An Evans function for the linearised 2D Euler equations using Hill's determinant
AAN EVANS FUNCTION FOR THE LINEARISED 2D EULEREQUATIONS USING HILL’S DETERMINANT
HOLGER R. DULLIN AND ROBERT MARANGELL
Abstract.
We study the point spectrum of the linearisation of Euler’s equation forthe ideal fluid on the torus about a shear flow. By separation of variables the problemis reduced to the spectral theory of a complex Hill’s equation. Using Hill’s determinantan Evans function of the original Euler equation is constructed. The Evans functionallows us to completely characterise the point spectrum of the linearisation, and tocount the isolated eigenvalues with non-zero real part. In particular this approach alsoworks in the case where complex eigenvalues appear. Introduction
Euler’s fluid equations describe the flow of an incompressible inviscid Newtonian fluid.It is a non-linear partial differential equation that is a limit of the incompressible Navier-Stokes equations for vanishing viscosity. The simplest solutions of this PDE are steadysolutions for which the velocity field is constant in time. In a shear flow adjacent layersof fluid move parallel to each other, with possibly different speeds. Given such anequilibrium solution the stability of this solution is studied by linearising the Eulerequations about this particular solution. When considering the equation with periodicboundary conditions in both directions it is natural to consider the linearised equationsin Fourier space. The linear operator typically has continuous spectrum on the imaginaryaxis, and may have discrete spectrum outside the imaginary axis. Due to the Hamiltoniansymmetry of the spectrum any spectrum with non-zero imaginary part leads to linearinstability through exponential growth.In particular we study the linearisation about the steady state with stream function ψ = cos( p x + p y ) for co-prime integers p = ( p , p ). For the linearised Euler equationsin Fourier space, an integer lattice points represent a spatial Fourier mode. The linearoperator obtained from linearising the Euler equations about ψ block-diagonalises intoso-called classes of modes that are located on lines parallel to the line with directionvector p .A theorem of Li [Li00] shows that classes which do not have an integer lattice point insidethe ‘unstable disk’ of radius (cid:112) p + p can not contribute to instability. A theorem ofLatushkin, Li and Stanislavova [LLS04] shows that twice the number of non-zero integerlattice points inside the unstable disk is an upper bound for the number of discreteeigenvalues with non-zero real part. In this paper we show that this upper bound issharp, so that the number of discrete eigenvalues is exactly equal to twice the numberof non-zero integer lattice points inside D . a r X i v : . [ m a t h . D S ] J a n HOLGER R. DULLIN AND ROBERT MARANGELL
This result has been difficult to achieve because the discrete spectrum typically con-tains complex eigenvalues. It has been shown by Worthington, Dullin and Marangell[DMW16], and Latushkin and Vasuvedan [LV19] that classes which contain a single inte-ger lattice point inside the unstable disk correspond to a pair of real eigenvalues. In thismanuscript we provide another proof of this fact, and we show that classes containinga pair of integer lattice points inside the unstable disk correspond to either two pairs ofreal eigenvalues or a complex quartet.These new results about the (in)stability of the Euler equations are obtained throughan approach different from [Li00, DMW16, LV19]. Instead of first going into Fourierspace, then linearising the equation, and then block-diagonalising the operator, we firstlinearise the PDE, then separate variables, and then consider the resulting boundaryvalue problem in (a different) Fourier space. The separation step is somewhat unusualbecause it uses a non-orthogonal coordinate system in order to preserve the periodicity.The resulting boundary value problem is a Hill’s equation with quasi-periodic boundaryconditions. The spectral theory of this equation is well developed and is based on Hill’sdeterminant, which is a Fredholm determinant. The Hill’s determinant is used to writedown an Evans function for a class. In the final step all relevant Evans functions comingfrom Hill’s equation are combined into a single Evans function of the linearised 2D Eulerequations. With the help of this Evans function we are able to prove that the upperbound from [LLS04] is actually always attained. The main ingredient in the proof of ourresult is the ability to differentiate