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 and phase velocity 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
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 and . Looking for a solution in the form and neglecting the non-linear terms we obtain the dispersion relation in the form
This equation determines the phase speed of small amplitude harmonic waves as a function of the wave number . The dispersion relation can also be considered as an equation for possible values of for a fixed . Real values of 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 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 where a long linear wave with , which exists for , becomes an unstable linear mode for . The numerical study [7] suggests that the values of parameters corresponding to ES form a system of isolated lines on the -plane which emerge from a sequence of points with and .
Looking for a travelling wave solution of the equation (1) in the form , integrating once and setting the integration constant to zero, we obtain an ordinary differential equation (ODE)
where the primes refer to derivatives with respect to the new independent variable . We also substituted to shorten the notation without loosing in generality as the equation depends only on the sum . 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 is a solution, then so is . Moreover, (3) can be rewritten as a Hamiltonian system with two degrees of freedom, where the trivial solution 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 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 , where satisfies the dispersion relation (2), therefore
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 or when for . 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 -plane.
The equation (3) admits no soliton solutions when the parameters lie in Region 4, since the characteristic exponents are purely imaginary and therefore there are no solutions that decay exponentially when . 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 with , 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 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 , the equation (3) with , called the inner equation,
can be used to define a Stokes constant . 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 , then there are no single-bump homoclinic solutions for parameter values near the point 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 , , generates a family of homoclinic orbits, which form a smooth curve in the -plane and starts at .
The Stokes constant 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 (or equivalently for large negative ), 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
where is a Stokes constant obtained from a parameterless problem described in § 4.4, . It follows that, if is sufficiently large, there is a zero of near every point
where the approximation for changes its sign. It is interesting to note that our numerical experiments show that every interval contains exactly one zero for every , and that there are no other zeroes. Moreover, the difference 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 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 , 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 for large , 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
which depends on five real parameters , , , , . We assume that the quadratic part is not degenerate, i.e., . (7) is an ODE of order four with a quadratic nonlinearity, which coincides with (3) when , , , and . In spite of the substantially larger number of parameters equation (7) is only a minor generalisation of (3). Indeed, if , then the substitution with ,
transforms (7) into a standard form
where , and . For example, for equation (3) we get , and . 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 and 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 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 and, equivalently, . The line belongs to the region where and, equivalently, . We denote the eigenvalues of the equilibrium by and assuming that . On the line , is a double eigenvalue and .
The scaling of the time variable changes the values of and without affecting the ratio . This freedom allows us to consider the corresponding equations in two equivalent ways.
First, an equation (7) with , and has the scaling invariants and and, consequently, it can be transformed into an equation of the form considered in [17],
with and . Indeed, computing the scaling invariants for this equation we get and . Thus we cover all possible values of the scaling invariants in the region where and .
In the case of equation (9), the characteristic equation takes the form
and consequently and . Thus in this formulation, the problem reduces to the study of a reversible Hamiltonian bifurcation for equation (9) which happens at .
Second, the standard scaling with transforms equation (9) into a singularly perturbed problem,
where the small parameter multiplies the highest derivative. In this formulation, we get and goes to when . In the limit equation (10) reduces to an equation of the second order,
In the singular perturbation theory, this equation is called an outer equation. The outer equation admits an explicit homoclinic solution
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 , such an orbit is generically destroyed when , see [14, 11]. To study the persistence of this solution for , we consider the stable and unstable solutions of equation (9). Since the unstable direction at the equilibrium is one dimensional, the unstable solution, as , is defined uniquely up to a constant shift in time. We eliminate this freedom by asking to be the first local maximum of . Then . We note that is also a maximum of , so , and the unstable solution admits the approximation
which holds uniformly for with any given . In the leading term of this estimate, comes from the standard scaling.
The stable solution is obtained using the reversible symmetry, . Since is even, admits approximation by the same function as but for .
Then we consider the difference
Since , it is odd and we get . The normalisation assumption implies that . Then, in the case , we can conclude that and satisfy the same initial conditions at and, since the solution to Cauchy problem is unique, that for all . Hence we have a symmetric soliton each time vanishes. This soliton is a single-bump one as it is approximated by the unperturbed homoclinic solution for all real .
A direct substitution shows that satisfies the variational equation:
around . The variational equation is a linear homogeneous ODE of order four. The space of its solution is spanned by four fundamental solutions. Let be a solution that satisfies the Cauchy problem for the variational equation: and . Then there is a real constant such that
We note that the existence of single-bump symmetric homoclinic solutions is equivalent to the vanishing of , i.e., it is a co-dimension one phenomenon.
