth non increasing function χ\chi such that 0≤χ≤10\leq\chi\leq1, equal to 11 on [0,3M4][0,\frac{3M}{4}] and 00 on (M,+∞)(M,+\infty). Let us add a small perturbation of fNappf_N^{\mathrm{app}} by considering

f~Napp=fNapp+(x−xλ)βhχh(x),(19)\widetilde{f}_N^{\mathrm{app}}=f_N^{\mathrm{app}}+(x-x_\lambda)\beta_h\chi_h(x), \tag*{(19)}

where βh=(f−fNapp)′(xλ)\beta_h=(f-f_N^{\mathrm{app}})'(x_\lambda) and χh(x)=χ(h−23x)\chi_h(x)=\chi(h^{-\frac{2}{3}}x). Provided MM is large enough, χh(xλ)=1\chi_h(x_\lambda)=1, and this function has derivative (f~Napp)′(xλ)=f′(xλ)(\widetilde{f}_N^{\mathrm{app}})'(x_\lambda)=f'(x_\lambda) at the boundary of the well.

Proposition 3.7. We let

ψ(x)={f(x)if x∈[0,xλ],χ(x)f~Napp(x)if x≥xλ,\psi(x)= \begin{cases} f(x) & \text{if }x\in[0,x_\lambda],\\ \chi(x)\widetilde{f}_N^{\mathrm{app}}(x) & \text{if }x\geq x_\lambda, \end{cases}

which is of class C1C^1 and belongs to the domain of L\mathcal{L} by construction. It satisfies in L2L^2

(L−λ)ψ=O(h2N+53)∥ψ∥.(\mathcal{L}-\lambda)\psi=O\left(h^{2N+\frac{5}{3}}\right)\|\psi\|.

Moreover,

∥ψ−f∥L2(R+)=O(h2N+1).\|\psi-f\|_{L^2(\mathbb{R}_+)}=O\left(h^{2N+1}\right).

Proof. We estimate the two terms separately in (19) to find

∥(L−λ)ψ∥=O(h2N+2αh)+O(h43βh)=O(h2N+53).\lVert(\mathcal{L}-\lambda)\psi\rVert=\mathcal{O}(h^{2N+2}\alpha_h)+\mathcal{O}(h^{\frac{4}{3}}\beta_h)=\mathcal{O}(h^{2N+\frac{5}{3}}).

Using the same arguments as in the proof of Proposition 3.2, the spectral theorem gives that

∥ψ−∥ψ∥f∥L2(R+)=O(h2N+1).\lVert\psi-\lVert\psi\rVert f\rVert_{L^2(\mathbb{R}_+)}=\mathcal{O}(h^{2N+1}).

Then, we notice that

∣1−∥ψ∥∣∥f∥L2(0,xλ)=∥f−∥ψ∥f∥L2(0,xλ)≤∥ψ−∥ψ∥f∥L2(R+).\left|1-\lVert\psi\rVert\right|\lVert f\rVert_{L^2(0,x_\lambda)}=\lVert f-\lVert\psi\rVert f\rVert_{L^2(0,x_\lambda)}\leq\lVert\psi-\lVert\psi\rVert f\rVert_{L^2(\mathbb{R}_+)}.

We use Proposition 3.2 to see that ∥f∥L2(0,xλ)≥c>0\lVert f\rVert_{L^2(0,x_\lambda)}\geq c>0. Thus,

∣1−∥ψ∥∣=O(h2N+1).\left|1-\lVert\psi\rVert\right|=\mathcal{O}(h^{2N+1}).

Approximation of the groundstate in exponentially weighted norm

In this section we improve the estimates on our quasimode ψ\psi, defined in Proposition 3.7. We show that ψ\psi is a good approximation of the true groundstate ff in exponentially weighted norm.

Lemma 3.8. We have

eΦ/h(L−λ)ψ=O(h2N+53),e^{\Phi/h}(\mathcal{L}-\lambda)\psi=\mathcal{O}(h^{2N+\frac{5}{3}}),

where Φ\Phi is defined in Proposition 3.1.

