betaFeedback

WorksMathematics

Singularly embedded solitons and Stokes phenomenon in the fifth order KdV equation

Abstract

We study embedded solitons (ES) in a fifth-order Korteweg-de Vries (KdV) equation near a reversible Hamiltonian saddle-centre bifurcation. The solitons correspond to homoclinic orbits to an equilibrium whose linearisation has a pair of real and a pair of purely imaginary eigenvalues. Existence of symmetric homoclinic near the bifurcation is a codimension-one phenomenon governed by a parameter-dependent Stokes constant associated with an inner equation. Using complex matching, we derive an oscillatory large-parameter asymptotic formula for the Stokes constant. This formula implies the existence of a countable sequence of zeros and gives their asymptotic distribution in parameter space. We also develop two complementary high-precision numerical procedures for evaluating the Stokes constant: one based on the late terms of the formal inner expansion and another on direct computation of the stable and unstable manifolds. The numerical results agree with each other and with the asymptotic prediction, confirming both the existence and the predicted locations of the embedded-soliton branches.

prediction, confirming both the existence and the predicted locations of the embedded-soliton branches.

Introduction

Embedded solitons in a fifth order KdV

A soliton is a solution of a non-linear partial differential equation (PDE) that is localised in space and moves with constant speed while preserving its shape. In some equations the soliton may coexist with linear harmonic waves that travel with the same velocity [7]. Those solutions are called embedded solitons (ES) to stress their relation with the spectrum of the linearised equation [18]. ES are observed in various models, for example, in models of non-linear optics, capillary-gravity surface wave and higher-order KdV type equations [6].

It is known that emergence of ES can be related to a bifurcation in the PDE’s dispersion relation that establishes connection between wave number kk and phase velocity cc of a small-amplitude harmonic solution to the problem linearised around an equilibrium state. Following [7] we illustrate this situation with the help of the 5th order KdV model

ut=[215uxxxx−buxx+au+32u2−12ux2+(uux)x]x.(1)u_t = \left[\frac{2}{15}u_{xxxx} - b u_{xx} + a u + \frac{3}{2}u^2 - \frac{1}{2}u_x^2 + (u u_x)_x\right]_x . \tag*{(1)}

This equation was derived from the Euler equation for water waves with surface tension using the regular Hamiltonian perturbation theory [9]. The equation depends on two real parameters aa and bb. Looking for a solution in the form u=Aeik(x−ct)u=Ae^{ik(x-ct)} and neglecting the non-linear terms we obtain the dispersion relation in the form

215k4+bk2+a+c=0.(2)\frac{2}{15}k^4 + bk^2 + a + c = 0 . \tag*{(2)}

This equation determines the phase speed cc of small amplitude harmonic waves as a function of the wave number kk. The dispersion relation can also be considered as an equation for possible values of kk for a fixed cc. Real values of kk correspond to small amplitude harmonic waves, while non-real ones correspond to exponentially increasing/decreasing solutions of the linearised equation and open a possibility for existence of solitons. It is easy to see that for a+c<0a+c<0 the equation (2) has a pair of real solutions and a pair of purely imaginary ones. In this region of parameters a soliton may co-exist with a small amplitude harmonic wave.

A bifurcation happens at a+c=0a+c=0 where a long linear wave with k≪1k\ll1, which exists for c>−ac>-a, becomes an unstable linear mode for c<−ac<-a. The numerical study [7] suggests that the values of parameters corresponding to ES form a system of isolated lines on the (b,a+c)(b,a+c)-plane which emerge from a sequence of points with b=bk<0b=b_k<0 and a+c=0a+c=0.

Looking for a travelling wave solution of the equation (1) in the form u=u(x−ct)u = u(x - ct), integrating once and setting the integration constant to zero, we obtain an ordinary differential equation (ODE)

215u(iv)−bu′′+cu+32u2−12(u′)2+(uu′)′=0(3)\frac{2}{15}u^{(iv)} - bu^{\prime\prime} + cu + \frac{3}{2}u^2 - \frac{1}{2}(u^\prime)^2 + (uu^\prime)^\prime= 0 \tag*{(3)}

where the primes refer to derivatives with respect to the new independent variable τ=x−ct\tau= x - ct. We also substituted a=0a = 0 to shorten the notation without loosing in generality as the equation depends only on the sum a+ca + c. Solitons correspond to solutions of this equation that converge to zero at infinity.

The equation (3) inherits reversibility from the original PDE: if a function u(τ)u(\tau) is a solution, then so is u(−τ)u(-\tau). Moreover, (3) can be rewritten as a Hamiltonian system with two degrees of freedom, where the trivial solution u≡0u \equiv0 corresponds to an equilibrium point. This reformulation of the problem is described in Appendix § 8.

In dynamical systems theory, a solution of an ODE that converges to the same equilibrium as τ→±∞\tau\to\pm\infty is called a homoclinic orbit. Consequently, homoclinic solutions to the origin correspond to solitons. The theory of dynamical systems provides a range of tools suitable for studying homoclinic solutions, including regular and singular perturbation theories, bifurcation theory, as well as various numerical methods (see [8] for a detailed discussion, as well as [15, 13]).

Bifurcation diagram

The linear stability of the equilibrium is determined by the characteristic exponents λ=ik\lambda= ik, where kk satisfies the dispersion relation (2), therefore

215λ4−bλ2+c=0.\frac{2}{15}\lambda^4 - b\lambda^2 + c = 0.

This is a biquadratic equation and, in general, admits four distinct solutions. As the parameters vary, two of these solutions may coincide. A straightforward calculation shows that this occurs when the determinant b2−815c=0b^2 - \frac{8}{15}c = 0 or when λ2=0\lambda^2 = 0 for c=0c = 0. These two lines partition the parameter plane into four regions sketched in Figure 1.

Figure 1. Classification of the different regions of the linearisation at the origin of the equation (3) on the (b,c)(b,c)-plane.

The equation (3) admits no soliton solutions when the parameters (b,c)(b,c) lie in Region 4, since the characteristic exponents are purely imaginary and therefore there are no solutions that decay exponentially when τ→±∞\tau\to\pm\infty. In contrast, solitons may exist in all other regions.

From a generic perspective, one expects solitons to occur for an open set of parameter values in Regions 1 and 2. This is because the stable and unstable manifolds are two-dimensional surfaces contained within a three-dimensional level set of the Hamiltonian [7, 6].

The situation is different in Region 3, where the stable and unstable manifolds of the equilibrium are one-dimensional and, consequently, the existence of homoclinic solutions is expected to be a codimension-one phenomenon. The embedded solitons arise for parameter values in Region 3 and correspond to solutions homoclinic to the saddle–centre equilibrium at the origin [7, 6].

Existence of solitons

The above arguments do not, however, establish the existence of solitons. When the parameters lie on the curve c=35(2b+1)(b−2)c = \frac{3}{5}(2b + 1)(b - 2) with −12<b<2-\frac{1}{2} < b < 2, homoclinic solutions can be expressed explicitly in terms of elementary functions [7]. Apart from this special case, no explicit soliton solutions of (3) are known. Additional soliton solutions can be obtained near the boundaries between regions using bifurcation theory [8]. The bifurcation theory employs normal forms to describe changes in the dynamics that occur when parameters cross a bifurcation curve. In our case, the normal forms are integrable and the resulting dynamics can be analysed explicitly. In particular, the normal forms predict the emergence of a homoclinic solution from the equilibrium when parameters move from Region 3 to Region 2, from Region 4 to Region 1, and from Region 4 to Region 3.

However, the normal forms theory provides only an approximate description of the original dynamics. In the first two cases, regular perturbation theory can be used to show that the existence of a homoclinic solution in the normal form implies the existence of a corresponding homoclinic solution in the full system.

The situation near S34S_{34} is different, and the analysis of persistence of homoclinic orbits requires singular perturbation theory [14, 12, 11, 17]. The main conclusion of the theory is that, when the parameters lie on the curve S34S_{34}, the equation (3) with c=0c=0, called the inner equation,

215u(iv)−bu′′+32u2−12(u′)2+(uu′)′=0,(4)\frac{2}{15}u^{(iv)}-bu''+\frac{3}{2}u^2-\frac{1}{2}(u')^2+(uu')'=0, \tag*{(4)}

can be used to define a Stokes constant S=S(b)S=S(b). This constant encodes the splitting of stable and unstable manifolds of the degenerate saddle–centre equilibrium and plays a central role in determining whether homoclinic orbits persist under perturbation.

If S(b)≠0S(b)\ne0, then there are no single-bump homoclinic solutions for parameter values near the point (b,0)(b,0) in Region 3. Consequently, a primary homoclinic solution can only exist near a zero of the Stokes constant. In the context of equation (3), Tovbis and Pelinovsky [17] showed that every simple zero b0b_0, S(b0)=0S(b_0)=0, generates a family of homoclinic orbits, which form a smooth curve in the (b,c)(b,c)-plane and starts at (b0,0)(b_0,0).

The Stokes constant SS is an analytic function of the parameter, although no explicit closed-form expression is known and usually its values are obtained numerically.

We show that, for large positive γ=−5b/2\gamma=-5b/2 (or equivalently for large negative bb), the inner problem (4) can itself be interpreted within the framework of singular perturbation theory. In this setting, two complex singularities of an outer problem contribute comparably, leading to oscillatory asymptotics

S0(γ)≈i 2(34)1/4∣K0∣γ−23/16sin⁡(π16+43γ)(5)S_0(\gamma)\approx\mathrm{i}\,2\left(\frac{3}{4}\right)^{1/4}|K_0|\gamma^{-23/16}\sin\left(\frac{\pi}{16}+4\sqrt{3\gamma}\right) \tag*{(5)}

where K0K_0 is a Stokes constant obtained from a parameterless problem described in § 4.4, K0=0.251346943975660732…⋅eiπ/16K_0=0.251346943975660732\ldots\cdot e^{\mathrm{i}\pi/16}. It follows that, if γ\gamma is sufficiently large, there is a zero of S0(γ)S_0(\gamma) near every point

γ0(n)=(πn+15π1643)2,(6)\gamma_0(n)=\left(\frac{\pi n+\frac{15\pi}{16}}{4\sqrt{3}}\right)^2, \tag*{(6)}

where the approximation for S0(γ)S_0(\gamma) changes its sign. It is interesting to note that our numerical experiments show that every interval (γ0(n),γ0(n+1))(\gamma_0(n),\gamma_0(n+1)) contains exactly one zero γn∗\gamma_n^* for every n∈Nn\in\mathbb{N}, and that there are no other zeroes. Moreover, the difference γn∗−γ0(n)\sqrt{\gamma_n^*}-\sqrt{\gamma_0(n)} monotonically converges to zero.

This is a rather curious example of the theory of asymptotic beyond all orders. Usually the theory is used to estimate exponentially small terms which contain a pre-factor with the imaginary part of a leading singularity in the exponent. In our case the leading singularity is on the real axis and, consequently, the “exponentially small” pre-factor equals to one.

To confirm our theoretical predictions, we numerically computed the values of S0S_0 and analysed the locations of its zeros using two complementary methods. The first is based on the analytic continuation of solutions, while the second relies on an analysis of the coefficients of formal solutions. In these numerical experiments, high-precision arithmetic is essential for obtaining reliable results, particularly for larger values of γ\gamma, where numerical evaluation becomes increasingly delicate.

The rest of the paper is organized in the following way.

  • Section 2: summary of the singular perturbation theory, inner and outer limits for the equation (3), reduction of persistence of solitons to zeroes of the Stokes constant

  • Section 3: analysis of the inner problem, definition and basic properties of the Stokes constant

  • Section 4: inner problem as a singularly perturbed problem, outer and inner equations for the inner equation, asymptotics of S0S_0 for large γ\gamma, existence of infinitely many zeroes.

  • Section 5: numerical methods for evaluation of the Stokes constant and finding its zeroes.