the Hill’s determinant and hence the Evans functionat vanishing spectral parameter. This allows us to treat questions about the number ofunstable eigenvalues as a kind of bifurcation problem.Our periodic Evans function for a class is a particular example of a periodic Evansfunction associated with Bloch-wave eigenvalue problems of PDEs in 1+1 dimensions.There are many other specific examples of such Evans functions constructed in the last15 years or so, for various PDEs of note, including the Kuramoto-Sivashinsky equa-tion [BNaKZ12], the Kawahara equation [TDK18], the sine and Klein-Gordon equations[CM20a, JMMP13, JMMP14], as well as linear and nonlinear Schr¨odinger equations[BR05, CM20b, GH07, JLM13, LBJM19] to name a few. For a description of periodicEvans functions along the lines of [Gar93], as well as for additional applications of theEvans function to periodic problems, see [KP13] and the references therein.Our Evans function differs from the aforementioned references in that, rather than lin-earising around a periodic solution to a 1+1 dimensional travelling wave and consideringarbitrary longitudinal perturbations, our ‘eigenvalue problem’ (equations (4) and (5))comes from the reduction of a 2+1 dimensional dispersion relation that is consideringperiodic perturbations in both directions. After separation of variables the temporalspectral parameter λ (or c ) is now a parameter within the potential of the Hill’s equa-tion, and what would typically be a range of Floquet exponents is reduced to a singleBloch mode, as it would be in the 1+1 dimensional case of periodic perturbations, butnow the ‘eigenvalue’ parameter (what we call µ = d ), is a function of the Fourier modeof the periodic perturbation in the transverse direction. The upshot is that we can con-struct an Evans function for a 2+1 dimensional problem by multiplying together theEvans functions for the relevant classes. The lack of continuity in the Floquet exponents ULER, HILL, EVANS 3 considered means that we cannot exploit the analyticity of the periodic Evans functionnear the origin for a stability index, as in [BJK11, BJK14, BR05, JMMP13, JMMP14].Instead we employ the Floquet-Fourier-Hill method of [DK06] (and [MW66]) in order torelate the Hill discriminant to the parameters in our original equation.Evans functions for systems with more than one independent spatial variable are notcommon. To the best of our knowledge they have only been determined numericallytwice: first, to examine the stability (or lack thereof) in large amplitude planar Navier-Stokes shocks [HLZ17] and then in the case of planar viscous strong detonation waves[BHLL18]. The first proofs of instability for the equilibria considered here were given in[DMW16] and [LV19, DLM + Separation of the linearised Euler equations
The stream function formulation of the incompressible 2-dimensional Euler equationsis(1) ∇ Ψ t = Ψ x ( ∇ Ψ) y − Ψ y ( ∇ Ψ) x , where Ψ( x, y ) is the stream function of the velocity field, that is u = Ψ y and v = − Ψ x are the x and y components of the velocity of an incompressible ideal fluid flow, andsubscripts denote partial derivatives. Specifically we are considering fluid flow on thetorus T = [ − π, π ] × [ − π, π ]. A simple steady shear flow in the x -direction is given byΨ( x, y ) = cos y , such that u = − sin y and v = 0. More generally, a doubly periodicsteady state solution to equation (1) is given by Ψ( x, y ) = ψ ( η ), where ψ ( η ) = ψ ( η + 2 π )is a real periodic function of the single variable η = p x + p y with p , p ∈ Z relativelyprime integers. Such a solution ψ ( η ) is steady state solution to equation (1).This particular solution is best described in a new coordinate system ( ξ, η ) with η as defined before, and ξ = q x + q y with relatively prime integers q , q ∈ Z , i.e.gcd( q , q ) = 1. When gcd( p , p ) = 1 then there exist relatively prime integers q , q ∈ Z with p q − p q = 1 (see, e.g., [HW79, Thm25]), such that the vectors q = ( q , q ) and p = ( p , p ) are a basis of the lattice Z . The vector q is unique up to adding multiplesof p . We choose the unique q that makes the angle between p and q as large as possible,equivalently | p · q | as small as possible. For given fixed p this basis of Z is as close as HOLGER R. DULLIN AND ROBERT MARANGELL possible to an orthogonal basis. The linear transformation from ( x, y ) to ( ξ, η ) preservesperiodicity and volume. The Laplacian in the ( ξ, η ) coordinates reads ∇ = q ∂ ξ + 2 p · q ∂ ξ ∂ η + p ∂ η , where we have defined p := p + p and q := q + q . We note that our choice of q implies that | p · q /p | ≤ /