Proof. We only need to consider the region x≥xλx\geq x_\lambda, where ψ=χf~Napp\psi=\chi\widetilde{f}^{\mathrm{app}}_N. Then,

eΦ/h(L−λ)(χf~Napp)=eΦ/h[−h2∂x2,χ]f~Napp+eΦ/hχ(L−λ)f~Napp=eΦ/h[−h2∂x2,χ]fNapp+eΦ/hχ(L−λ)(fNapp+βh(x−xλ)χh),\begin{aligned} e^{\Phi/h}(\mathcal{L}-\lambda)(\chi\widetilde{f}^{\mathrm{app}}_N)=e^{\Phi/h}[-h^2\partial_x^2,\chi]\widetilde{f}^{\mathrm{app}}_N+e^{\Phi/h}\chi(\mathcal{L}-\lambda)\widetilde{f}^{\mathrm{app}}_N \\ &=e^{\Phi/h}[-h^2\partial_x^2,\chi]f^{\mathrm{app}}_N+e^{\Phi/h}\chi(\mathcal{L}-\lambda)(f^{\mathrm{app}}_N+\beta_h(x-x_\lambda)\chi_h), \end{aligned}

where the first ‘tilde’ was removed by support considerations. The last term can be estimated using that Φ≤Ch\Phi\leq Ch on the support of χh\chi_h.

eΦ/h(L−λ)(χf~Napp)=eΦ/h[−h2∂x2,χ]fNapp+eΦ/hχ(L−λ)fNapp+O(h43∣βh∣).e^{\Phi/h}(\mathcal{L}-\lambda)(\chi\widetilde{f}^{\mathrm{app}}_N)=e^{\Phi/h}[-h^2\partial_x^2,\chi]f^{\mathrm{app}}_N+e^{\Phi/h}\chi(\mathcal{L}-\lambda)f^{\mathrm{app}}_N+\mathcal{O}(h^{\frac{4}{3}}|\beta_h|).

Let us notice that

∥eΦ/hχ(L−λ)fNapp∥≤Ch2N+2∥χeΦ/hfNapp∥≤Ch2N+53∥χeΦ/hAi⁡(h−23φ(x))∥,\lVert e^{\Phi/h}\chi(\mathcal{L}-\lambda)f^{\mathrm{app}}_N\rVert\leq Ch^{2N+2}\lVert\chi e^{\Phi/h}f^{\mathrm{app}}_N\rVert\leq Ch^{2N+\frac{5}{3}}\lVert\chi e^{\Phi/h}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x))\rVert,

where we used Proposition 3.5 and Corollary 3.3. Then bound the Airy function, for z∈Rz\in\mathbb{R},

∣Ai⁡(z)∣≤Ce−23z32,|\operatorname{Ai}(z)|\leq Ce^{-\frac{2}{3}z^{\frac{3}{2}}},

to get that

∥χeΦ/hAi⁡(h−23φ(x))∥≤C∥χeΦ/he−23φ32/h∥≤C∥χeS/he−23φ32/h∥\lVert\chi e^{\Phi/h}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x))\rVert\leq C\lVert\chi e^{\Phi/h}e^{-\frac{2}{3}\varphi^{\frac{3}{2}}/h}\rVert\leq C\lVert\chi e^{S/h}e^{-\frac{2}{3}\varphi^{\frac{3}{2}}/h}\rVert

where we used that Φ≤S\Phi\le S. Then, by construction we have, for all K>x≥xλK>x\ge x_\lambda,

φ(x)=(32∫xλxV−λ)23+O(h2),\varphi(x)=\left(\frac{3}{2}\int_{x_\lambda}^{x}\sqrt{V-\lambda}\right)^{\frac{2}{3}}+O(h^2),

and by splitting the integral into two parts x≥xλx\ge x_\lambda and x≤xλx\le x_\lambda, we deduce

∥χeΦ/hAi⁡(h−23φ(x))∥≤C.\left\|\chi e^{\Phi/h}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x))\right\|\le C.

There remains to estimate the commutator

eΦ/h[−h2∂x2,χ]fNapp=−h2χ′′eΦ/hfNapp−2h2χ′eΦ/h(fNapp)′,e^{\Phi/h}[-h^2\partial_x^2,\chi]f_N^{\mathrm{app}}=-h^2\chi^{\prime\prime}e^{\Phi/h}f_N^{\mathrm{app}}-2h^2\chi^{\prime}e^{\Phi/h}(f_N^{\mathrm{app}})',

which is supported in (3M4,M)\left(\frac{3M}{4},M\right). The first term can be estimated by