Hamiltonian bifurcation and singular perturbations

Scaling invariants of the equation

The study of solitons emerging at the bifurcation relies on various changes of scales in the ODE. Since these substitutions affect the coefficients without changing the form of the equation, it is convenient to consider a more general equation

a2u(iv)+a1u′′+a0u+b0u2+b1(u′2+2uu′′)=0(7)a_2u^{(iv)} + a_1u'' + a_0u + b_0u^2 + b_1(u'^2 + 2uu'') = 0 \tag*{(7)}

which depends on five real parameters a0a_0, a1a_1, a2a_2, b0b_0, b1b_1. We assume that the quadratic part is not degenerate, i.e., b0,b1≠0b_0,b_1 \ne0. (7) is an ODE of order four with a quadratic nonlinearity, which coincides with (3) when a2=215a_2 = \frac{2}{15}, a1=−ba_1 = -b, a0=ca_0 = c, b0=32b_0 = \frac{3}{2} and b1=12b_1 = \frac{1}{2}. In spite of the substantially larger number of parameters equation (7) is only a minor generalisation of (3). Indeed, if a2,b0,b1≠0a_2,b_0,b_1 \ne0, then the substitution u(t)=αu~(τ)u(t)=\alpha\tilde{u}(\tau) with t=βτt=\beta\tau,

α=a2b0b12andβ=∣b0b1∣1/2\alpha=\frac{a_2b_0}{b_1^2}\quad\text{and}\quad\beta=\left|\frac{b_0}{b_1}\right|^{1/2}

transforms (7) into a standard form

u(iv)+p1u′′+p2u+u2+σ0(u′2+2uu′′)=0(8)u^{(iv)}+p_1u''+p_2u+u^2+\sigma_0(u'^2+2uu'')=0 \tag*{(8)}

where σ0=sign⁡(b0b1)\sigma_0=\operatorname{sign}(b_0b_1), p1=a1∣b1∣a2∣b0∣p_1=\frac{a_1|b_1|}{a_2|b_0|} and p2=a0b12a2b02p_2=\frac{a_0b_1^2}{a_2b_0^2}. For example, for equation (3) we get p1=−5b/2p_1=-5b/2, p2=5c/6p_2=5c/6 and σ0=1\sigma_0=1. In equation (8), we have dropped the tildes for brevity. Here and later, we use primes to denote derivatives, regardless of the letter used for the function’s argument.

We call parameters p1p_1 and p2p_2 scaling invariants, since any two equations (7) with non-vanishing coefficients can be transformed into one another if and only if their scaling invariants coincide. These invariants provide a convenient criterion for establishing the equivalence of two equations without explicitly constructing the transformation between them.

Hamiltonian bifurcation and singular perturbations

In the search for embedded solitons, we consider the reversible Hamiltonian bifurcation which happens when the linear stability of the equilibrium changes along the line S34S_{34} of Figure 1. An embedded soliton can exists only in the area below the line, where the equilibrium is of the saddle–centre type. That means a2a0<0a_2a_0<0 and, equivalently, p2<0p_2<0. The line S34S_{34} belongs to the region where a2a1>0a_2a_1>0 and, equivalently, p1>0p_1>0. We denote the eigenvalues of the equilibrium by ±λ\pm\lambda and ±iω\pm i\omega assuming that ω,λ>0\omega,\lambda>0. On the line S34S_{34}, λ=0\lambda=0 is a double eigenvalue and ω>0\omega>0.

The scaling of the time variable changes the values of ω\omega and λ\lambda without affecting the ratio λ/ω\lambda/\omega. This freedom allows us to consider the corresponding equations in two equivalent ways.

First, an equation (7) with a2a1>0a_2a_1>0, a2a0≤0a_2a_0\le0 and b0b1≠0b_0b_1\ne0 has the scaling invariants p1>0p_1>0 and p2≤0p_2\le0 and, consequently, it can be transformed into an equation of the form considered in [17],

u(iv)+(1−ε2)u′′−ε2u+(1−ε2)u2+γ(u′2+2uu′′)=0(9)u^{(iv)}+(1-\varepsilon^2)u''-\varepsilon^2u+(1-\varepsilon^2)u^2+\gamma(u'^2+2uu'')=0 \tag*{(9)}

with ε∈[0,1)\varepsilon\in[0,1) and γ≠0\gamma\ne0. Indeed, computing the scaling invariants for this equation we get p1=∣γ∣p_1=|\gamma| and p2=−ε2γ2/(1−ε2)2p_2=-\varepsilon^2\gamma^2/(1-\varepsilon^2)^2. Thus we cover all possible values of the scaling invariants in the region where p1>0p_1>0 and p2≤0p_2\le0.

In the case of equation (9), the characteristic equation takes the form

λ4+(1−ε2)λ2−ε2=0\lambda^4 + (1-\varepsilon^2)\lambda^2-\varepsilon^2=0

and consequently λ=ε\lambda=\varepsilon and ω=1\omega=1. Thus in this formulation, the problem reduces to the study of a reversible Hamiltonian (02,±iω)(0^2,\pm i\omega) bifurcation for equation (9) which happens at ε=0\varepsilon=0.

Second, the standard scaling u(t)=ε2v(τ)u(t)=\varepsilon^2v(\tau) with τ=εt\tau=\varepsilon t transforms equation (9) into a singularly perturbed problem,

ε2v(iv)+(1−ε2)v′′−v+(1−ε2)v2+ε2γ(v′2+2vv′′)=0,(10)\varepsilon^2v^{(iv)}+(1-\varepsilon^2)v''-v+(1-\varepsilon^2)v^2+\varepsilon^2\gamma(v'^2+2vv'')=0, \tag*{(10)}

where the small parameter multiplies the highest derivative. In this formulation, we get λ=1\lambda=1 and ω=ε−1\omega=\varepsilon^{-1} goes to +∞+\infty when ε→0\varepsilon\to0. In the limit ε→0\varepsilon\to0 equation (10) reduces to an equation of the second order,

v′′−v+v2=0.v''-v+v^2=0.

In the singular perturbation theory, this equation is called an outer equation. The outer equation admits an explicit homoclinic solution

v0(τ)=32cosh⁡−2(τ2).(11)v_0(\tau)=\frac{3}{2}\cosh^{-2}\left(\frac{\tau}{2}\right). \tag*{(11)}

After inverting the scalings, we obtain an approximation for a soliton of the original equation (3) provided it survives the influence of the singular perturbation.

Persistence of a soliton is of co-dimension one.

Despite the existence of the homoclinic orbit in equation (10) with ε=0\varepsilon=0, such an orbit is generically destroyed when ε>0\varepsilon>0, see [14, 11]. To study the persistence of this solution for ε≠0\varepsilon\ne0, we consider the stable and unstable solutions of equation (9). Since the unstable direction at the equilibrium is one dimensional, the unstable solution, u−(t)→0u_-(t)\to0 as t→−∞t\to-\infty, is defined uniquely up to a constant shift in time. We eliminate this freedom by asking t=0t=0 to be the first local maximum of u−u_-. Then u−′(0)=0u_-'(0)=0. We note that t=0t=0 is also a maximum of v0v_0, so v0′(0)=0v_0'(0)=0, and the unstable solution admits the approximation

u−(t)=ε2v0(εt)+O(ε4)(12)u_-(t)=\varepsilon^2v_0(\varepsilon t)+O(\varepsilon^4) \tag*{(12)}

which holds uniformly for t≤t0t\le t_0 with any given t0t_0. In the leading term of this estimate, ε\varepsilon comes from the standard scaling.

The stable solution is obtained using the reversible symmetry, u+(t)=u−(−t)u_+(t)=u_-(-t). Since v0v_0 is even, u+u_+ admits approximation by the same function as u−u_- but for t≥−t0t\ge-t_0.

Then we consider the difference

δ(t)=u+(t)−u−(t).\delta(t) = u_{+}(t) - u_{-}(t).

Since δ(t)=u−(−t)−u−(t)\delta(t) = u_{-}(-t) - u_{-}(t), it is odd and we get δ(0)=δ′′(0)=0\delta(0) = \delta''(0) = 0. The normalisation assumption u−′(0)=0u_{-}'(0) = 0 implies that δ′(0)=0\delta'(0) = 0. Then, in the case δ′′′(0)=0\delta'''(0) = 0, we can conclude that u+u_{+} and u−u_{-} satisfy the same initial conditions at t=0t = 0 and, since the solution to Cauchy problem is unique, that u+(t)=u−(t)u_{+}(t) = u_{-}(t) for all tt. Hence we have a symmetric soliton each time δ′′′(0)\delta'''(0) vanishes. This soliton is a single-bump one as it is approximated by the unperturbed homoclinic solution ε2v0(εt)\varepsilon^{2}v_{0}(\varepsilon t) for all real tt.

A direct substitution shows that δ\delta satisfies the variational equation:

δ(iv)+(1−ε2)δ′′−ε2δ+2(1−ε2)uδ+2γ(uδ′′+u′δ′+u′′δ)=0(13)\delta^{(iv)} + (1-\varepsilon^{2})\delta'' - \varepsilon^{2}\delta+ 2(1-\varepsilon^{2})u\delta+ 2\gamma(u\delta'' + u'\delta' + u''\delta) = 0 \tag*{(13)}

around u=uc=12(u++u−)u = u_{c} = \frac{1}{2}(u_{+} + u_{-}). The variational equation is a linear homogeneous ODE of order four. The space of its solution is spanned by four fundamental solutions. Let δ1\delta_{1} be a solution that satisfies the Cauchy problem for the variational equation: δ1(0)=δ1′(0)=δ1′′(0)=0\delta_{1}(0) = \delta_{1}'(0) = \delta_{1}''(0) = 0 and δ1′′′(0)=1\delta_{1}'''(0) = 1. Then there is a real constant C=C(ε,γ)C = C(\varepsilon,\gamma) such that

δ(t)=C(ε,γ)δ1(t).\delta(t) = C(\varepsilon,\gamma)\delta_{1}(t).

We note that the existence of single-bump symmetric homoclinic solutions is equivalent to the vanishing of C(ε,γ)C(\varepsilon,\gamma), i.e., it is a co-dimension one phenomenon.

Inner equation and Stokes phenomenon

Determining the zeros of C(ε,γ)C(\varepsilon,\gamma) is a non-trivial problem, since C(ε,γ)C(\varepsilon,\gamma) is exponentially small in ε\varepsilon. The complex matching method suggests studying the analytic continuation of the solutions into the complex plane in order to obtain an approximation for C(ε,γ)C(\varepsilon,\gamma).

The leading order of this approximations comes from the study of stable and unstable solutions of the inner equation that is obtained from (9) by setting ε=0\varepsilon= 0,

u(iv)+u′′+u2+γ(u′2+2uu′′)=0.(14)u^{(iv)} + u'' + u^{2} + \gamma(u'^{2} + 2uu'') = 0. \tag*{(14)}

The equilibrium solution u=0u = 0 is a degenerate saddle-centre: the linearisation at the origin has two purely imaginary eigenvalues ±i\pm i and a double zero eigenvalue. The inner equation admits two distinguished solutions, referred to as the stable and unstable solutions, which are uniquely characterised by their asymptotic behaviour at the infinity:

u±(x)=−6x2+O(x−4),x→±∞.(15)u^{\pm}(x) = -\frac{6}{x^2} + O(x^{-4}), \qquad x \to\pm\infty. \tag*{(15)}

respectively, where we introduce yet a new time, x=(τ−iπ)/εx = (\tau- \mathrm{i}\pi)/\varepsilon (see Remark 1 below). Similarly to u±u_{\pm}, the stable and unstable solutions of (9), the solutions u±u^{\pm} belong to the stable and unstable manifolds of the equilibrium for equation (14), although the convergence rate is much slower.

Remark 1 The complex matching method uses the observation that the approximation (12) fails near the singularity of v0(τ)v_0(\tau) located at τ=iπ\tau= \mathrm{i}\pi, and that in this region a better approximation is obtained using the inner equation. The method uses the change of time

τ=εt=iπ+εx(16)\tau= \varepsilon t = \mathrm{i}\pi+ \varepsilon x \tag*{(16)}

to study stable and unstable solutions in the region where ∣τ−iπ∣|\tau- \mathrm{i}\pi| is small and ∣x∣|x| is large. The application of the standard scaling to the Laurent series v0(τ)=−6(τ−iπ)−2+O(1)v_0(\tau) = -6(\tau- \mathrm{i}\pi)^{-2} + O(1) suggests to use (15) as an asymptotic condition for the inner equation at infinity. Then it can be shown that u±(x)u^{\pm}(x) approximate the stable and unstable solutions u±(t)u_{\pm}(t) with t=iπε+xt = \mathrm{i}\frac{\pi}{\varepsilon} + x.

Since equation (14) is reversible and all its coefficients are real, the involution

(Ru)(x)=u(−x‾)‾(17)(Ru)(x) = \overline{u(-\overline{x})} \tag*{(17)}

maps a solution uu of (14) to another solution of the same equation. Under this transformation, the domain of uu is reflected in the imaginary axis. Since the solutions u±u^{\pm} are uniquely defined by the asymptotic condition (15) which is invariant under the involution RR, we conclude that u+=Ru−u^{+} = Ru^{-}, or equivalently

u+(x)=u−(−x‾)‾.u^{+}(x) = \overline{u^{-}(-\overline{x})}.

All points on the imaginary axis are fixed under the involution x↦−x‾x \mapsto-\overline{x}. We consider the difference

δu(x)=u+(x)−u−(x).\delta_u(x) = u^{+}(x) - u^{-}(x).

on the imaginary axis for large negative values of Im⁡(x)\operatorname{Im}(x).

First we look for a formal solution of the inner equation (14). The reversibility of the equation allows us to use an ansatz that contains only even powers:

uformal=∑k≥1akx2k.(18)u_{\mathrm{formal}} = \sum_{k\geq1}\frac{a_k}{x^{2k}}. \tag*{(18)}

After substituting the series into (14) and collecting terms of equal order in xx we obtain a recurrent system for the coefficients aka_k. The leading order gives a1=−6a_1=-6 and equation (50) provides an explicit recursive formula for aka_k with k>1k>1.