Inner equation and Stokes phenomenon
Determining the zeros of is a non-trivial problem, since is exponentially small in . The complex matching method suggests studying the analytic continuation of the solutions into the complex plane in order to obtain an approximation for .
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 ,
The equilibrium solution is a degenerate saddle-centre: the linearisation at the origin has two purely imaginary eigenvalues 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:
respectively, where we introduce yet a new time, (see Remark 1 below). Similarly to , the stable and unstable solutions of (9), the solutions 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 located at , and that in this region a better approximation is obtained using the inner equation. The method uses the change of time
to study stable and unstable solutions in the region where is small and is large. The application of the standard scaling to the Laurent series suggests to use (15) as an asymptotic condition for the inner equation at infinity. Then it can be shown that approximate the stable and unstable solutions with .
Since equation (14) is reversible and all its coefficients are real, the involution
maps a solution of (14) to another solution of the same equation. Under this transformation, the domain of is reflected in the imaginary axis. Since the solutions are uniquely defined by the asymptotic condition (15) which is invariant under the involution , we conclude that , or equivalently
All points on the imaginary axis are fixed under the involution . We consider the difference
on the imaginary axis for large negative values of .
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:
After substituting the series into (14) and collecting terms of equal order in we obtain a recurrent system for the coefficients . The leading order gives and equation (50) provides an explicit recursive formula for with .
It can be shown that the formal solution provides an asymptotic expansion for both the stable and unstable solutions,
as in two large sectors (with opening angles greater than ) centred respectively around the positive and negative real semi-axes sketched in Figure 2.
Figure 2. Sketch of the sectors and where are defined.
We emphasize that, although the same formal series arises in both cases, the domains in which the partial sums of approximate are different. On the intersection of these domains, the difference is beyond all algebraic orders, in the sense that for every . 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 , we consider the variational equation
which is obtained by linearising the inner equation (14) around a given solution . A direct substitution shows that satisfies (20) with , where
Note that equation (20) coincides with (13) with and replaced by .
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 , , and as along . Of these four solutions, only two decay along the negative imaginary axis, and only one is .
To construct an approximation to the fast decaying solution, we observe that the variational equation (20) with admits a formal solution of the form
Substituting the formal series and into (20) and collecting terms of equal order in yields a recursive system for the coefficients . This system determines the coefficients uniquely up to a multiplicative constant, which we fix by setting . Then the coefficients with even are real and the ones with odd are purely imaginary.
Since both and share the same asymptotic expansion on the intersection of their domains, it follows that . 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 admits a unique solution,
as in a sector centred on the negative imaginary semi-axis.
We note that (19) implies that decays faster than any power along the imaginary axis. Therefore, since is a solution of (20), it must be proportional to .
Definition 1 (Stokes constant 1) The complex constant such that in a neighbourhood of infinity at the negative imaginary semi-axis
is called a Stokes constant.
Since we get an equivalent description for .
Proposition 1 The Stokes constant is equal to the following limit:
Proposition 2. The Stokes constant is purely imaginary.
This property follows from the reversibility and real-analytic symmetry of the inner equation. Indeed, is antisymmetric with respect to involution defined in (17),
On the other hand, it can be checked using induction that for all . Therefore the formal expansion is symmetric, . Since both the equation and asymptotic condition are invariant, the uniquely defined solution is also symmetric, i.e.
Applying to the identity (23) therefore yields , implying that the Stokes constant is purely imaginary.
Remark 2. If for some , we get an example of the non-linear Stokes phenomenon. Informally we can say that the unstable solution changes its asymptotic behavior at infinity from to when 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 , and denoting the corresponding partial sum of the formal series by , we visualize the Stokes phenomenon in Figure 3: the difference transitions from monotone to oscillatory behaviour as crosses the imaginary axis following a line .
Figure 3. Stokes phenomenon: deviation of the unstable solution from the truncated asymptotic series. Plot of with versus , real part (left) and imaginary part (right). Parameters , , .
Proposition 3. The Stokes constant if, and only if, the formal series converge.
Suppose that for some . Then the equation (23) implies that near infinity on the negative imaginary semi-axis. The principle of analytic continuation implies that for all where the functions are defined. The union of the domains of and includes a complete neighbourhood of infinity and is not simply connected. The real-analytic symmetry can be used to show that is single-valued in a neighbourhood of infinity. Thus its Laurent series converge and, due to the uniqueness, coincide with the formal solution . Thus the series 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 and . Thus and .