∥−h2χ′′eΦ/hfNapp∥≤Ch2∥eΦ/hAi⁡(h−23φ)∥L2(3M4,M)≤Ch2∥eΦ/he−23hφ32∥L2(3M4,M)≤Ch2∥e(S(M/2)−S)/h∥L2(3M4,M).\begin{aligned} \left\|-h^2\chi^{\prime\prime}e^{\Phi/h}f_N^{\mathrm{app}}\right\|\le Ch^2\left\|e^{\Phi/h}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi)\right\|_{L^2(\frac{3M}{4},M)} \\ &\le Ch^2\left\|e^{\Phi/h}e^{-\frac{2}{3h}\varphi^{\frac{3}{2}}}\right\|_{L^2(\frac{3M}{4},M)} \\ &\le Ch^2\left\|e^{(S(M/2)-S)/h}\right\|_{L^2(\frac{3M}{4},M)}. \end{aligned}

Thus,

∥−h2χ′′eΦ/hfNapp∥≤Ch2e(S(M/2)−S(3M/4))/h=O(h∞),\left\|-h^2\chi^{\prime\prime}e^{\Phi/h}f_N^{\mathrm{app}}\right\|\le Ch^2e^{(S(M/2)-S(3M/4))/h}=O(h^\infty),

and we can deal with the other term in a similar way. We deduce that

eΦ/h(L−λ)(χf~Napp)=O(h2N+53).e^{\Phi/h}(\mathcal{L}-\lambda)(\chi\widetilde{f}_N^{\mathrm{app}})=O(h^{2N+\frac{5}{3}}).

Since ψ\psi is C1C^1 and (L−λ)f=0(\mathcal{L}-\lambda)f=0, the result follows. □

Proof of Theorem 1.2

We apply the Agmon estimates from Proposition 3.1 to the function u=ψ−fu=\psi-f. We use Lemma 3.8 and Proposition 3.7 to estimate the right-hand-side.

∥eΦ/hu∥+∥eΦ/hh2∂x2u∥L2(0,M/2)=O(h2N+13).(20)\left\|e^{\Phi/h}u\right\|+\left\|e^{\Phi/h}h^2\partial_x^2u\right\|_{L^2(0,M/2)}=O\left(h^{2N+\frac{1}{3}}\right). \tag*{(20)}

and

∥h∂x(eΦ/hu)∥=O(h2N+1).(21)\left\|h\partial_x(e^{\Phi/h}u)\right\|=O(h^{2N+1}). \tag*{(21)}

Then, by the Agmon inequality we can estimate the supremum norm,

∥eΦ/h(ψ−f)∥∞≤C∥eΦ/h(ψ−f)∥L212∥eΦ/h(ψ−f)∥H112=O(h2N).\left\|e^{\Phi/h}(\psi-f)\right\|_\infty\le C\left\|e^{\Phi/h}(\psi-f)\right\|_{L^2}^{\frac{1}{2}}\left\|e^{\Phi/h}(\psi-f)\right\|_{H^1}^{\frac{1}{2}}=O(h^{2N}).

Therefore, given a fixed x>0x>0 we have f(x)=fNapp(x)+O(h2N)e−Φ(x)/hf(x)=f_N^{\mathrm{app}}(x)+O(h^{2N})e^{-\Phi(x)/h}, with

fNapp(x)=chφ′(x)Ai⁡(h−23φ(x))∼c0V′(0)162πh−16φ′(x)φ(x)14e−23hφ(x)32,f_N^{\mathrm{app}}(x)=\frac{c_h}{\sqrt{\varphi'(x)}}\operatorname{Ai}(h^{-\frac{2}{3}}\varphi(x))\sim\frac{c_0V'(0)^{\frac{1}{6}}}{2\sqrt{\pi}}\frac{h^{-\frac{1}{6}}}{\sqrt{\varphi'(x)}\varphi(x)^{\frac{1}{4}}}e^{-\frac{2}{3h}\varphi(x)^{\frac{3}{2}}},

so that

fNapp(x)∼c0V′(0)162πh−16φ0′(x)φ0(x)14e−23hφ0(x)32=c0V′(0)162πh−16φ0′(x)φ0(x)14e−S(x)/h.f_N^{\mathrm{app}}(x)\sim\frac{c_0V'(0)^{\frac{1}{6}}}{2\sqrt{\pi}}\frac{h^{-\frac{1}{6}}}{\sqrt{\varphi_0'(x)}\varphi_0(x)^{\frac{1}{4}}}e^{-\frac{2}{3h}\varphi_0(x)^{\frac{3}{2}}}=\frac{c_0V'(0)^{\frac{1}{6}}}{2\sqrt{\pi}}\frac{h^{-\frac{1}{6}}}{\sqrt{\varphi_0'(x)}\varphi_0(x)^{\frac{1}{4}}}e^{-S(x)/h}.