It can be shown that the formal solution provides an asymptotic expansion for both the stable and unstable solutions,

u±(x)≃uformal,(19)u^{\pm}(x)\simeq u_{\mathrm{formal}}, \tag*{(19)}

as x→∞x\to\infty in two large sectors (with opening angles greater than π\pi) centred respectively around the positive and negative real semi-axes sketched in Figure 2.

Figure 2. Sketch of the sectors D−D^{-} and D+=R(D−)D^{+}=R(D^{-}) where u±u^{\pm} are defined.

We emphasize that, although the same formal series arises in both cases, the domains in which the partial sums of uformalu_{\mathrm{formal}} approximate u±u^{\pm} are different. On the intersection of these domains, the difference δu(x)=u+(x)−u−(x)\delta_u(x)=u^{+}(x)-u^{-}(x) is beyond all algebraic orders, in the sense that δu(x)=O(x−n)\delta_u(x)=O(x^{-n}) for every n∈Nn\in\mathbb{N}. This common domain of validity contains two sectorial neighbourhoods of infinity centred around the positive and negative imaginary semi-axes (see Figure 2).

In order to obtain an estimate for δu\delta_u, we consider the variational equation

δ(iv)+δ′′+2uδ+2γ(uδ′′+u′δ′+u′′δ)=0,(20)\delta^{(iv)}+\delta''+2u\delta+2\gamma\left(u\delta''+u'\delta'+u''\delta\right)=0, \tag*{(20)}

which is obtained by linearising the inner equation (14) around a given solution uu. A direct substitution shows that δu\delta_u satisfies (20) with u=ucu=u^c, where

uc(x)=12(u+(x)+u−(x)).u^c(x)=\frac{1}{2}\left(u^{+}(x)+u^{-}(x)\right).

Note that equation (20) coincides with (13) with ε=0\varepsilon= 0 and ucu_c replaced by ucu^c.

Equation (20) is a fourth-order linear homogeneous ODE, so the space of its solutions is spanned by four linearly independent solutions. Taking into account the asymptotic behaviour of the coefficients in (20), one can show that these solutions may be chosen to behave, respectively, as x−3x^{-3}, x4x^4, and e±ixe^{\pm ix} as x→−i∞x \to-i\infty along iRi\mathbb{R}. Of these four solutions, only two decay along the negative imaginary axis, and only one is o(x−4)o(x^{-4}).

To construct an approximation to the fast decaying solution, we observe that the variational equation (20) with u=uformalu = u_{\mathrm{formal}} admits a formal solution of the form

δformal=e−ix∑k≥0bkxk.(21)\delta_{\mathrm{formal}}=e^{-ix}\sum_{k\geq0}\frac{b_k}{x^k}. \tag*{(21)}

Substituting the formal series uformalu_{\mathrm{formal}} and δformal\delta_{\mathrm{formal}} into (20) and collecting terms of equal order in xx yields a recursive system for the coefficients bkb_k. This system determines the coefficients uniquely up to a multiplicative constant, which we fix by setting b0=1b_0=1. Then the coefficients bkb_k with even kk are real and the ones with odd kk are purely imaginary.

Since both u−u^- and u+u^+ share the same asymptotic expansion on the intersection of their domains, it follows that uc∼uformalu^c \sim u_{\mathrm{formal}}. It can be shown that truncated formal series provide approximation to an analytic solutions of the variational equation. In particular, one can show that the variational equation with u=ucu=u^c admits a unique solution,

δWKB(x)∼δformal(22)\delta_{\mathrm{WKB}}(x)\sim\delta_{\mathrm{formal}} \tag*{(22)}

as x→−i∞x\to-i\infty in a sector centred on the negative imaginary semi-axis.

We note that (19) implies that δu\delta_u decays faster than any power along the imaginary axis. Therefore, since δu\delta_u is a solution of (20), it must be proportional to δWKB\delta_{\mathrm{WKB}}.

Definition 1 (Stokes constant 1) The complex constant S0=S0(γ)S_0=S_0(\gamma) such that in a neighbourhood of infinity at the negative imaginary semi-axis

u+(x)−u−(x)=S0δWKB(x)(23)u^+(x)-u^-(x)=S_0\delta_{\mathrm{WKB}}(x) \tag*{(23)}

is called a Stokes constant.

Since δWKB(x)=e−ix(1+O(x−1))\delta_{\mathrm{WKB}}(x)=e^{-ix}(1+O(x^{-1})) we get an equivalent description for S0S_0.

Proposition 1 The Stokes constant S0S_0 is equal to the following limit:

S0=lim⁡Im⁡x→−∞eix(u+(x)−u−(x)).(24)S_0=\lim_{\operatorname{Im}x\to-\infty}e^{ix}\left(u^+(x)-u^-(x)\right). \tag*{(24)}

Proposition 2. The Stokes constant S0S_0 is purely imaginary.

This property follows from the reversibility and real-analytic symmetry of the inner equation. Indeed, δu\delta_u is antisymmetric with respect to involution RR defined in (17),

(Rδu)(x)=(Ru+)(x)−(Ru−)(x)=u−(x)−u+(x)=−δu(x).(R\delta_u)(x)=(Ru^+)(x)-(Ru^-)(x)=u^-(x)-u^+(x)=-\delta_u(x).

On the other hand, it can be checked using induction that bˉk=(−1)kbk\bar{b}_k=(-1)^k b_k for all k≥0k \ge0. Therefore the formal expansion δformal\delta_{\mathrm{formal}} is symmetric, R(δformal)=δformalR(\delta_{\mathrm{formal}})=\delta_{\mathrm{formal}}. Since both the equation and asymptotic condition are invariant, the uniquely defined solution δWKB\delta_{\mathrm{WKB}} is also symmetric, i.e.

δWKB(x)=R(δWKB)(x).\delta_{\mathrm{WKB}}(x)=R(\delta_{\mathrm{WKB}})(x).

Applying RR to the identity (23) therefore yields S0=−S0‾S_0=-\overline{S_0}, implying that the Stokes constant S0S_0 is purely imaginary.

Remark 2. If S0(γ)≠0S_0(\gamma)\ne0 for some γ\gamma, we get an example of the non-linear Stokes phenomenon. Informally we can say that the unstable solution u−u^- changes its asymptotic behavior at infinity from uformalu_{\mathrm{formal}} to uformal−S0(γ)δformalu_{\mathrm{formal}}-S_0(\gamma)\delta_{\mathrm{formal}} when xx crosses the Stokes line located on the negative imaginary semi-axis. We note that a special care is needed in the interpretation of this statement as it involves divergent series. Truncating the formal series at the optimal order, i.e. at the least term for ∣x∣=σ|x|=\sigma, and denoting the corresponding partial sum of the formal series by unu_n, we visualize the Stokes phenomenon in Figure 3: the difference u−(x)−un(x)u^-(x)-u_n(x) transitions from monotone to oscillatory behaviour as xx crosses the imaginary axis following a line Im⁡x=−σ\operatorname{Im}x=-\sigma.

Figure 3. Stokes phenomenon: deviation of the unstable solution from the truncated asymptotic series. Plot of eσ(u−(x)−un(x))e^{\sigma}(u^{-}(x)-u_n(x)) with x=t−iσx=t-i\sigma versus tt, real part (left) and imaginary part (right). Parameters γ=3\gamma=3, n=20n=20, σ=40\sigma=40.

Proposition 3. The Stokes constant S0=0S_0=0 if, and only if, the formal series uformalu_{\mathrm{formal}} converge.

Suppose that S0(γ)=0S_0(\gamma)=0 for some γ\gamma. Then the equation (23) implies that u+(x)−u−(x)≡0u^+(x)-u^-(x)\equiv0 near infinity on the negative imaginary semi-axis. The principle of analytic continuation implies that u+(x)=u−(x)u^+(x)=u^-(x) for all xx where the functions are defined. The union of the domains of u+u^+ and u−u^- includes a complete neighbourhood of infinity and is not simply connected. The real-analytic symmetry can be used to show that u−=u+u^-=u^+ is single-valued in a neighbourhood of infinity. Thus its Laurent series converge and, due to the uniqueness, coincide with the formal solution uformalu_{\mathrm{formal}}. Thus the series uformalu_{\mathrm{formal}} have a finite radius of convergence.

The reverse statement also holds: if the formal series converge, the sum satisfies the inner equation and both asymptotic conditions. Consequently the sum coincides both with u+u^{+} and u−u^{-}. Thus u+≡u−u^{+}\equiv u^{-} and S0(γ)=0S_0(\gamma)=0.

These arguments imply that the Stokes constant is related to the growth rate of the coefficients in the formal solution.

Definition 2 (Stokes constant 2) The Stokes constant S1=S1(γ)S_1=S_1(\gamma) is the limit

S1=lim⁡k→∞(−1)kak(2k−1)!(25)S_1=\lim_{k\to\infty}\frac{(-1)^k a_k}{(2k-1)!} \tag*{(25)}

where aka_k are coefficients of the formal solution (18).

Remark 3 Although it is not obvious, it can be shown that

S0(γ)=−πiS1(γ).S_0(\gamma)=-\pi i S_1(\gamma).

A sketch of the proof is in Appendix B. This equality provides an efficient way for computing the Stokes constant numerically. The equation (25) can also be used to prove that the Stokes constant depends analytically on the parameter γ\gamma. Indeed, applying the induction to recursion (50), it is easy to check that ak=ak(γ)a_k=a_k(\gamma) are polynomial of degree k−1k-1 with sign-alternating coefficients. In order to show that S1(γ)S_1(\gamma) is analytic in γ\gamma for ∣γ∣<γ0|\gamma|<\gamma_0 it is sufficient to prove uniform convergence to the limit in (25).

The Stokes constant is an analytic function of the real parameter γ\gamma appearing in the inner equation (14). Since Re⁡S0(γ)≡0\operatorname{Re} S_0(\gamma)\equiv0, any change of sign of Im⁡S0(γ)\operatorname{Im} S_0(\gamma) implies the existence of a zero. By analyticity, all zeros are isolated and have finite multiplicity. Our numerical experiments suggest that all zeros are simple.

In the next section, we show that there are infinitely many zeros for γ>1\gamma> 1 and derive an asymptotic description of their location in the limit γ→+∞\gamma\to+\infty.

Asymptotics for the Stokes constant

Inner equation as a singularly perturbed problem

In order to study the dependence of the Stokes constant on the parameter γ\gamma for γ≫1\gamma\gg1 we introduce the new small parameter ϵ=γ−1/2\epsilon= \gamma^{-1/2}. The substitution

u(x)=ϵ2v(y)andy=ϵx(26)u(x) = \epsilon^2 v(y) \quad\text{and} \quad y = \epsilon x \tag*{(26)}

transforms the inner equation (14) into a singularly perturbed problem,

ϵ2v(iv)+v′′+v2+v′2+2vv′′=0,(27)\epsilon^2 v^{(iv)} + v'' + v^2 + v'^2 + 2vv'' = 0, \tag*{(27)}

where the small parameter appears in front of the highest derivative. We stress that this equation depends on γ=ϵ−2\gamma= \epsilon^{-2} only and is completely independent of small parameter ϵ\epsilon introduced in (9). From the viewpoint of singular perturbation theory, both parameters play a similar role due to the apparent similarity in the structure of equations (10) and (27).

The unstable and stable solutions of the equation (27) can be obtained by applying the scaling to the corresponding solutions of equation (14). In particular, after the substitution (26), the asymptotic condition (15) takes the form

v±(y)=−6y2+O(y−4)(28)v^{\pm}(y) = -\frac{6}{y^2} + O(y^{-4}) \tag*{(28)}

when y→∞y \to\infty in sectors centred on the positive and negative real semi-axes respectively. We use the method of complex matching to approximate the difference

δv(y)=v+(y)−v−(y)\delta_v(y) = v^+(y) - v^-(y)

and to derive the asymptotic expression (5) for the Stokes constant in the limit of small ϵ\epsilon.

Equation (27) admits a formal solution vformalv_{\mathrm{formal}}, obtained from (18) using the scaling (26). Applying induction to the recurrence relations (50), one verifies that aka_k is a polynomial of degree k−1k-1 in γ=ϵ−2\gamma= \epsilon^{-2}. Hence vformalv_{\mathrm{formal}} can be written as a double series

vformal=∑k≥00≤j≤k−1akjϵ2jy2k(29)v_{\mathrm{formal}} = \sum_{\substack{k \ge0 \\ 0 \le j \le k-1}} \frac{a_{kj}\epsilon^{2j}}{y^{2k}} \tag*{(29)}

where akj∈Ra_{kj} \in\mathbb{R}. Since only even powers of ϵ\epsilon occur, we look for a formal solution of (27) in the form

v=∑k≥0ϵ2kvk(y).v=\sum_{k\geq0}\epsilon^{2k}v_k(y).