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 is the limit
where are coefficients of the formal solution (18).
Remark 3 Although it is not obvious, it can be shown that
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 . Indeed, applying the induction to recursion (50), it is easy to check that are polynomial of degree with sign-alternating coefficients. In order to show that is analytic in for it is sufficient to prove uniform convergence to the limit in (25).
The Stokes constant is an analytic function of the real parameter appearing in the inner equation (14). Since , any change of sign of 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 and derive an asymptotic description of their location in the limit .
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 for we introduce the new small parameter . The substitution
transforms the inner equation (14) into a singularly perturbed problem,
where the small parameter appears in front of the highest derivative. We stress that this equation depends on only and is completely independent of small parameter 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
when in sectors centred on the positive and negative real semi-axes respectively. We use the method of complex matching to approximate the difference
and to derive the asymptotic expression (5) for the Stokes constant in the limit of small .
Equation (27) admits a formal solution , obtained from (18) using the scaling (26). Applying induction to the recurrence relations (50), one verifies that is a polynomial of degree in . Hence can be written as a double series
where . Since only even powers of occur, we look for a formal solution of (27) in the form
Substitution into (27) shows that the leading term satisfies the ODE
which is the equation obtained from (27) in the limit . In view of the asymptotics (28), we select the solution satisfying
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 , integrating and setting the integration constant to zero we get
We look for a solution of this equation in the form . We fix the branch of by the asymptotic condition
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 , we get
Dividing the equality by and integrating we get an explicit expression for the inverse of :
The inverse function theorem implies that this equation has a unique solution 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 , 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
In particular, the difference satisfies (36) with . The WKB theory suggests to look for a solution in the form
where the dots stand for terms of higher orders in . Substituting this ansatz into the equation (36), taking into account that and collecting the leading terms in we get (after simplifying the expressions) that
These equations can be solved explicitly in terms of :
Taking into account the properties of we conclude that and are both single-valued in a neighbourhood of infinity. Then (33) implies
Finally, we note that both and are solutions of (36) of the fastest decay along the imaginary axis. Comparing (37) with (22), we conclude that as their ratio is constant and converges to 1 when .
Singularities of the outer problem
We note that the asymptotic (31) holds in a full neighbourhood of infinity; hence provides a common approximation to both and . Complex matching suggests to consider analytic continuation of solutions into neighbourhoods of the singularities of . Therefore we need information on the location and nature of the singularities of .
Equation (34) is a non-linear first-order ODE and the corresponding vector field grows as at the infinity. Consequently, admits analytic continuation along any path in the complex -plane such that the values of along this path avoid the singularities of the vector field located at . Using that
we conclude that equation (35) implies that the singular values of the function are located at the points
Consequently, admits analytic continuation along any path which avoids .
For the principal branch of the inverse sine function, the left-hand side of (35) is a monotone function mapping onto , where . Consequently, is analytic on . Since is odd, it is also analytic on . Hence the principal branch of extends analytically to , i.e. the complex plane cut along the interval .
In order to study near the singularity at , we recall the trigonometric identity that holds for all :
Then (35) implies that for
Recalling the Maclaurin series and applying the inverse function theorem we get that
where the remainder is a converging series in odd powers of . Taking the square and resolving the result with respect to we get that
where the remainder is a converging power series in . Therefore is a rational branching point with .
Finally we note that the uniqueness of the initial value problem for the differential equation (34) implies that does not vanish. We conclude that the functions , and are analytic on . Moreover
The value of will play an important role in the subsequent analysis.
Inner problem for the inner problem
We have established that is a branching point of and . In order to study in a neighbourhood of this point, we make the substitution
In the new variables the equation (27) takes the form
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 we get the inner equation for the inner equation:
We look for solutions of (45) whose behaviour at large matches that of near , as described by (40). This asymptotic form suggests a formal solution of the form
Substitution of this series into (45) shows that the coefficients for are uniquely determined once is fixed. Comparison with (40) then leads to the choice .
The variational equation associated with (45) is given by
This is a linear homogeneous ODE of order four and, for , the space of its solutions is spanned by four independent solutions that respectively behave as , and for large . The formal exponentially decaying solution has the form
The coefficients for are uniquely determined by the equation once is fixed. We set . The exponential factor decays in the sector , where and . The equation (45) admits two analytic solutions such that
as , respectively. Inside the sector between these directions, the difference
is superalgebraically small, i.e. it decays faster than for any . A direct computation shows that satisfies the variational equation (46) with replaced by .