2. The mixed term in the Laplacian appears because thechosen coordinate system is not in general orthogonal. For our purposes it is crucial topreserve periodicity and volume, so that instead of orthogonality we just require thatthe determinant of the linear map from ( x, y ) to ( ξ, η ) has determinant 1.Now consider the linearisation of equation (1) about ψ ( p x + p y ) written in the newcoordinate system. The perturbed solution is ψ ( η ) + εφ ( ξ, η, t ) and considering only thefirst order terms in ε we obtain the linear governing equation for the perturbation φ as(2) ∇ φ t = p ψ (cid:48)(cid:48)(cid:48) φ ξ − ψ (cid:48) ∇ φ ξ where we have abbreviated ddη by (cid:48) . Two additional terms of the linearised equationvanish because ψ ξ = 0. This is a linear PDE for φ ( ξ, η, t ) with coefficients that de-pend periodically on η and periodic boundary conditions φ ( ξ + 2 π, η ) = φ ( ξ, η + 2 π ) = φ ( ξ, η ).Consider a single mode φ = f ( η ) exp( ikξ + λt )) with wave number k ∈ Z and temporaleigenvalue λ ∈ C where f ( η ) is a 2 π -periodic function. Then ∇ φ = ( − k q f + 2 ik p · q f (cid:48) + p f (cid:48)(cid:48) ) φ/f and equation (2) separates into(3) f (cid:48)(cid:48) + 2 ik p · q p f (cid:48) + (cid:18) kψ (cid:48)(cid:48)(cid:48) ( η ) iλ − kψ (cid:48) ( η ) − k q p (cid:19) f = 0 . For p · q = 0 and p = q = 1 this reduces to the usual equation for linearisation about a2D shear flow, see, e.g., [Ach91], except that there the velocity U = ψ η is used to writethe coefficient. The term proportional to f (cid:48) in this second order linear equation can betransformed away by redefining the independent variable f . Defining θ = k p · q /p mod 1and f ( η ) = e − iθη g ( η ) we arrive at Proposition 2.1.
The linearisation of Euler’s equation for the ideal incompressible fluidflow equation (1) on the torus ( x, y ) ∈ T around the steady state with periodic streamfunction ψ ( η ) where η = p x + p y leads to the Hill’s equation (4) g (cid:48)(cid:48) + Q ( η ) g = µg, Q ( η ) := ψ (cid:48)(cid:48)(cid:48) c − ψ (cid:48) where µ is defined as µ := k /p , and c := iλ/k with quasi-periodic boundary conditions g (2 π ) = exp(2 πiθ ) g (0) ,g (cid:48) (2 π ) = exp(2 πiθ ) g (cid:48) (0) . (5)We have thus rewritten the dispersion relation for the linearised equation (2) on the torusas the quasi-periodic boundary value problem equation (4) with boundary conditions (5).There will be a temporal eigenvalue λ = − ick of the linearised Euler equation about the ULER, HILL, EVANS 5 shear flow ψ ( η ) when we can find a quasi-periodic solution to equation (4) satisfying theboundary condition in equation (5).The spatial mode exp( i ( lη + kξ )) written in the original coordinates is exp( i ( lp + kq ) x +( lp + kq ) y ) = exp( i ( a x + a y )), such that a = k q + l p ∈ Z is the Fourier mode inthe original coordinates. It is convenient to introduce yet another coordinates system indual Fourier-mode space with orthogonal basis ( p ⊥ , p ). The vector p ⊥ = − J p where J is a rotation by π/
2. The vector a in this coordinate system has coordinates which wedenote ( d, θ ) such that a = k q + l p = d p ⊥ + θ p implies(6) d = kp , θ = k p · q p . This is obtained by taking scalar products with p and p ⊥ and using q · p ⊥ = 1. Noticethat in the boundary condition equation (5) any integer part in θ is irrelevant, and thisis the reason that θ can be considered as an angle mod 1. In this coordinate systemthe linearisation of equation (1) about ψ ( p x + p y ) directly leads to equation (4). Thequantity θ measures the deviation of the basis ( q , p ) from orthogonality. Conversely,when using the orthogonal basis ( p ⊥ , p ) the quantity θ determines the quasi-periodicboundary conditions.For later reference we translate Li’s unstable disk theorem [Li00] into our setting. Listudied the case ψ ( η ) = cos( η ) and decomposed the Fourier modes a = k q + l p intosubsets called “classes” with fixed integer k and varying integer l . The unstable disktheorem states that the class with given k has no discrete spectrum when all modes a inthis class are outside the closed disk centred at the origin with radius p = | p | = (cid:112) p + p .The mode in class k that is closest to the origin has the smallest projection onto the p -direction. Hence k q · p + l p · p is minimal and dividing by p = p · p we see that this occurswhen θ − l is minimal. Thus we choose the angle θ to be such that θ ∈ ( − / , / k for which l is uniquelydetermined. Thus in our notation Li’s unstable disk theorem [Li00] reads Theorem 2.2.