Substitution into (27) shows that the leading term v0v_0 satisfies the ODE

v0′′+v02+v0′2+2v0v0′′=0,(30)v_0^{\prime\prime}+v_0^2+v_0^{\prime2}+2v_0v_0^{\prime\prime}=0, \tag*{(30)}

which is the equation obtained from (27) in the limit ϵ→0\epsilon\to0. In view of the asymptotics (28), we select the solution satisfying

v0(y)=−6y−2+O(y−4),y→∞.(31)v_0(y)=-6y^{-2}+O(y^{-4}),\qquad y\to\infty. \tag*{(31)}

Outer problem for the inner problem

In the terminology of the complex matching, equation (30) plays the role of an outer equation. It is a reversible non-linear differential equation of the second order. The equation is integrable. Indeed, multiplying (30) by v0′v_0', integrating and setting the integration constant to zero we get

(12+v0)v0′2+13v03=0.(32)\left(\frac{1}{2}+v_0\right)v_0^{\prime2}+\frac{1}{3}v_0^3=0. \tag*{(32)}

We look for a solution of this equation in the form v0=−12q02v_0=-\frac{1}{2}q_0^2. We fix the branch of q0q_0 by the asymptotic condition

q0(y)=23y+O(y−3).(33)q_0(y)=\frac{2\sqrt{3}}{y}+O(y^{-3}). \tag*{(33)}

After taking the square root of (32) and choosing branches of the square root to be consistent with the asymptotic condition (33) for large positive yy, we get

1−q02 q0′=−123q02.(34)\sqrt{1-q_0^2}\,q_0'=-\frac{1}{2\sqrt{3}}q_0^2. \tag*{(34)}

Dividing the equality by q02q_0^2 and integrating we get an explicit expression for the inverse of q0q_0:

1−q02q0+arcsin⁡q0=y23,(35)\frac{\sqrt{1-q_0^2}}{q_0}+\arcsin q_0=\frac{y}{2\sqrt{3}}, \tag*{(35)}

The inverse function theorem implies that this equation has a unique solution q0q_0 that maps a neighbourhood of infinity to a neighbourhood of zero. This solution is an odd analytic function and satisfies (33).

In addition to the functions u±u^{\pm}, the definition of the Stokes constant (23) involves an oscillatory solution of the variational equation (20). After applying the scaling (26), this equation takes the form

ε2η(iv)+η′′+2vη+2(vη′′+v′η′+v′′η)=0.(36)\varepsilon^{2}\eta^{(\mathrm{iv})}+\eta''+2v\eta+2(v\eta''+v'\eta'+v''\eta)=0. \tag*{(36)}

In particular, the difference δv(y)=v+(y)−v−(y)\delta_v(y)=v^{+}(y)-v^{-}(y) satisfies (36) with v=vc(y):=12(v+(y)+v−(y))v=v_c(y):=\frac{1}{2}(v^{+}(y)+v^{-}(y)). The WKB theory suggests to look for a solution in the form

ηWKB(y)=e−iε−1ϕ0(y)(η0(y)+…)(37)\eta_{\mathrm{WKB}}(y)=e^{-\mathrm{i}\varepsilon^{-1}\phi_0(y)}\left(\eta_0(y)+\ldots\right) \tag*{(37)}

where the dots stand for terms of higher orders in ε\varepsilon. Substituting this ansatz into the equation (36), taking into account that vc∼v0v_c\sim v_0 and collecting the leading terms in ε\varepsilon we get (after simplifying the expressions) that

ϕ0′2=1+2v0,2(1+2v0)η0′+3v0′η0=0.\phi_0'^2=1+2v_0,\qquad2(1+2v_0)\eta_0'+3v_0'\eta_0=0.

These equations can be solved explicitly in terms of q0q_0:

ϕ0=23(q0+q0−1),η0=(1−q02)−3/4.(38)\phi_0=2\sqrt{3}\left(q_0+q_0^{-1}\right),\qquad\eta_0=\left(1-q_0^2\right)^{-3/4}. \tag*{(38)}

Taking into account the properties of q0q_0 we conclude that ϕ0\phi_0 and η0\eta_0 are both single-valued in a neighbourhood of infinity. Then (33) implies

ϕ0(y)=y+O(y−1)andη0(y)=1+O(y−2).(39)\phi_0(y)=y+O(y^{-1})\qquad\text{and}\qquad\eta_0(y)=1+O(y^{-2}). \tag*{(39)}

Finally, we note that both ηWKB(y)\eta_{\mathrm{WKB}}(y) and δWKB(ε−1y)\delta_{\mathrm{WKB}}(\varepsilon^{-1}y) are solutions of (36) of the fastest decay along the imaginary axis. Comparing (37) with (22), we conclude that ηWKB(y)=δWKB(ε−1y)\eta_{\mathrm{WKB}}(y)=\delta_{\mathrm{WKB}}(\varepsilon^{-1}y) as their ratio is constant and converges to 1 when y→−i∞y\to-\mathrm{i}\infty.

Singularities of the outer problem

We note that the asymptotic (31) holds in a full neighbourhood of infinity; hence v0v_0 provides a common approximation to both v−v^{-} and v+v^{+}. Complex matching suggests to consider analytic continuation of solutions into neighbourhoods of the singularities of v0v_0. Therefore we need information on the location and nature of the singularities of v0=−12q02v_0=-\frac{1}{2}q_0^2.

Equation (34) is a non-linear first-order ODE and the corresponding vector field grows as O(q0)O(q_0) at the infinity. Consequently, q0q_0 admits analytic continuation along any path in the complex yy-plane such that the values of q0q_0 along this path avoid the singularities of the vector field located at q0=±1q_0=\pm1. Using that

arcsin⁡q=−ilog⁡(iq+1−q2),\arcsin q=-\mathrm{i}\log\left(\mathrm{i}q+\sqrt{1-q^2}\right),

we conclude that equation (35) implies that the singular values of the function y(q0)y(q_0) are located at the points

yk=π3(2k+1),k∈Z.y_k = \pi\sqrt{3}(2k+1), \qquad k \in\mathbb{Z}.

Consequently, q0(y)q_0(y) admits analytic continuation along any path which avoids yky_k.

For the principal branch of the inverse sine function, the left-hand side of (35) is a monotone function mapping (0,1)(0,1) onto (y∗,+∞)(y_*,+\infty), where y∗:=y0=π3y_* := y_0 = \pi\sqrt{3}. Consequently, q0q_0 is analytic on (y∗,+∞)(y_*,+\infty). Since q0q_0 is odd, it is also analytic on (−∞,−y∗)(-\infty,-y_*). Hence the principal branch of q0q_0 extends analytically to C∖[−y∗,y∗]\mathbb{C}\setminus[-y_*,y_*], i.e. the complex plane cut along the interval [−y∗,y∗][-y_*,y_*].

In order to study q0(y)q_0(y) near the singularity at y=y∗y=y_*, we recall the trigonometric identity that holds for all q∈(0,1)q\in(0,1):

arcsin⁡q=arctan⁡q1−q2=π2−arctan⁡1−q2q.\arcsin q=\arctan\frac{q}{\sqrt{1-q^2}}=\frac{\pi}{2}-\arctan\frac{\sqrt{1-q^2}}{q}.

Then (35) implies that for q0∈(0,1)q_0\in(0,1)

1−q02q0−arctan⁡1−q02q0=y23−π2.\frac{\sqrt{1-q_0^2}}{q_0}-\arctan\frac{\sqrt{1-q_0^2}}{q_0}=\frac{y}{2\sqrt{3}}-\frac{\pi}{2}.

Recalling the Maclaurin series arctan⁡x=x−x33+x55+…\arctan x=x-\frac{x^3}{3}+\frac{x^5}{5}+\ldots and applying the inverse function theorem we get that

1−q02q0=31/3(y23−π2)1/3+O(y−y∗).\frac{\sqrt{1-q_0^2}}{q_0}=3^{1/3}\left(\frac{y}{2\sqrt{3}}-\frac{\pi}{2}\right)^{1/3}+O(y-y_*).

where the remainder is a converging series in odd powers of (y−y∗)1/3(y-y_*)^{1/3}. Taking the square and resolving the result with respect to q0q_0 we get that

q0(y)=1−12(34)1/3(y−y∗)2/3+O(y−y∗)4/3.q_0(y)=1-\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}(y-y_*)^{2/3}+O(y-y_*)^{4/3}.

where the remainder is a converging power series in (y−y∗)2/3(y-y_*)^{2/3}. Therefore y=y∗y=y_* is a rational branching point with q0(y∗)=1q_0(y_*)=1.

Finally we note that the uniqueness of the initial value problem for the differential equation (34) implies that q0q_0 does not vanish. We conclude that the functions v0v_0, ϕ0\phi_0 and η0\eta_0 are analytic on C∖[−y∗,y∗]\mathbb{C}\setminus[-y_*,y_*]. Moreover

v0(y)=−12q0(y)2=−12+12(34)1/3(y−y∗)2/3+O(y−y∗)4/3(40)v_0(y)=-\frac{1}{2}q_0(y)^2=-\frac{1}{2}+\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}(y-y_*)^{2/3}+O(y-y_*)^{4/3} \tag*{(40)}
ϕ0(y)=23(q0(y)+q0(y)−1)=43+(34)7/6(y−y∗)4/3+O(y−y∗)2,(41)\phi_0(y)=2\sqrt{3}\left(q_0(y)+q_0(y)^{-1}\right) =4\sqrt{3}+\left(\frac{3}{4}\right)^{7/6}(y-y_*)^{4/3}+O(y-y_*)^2, \tag*{(41)}
η0(y)=(1−q02(y))−3/4=(43)1/4(y−y∗)−1/2(1+O(y−y∗)2/3).(42)\eta_0(y)=\left(1-q_0^2(y)\right)^{-3/4} =\left(\frac{4}{3}\right)^{1/4}(y-y_*)^{-1/2}\left(1+O(y-y_*)^{2/3}\right). \tag*{(42)}

The value of ϕ0(y∗)=43\phi_0(y_*) = 4\sqrt{3} will play an important role in the subsequent analysis.

Inner problem for the inner problem

We have established that y∗y_* is a branching point of v0v_0 and v0(y∗)=−12v_0(y_*)=-\frac{1}{2}. In order to study v+v^+ in a neighbourhood of this point, we make the substitution

v(y)=−12+ϵ1/2w(z),y=y∗+ϵ3/4z.(43)v(y)=-\frac{1}{2}+\epsilon^{1/2}w(z), \qquad y=y_*+\epsilon^{3/4}z. \tag*{(43)}

In the new variables the equation (27) takes the form

w(iv)+ϵ1/2(−12+ϵ1/2w)2+w′2+2ww′′=0.(44)w^{(iv)}+\epsilon^{1/2}\left(-\frac{1}{2}+\epsilon^{1/2}w\right)^2+w^{\prime2}+2ww^{\prime\prime}=0. \tag*{(44)}

We note that this equation represents a regular perturbation problem as the derivative of order four is no longer multiplied by the small parameter. For ϵ=0\epsilon=0 we get the inner equation for the inner equation:

w0(iv)+w0′2+2w0w0′′=0.(45)w_0^{(iv)}+w_0^{\prime2}+2w_0w_0^{\prime\prime}=0. \tag*{(45)}

We look for solutions w0±(z)w_0^\pm(z) of (45) whose behaviour at large zz matches that of v0(y)v_0(y) near y∗y_*, as described by (40). This asymptotic form suggests a formal solution of the form

wformal=z2/3∑k≥0ckz8k/3.w_{\mathrm{formal}}=z^{2/3}\sum_{k\geq0}\frac{c_k}{z^{8k/3}}.

Substitution of this series into (45) shows that the coefficients ckc_k for k≥1k\geq1 are uniquely determined once c0c_0 is fixed. Comparison with (40) then leads to the choice c0=12(34)1/3c_0=\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}.

The variational equation associated with (45) is given by

ξ(iv)+2(w0ξ′′+w0′ξ′+w0′′ξ)=0.(46)\xi^{(iv)}+2\left(w_0\xi^{\prime\prime}+w_0^{\prime}\xi^{\prime}+w_0^{\prime\prime}\xi\right)=0. \tag*{(46)}

This is a linear homogeneous ODE of order four and, for w0=wformalw_0=w_{\mathrm{formal}}, the space of its solutions is spanned by four independent solutions that respectively behave as z−1/3z^{-1/3}, z2/3z^{2/3} and z−1/2e±icz4/3z^{-1/2}e^{\pm icz^{4/3}} for large zz. The formal exponentially decaying solution has the form

ξformal=exp⁡(−i(34)7/6z4/3)z−1/2∑k≥0dkz4k/3.(47)\xi_{\mathrm{formal}}=\exp\left(-i\left(\frac{3}{4}\right)^{7/6}z^{4/3}\right)z^{-1/2}\sum_{k\geq0}\frac{d_k}{z^{4k/3}}. \tag*{(47)}

The coefficients dkd_k for k≥1k\geq1 are uniquely determined by the equation once d0d_0 is fixed. We set d0=1d_0=1. The exponential factor decays in the sector arg⁡z∈(α−,α+)\arg z\in(\alpha_-,\alpha_+), where α+=0\alpha_+=0 and α−=−3π4\alpha_-=-\frac{3\pi}{4}. The equation (45) admits two analytic solutions w0±w_0^\pm such that

w0±(z)∼wformal(z),(48)w_0^\pm(z) \sim w_{\mathrm{formal}}(z), \tag*{(48)}

as z→eiα±∞z \to e^{i\alpha^\pm}\infty, respectively. Inside the sector between these directions, the difference

δw(z)=w0+(z)−w0−(z)\delta_w(z) = w_0^+(z) - w_0^-(z)

is superalgebraically small, i.e. it decays faster than z−nz^{-n} for any n>0n > 0. A direct computation shows that δw\delta_w satisfies the variational equation (46) with w0w_0 replaced by wc:=12(w0++w0−)w_c := \frac{1}{2}(w_0^+ + w_0^-).