Since (46) with is a linear ODE of order four, its solution space is four-dimensional. Among its solutions, only one,
decays faster than along the bisecting direction of the sector. It follows that there exists a complex constant such that
The equation (46) has a symmetry which involves the reflection of the -plane in the line , the bi-section of the sector, and the reflection of the -plane in the line : if is a solution then also is. This symmetry swaps and and multiplies and, consequently, by . Applying this symmetry to (49) we conclude that is real. Evaluating and numerically we find that
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 . First we note that the functions and 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
vanishes as in the limit .
We recall that is an exact solution of (27) and while is an approximate one. Since the growth of small errors is governed by the variational equation it is natural to study along a line
as the solutions of the variational equation are oscillatory and do not grow exponentially along this line. Since for large , it is easy to see that with a large is approximately horizontal. Solving equation (38) with respect to we get
Substituting this expression into (35) and simplifying the result we get the parametrization for ,
Let for some . The corresponding curve lies in the lower half-plane. In the limit , it converges to a contour consisting of two rays, and , joined by an arc connecting and .
Let , , be the intersection points of with the circle . These points lie in a neighbourhood of on opposite sides of the Stokes line defined by , so that
We use these points for complex matching. By (43), the circle corresponds to . Hence, it represents a neighbourhood of in the -plane and a neighbourhood of infinity in the -plane. The scaling is chosen so that the errors of the inner and outer approximations are of the same order.
For to the right of , the function remains sufficiently far from the singular value , so that the outer approximation is valid and . In a neighbourhood of , we compare (40) with the leading term in (48) for and conclude that
Consequently near
When lies on between and , the solution remains close to , and the inner approximation provides a more accurate description than the outer one. In a neighbourhood of , we use (49) to switch from the stable solution to the unstable solution :
Then we compare (40) and the first term of (48) for and conclude that
Then we compare the approximation of by the first term of (47) and the approximation of by (37), (40) and (42). We get that
Consequently, near
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 with the imaginary axis. Then we apply the reversor to get the unstable solution. For a point on the imaginary axis and, consequently,
Taking the difference of the last two approximations and using , we obtain that on the imaginary axis
Comparing the last estimate with the definition (23) of the Stokes constant, and recalling that and , we obtain
Finally, recalling that , 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 of the formal solution (18) to the inner equation (14). These coefficients are defined by the recurrence relation with initial condition and
where the first of the sums at the rhs of (50) must be omitted for . The induction in implies that this equation defines which are polynomial of degree in with sign-alternating coefficients, and satisfy for all .
For large , the coefficients exhibit Gevrey-one growth with a sign-alternating tail. We introduce the normalized coefficients
and aim to compute the Stokes constant as the limit
A typical behaviour of for is shown in Figure 4 (left), where we take as an example. The tail of the sequence is monotone and decreases in absolute value. Plotting as a function of (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 .
Figure 4. Normalized coefficients of the formal solution for : for in log–log scale (left) and versus (right); the thin discontinuous line shows an extrapolation to zero.
To this end, we proceed as follows. We identify two consecutive intervals and on which the sequence is monotone. On each interval, we construct interpolating polynomials and of degree using the data . We take as an estimate of , provided that the predicted values are in good agreement:
If this criterion is not satisfied, we increase and repeat the procedure. We use the difference (52) as a proxy for the error in evaluation of .
The computation time depends on three factors: the starting point that determines the number of coefficients to be computed, the number of correct digits in numerical values of and the degree of the interpolating polynomial.
As an example, we consider . Using the tail of 4000 coefficients and an interpolating polynomial of degree 80, we obtain an estimate of with an error below . 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 , respectively, and compare the estimates for obtained using interpolating polynomials of varying degree. The deviation from the reference value for is shown in Figure 5. We observe that the error initially decreases rapidly as the polynomial degree increases, but eventually begins to grow.
Figure 5. Accuracy of the extrapolation to zero as a function of the extrapolating polynomial degree 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 , we obtain an estimate of the Stokes constant which deviates by less than from the reference value. The experiment confirms the numerical stability of the extrapolation procedure, with estimates of 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 and , we observe that approximately 180 digits are lost in the evaluation of for and . 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 to find suitable for numerically reliable extrapolation. Then we use to get a high precision approximation of .