This finishes the proof of Theorem 1.2 since

φ0(x)14=(32∫xλxV−λ)16,φ0′(x)=(32∫xλxV−λ)−16(V(x)−λ)14.\varphi_0(x)^{\frac{1}{4}}=\left(\frac{3}{2}\int_{x_\lambda}^{x}\sqrt{V-\lambda}\right)^{\frac{1}{6}},\qquad\sqrt{\varphi_0'(x)}=\left(\frac{3}{2}\int_{x_\lambda}^{x}\sqrt{V-\lambda}\right)^{-\frac{1}{6}}(V(x)-\lambda)^{\frac{1}{4}}.

Note that N=0N=0 is enough in the quasimode construction, for the error terms to be smaller than the leading order.

Remark 3.9. Note that we can similarly use (20) to estimate the derivative of ff, on the interval (0,M/2)(0,M/2):

∥eΦ/h(ψ′−f′)∥∞≤C∥eΦ/h(ψ′−f′)∥L212∥eΦ/h(ψ′−f′)∥H112=O(h2N−1),\left\|e^{\Phi/h}(\psi'-f')\right\|_{\infty}\leq C\left\|e^{\Phi/h}(\psi'-f')\right\|_{L^2}^{\frac{1}{2}}\left\|e^{\Phi/h}(\psi'-f')\right\|_{H^1}^{\frac{1}{2}}=O(h^{2N-1}),

using the control on the H2H^2 norm.

Applications to tunneling

Double linear wells

Once we have proven sharp exponential decay estimates on the single-well eigenfunctions, together with pointwise asymptotic formulas, we have all the necessary ingredients to obtain tunneling formulas for the double-well problem.

In this section we briefly explain how to obtain a tunneling formula between linear wells using our single well results. We will only recall the main ideas, since they follow a standard strategy.

Let us consider a smooth non-negative electric potential WW which is an even function on I=[−a,a]I=[-a,a]. We assume that WW has exactly two minima at ±a\pm a such that W(±a)=0W(\pm a)=0 and W′(a)=−W′(−a)<0W'(a)=-W'(-a)<0. We consider the operator

L=−h2∂x2+W,\mathcal{L}=-h^2\partial_x^2+W,

acting on the domain H2(I)∩H01(I)H^2(I)\cap H_0^1(I). A tunneling effect between the two wells yields the following spectral gap.

Proposition 4.1. There exists a universal constant κ>0\kappa>0 such that

λ2(h)−λ1(h)=κh23∣W′(a)∣23e−Sh/h(1+o(1)),Sh=∫−aa(W−h23∣z1∣∣W′(a)∣23)+.\lambda_2(h)-\lambda_1(h)=\kappa h^{\frac{2}{3}}\lvert W'(a)\rvert^{\frac{2}{3}}e^{-S_h/h}(1+o(1)),\qquad S_h=\int_{-a}^{a}\sqrt{\left(W-h^{\frac{2}{3}}\lvert z_1\rvert\lvert W'(a)\rvert^{\frac{2}{3}}\right)_+}.

The tunneling estimate of Proposition 4.1 does not seem to be already known. To the authors’ knowledge, the only similar result appears in [18 Section 1.5] in the special case when W(x)=1−∣x∣W(x)=1-|x| and a=1a=1. There the author describes the eigenfunctions explicitly by means of the Airy functions Ai and Bi. Note that, in this case,

Sh=43−2∣z1∣h23+O(h).S_h=\frac{4}{3}-2|z_1|h^{\frac{2}{3}}+O(h).

Of course, we cannot directly apply our result due to the peak at x=0x=0, but our estimate agrees with [18 Corollaire 1.18]. Note also that a similar formula is obtained in [14 Corollaire 3.2] in the case of degenerate wells (with a behavior $x^{2 \sigma near the minima, with σ≥1\sigma\ge11 and it is consistent with ours if one were allowed to take σ=12\sigma= \frac{1}{2}, but their Agmon distance does not include a perturbation like our h23h^{\frac{2}{3}}.

Proof. Let us only sketch the proof by explaining how to use our estimate of the one-well groundstate in Theorem 1.2. The details of the computations can be found in [5 Chapter 6] and also in recent articles, see for instance [6 Section 5] or [2 Section 5].