Since (46) with w0=wcw_0 = w_c is a linear ODE of order four, its solution space is four-dimensional. Among its solutions, only one,

ξWKB(z)∼ξformal(z),\xi_{\mathrm{WKB}}(z) \sim\xi_{\mathrm{formal}}(z),

decays faster than z−1z^{-1} along the bisecting direction of the sector. It follows that there exists a complex constant K0K_0 such that

w0+(z)−w0−(z)=δw(z)=K0ξWKB(z).(49)w_0^+(z) - w_0^-(z) = \delta_w(z) = K_0 \xi_{\mathrm{WKB}}(z). \tag*{(49)}

The equation (46) has a symmetry which involves the reflection of the zz-plane in the line arg⁡z=3π8\arg z = \frac{3\pi}{8}, the bi-section of the sector, and the reflection of the w0w_0-plane in the line arg⁡w0=3π4\arg w_0 = \frac{3\pi}{4}: if w0(z)w_0(z) is a solution then e−i3π/2w0(e−i3π/4z‾)‾e^{-i3\pi/2}\overline{w_0(e^{-i3\pi/4}\overline{z})} also is. This symmetry swaps w0+w_0^+ and w0−w_0^- and multiplies ξformal\xi_{\mathrm{formal}} and, consequently, ξWKB\xi_{\mathrm{WKB}} by ei15π/8e^{i15\pi/8}. Applying this symmetry to (49) we conclude that K0e−iπ/16K_0e^{-i\pi/16} is real. Evaluating δw\delta_w and ξWKB\xi_{\mathrm{WKB}} numerically we find that

∣K0∣=0.251346943975660732…|K_0| = 0.251346943975660732\ldots

Complex matching for the inner problem

Now we are ready to apply the complex matching to estimate the Stokes constant for equation (27) with small ϵ>0\epsilon> 0. First we note that the functions v+v^+ and v0v_0 have the same asymptotic behaviour (given by (28) and (31) respectively) at infinity in a large sector centered around the positive semi-axis. Consequently, the difference

δ+(y)=v+(y)−v0(y)\delta^+(y) = v^+(y) - v_0(y)

vanishes as O(y−4)O(y^{-4}) in the limit Re⁡(y)→+∞\operatorname{Re}(y) \to+\infty.

We recall that v+v^+ is an exact solution of (27) and while v0v_0 is an approximate one. Since the growth of small errors is governed by the variational equation it is natural to study δ+\delta^+ along a line

ℓσ={y∈C:Im⁡ϕ0(y)=σ}\ell_\sigma= \left\{ y \in\mathbb{C} : \operatorname{Im}\phi_0(y) = \sigma\right\}

as the solutions of the variational equation are oscillatory and do not grow exponentially along this line. Since ϕ0(y)≈y\phi_0(y) \approx y for large yy, it is easy to see that ℓσ\ell_\sigma with a large σ\sigma is approximately horizontal. Solving equation (38) with respect to ϕ0\phi_0 we get

q0=ϕ048(1−1−48ϕ02).q_0 = \frac{\phi_0}{\sqrt{48}}\left(1-\sqrt{1-\frac{48}{\phi_0^2}}\right).

Substituting this expression into (35) and simplifying the result we get the parametrization for ℓσ\ell_\sigma,

y=ϕ012+121−48ϕ02(1−48ϕ02)1/4+23arcsin⁡(ϕ048(1−1−48ϕ02)).y=\phi_0\sqrt{\frac{1}{2}+\frac{1}{2}\sqrt{1-\frac{48}{\phi_0^2}}\left(1-\frac{48}{\phi_0^2}\right)^{1/4}}+2\sqrt{3}\arcsin\left(\frac{\phi_0}{\sqrt{48}}\left(1-\sqrt{1-\frac{48}{\phi_0^2}}\right)\right).

Let σ=−σ0ϵ\sigma=-\sigma_0\epsilon for some σ0>0\sigma_0>0. The corresponding curve lies in the lower half-plane. In the limit ϵ→0\epsilon\to0, it converges to a contour consisting of two rays, (−∞,−π3](-\infty,-\pi\sqrt{3}] and [π3,+∞)[\pi\sqrt{3},+\infty), joined by an arc connecting −π3-\pi\sqrt{3} and π3\pi\sqrt{3}.

Let yk∈ℓσy_k\in\ell_\sigma, k=1,2k=1,2, be the intersection points of ℓσ\ell_\sigma with the circle ∣y−y∗∣=ϵ1/2|y-y_*|=\epsilon^{1/2}. These points lie in a neighbourhood of y∗y_* on opposite sides of the Stokes line defined by Re⁡ϕ0(y)=Re⁡ϕ0(y∗)\operatorname{Re}\phi_0(y)=\operatorname{Re}\phi_0(y_*), so that

Re⁡ϕ0(y1)<Re⁡ϕ0(y∗)<Re⁡ϕ0(y2).\operatorname{Re}\phi_0(y_1)<\operatorname{Re}\phi_0(y_*)<\operatorname{Re}\phi_0(y_2).

We use these points for complex matching. By (43), the circle ∣y−y∗∣=ϵ1/2|y-y_*|=\epsilon^{1/2} corresponds to ∣z∣=ϵ−1/4|z|=\epsilon^{-1/4}. Hence, it represents a neighbourhood of y∗y_* in the yy-plane and a neighbourhood of infinity in the zz-plane. The scaling is chosen so that the errors of the inner and outer approximations are of the same order.

For y∈ℓσy\in\ell_\sigma to the right of y2y_2, the function v+(y)v^+(y) remains sufficiently far from the singular value −12-\frac{1}{2}, so that the outer approximation is valid and v+(y)≈v0(y)v^+(y)\approx v_0(y). In a neighbourhood of y2y_2, we compare (40) with the leading term in (48) for w0+w_0^+ and conclude that

v0(y)≈−12+12(34)1/3(y−y∗)2/3=−12+ϵ1/212(34)1/3z2/3≈−12+ϵ1/2w0+(z).v_0(y)\approx-\frac{1}{2}+\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}(y-y_*)^{2/3}=-\frac{1}{2}+\epsilon^{1/2}\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}z^{2/3}\approx-\frac{1}{2}+\epsilon^{1/2}w_0^+(z).

Consequently near y=y2y=y_2

v+(y)≈v0(y)≈−12+ϵ1/2w0+(z).v^+(y)\approx v_0(y)\approx-\frac{1}{2}+\epsilon^{1/2}w_0^+(z).

When yy lies on ℓσ\ell_\sigma between y1y_1 and y2y_2, the solution v+(y)v^+(y) remains close to −12-\frac{1}{2}, and the inner approximation provides a more accurate description than the outer one. In a neighbourhood of y2y_2, we use (49) to switch from the stable solution w0+w_0^+ to the unstable solution w0−w_0^-:

v+(y)≈−12+ϵ1/2(w0−(z)+K0ξWKB(z)).v^+(y) \approx-\frac{1}{2}+\epsilon^{1/2}\left(w_0^-(z)+K_0\xi_{\mathrm{WKB}}(z)\right).

Then we compare (40) and the first term of (48) for w0−w_0^- and conclude that

v0(y)≈−12+12(34)1/3(y−y∗)2/3=−12+ϵ1/212(34)1/3z2/3≈−12+ϵ1/2w0−(z).v_0(y) \approx-\frac{1}{2}+\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}(y-y_*)^{2/3}=-\frac{1}{2}+\epsilon^{1/2}\frac{1}{2}\left(\frac{3}{4}\right)^{1/3}z^{2/3}\approx-\frac{1}{2}+\epsilon^{1/2}w_0^-(z).

Then we compare the approximation of ξWKB\xi_{\mathrm{WKB}} by the first term of (47) and the approximation of ηWKB\eta_{\mathrm{WKB}} by (37), (40) and (42). We get that

ξWKB(z)≈exp⁡(−i(34)7/6z4/3)z−1/2=exp⁡(−i(34)7/6ϵ−1(y−y∗)4/3)ϵ3/8(y−y∗)−1/2≈(34)1/4ϵ3/8exp⁡(−i−43+ϕ0(y)ϵ)η0(y).\begin{aligned} \xi_{\mathrm{WKB}}(z)&\approx\exp\left(-i\left(\frac{3}{4}\right)^{7/6}z^{4/3}\right)z^{-1/2}\\ &=\exp\left(-i\left(\frac{3}{4}\right)^{7/6}\epsilon^{-1}(y-y_*)^{4/3}\right)\epsilon^{3/8}(y-y_*)^{-1/2}\\ &\approx\left(\frac{3}{4}\right)^{1/4}\epsilon^{3/8}\exp\left(-i\frac{-4\sqrt{3}+\phi_0(y)}{\epsilon}\right)\eta_0(y). \end{aligned}

Consequently, near y2y_2

v+(y)≈v0(y)+ϵ7/8K0(34)1/4exp⁡(i43ϵ)ηWKB(y).v^+(y)\approx v_0(y)+\epsilon^{7/8}K_0\left(\frac{3}{4}\right)^{1/4}\exp\left(i\frac{4\sqrt{3}}{\epsilon}\right)\eta_{\mathrm{WKB}}(y).

Taking into account that small errors in initial conditions follow solutions of the variational equation we get that this approximations extends to a neighbourhood of intersection of ℓσ\ell_\sigma with the imaginary axis. Then we apply the reversor to get the unstable solution. For a point yy on the imaginary axis y‾=−y\overline{y}=-y and, consequently,

v−(y)=v+(−y‾)‾=v+(y)‾≈v0(y)+ϵ7/8K0‾(34)1/4exp⁡(−i43ϵ)ηWKB(y).\begin{aligned} v^-(y)=\overline{v^+(-\overline{y})}=\overline{v^+(y)} \\ &\approx v_0(y)+\epsilon^{7/8}\overline{K_0}\left(\frac{3}{4}\right)^{1/4}\exp\left(-i\frac{4\sqrt{3}}{\epsilon}\right)\eta_{\mathrm{WKB}}(y). \end{aligned}

Taking the difference of the last two approximations and using K0=eiπ16∣K0∣K_0=e^{i\frac{\pi}{16}}|K_0|, we obtain that on the imaginary axis

v+(y)−v−(y)≈ϵ7/8∣K0∣(34)1/42isin⁡(π16+43ϵ)ηWKB(y).v^+(y)-v^-(y)\approx\epsilon^{7/8}|K_0|\left(\frac{3}{4}\right)^{1/4}2i\sin\left(\frac{\pi}{16}+\frac{4\sqrt{3}}{\epsilon}\right)\eta_{\mathrm{WKB}}(y).

Comparing the last estimate with the definition (23) of the Stokes constant, and recalling that u(x)=ϵ2v(y)u(x)=\epsilon^2v(y) and ξWKB(x)=ηWKB(y)\xi_{\mathrm{WKB}}(x)=\eta_{\mathrm{WKB}}(y), we obtain

S0≈ϵ2+7/8∣K0∣(34)1/42isin⁡(π16+43ϵ).S_0\approx\epsilon^{2+7/8}|K_0|\left(\frac{3}{4}\right)^{1/4}2i\sin\left(\frac{\pi}{16}+\frac{4\sqrt{3}}{\epsilon}\right).

Finally, recalling that ϵ=γ−1/2\epsilon=\gamma^{-1/2}, we recover the asymptotic expression (5).

Numerical methods for evaluation of the Stokes constant

Methods based on analysis of the formal solution

The most efficient method for computing the Stokes constant is based on the analysis of the coefficients aka_k of the formal solution (18) to the inner equation (14). These coefficients are defined by the recurrence relation with initial condition a1=−6a_1=-6 and

ak=−12(2k2+k−6)((2k−2)(2k−1)(2k)(2k+1)ak−1+∑j=2k−1ajak+1−j+4γ∑j=1k−1j(k+1+j)ajak−j),for k≥2,(50)a_k = \frac{-1}{2(2k^2+k-6)}\left((2k-2)(2k-1)(2k)(2k+1)a_{k-1}+\sum_{j=2}^{k-1}a_j a_{k+1-j}+4\gamma\sum_{j=1}^{k-1}j(k+1+j)a_j a_{k-j}\right),\quad\text{for } k\ge2, \tag*{(50)}

where the first of the sums at the rhs of (50) must be omitted for k=2k=2. The induction in kk implies that this equation defines aka_k which are polynomial of degree k−1k-1 in γ\gamma with sign-alternating coefficients, and satisfy ak(5/4)=0a_k(5/4)=0 for all k>1k>1.