The results of these computations are presented in Figure 6. They show that is positive and monotonically decreasing for . It has its first zero at . Beyond this point, the behaviour becomes oscillatory.
Figure 6. Stokes constant as a function of (left), and the normalized value as a function of on the interval which contains the first 26 zeroes of (right).
To illustrate the oscillatory behavior, we adopt the scaling suggested by the asymptotic estimate (5) and plot as a function of . For large , 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 to study the dependence of the Stokes constant on . It is easy to see that is polynomial of degree . For example,
The polynomials with share the common root at . For this special value of , solves the inner equation (14) exactly. Figure 7 illustrates convergence of to the limit.
Figure 7. Normalized coefficients as functions of for (green) and (blue), and the limiting curve (red).
We observe that the zeros of the Stokes constant can be located without relying on the extrapolation procedure. Indeed, suppose that for some . Then there exists such that
Therefore, if
for some , then the interval contains a zero of . In our computations we refine such intervals until .
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 root of . As an initial approximation (), we take suggested by the asymptotic formula (6). This approximation works well for all . At each subsequent step, we employ the recurrence relation (50) and its derivative with respect to to compute and till we detect the onset of the monotone tail for with . Then we perform one step of Newton’s method:
in order to find with a smaller value of . This procedure is iterated until , which is typically achieved in less than 10 steps. We let . As a final check, we verify that
for in the tail region.
Figure 8 shows the plot of the normalized deviation of the computed zeroes of from their theoretical prediction,
Figure 8. The first 440 zeroes of : deviation from the predicted value, , in the units of the limit root spacing vs .
The difference apparently converge to zero as . 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 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
Then we fix constants and set the initial point at . Then we solve numerically the Cauchy problem for the inner equation (14) with initial conditions
where is chosen according to the rule of truncation at the least term. The approximation to the unstable solution is numerically computed along the line in the complex plane and provides a good approximation to for .
The stable solution is defined using the reversible symmetry, . On the imaginary axis the difference is purely imaginary:
Then we find coefficients of the formal solution (21) of the variational equation (20) with and define a partial sum
Choosing from the optimal truncation rule for we obtain an approximation for the Stokes constant
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
Our numerical experiments confirm this relation with an error below . 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 decays exponentially when 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 increases and is reliable only for sufficiently large values of .
Method based on the limit behaviour
In this section, we describe an alternative method based on the definition of 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:
For a given value of we fix and consider a large and define . Then we numerically solve the Cauchy problem for the inner equation (14) with initial conditions (54): integrating the equation during time , we obtain a very accurate approximation for and its derivatives at this point.
Then we choose and integrate the inner equation (14) in the purely imaginary direction substituting and using as the new independent variable. We use the results of the previous step to generate the initial condition for the Cauchy problem with and integrate the equation numerically till .
Once performed the numerical integration we use equation (24) to get an approximation for the Stokes constant,
The integration path used in Steps 1 and 2 is sketched in Figure 9.

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
the number of terms retained in the approximation of the initial conditions (53) (thus, the parameters associated with the first path are , , and ), and the final imaginary time 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 , which our tests indicate is sufficiently large, using either 200 or 1000 decimal digits of precision and several values of and . As 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 for using arithmetics with 200 and 1000 decimal digits and different values of and .
Figure 11. Approximation of the Stokes constant approaching the limit value for (left) and (right) using different precision levels (digits = ) and different initial times () for the numerical evaluation.
For , computations with 25 decimal digits exhibit strong oscillations when the imaginary time along the second path approaches , resulting in a highly inaccurate approximation of the Stokes constant. Increasing the precision to 50 digits delays the onset of the oscillations until , whereas 200 digits are required to maintain adequate accuracy over the range considered. For , even higher precision is necessary, together with a sufficiently large value of the initial real displacement .
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 , 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, , for several values of , , , 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 digits | correct digits | |||
| 200 | 1000 | 70 | 134 | |
| 200 | 80 | 1000 | 122 | |
| 200 | 80 | 1000 | 4 | |
| 200 | 80 | 9 | ||
| 200 | 60 | 1000 | 70 | 173 |
| 5000 | 70 | 134 | ||
| 60 | 1000 | 53 | ||
| 60 | 1000 | 3 |
Table 1. Number of correct digits in the numerical determination of the 21st zero of the Stokes constant () for various values of , , and precision digits.