Given a/2>η>0a/2 > \eta> 0, we consider a smooth single-well potential WℓW_\ell on [−a,+∞)[-a,+\infty) that coincides with WW on [−a,a−η][-a,a-\eta]. We also assume that WℓW_\ell is bounded (with all its derivatives) and that, on [a−η,+∞)[a-\eta,+\infty), Wℓ≥c>0W_\ell\ge c > 0. We denote by fℓf_\ell the positive, L2L^2-normalized groundstate of the associated operator Lℓ=−h2∂x2+Wℓ\mathcal{L}_\ell=-h^2\partial_x^2+W_\ell and we denote by μ\mu the corresponding eigenvalue. We consider a smooth cutoff function 0≤χℓ≤10\le\chi_\ell\le1 supported in [−a,a−η)[-a,a-\eta) and equal to 1 on [−a,a−2η)[-a,a-2\eta).

Thanks to our decay estimate on fℓf_\ell and support considerations, we have

(L−μ)(χℓfℓ)=O(e−(S0−ε)/h)(\mathcal{L}-\mu)(\chi_\ell f_\ell)=O(e^{-(S_0-\varepsilon)/h})

where S0=∫−aaW(s) dsS_0=\int_{-a}^{a}\sqrt{W(s)}\,\mathrm{d}s, and ε>0\varepsilon>0 can be made as small as we want by taking η>0\eta>0 small enough. By symmetry, we also define the right-well potential WrW_r, and using self-evident notation, we have

(L−μ)(χrfr)=O(e−(S0−ε)/h).(\mathcal{L}-\mu)(\chi_r f_r)=O(e^{-(S_0-\varepsilon)/h}).

We also notice that

⟨χℓfℓ,χrfr⟩=O(e−(S0−ε)/h).\langle\chi_\ell f_\ell,\chi_r f_r\rangle=O(e^{-(S_0-\varepsilon)/h}).

Then, the spectral theorem shows that there are at least two eigenvalues near μ\mu at a distance of order at most O(e−(S0−ε)/h)O(e^{-(S_0-\varepsilon)/h}). In fact there are exactly two: we can show that the eigenfunctions of L\mathcal{L} are localized near ±a\pm a and that the low-lying spectra of L\mathcal{L} and Lℓ⊕Lr\mathcal{L}_\ell\oplus\mathcal{L}_r coincide modulo O(h∞)O(h^\infty). Due to the spectral asymptotics of Lℓ/r\mathcal{L}_{\ell/r} (see Theorem 1.2), we deduce that

λ2(h)−λ1(h)=O(e−(S0−ε)/h)\lambda_2(h)-\lambda_1(h)=O(e^{-(S_0-\varepsilon)/h})

and that λ3(h)−λ2(h)≥ch23\lambda_3(h)-\lambda_2(h)\ge ch^{\frac{2}{3}}.

There remains to get the accurate asymptotics of the spectral gap. We denote by Π\Pi the spectral projector onto Ker⁡(L−λ1)⊕Ker⁡(L−λ2)\operatorname{Ker}(\mathcal{L}-\lambda_1)\oplus\operatorname{Ker}(\mathcal{L}-\lambda_2). The above discussion and the spectral theorem imply that

(Id⁡−Π)(χ⋆f⋆)=O(e−(S0−ε)/h).(22)(\operatorname{Id}-\Pi)(\chi_\star f_\star)=O(e^{-(S_0-\varepsilon)/h}). \tag*{(22)}

We let

g⋆=Π(χ⋆f⋆).g_\star=\Pi(\chi_\star f_\star).

The vectors (gℓ,gr)(g_{\ell},g_{r}) form a basis of Ran⁡Π\operatorname{Ran}\Pi. However, it is not orthonormal. Therefore, we consider the associated Gram matrix of the family (gℓ,gr)(g_{\ell},g_{r}), namely the 2×22 \times2 matrix G=(⟨gj,gk⟩)G=\left(\langle g_j,g_k\rangle\right). Using (22) we can check that

G=Id⁡+T1+O(e−2(S0−ε)/h),G=\operatorname{Id}+T_1+O\left(e^{-2(S_0-\varepsilon)/h}\right),

with

T1=(0⟨χℓfℓ,χrfr⟩⟨χrfr,χℓfℓ⟩0)=O(e−(S0−ε)/h),T_1=\begin{pmatrix} 0 & \langle\chi_{\ell}f_{\ell},\chi_{r}f_{r}\rangle\\ \langle\chi_{r}f_{r},\chi_{\ell}f_{\ell}\rangle& 0 \end{pmatrix}=O\left(e^{-(S_0-\varepsilon)/h}\right),

which shows that the family (gℓ,gr)(g_{\ell},g_{r}) is almost orthonormal. The family G−1/2(gℓ,gr)G^{-1/2}(g_{\ell},g_{r}) is orthonormalized and the matrix of L\mathcal{L} in this basis is given by

L:=G−1/2(⟨Lgℓ,gℓ⟩⟨Lgℓ,gr⟩⟨Lgr,gℓ⟩⟨Lgr,gr⟩)G−1/2.L:=G^{-1/2}\begin{pmatrix} \langle\mathcal{L}g_{\ell},g_{\ell}\rangle& \langle\mathcal{L}g_{\ell},g_{r}\rangle\\ \langle\mathcal{L}g_{r},g_{\ell}\rangle& \langle\mathcal{L}g_{r},g_{r}\rangle\end{pmatrix}G^{-1/2}.