For large kk, the coefficients exhibit Gevrey-one growth with a sign-alternating tail. We introduce the normalized coefficients

a^k=(−1)kak(2k−1)!,(51)\hat{a}_k=\frac{(-1)^k a_k}{(2k-1)!}, \tag*{(51)}

and aim to compute the Stokes constant as the limit

S1=lim⁡k→∞a^k.S_1=\lim_{k\to\infty}\hat{a}_k.

A typical behaviour of a^k\hat{a}_k for γ>1\gamma>1 is shown in Figure 4 (left), where we take γ=42\gamma=4^2 as an example. The tail of the sequence is monotone and decreases in absolute value. Plotting a^k\hat{a}_k as a function of k−1k^{-1} (Figure 4, right), we observe apparent convergence to a finite limit. The last of the computed terms of the series can be taken as an approximate value of the limit. Much higher accuracy can be achieved by extrapolation in powers of k−1k^{-1}.

Figure 4. Normalized coefficients of the formal solution for γ=42\gamma= 4^2: ∣a^k∣|\hat{a}_k| for 1≤k≤40001 \le k \le4000 in log–log scale (left) and −a^k-\hat{a}_k versus k−1k^{-1} (right); the thin discontinuous line shows an extrapolation to zero.

To this end, we proceed as follows. We identify two consecutive intervals [k0,k0+m][k_0,k_0+m] and [k0+m,k0+2m][k_0+m,k_0+2m] on which the sequence a^k\hat{a}_k is monotone. On each interval, we construct interpolating polynomials pm0p_{m0} and pm1p_{m1} of degree mm using the data (k−1,a^k)(k^{-1},\hat{a}_k). We take pm1(0)p_{m1}(0) as an estimate of S1S_1, provided that the predicted values are in good agreement:

∣pm1(0)−pm0(0)∣<10−50.(52)\left|p_{m1}(0)-p_{m0}(0)\right|<10^{-50}. \tag*{(52)}

If this criterion is not satisfied, we increase k0k_0 and repeat the procedure. We use the difference (52) as a proxy for the error in evaluation of S1S_1.

The computation time depends on three factors: the starting point k0k_0 that determines the number of coefficients a^k\hat{a}_k to be computed, the number of correct digits NN in numerical values of a^k\hat{a}_k and the degree mm of the interpolating polynomial.

As an example, we consider γ=42\gamma= 4^2. Using the tail of 4000 coefficients and an interpolating polynomial of degree 80, we obtain an estimate of S1(42)S_1(4^2) with an error below 10−18010^{-180}. This value is then used as a reference in subsequent, less accurate computations.

We perform two series of experiments based on the tails of the first 400 and 4000 coefficients a^k\hat{a}_k, respectively, and compare the estimates for S1(γ)S_1(\gamma) obtained using interpolating polynomials of varying degree. The deviation from the reference value for S1(42)S_1(4^2) is shown in Figure 5. We observe that the error initially decreases rapidly as the polynomial degree mm increases, but eventually begins to grow.

Figure 5. Accuracy of the extrapolation to zero as a function of the extrapolating polynomial degree mm using the tail of the first 4 000 coefficients (blue) and 400 coefficients (red).

In particular, using the tail of the first 400 coefficients and an interpolating polynomial of degree m=100m = 100, we obtain an estimate of the Stokes constant which deviates by less than 10−10010^{-100} from the reference value. The experiment confirms the numerical stability of the extrapolation procedure, with estimates of S1S_1 based on different intervals and different degrees of extrapolating polynomials agreeing up to over 100 decimal digits.

In this experiment we also check that (52) provides a reliable indicator of the extrapolation error.

We emphasize that the use of high-precision arithmetic is essential for the reliable implementation of this method, for two closely related reasons.

First, repeated applications of the recurrence relation (50) lead to a significant loss of numerical precision. For instance, comparing computations performed with working precision N=500N = 500 and N=1000N = 1000, we observe that approximately 180 digits are lost in the evaluation of a^k\hat{a}_k for k≤1000k \le1000 and γ=1002\gamma= 100^2. Second, the extrapolation using a high-degree polynomial is numerically stable only when the input data is computed with sufficiently high accuracy.

In the majority of our computation we use m=50m = 50 to find k0k_0 suitable for numerically reliable extrapolation. Then we use m=100m = 100 to get a high precision approximation of S1S_1.

The results of these computations are presented in Figure 6. They show that S1(γ)S_1(\gamma) is positive and monotonically decreasing for γ<54\gamma< \frac{5}{4}. It has its first zero at γ=5/4\gamma= 5/4. Beyond this point, the behaviour becomes oscillatory.

Figure 6. Stokes constant S1S_1 as a function of γ\gamma (left), and the normalized value g23/8S1(g2)g^{23/8}S_1(g^2) as a function of g=γ1/2g=\gamma^{1/2} on the interval which contains the first 26 zeroes of S1S_1 (right).

To illustrate the oscillatory behavior, we adopt the scaling suggested by the asymptotic estimate (5) and plot g23/8S1(g2)g^{23/8}S_1(g^2) as a function of g=γ−1/2g = \gamma^{-1/2}. For large gg, the data clearly indicate convergence to a periodic function. A more accurate analysis of maximal and minimal values confirms correctness of our asymptotic formula for the Stokes constant.

Zeroes of the Stokes constant

We use the behavior of the normalized coefficients a^k\hat{a}_k to study the dependence of the Stokes constant S1S_1 on γ\gamma. It is easy to see that a^k\hat{a}_k is polynomial of degree k−1k - 1. For example,

a^1=6,a^2=−3(4γ−5),a^3=320(64γ2−204γ+155).\hat{a}_1 = 6,\qquad\hat{a}_2 = -3(4\gamma- 5),\qquad\hat{a}_3 = \frac{3}{20}(64\gamma^2 - 204\gamma+ 155).

The polynomials with k>1k > 1 share the common root at γ=5/4\gamma= 5/4. For this special value of γ\gamma, u(x)=−6x−2u(x) = -6x^{-2} solves the inner equation (14) exactly. Figure 7 illustrates convergence of aka_k to the limit.

Figure 7. Normalized coefficients a^k\hat{a}_k as functions of g=γ1/2g=\gamma^{1/2} for k=30k=30 (green) and k=100k=100 (blue), and the limiting curve S1(g2)S_1(g^2) (red).

We observe that the zeros of the Stokes constant S1(γ)S_1(\gamma) can be located without relying on the extrapolation procedure. Indeed, suppose that S1(γ)≠0S_1(\gamma) \ne0 for some γ\gamma. Then there exists k0(γ)k_0(\gamma) such that

sign⁡S1(γ)=sign⁡a^k(γ)for all k≥k0(γ).\operatorname{sign} S_1(\gamma) = \operatorname{sign} \hat{a}_k(\gamma) \quad\text{for all } k \ge k_0(\gamma).

Therefore, if

a^k(γ1)a^k(γ2)<0\hat{a}_k(\gamma_1)\hat{a}_k(\gamma_2) < 0

for some k>min⁡(k0(γ1),k0(γ2))k > \min\left(k_0(\gamma_1), k_0(\gamma_2)\right), then the interval (γ1,γ2)(\gamma_1,\gamma_2) contains a zero of S1(γ)S_1(\gamma). In our computations we refine such intervals until ∣γ2−γ1∣<10−100|\gamma_2-\gamma_1| < 10^{-100}.

This approach has several advantages. First, it does not depend on the stability of the extrapolation procedure. Second, it is computationally more efficient, as it requires significantly fewer coefficients.

We use the following procedure to determine the nthn^{\mathrm{th}} root of S1(γ)S_1(\gamma). As an initial approximation (i=0i=0), we take γ0=γ0(n)\gamma_0=\gamma_0(n) suggested by the asymptotic formula (6). This approximation works well for all n≥1n\geq1. At each subsequent step, we employ the recurrence relation (50) and its derivative with respect to γ\gamma to compute ak(γi)a_k(\gamma_i) and ak′(γi)a'_k(\gamma_i) till we detect the onset of the monotone tail for a^k(γi)\hat{a}_k(\gamma_i) with k≥k0(γi)k\geq k_0(\gamma_i). Then we perform one step of Newton’s method:

γi+1=γi−ak(γi)ak′(γi)\gamma_{i+1}=\gamma_i-\frac{a_k(\gamma_i)}{a'_k(\gamma_i)}

in order to find γ\gamma with a smaller value of aka_k. This procedure is iterated until ∣γi+1−γi∣<10−100|\gamma_{i+1}-\gamma_i|<10^{-100}, which is typically achieved in less than 10 steps. We let γ∗(n):=γi+1\gamma_*(n):=\gamma_{i+1}. As a final check, we verify that

ak(γ∗(n)+10−101)ak(γ∗(n)−10−101)<0a_k\left(\gamma_*(n)+10^{-101}\right)a_k\left(\gamma_*(n)-10^{-101}\right)<0

for kk in the tail region.

Figure 8 shows the plot of the normalized deviation of the computed zeroes of S1(γ)S_1(\gamma) from their theoretical prediction,

Figure 8. The first 440 zeroes of S1S_1: deviation from the predicted value, Δn\Delta_n, in the units of the limit root spacing vs nn.

Δn=(γ∗(n)1/2−γ0(n)1/2)48π.\Delta_n=\left(\gamma_*(n)^{1/2}-\gamma_0(n)^{1/2}\right)\sqrt{\frac{48}{\pi}}.

The difference apparently converge to zero as n−1/2n^{-1/2}. This conclusion confirms our theoretical prediction for the position of zeroes given by equation (6).

Method based on the variational equation

A more direct computation of the Stokes constant is based on numerical evaluation of the difference between the stable and unstable solutions of the inner equation on the negative imaginary axis and expressing this difference in terms of the exponentially decaying solution of the variational equation as defined in (23).

We construct an approximation for the unstable solution u−u^- to the inner equation using a method which is routinely used in dynamical system for construction of unstable manifolds. First we compute the coefficients of the formal solution (18) to the inner equation (14) using the recurrence relation (50), and define the truncated series

un(x)=∑k=1nakx2k.(53)u_n(x)=\sum_{k=1}^{n}\frac{a_k}{x^{2k}}. \tag*{(53)}

Then we fix constants xR>σ0>0x_R > \sigma_0 > 0 and set the initial point at x0=−xR−iσ0x_0=-x_R-i\sigma_0. Then we solve numerically the Cauchy problem for the inner equation (14) with initial conditions

u(x0)=un(x0),u′(x0)=un′(x0),u′′(x0)=un′′(x0),u′′′(x0)=un′′′(x0),(54)u(x_0)=u_n(x_0),\quad u'(x_0)=u'_n(x_0),\quad u''(x_0)=u''_n(x_0),\quad u'''(x_0)=u'''_n(x_0), \tag*{(54)}

where nn is chosen according to the rule of truncation at the least term. The approximation to the unstable solution u−u^- is numerically computed along the line Im⁡(x)=−σ0\operatorname{Im}(x)=-\sigma_0 in the complex plane and provides a good approximation to u−(x)u^-(x) for −xR≤Re⁡x≤xR-x_R \le\operatorname{Re} x \le x_R.

The stable solution is defined using the reversible symmetry, u+(x)=−u−(−x‾)u^+(x)=-u^-(-\overline{x}). On the imaginary axis the difference δ=u+−u−\delta=u^+-u^- is purely imaginary:

δ(−iσ0)=u−(−iσ0)−u−(−iσ0)‾=2iIm⁡u−(−iσ0).\delta(-i\sigma_0)=u^-(-i\sigma_0)-\overline{u^-(-i\sigma_0)}=2i\operatorname{Im}u^-(-i\sigma_0).

Then we find coefficients of the formal solution (21) of the variational equation (20) with u=unu=u_n and define a partial sum

ξm(x)=∑k=1mbke−ixxk.(55)\xi_m(x)=\sum_{k=1}^{m}\frac{b_k e^{-ix}}{x^k}. \tag*{(55)}

Choosing mm from the optimal truncation rule for x=−iσ0x=-i\sigma_0 we obtain an approximation for the Stokes constant

S0≈δ(−iσ0)ξm(−iσ0).(56)S_0\approx\frac{\delta(-i\sigma_0)}{\xi_m(-i\sigma_0)}. \tag*{(56)}

This method yields results in excellent agreement with those obtained from the analysis of the coefficients. The theory predicts that the two definitions of the Stokes constant are related by

S0(γ)=−πiS1(γ).S_0(\gamma) = -\pi i S_1(\gamma).