This also requires a suitable choice of and , since the initial point must be sufficiently far from the origin. Moreover, choosing a large value of 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 .
Based on the preceding tests, and recognizing the need to balance computational cost against numerical accuracy, we selected the following parameter values: digits of precision, , , and . 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 digits of precision, , , and .
| 1 | |
| 2 | |
| 3 | |
| 4 | |
| 5 | |
| 6 | |
| 7 | |
| 8 | |
| 9 | |
| 10 | |
| 11 | |
| 12 | |
| 13 | |
| 14 | |
| 15 | |
| 16 | |
| 17 | |
| 18 | |
| 19 | |
| 20 | |
| 21 | |
| 22 |
Table 2. The first 22 zeros of the Stokes constant with 80 correct digits (the numerical process has been done with 200 precision digits, , and ).
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.
As shown in Figure 12, the quadratic approximation provides an excellent fit, with the error decreasing as increases.
Finally, the Stokes constant vanishes if and only if
For such parameter values, 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 changes sign across the endpoints of a sufficiently small interval in 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 , 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 . 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 satisfies the equation, then 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 the -derivative of , we rewrite the equation (7) in the form of a system of four first order equations:
This system is hamiltonian with the Hamilton function
and the non-standard symplectic form
Indeed, let denote the vector field in the right hand side of the system (58). A direct substitution shows that satisfies the identity for all vectors where denotes the differentiation with respect to the variable . Recall that where is the Kronecker delta.
Soliton solutions belong to the zero level set of the first integral . Note that the equality 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 in the definition of as well as its connection to deserve an explanation. The factorial growth rate for can be easily proved with the help of the induction and recurrent formulae (50). The relation between and 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 and coincide with Borel-Laplace sums of the formal series . In this interpretation the difference between and 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 ,
where the integration path is a ray that forms the angle with the positive real axis. We chose for and for and assume that is the same analytic functions in both cases.
Suppose that near the origin
The growth rate of implies that the series in the right hand side converge for . Let be the convolution of functions. The Laplace transforms maps the convolution into the product and multiplication by into . Therefore the Taylor series formally satisfy the integral equation
and, since the series converge, is an analytic solution of the same equation.
The integral equation can be used to prove that the analytic continuation of along any ray with does not grow faster than an exponential function. We conclude that the Laplace integrals converge and satisfies equation (14). Then integrating by parts we can easily check that as the asymptotic behaviour of the Laplace integral for large is determined by the Taylor series of around . The domain of validity of the asymptotic relation depends on , and we conclude that our choices of indeed lead to representation of the stable and unstable solution in the form of the Laplace integrals. Then for all sufficiently large located in a neighbourhood of the negative imaginary semi-axis we can write
for some .
The upper bound on the growth rate of shows that the series converge for . In general the radius of convergence is equal to one. Indeed, the integral equation can be rewritten in the form
and consequently goes to infinity when (unless the integrals in the numerator vanish for some ). Moreover, the dominant part of the singularity is a simple pole. Ecalle’s theory suggests that the singularity should be of the form
Then deforming the path of (62) into the upper half plane we conclude that
Comparing this expansion with (24) we conclude that
Proposition 4 .
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 is real analytic, its poles are symmetric. Moreover, we note that it is an odd function with real Taylor coefficients. It follows that is purely imaginary on the imaginary axis. Consequently, . Then we note that
Comparing this expansion with the Taylor series for and using equation (63) we get
Consequently,
References
- [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]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]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]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]R.P. Brent. An algorithm with guaranteed convergence for finding a zero of a function. Comput. J., 14:422–425, 1971.DOI
- [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]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]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]W. Craig and M.D. GrovesSegur. Hamiltonian long-wave approxima- tions to the water-wave problem. Wave Motion, 19:367–389, 1994.DOI
- [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]V. Gelfreich and L. Lerman. Separatrix splitting at a Hamiltonian 0²iω bifurcation. Regul. Chaotic Dyn., 19(6):635–655, 2014.
- [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]Y.A. Kuznetsov. Elements of applied bifurcation theory, volume 112 of Applied Mathematical Sciences. Springer, Cham, fourth edition, 2023.
- [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]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]G. Torbjörn et al. GNU multiple precision arithmetic library 4.1.2, December 2002.
- [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]J. Yang, B.A. Malomed, and D.J. Kaup. Embedded solitons in second- harmonic-generating systems. Phys. Rev. Lett., 83:1958–1961, 1999.DOI