We have

(⟨Lgℓ,gℓ⟩⟨Lgℓ,gr⟩⟨Lgr,gℓ⟩⟨Lgr,gr⟩)=μId⁡+T2+O(e−2(S0−ε)/h),\begin{pmatrix} \langle\mathcal{L}g_{\ell},g_{\ell}\rangle& \langle\mathcal{L}g_{\ell},g_{r}\rangle\\ \langle\mathcal{L}g_{r},g_{\ell}\rangle& \langle\mathcal{L}g_{r},g_{r}\rangle\end{pmatrix}=\mu\operatorname{Id}+T_2+O\left(e^{-2(S_0-\varepsilon)/h}\right),

with

T2=(0⟨L(χℓfℓ),(χrfr)⟩⟨L(χℓfℓ),(χrfr)⟩0).T_2=\begin{pmatrix} 0 & \langle\mathcal{L}(\chi_{\ell}f_{\ell}),(\chi_{r}f_{r})\rangle\\ \langle\mathcal{L}(\chi_{\ell}f_{\ell}),(\chi_{r}f_{r})\rangle& 0 \end{pmatrix}.

Therefore,

L=(μwwμ)+O(e−2(S0−ε)/h),w=⟨(L−μ)(χℓfℓ),χrfr⟩,L=\begin{pmatrix} \mu& w \\ w & \mu\end{pmatrix}+O\left(e^{-2(S_0-\varepsilon)/h}\right),\qquad w=\langle(\mathcal{L}-\mu)(\chi_{\ell}f_{\ell}),\chi_{r}f_{r}\rangle,

and the spectral gap of this matrix is

λ2−λ1=2∣w∣+O(e−2(S0−ε)/h).\lambda_2-\lambda_1=2|w|+O\left(e^{-2(S_0-\varepsilon)/h}\right).

Then, we have, by support considerations and by using the definition of fℓf_{\ell},

w=⟨(Lℓ−μ)(χℓfℓ),χrfr⟩=−h2⟨[∂x2,χℓ]fℓ,fr⟩=−h2⟨2χℓ′fℓ′+χℓ′′fℓ,fr⟩.w=\langle(\mathcal{L}_{\ell}-\mu)(\chi_{\ell}f_{\ell}),\chi_{r}f_{r}\rangle=-h^2\langle[\partial_x^2,\chi_{\ell}]f_{\ell},f_{r}\rangle=-h^2\langle2\chi_{\ell}'f_{\ell}'+\chi_{\ell}''f_{\ell},f_{r}\rangle.

By integration by parts, we deduce that

⟨2χℓ′fℓ′+χℓ′′fℓ,fr⟩=⟨2χℓ′fℓ′,fr⟩−⟨χℓ′fℓ,fr⟩−⟨χℓ′fℓ,fr′⟩=∫0a−ηχℓ′(fℓ′fr−fℓfr′) ds,\begin{aligned} \langle2\chi_{\ell}'f_{\ell}'+\chi_{\ell}''f_{\ell},f_{r}\rangle=\langle2\chi_{\ell}'f_{\ell}',f_{r}\rangle-\langle\chi_{\ell}'f_{\ell},f_{r}\rangle-\langle\chi_{\ell}'f_{\ell},f_{r}'\rangle\\ &=\int_0^{a-\eta}\chi_{\ell}'(f_{\ell}'f_{r}-f_{\ell}f_{r}')\,\mathrm{d}s, \end{aligned}

where 00 is the middle point between the two wells. Since, on (0,a−η)(0,a-\eta)

(fℓ′fr−fℓfr′)′=fℓ′′fr−fℓfr′′=h−2(Wℓ−Wr)fℓfr=0,(f_{\ell}'f_{r}-f_{\ell}f_{r}')'=f_{\ell}''f_{r}-f_{\ell}f_{r}''=h^{-2}(W_{\ell}-W_{r})f_{\ell}f_{r}=0,

a second integration by parts gives

w=h2(fℓ′(0)fr(0)−fℓ(0)fr′(0))=2h2fℓ′(0)fℓ(0).w=h^2(f_{\ell}'(0)f_{r}(0)-f_{\ell}(0)f_{r}'(0))=2h^2f_{\ell}'(0)f_{\ell}(0).