Our numerical experiments confirm this relation with an error below 10−1410^{-14}. The method described in this section has a more direct and transparent connection to the Stokes phenomenon. However, achieving high precision is more challenging, as it requires the computation of highly accurate solutions to the Cauchy problem for the underlying ODE, which is computationally expensive.

The best precision achievable by this method is governed by the balance between two competing effects. On the one hand, the quantity δ\delta decays exponentially when σ0\sigma_0 increases, while the accuracy of its evaluation is limited by the numerical precision of the solution to the Cauchy problem. On the other hand, the truncated-series approximation of the solution to the variational equation improves in accuracy as σ0\sigma_0 increases and is reliable only for sufficiently large values of σ0\sigma_0.

Method based on the limit behaviour

In this section, we describe an alternative method based on the definition of S0S_0 as the limit (24), rather than on the ratio (56). An important advantage of this method is that it does not require knowledge of the solution to the variational equation. However, it requires a very-high-precision Taylor-series integrator to reduce errors in solving the Cauchy problem in the complex time plane to an acceptable level.

Algorithm:

  1. For a given value of γ\gamma we fix σ0>0\sigma_0 > 0 and consider a large xR>0x_R > 0 and define x0=−xR−iσ0x_0 = -x_R - i\sigma_0. Then we numerically solve the Cauchy problem for the inner equation (14) with initial conditions (54): integrating the equation during time tRt_R, we obtain a very accurate approximation for u−(−iσ0)u^{-}(-i\sigma_0) and its derivatives at this point.

  1. Then we choose σf>σ0\sigma_f > \sigma_0 and integrate the inner equation (14) in the purely imaginary direction substituting x=−iσx = -i\sigma and using σ\sigma as the new independent variable. We use the results of the previous step to generate the initial condition for the Cauchy problem with σ=σ0\sigma= \sigma_0 and integrate the equation numerically till σ=σf\sigma= \sigma_f.

  1. Once performed the numerical integration we use equation (24) to get an approximation for the Stokes constant,

S0(γ)=lim⁡Im⁡x→−∞eix(u+(x)−u−(x))=2lim⁡σ→+∞eσIm⁡(u+(−iσ)−u−(−iσ))≈2eσfIm⁡(u−(−iσf)).(57)\begin{aligned} S_0(\gamma) &= \lim_{\operatorname{Im} x \to-\infty} e^{ix}\left(u^+(x)-u^-(x)\right) \\ &= 2\lim_{\sigma\to+\infty} e^\sigma\operatorname{Im}\left(u^+(-i\sigma)-u^-(-i\sigma)\right) \\ &\approx2e^{\sigma_f}\operatorname{Im}\left(u^-(-i\sigma_f)\right). \tag*{(57)} \end{aligned}

The integration path used in Steps 1 and 2 is sketched in Figure 9.

Integration paths in the complex plane

Figure 9. Integration paths in the complex plane.

The algorithm has been implemented in multiple precision. The numerical tests have been done using an optimized variable-step size variable-order Taylor series integrator (the free software TIDES 12 [1, 4]) in combination with the GNU multiple precision libraries GMP [16] and MPFR [10]. Note that Taylor methods are the only ODE solvers able to reach very high precision levels [2, 3].

The above algorithm involves several parameters that must be chosen carefully: the displacement used to determine the initial condition

x0=−xR−iσ0x_0=-x_R-i\sigma_0

the number of terms nn retained in the approximation of the initial conditions (53) (thus, the parameters associated with the first path are xRx_R, σ0\sigma_0, and nn), and the final imaginary time σf\sigma_f along the second path. Maintaining high accuracy requires a careful choice of all these parameters and, consequently, extensive numerical testing.

In Figure 10, we show the estimates for the Stokes constant computed with σ0=15\sigma_0 = 15, which our tests indicate is sufficiently large, using either 200 or 1000 decimal digits of precision and several values of σf\sigma_f and xRx_R. As γ\gamma increases, obtaining an accurate value of the Stokes constant becomes considerably more difficult. This is due both to the loss of accuracy in the numerical integration and/or the initial data, unless sufficiently high precision is used, and to the increasingly slow convergence, as illustrated in Figure 11.

Figure 10. The estimates of Stokes constant S0(γ)S_0(\gamma) for σ0=15\sigma_0 = 15 using arithmetics with 200 and 1000 decimal digits and different values of σf\sigma_f and xRx_R.

Figure 11. Approximation of the Stokes constant approaching the limit value for γ=20\gamma= 20 (left) and γ=80\gamma= 80 (right) using different precision levels (digits = dd) and different initial times (xRx_R) for the numerical evaluation.

For γ=20\gamma= 20, computations with 25 decimal digits exhibit strong oscillations when the imaginary time σ\sigma along the second path approaches σf=36\sigma_f = 36, resulting in a highly inaccurate approximation of the Stokes constant. Increasing the precision to 50 digits delays the onset of the oscillations until σf≃62\sigma_f \simeq62, whereas 200 digits are required to maintain adequate accuracy over the range considered. For γ=80\gamma= 80, even higher precision is necessary, together with a sufficiently large value of the initial real displacement xRx_R.

High precision in the initial condition (IC) is crucial for obtaining an accurate value of the Stokes constant, although it is less important for locating its zeros γnum(i)\gamma_{\mathrm{num}}^{(i)}, as shown in Figure 10. Table 1 presents the number of correct digits in the numerical determination of the 21st zero of the Stokes constant, γnum(21)\gamma_{\mathrm{num}}^{(21)}, for several values of σf\sigma_f, xRx_R, nn, and the working precision. In each case, the quantity marked with an asterisk is the factor that limits the number of correct digits. These results show that an appropriate balance between the working precision and the accuracy of the initial condition is essential.

precision digitsσf\sigma_fxRx_RJJcorrect digits γnum(21)\gamma_{\mathrm{num}}^{(21)}
20080∗80^{*}100070134
20080100040∗40^{*}122
20080100020∗20^{*}4
20080100∗100^{*}70∗70^{*}9
20060100070173
200∗200^{*}80∗80^{*}500070134
100∗100^{*}60100035∗35^{*}53
50∗50^{*}60100020∗20^{*}3

Table 1. Number of correct digits in the numerical determination of the 21st zero of the Stokes constant (γnum(21)\gamma_{\mathrm{num}}^{(21)}) for various values of σf\sigma_f, xRx_R, JJ and precision digits.

This also requires a suitable choice of nn and xRx_R, since the initial point must be sufficiently far from the origin. Moreover, choosing a large value of σf\sigma_f to assess convergence to the Stokes constant requires correspondingly higher precision; otherwise, numerical errors grow and may produce less accurate results than those obtained with a smaller value of σf\sigma_f.

Based on the preceding tests, and recognizing the need to balance computational cost against numerical accuracy, we selected the following parameter values: d=200d = 200 digits of precision, n=70n = 70, xR=1000x_R = 1000, and σf=80\sigma_f = 80. Table 2 lists the first 22 zeros of the Stokes constant, each computed to 80 correct digits. These values were independently verified using a higher-precision computation with d=500d = 500 digits of precision, n=100n = 100, xR=5000x_{\mathrm{R}} = 5000, and σf=90\sigma_f = 90.

iiγnum(i)\gamma_{\mathrm{num}}^{(i)}
11.251.25
22.401049919639515487394205386216120634813759442477543869728648711149740720627580e+002.401049919639515487394205386216120634813759442477543869728648711149740720627580\mathrm{e}{+00}
33.930781055958246541778147654830619930170526589023534774773406052607795653141820e+003.930781055958246541778147654830619930170526589023534774773406052607795653141820\mathrm{e}{+00}
45.852175445170598283928616134085791767270340934872156470428868607330389222422947e+005.852175445170598283928616134085791767270340934872156470428868607330389222422947\mathrm{e}{+00}
58.171452010273213282253122677012053565621676960419593282409730035589907076181711e+008.171452010273213282253122677012053565621676960419593282409730035589907076181711\mathrm{e}{+00}
61.089222463289445398270765765983813970333283788702073647715940847811222347356477e+011.089222463289445398270765765983813970333283788702073647715940847811222347356477\mathrm{e}{+01}
71.401680066326203665412939722195407894879580483167280665124849956936106476459307e+011.401680066326203665412939722195407894879580483167280665124849956936106476459307\mathrm{e}{+01}
81.754674662406976296481029163323071683210209381245467456232394192894622165194123e+011.754674662406976296481029163323071683210209381245467456232394192894622165194123\mathrm{e}{+01}
92.148317536108186990216898283507374638983245347717053332798303126444890472248817e+012.148317536108186990216898283507374638983245347717053332798303126444890472248817\mathrm{e}{+01}
102.582690598464499793658543836930884706796068421901201419110303533400660368996520e+012.582690598464499793658543836930884706796068421901201419110303533400660368996520\mathrm{e}{+01}
113.057855900716822187495640835544144125961522482541362008239185246549622460852201e+013.057855900716822187495640835544144125961522482541362008239185246549622460852201\mathrm{e}{+01}
123.573861586739780457409126844937310873553227634093720444728233382699098112907917e+013.573861586739780457409126844937310873553227634093720444728233382699098112907917\mathrm{e}{+01}
134.130745772377472838094582156638522661775479887634910347718985055263149623896948e+014.130745772377472838094582156638522661775479887634910347718985055263149623896948\mathrm{e}{+01}
144.728539161014315214863834028085565871951097889866300564301068908172230572247931e+014.728539161014315214863834028085565871951097889866300564301068908172230572247931\mathrm{e}{+01}
155.3672668586848729956826334696090521762548476699868734674493271657528979338213281e+015.3672668586848729956826334696090521762548476699868734674493271657528979338213281\mathrm{e}{+01}
166.046949665426738133424415888966264863522612829910863019332850383006651406152488e+016.046949665426738133424415888966264863522612829910863019332850383006651406152488\mathrm{e}{+01}
176.76760501417812051439092310398722795062622176380835836981450981438083384645476685e+016.76760501417812051439092310398722795062622176380835836981450981438083384645476685\mathrm{e}{+01}
187.529247666606744212098373234109673095059304533634144776289061390477385198353880e+017.529247666606744212098373234109673095059304533634144776289061390477385198353880\mathrm{e}{+01}
198.331890237638685534133062359466256029977746734129818523264416732222084919042499e+018.331890237638685534133062359466256029977746734129818523264416732222084919042499\mathrm{e}{+01}
209.175543596915918363502484254063757642763319622100450702561676249884517762602421e+019.175543596915918363502484254063757642763319622100450702561676249884517762602421\mathrm{e}{+01}
211.006021718029178275020516133573017787027302673555997412932160876185711739569063e+021.006021718029178275020516133573017787027302673555997412932160876185711739569063\mathrm{e}{+02}
221.098591923453379924653981838458246921764827176552002828979475464931035514144293e+021.098591923453379924653981838458246921764827176552002828979475464931035514144293\mathrm{e}{+02}

Table 2. The first 22 zeros of the Stokes constant with 80 correct digits (the numerical process has been done with 200 precision digits, n=70n = 70, xR=1000x_{\mathrm{R}} = 1000 and σf=80\sigma_f = 80).

The zeros were determined using Brent’s method (see [5]), which combines bisection, the secant method, and inverse quadratic interpolation. A quadratic least-squares fit to the computed zeros (Figure 12) yields

Figure 12. First 22 zeros of the Stokes constant and a quadratic fit of the points in log-log scale.

γnum(i)≈0.7391+0.4646i+0.20431i2,∥error∥2=0.13687.\gamma_{\mathrm{num}}^{(i)} \approx0.7391 + 0.4646i + 0.20431i^2,\qquad\lVert\mathrm{error}\rVert_2 = 0.13687.

As shown in Figure 12, the quadratic approximation provides an excellent fit, with the error decreasing as ii increases.

Finally, the Stokes constant vanishes if and only if

Im⁡u−(−iσ)=0for all σ≥σ0.\operatorname{Im} u^{-}(-i\sigma) = 0 \qquad\text{for all } \sigma\ge\sigma_0.

For such parameter values, u−u^{-} is real along the imaginary axis. Therefore, if the aim is solely to locate the zeros, it suffices to integrate along the first path. The accuracy of a computed zero can then be verified by confirming that Im⁡u−(−iσ0)\operatorname{Im} u^{-}(-i\sigma_{0}) changes sign across the endpoints of a sufficiently small interval in γ\gamma containing the zero.

Conclusions

We have shown that the existence of a symmetric embedded soliton is determined by the zeros of a single parameter-dependent constant. This constant is the proportionality coefficient in the relation expressing the difference between the stable and unstable solutions in terms of a normalized solution of the variational equation. Using complex matching, we approximated this coefficient in terms of a Stokes constant associated with the inner problem. For large values of the parameter γ\gamma, the inner equation itself becomes singularly perturbed; we therefore applied the complex matching procedure a second time to derive an oscillatory asymptotic formula for the Stokes constant. We proved that this constant has countably many zeros that are asymptotically periodic in γ1/2\gamma^{1/2}. Finally, we verified these asymptotic predictions by means of high-precision numerical computations.