Finally, we can use the pointwise estimate from Theorem 1.2, for the single-well groundstate:

fℓ(0)∼∣W′(a)∣132π∣Ai⁡′(z1)∣e−1hS(0)∣W(0)∣14h16f_\ell(0) \sim\frac{|W'(a)|^{\frac{1}{3}}}{2\sqrt{\pi}|\operatorname{Ai}'(z_1)|}\frac{e^{-\frac{1}{h}S(0)}}{|W(0)|^{\frac{1}{4}}h^{\frac{1}{6}}}
fℓ′(0)∼−W(0)h∣W′(a)∣132π∣Ai⁡′(z1)∣e−1hS(0)∣W(0)∣14h16.f'_\ell(0) \sim-\frac{\sqrt{W(0)}}{h}\frac{|W'(a)|^{\frac{1}{3}}}{2\sqrt{\pi}|\operatorname{Ai}'(z_1)|}\frac{e^{-\frac{1}{h}S(0)}}{|W(0)|^{\frac{1}{4}}h^{\frac{1}{6}}}.

We conclude that

λ2(h)−λ1(h)∼h→0h∣W′(a)∣2/3πAi⁡′(z1)2h1/3e−Sh/h.\lambda_2(h)-\lambda_1(h)\mathrel{\sim}_{h\to0}\frac{h|W'(a)|^{2/3}}{\pi\operatorname{Ai}'(z_1)^2h^{1/3}}e^{-S_h/h}.

Adaptation to double V-shaped wells

In this final subsection we discuss how to adapt the results of this article to VV-shaped wells. Take a>0a>0 and consider an even continuous non-negative electric potential WW that is piecewise smooth on R∖{±a}\mathbb{R}\setminus\{\pm a\} with exactly two minima at ±a\pm a such that W(±a)=0W(\pm a)=0. Assume that A=W′(a+)>0A=W'(a^+)>0 and −B=W′(a−)<0-B=W'(a^-)<0. Also assume that WW is confining, i.e. lim inf⁡∣x∣→∞W>0\liminf_{|x|\to\infty}W>0. We consider the corresponding operator

L=−h2∂x2+W.\mathcal{L}=-h^2\partial_x^2+W.

The following operator plays an important role in the analysis:

M=−h2∂x2−Bx1R−(x)+Ax1R+(x).\mathcal{M}=-h^2\partial_x^2-Bx\mathbf{1}_{\mathbb{R}_-}(x)+Ax\mathbf{1}_{\mathbb{R}_+}(x).

It appears as a first order approximation of L\mathcal{L} near the well at aa. This selfadjoint operator has compact resolvent and its eigenvalues and eigenfunctions can be expressed in terms of the Airy functions. By homogeneity, its groundstate eigenvalue z(h)z(h) satisfies z(h)=h23z(1)z(h)=h^{\frac{2}{3}}z(1).

Proposition 4.2. There exists a 2/32/3 homogeneous positive function κ(A,B)\kappa(A,B) such that

λ2(h)−λ1(h)=κ(A,B)h23e−Sh/h(1+o(1)),Sh=∫−aa(W−h23z(1))+.\lambda_2(h)-\lambda_1(h)=\kappa(A,B)h^{\frac{2}{3}}e^{-S_h/h}(1+o(1)),\qquad S_h=\int_{-a}^{a}\sqrt{\left(W-h^{\frac{2}{3}}z(1)\right)_+}.

The κ\kappa function is obtained by solving an implicit equation involving Airy functions. There does not seem to be a compact formula for it.

Proof idea. The proof follows the same lines as for Proposition 4.1. We start with a piecewise linear approximation near the bottom of a single well, which gives an L2L^2 approximation of the normalized eigenfunctions in terms of Airy functions (at leading order), and permits the computation of the relevant constants. Then a quasimode construction and Agmon estimates propagate the Airy approximation inside the forbidden region between the two wells. □

Declaration of AI use

AI tools were used solely to check the mathematical calculations and reasoning in the final version of the manuscript. They were not used to write the manuscript, develop proofs, or generate mathematical ideas.

(Y. Guedes Bonthonneau) DMA, ENS. 45 rue d’Ulm, Paris. Email address: [email protected]

(L. Morin) Laboratoire de Mathématiques Jean Leray (UMR CNRS 6629), Nantes Université, 2 rue de la Houssinière, 44322 Nantes Cedex 03, France Email address: [email protected]

(N. Raymond) Univ Angers, CNRS, LAREMA, Institut Universitaire de France, F-49000 Angers, France Email address: [email protected]

References

  1. [1]S. Agmon. Lectures on exponential decay of solutions of second-order elliptic equations: bounds on eigenfunctions of N-body Schrödinger operators, volume 29 of Mathematical Notes. Princeton University Press, Princeton, NJ; University of Tokyo Press, Tokyo, 1982.
  2. [2]M. Averseng, N. Frantz, F. Hérau, and N. Raymond. Semiclassical tunneling for some 1D Schrödinger operators with complex-valued potentials. J. Spectr. Theory, 2026. Published online first, 3 September 2026.
  3. [3]H. Cornean, D. Krejčiřík, T. G. Pedersen, N. Raymond, and E. Stockmeyer. On the two-dimensional quantum confined Stark effect in strong electric fields. SIAM J. Math. Anal., 54(2):2114–2127, 2022.DOI
  4. [4]M. Dauge and N. Raymond. Plane waveguides with corners in the small angle limit. J. Math. Phys., 53(12):123529, 2012.DOI
  5. [5]M. Dimassi and J. Sjöstrand. Spectral asymptotics in the semi-classical limit, volume 268 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1999.DOI
  6. [6]A. Duraffour and N. Raymond. An example of accurate microlocal tunneling in one dimension. Math. Res. Lett., 32(5):1421–1463, 2025.
  7. [7]A. Erdélyi. Asymptotic solutions of differential equations with transition points or singularities. J. Math. Phys., 1(1):16–26, 1960.DOI
  8. [8]E. M. Harrell. Double wells. Commun. Math. Phys., 75:239–261, 1980.DOI
  9. [9]B. Helffer. Semi-classical analysis for the Schrödinger operator and applications, volume 1336 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1988.DOI
  10. [10]B. Helffer and J. Sjöstrand. Multiple wells in the semiclassical limit. I. Comm. Partial Differential Equations, 9(4):337–408, 1984.DOI
  11. [11]B. Helffer and J. Sjöstrand. Puits multiples en limite semi-classique. II: Interaction moléculaire. Symétries. Perturbation. Ann. Inst. Henri Poincaré, Phys. Théor., 42(2):127–212, 1985.
  12. [12]R. E. Langer. On the asymptotic solutions of ordinary differential equations, with an application to the Bessel functions of large order. Trans. Amer. Math. Soc., 33:23–64, 1931.DOI
  13. [13]J. Liouville. Second mémoire sur le développement des fonctions ou parties de fonctions en séries dont les divers termes sont assujettis à satisfaire à une même équation différentielle du second ordre, contenant un paramètre variable. Journal de Mathématiques Pures et Appliquées, 1e série, 2:16–35, 1837.
  14. [14]A. Martinez and M. Rouleux. Effet tunnel entre puits dégénérés. Commun. Partial Differ. Equations, 13(9):1157–1187, 1988.DOI
  15. [15]F. W. J. Olver. Error bounds for first approximations in turning-point problems. J. Soc. Ind. Appl. Math., 11(3):748–772, 1963.
  16. [16]F. W. J. Olver. Asymptotics and Special Functions. Academic Press, 1974.
  17. [17]T. Ourmières-Bonafos. Dirichlet eigenvalues of asymptotically flat triangles. Asymptotic Anal., 92(3-4):279–312, 2015.
  18. [18]N. Popoff. Sur le spectre de l’opérateur de Schrödinger magnétique dans un domaine diédral. Ph.D. thesis, Université Rennes 1, Nov. 2012.
  19. [19]D. Robert. Analyse semi-classique de l’effet tunnel [d’après B. Helfer et J. Sjöstrand]. In Séminaire Bourbaki, Vol. 1985/86, number 145–146 in Astérisque, pages 257–281. Société Mathématique de France, 1987. Exp. no. 665.
  20. [20]B. Simon. Semiclassical analysis of low lying eigenvalues. I. Non-degenerate minima: asymptotic expansions. Ann. Inst. Henri Poincaré, Phys. Théor., 38(3):295–308, 1983. Erratum: ibid. 40 (1984), no. 2, 224.
  21. [21]B. Simon. Semiclassical analysis of low lying eigenvalues. II: Tunneling. Ann. Math. (2), 120(1):89–118, 1984.

Paper details

vlnc:W_nvd9bfxw
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

AI tools

View statement
Looking for related papers…
Contents