Acknowledgements

The authors are grateful to Professor Alan Champneys, who introduced us to the problem of embedded solitons and brought together the team that completed the work presented in this paper.

RB has been supported by the project PID2024-156032NB-I0 funded by MICIU/AEI/10.13039/501100011033 and by ERDF, EU. JTL has been supported by the grant PID2024-158570NB-I00 funded by MICIU/AEI/ and ERDF, UE, and also by the Spanish State Research Agency, through the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D (CEX2020-001084-M).

Hamiltonian structure

The equation (7) has two properties important for our analysis. First, the equation has a time-reversing symmetry, i.e., if a function u(t)u(t) satisfies the equation, then u(−t)u(-t) is also a solution of the same equation. Second, the equation has a first integral and is Hamiltonian with a non-standard symplectic form.

Let us describe the Hamiltonian structure in more details. Denoting by uku_k the kthk^{\mathrm{th}}-derivative of uu, we rewrite the equation (7) in the form of a system of four first order equations:

u0′=u1,u1′=u2,u2′=u3,u'_0 = u_1,\qquad u'_1 = u_2,\qquad u'_2 = u_3,
a2u3′+a1u2+a0u+b0u2+b1(u12+2u0u2)=0.(58)a_2u'_3 + a_1u_2 + a_0u + b_0u^2 + b_1(u_1^2 + 2u_0u_2) = 0. \tag*{(58)}

This system is hamiltonian with the Hamilton function

h=a22(2u3u1−u22)+a12u12+a02u02+b03u03+b1u0u12(59)h = \frac{a_2}{2}(2u_3u_1-u_2^2) + \frac{a_1}{2}u_1^2 + \frac{a_0}{2}u_0^2 + \frac{b_0}{3}u_0^3 + b_1u_0u_1^2 \tag*{(59)}

and the non-standard symplectic form

ω=a2(du3∧du0+du1∧du2)+(a1+2b1u0)du1∧du0.(60)\omega= a_2(du_3 \wedge du_0 + du_1 \wedge du_2) + (a_1 + 2b_1u_0)du_1 \wedge du_0. \tag*{(60)}

Indeed, let XhX_h denote the vector field in the right hand side of the system (58). A direct substitution shows that XhX_h satisfies the identity dh(ξ)=ω(ξ,Xh)dh(\xi) = \omega(\xi, X_h) for all vectors ξ=∑k=14ξk∂uk\xi= \sum_{k=1}^{4}\xi_k\partial_{u_k} where ∂uk\partial_{u_k} denotes the differentiation with respect to the variable uku_k. Recall that duk(∂uj)=δkjdu_k(\partial_{u_j}) = \delta_{kj} where δkj\delta_{kj} is the Kronecker delta.

Soliton solutions belong to the zero level set of the first integral hh. Note that the equality h=0h = 0 can be interpreted as an ordinary differential equation of order three. The equation has to be satisfied by every solution of (7) that converges to zero at infinity.

We note that the symplectic form can be easily transformed into the canonical form usually used in applications of the classical mechanics.

Borel-Laplace transform and the Stokes constant

The choice of normalisation for the coefficients aka_k in the definition of S1S_1 as well as its connection to S0S_0 deserve an explanation. The factorial growth rate for aka_k can be easily proved with the help of the induction and recurrent formulae (50). The relation between S0S_0 and S1S_1 can be explained provided with the help from Ecalle’s theory of resurgent functions. At the centre of this argument is the observation that the functions u+u^+ and u−u^- coincide with Borel-Laplace sums of the formal series uformalu_{\mathrm{formal}}. In this interpretation the difference between u+u^{+} and u−u^{-} is due to different choices of directions for Laplace integrals.

Let us discuss this point in more details. It is possible to look for a solution of the equation (14) in the form of a Laplace integral in a direction α\alpha,

uα(x)=∫0eiα∞e−sxv(s) ds(61)u_{\alpha}(x)=\int_{0}^{e^{i\alpha}\infty}e^{-sx}v(s)\,\mathrm{d}s \tag*{(61)}

where the integration path is a ray that forms the angle α\alpha with the positive real axis. We chose α=0\alpha=0 for u+u^{+} and α=π\alpha=\pi for u−u^{-} and assume that vv is the same analytic functions in both cases.

Suppose that near the origin

v(s)=∑k=1∞aks2k−1(2k−1)!.v(s)=\sum_{k=1}^{\infty}\frac{a_k s^{2k-1}}{(2k-1)!}.

The growth rate of aka_k implies that the series in the right hand side converge for ∣s∣<1|s|<1. Let (u∗v)(s)=∫0su(t)v(s−t) dt(u*v)(s)=\int_{0}^{s}u(t)v(s-t)\,\mathrm{d}t be the convolution of functions. The Laplace transforms maps the convolution into the product and multiplication by ss into ∂x\partial_x. Therefore the Taylor series formally satisfy the integral equation

(s4+s2)v(s)+(v∗v)(s)+γ((sv)∗(sv))(s)+2γ(v∗(s2v))(s)=0(s^4+s^2)v(s)+(v*v)(s)+\gamma((sv)*(sv))(s)+2\gamma(v*(s^2v))(s)=0

and, since the series converge, vv is an analytic solution of the same equation.

The integral equation can be used to prove that the analytic continuation of vv along any ray with α≠±π2(mod2π)\alpha\ne\pm\frac{\pi}{2}\pmod{2\pi} does not grow faster than an exponential function. We conclude that the Laplace integrals converge and uαu_{\alpha} satisfies equation (14). Then integrating by parts we can easily check that uα∼uformalu_{\alpha}\sim u_{\mathrm{formal}} as the asymptotic behaviour of the Laplace integral for large xx is determined by the Taylor series of vv around s=0s=0. The domain of validity of the asymptotic relation depends on α\alpha, and we conclude that our choices of α\alpha indeed lead to representation of the stable and unstable solution u±u^{\pm} in the form of the Laplace integrals. Then for all sufficiently large xx located in a neighbourhood of the negative imaginary semi-axis we can write

δ(x)=u+(x)−u−(x)=uα(x)−uπ−α(x)=∫e−iα∞eiα∞e−sxv(s) ds(62)\delta(x)=u^{+}(x)-u^{-}(x)=u_{\alpha}(x)-u_{\pi-\alpha}(x)=\int_{e^{-i\alpha}\infty}^{e^{i\alpha}\infty}e^{-sx}v(s)\,\mathrm{d}s \tag*{(62)}

for some α>0\alpha>0.

The upper bound on the growth rate of aka_k shows that the series converge for ∣s∣<1|s|<1. In general the radius of convergence is equal to one. Indeed, the integral equation can be rewritten in the form

v=−v∗v+γ(sv)∗(sv)+2γv∗(s2v)s2(s2+1)v=-\frac{v\ast v+\gamma(sv)\ast(sv)+2\gamma v\ast(s^{2}v)}{s^{2}(s^{2}+1)}

and consequently vv goes to infinity when s→±is\to\pm i (unless the integrals in the numerator vanish for some γ\gamma). Moreover, the dominant part of the singularity is a simple pole. Ecalle’s theory suggests that the singularity should be of the form

v(s)=C02πi(s−i)+log⁡(s−i)2πi∑k≥0ck(s−i)k+a regular function(s−i).(63)v(s)=\frac{C_{0}}{2\pi i(s-i)}+\frac{\log(s-i)}{2\pi i}\sum_{k\geq0}c_{k}(s-i)^{k}+a\ \text{regular function}(s-i). \tag*{(63)}

Then deforming the path of (62) into the upper half plane we conclude that

δ≃vf=e−ix(C0+∑k≥0ckk!xk+1).\delta\simeq v_{f}=e^{-ix}\left(C_{0}+\sum_{k\geq0}\frac{c_{k}k!}{x^{k+1}}\right).

Comparing this expansion with (24) we conclude that

S0=C0=2πilim⁡s→λ0(s−i)v(s).S_{0}=C_{0}=2\pi i\lim_{s\to\lambda_{0}}(s-i)v(s).

Proposition 4 S0=πiS1S_{0}=\pi iS_{1}.

A proof is based on the observation that the nearest singularity of a function is determined by the growth rate of coefficients of its Taylor series. Since the function vv is real analytic, its poles are symmetric. Moreover, we note that it is an odd function with real Taylor coefficients. It follows that vv is purely imaginary on the imaginary axis. Consequently, Re⁡C0=0\operatorname{Re}C_{0}=0. Then we note that

C02πi(s−i)+C02πi(s+i)=C0sπi(s2+1)=C0πi∑k≥0(−1)ks2k+1.\frac{C_{0}}{2\pi i(s-i)}+\frac{C_{0}}{2\pi i(s+i)}=\frac{C_{0}s}{\pi i(s^{2}+1)}=\frac{C_{0}}{\pi i}\sum_{k\geq0}(-1)^{k}s^{2k+1}.

Comparing this expansion with the Taylor series for vv and using equation (63) we get

C0(−1)kπi:ak(2k−1)!⟶1.\frac{C_{0}(-1)^{k}}{\pi i}:\frac{a_{k}}{(2k-1)!}\longrightarrow1.

Consequently,

S1=lim⁡k→∞(−1)kak(2k−1)!=C0πi.S_{1}=\lim_{k\to\infty}\frac{(-1)^{k}a_{k}}{(2k-1)!}=\frac{C_{0}}{\pi i}.

References

  1. [1]A. Abad, R. Barrio, F. Blesa, and M. Rodríguez. Algorithm 924: TIDES, a Taylor series integrator for differential equations. ACM Trans. Math. Softw., 39(1):5:1–5:28, November 2012.DOI
  2. [2]D.H. Bailey, R. Barrio, and J.M. Borwein. High-precision computation: Mathematical physics and dynamics. Applied Mathematics and Computation, 218(20):10106–10121, 2012.
  3. [3]R. Barrio. Taylor series methods. In Björn Engquist, editor, Encyclopedia of Applied and Computational Mathematics, pages 1465–1470. Springer Berlin Heidelberg, Berlin, Heidelberg, 2015.DOI
  4. [4]R. Barrio, M. Rodriguez, A. Abad, and F. Blesa. Breaking the limits: The Taylor series method. Applied Mathematics and Computation, 217(20):7940–7954, 2011.DOI
  5. [5]R.P. Brent. An algorithm with guaranteed convergence for finding a zero of a function. Comput. J., 14:422–425, 1971.DOI
  6. [6]A. R. Champneys, B.A. Malomed, J. Yang, and D. J. Kaup. Embedded solitons: solitary waves in resonance with the linear spectrum. Phys. D, 152/153:340–354, 2001.DOI
  7. [7]A.R. Champneys and M.D. Groves. A global investigation of solitary-wave solutions to a two-parameter model for water waves. J. Fluid Mech., 342:199–229, 1997.DOI
  8. [8]A.R. Champneys, B.A. Malomed, J. Yang, and D.J. Kaup. Embedded solitons: solitary waves in resonance with the linear spectrum. Physica D, 152:340–354, 2001.DOI
  9. [9]W. Craig and M.D. GrovesSegur. Hamiltonian long-wave approxima- tions to the water-wave problem. Wave Motion, 19:367–389, 1994.DOI
  10. [10]L. Fousse, G. Hanrot, V. Lefèvre, P. Pélissier, and P. Zimmermann. Mpfr: A multiple-precision binary floating-point library with correct rounding. ACM Trans. Math. Softw., 33(2), June 2007.
  11. [11]V. Gelfreich and L. Lerman. Separatrix splitting at a Hamiltonian 0²iω bifurcation. Regul. Chaotic Dyn., 19(6):635–655, 2014.
  12. [12]G. Iooss and E. Lombardi. Normal forms with exponentially small re- mainder: application to homoclinic connections for the reversible 0²+iω resonance. C. R. Math. Acad. Sci. Paris, 339(12):831–838, 2004.DOI
  13. [13]Y.A. Kuznetsov. Elements of applied bifurcation theory, volume 112 of Applied Mathematical Sciences. Springer, Cham, fourth edition, 2023.
  14. [14]E. Lombardi. Oscillatory integrals and phenomena beyond all algebraic orders, volume 1741 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2000. With applications to homoclinic orbits in reversible sys- tems.
  15. [15]K.R. Meyer and D.C. Offin. Introduction to Hamiltonian dynamical systems and the N-body problem, volume 90 of Applied Mathematical Sciences. Springer, Cham, third edition, 2017.
  16. [16]G. Torbjörn et al. GNU multiple precision arithmetic library 4.1.2, December 2002.
  17. [17]A. Tobvis and D. Pelinovsky. Exact conditions for existence of homo- clinic orbits in the fifth-order KdV model. Nonlinearity, 19(10):2277– 2312, 2006.
  18. [18]J. Yang, B.A. Malomed, and D.J. Kaup. Embedded solitons in second- harmonic-generating systems. Phys. Rev. Lett., 83:1958–1961, 1999.DOI

Paper details

vlnc:W_h6f5rnhq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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