Consider the case where ψ ( η ) = cos( η ) . If θ + d > then there isno discrete eigenvalue c of the linearised Euler’s equations and the solution is stable.Equivalently, if µ = d > − θ then there is no θ -quasiperiodic spectrum of Hill’sequation (4) for any c . The spectral parameter µ of Hill’s equation (4) satisfies µ = d . Coming from the Eulerequation the parameters θ and µ = d are discrete for fixed given p , labelled by thewave number k in the ξ direction in the mode ansatz for ψ . Whenever convenient in thefollowing we are going to consider ( d, θ ) as continuous parameters.In the linearised Euler equation the spectral parameter that determines instability is λ respectively c = iλ/k with wave number k , while the parameters of the equation are θ mod 1 and d determined by p and k . In the Hill’s equation (4) the spectral parameteris µ = d , the spectral parameter of the Hill’s equation, while the parameter of theequation is c , and the boundary conditions are determined by θ . Ultimately what we HOLGER R. DULLIN AND ROBERT MARANGELL are interested in is the relation between c , d , and θ , and this relation will be determinedby the spectral theory of the complex Hill’s equation.By multiplying equation (4) by ¯ g and integrating, we recover Rayleigh’s criterion (inte-grating by parts once):(7) 0 = (cid:60) ( λ ) (cid:90) | g | ψ (cid:48)(cid:48)(cid:48) | c − ψ (cid:48) | dη . Thus for instability where (cid:61) ( c ) (cid:54) = 0 it is necessary that ψ (cid:48)(cid:48)(cid:48) is positive somewhere andnegative somewhere else, hence somewhere ψ (cid:48)(cid:48)(cid:48) = 0. Usually this is written in terms ofvelocity U decreasing the number of derivatives by one, and hence the name inflectionpoint criterion. In the periodic setting this criterion is not effective, since the derivativeof a periodic function has zero average, and hence U (cid:48)(cid:48) will always change sign. Thispaper is not meant as an elaboration on this criterion, but instead we give a completedescription of the point spectrum of the linearised Euler equations for the case that ψ ( η ) = cos( η ). 3. A complex Hill’s Equation
We now study equation (4) in detail for the specific example where ψ ( η ) = cos( η ).Substituting this into equation (4) we obtain(8) g (cid:48)(cid:48) + Q ( η ) g = µg, Q ( η ) = sin ηc + sin η . Since c ∈ C this is a Hill’s equation with complex valued periodic potential Q ( η ) of period2 π . Much of the Hill’s equation theory (for example from [MW66]) will be applicable,though a few key facts will change. For example, the spectrum of the Hill’s equation,that is, the values µ where there exists a bounded solution (on all of R ) to equation (8),will no longer in general be real.Writing Hill’s equation (8) as a first order system the general solution is given by theprincipal fundamental solution matrix M ( η ). After one period the matrix M (2 π ) iscalled the monodromy matrix, and the trace of the monodromy matrix is called Hill’sdiscriminant ∆( µ ). Sometimes we will emphasise the dependence of the parameter c that changes the function Q ( η ) and will write ∆( µ ; c ). The boundaries of the intervalswhere bounded solutions exist are given by ∆( µ ) = 2 with Floquet multipliers +1 and∆( µ ) = − −
1. We are looking for solutions of equation (5)with quasi-periodic boundary conditions with Floquet multipliers e ± πiθ , such that werequire(9) ∆( µ ) = 2 cos(2 πθ ) . In figure 1 we show the function ∆( µ ) for real and purely imaginary values of c .We are now going to discuss some properties of the spectrum of the complex Hill’sequation (8). First we recall the classical case with real c >
1. Then we consider thecase of purely imaginary c , and finally the case of complex c . ULER, HILL, EVANS 7 - - - - - - μ - - - Δ - - - - - μ - - - Δ Figure 1.
Hill discriminant ∆( µ ) for equation (8); dashed lines at ± πθ . Left: c = 2; For real c > µ ≥
0. Right: c = 0 . i ; For pureley imaginary c there maybe spectrum for µ = d > Real c . When c ∈ R then Q ( η ) is a real valued 2 π -periodic function and thestandard theory of the Hill’s equation applies when c > c < −
1) such that Q ( η ) issmooth. In particular we have (see, e.g., [MW66, DLMF])
Theorem 3.1.
For given θ the equation ∆( µ ) = 2 cos 2 πθ has infinitely many solutions.There is a largest solution µ ∗ such that for µ > µ ∗ there are no solutions. For analytic Q ( η ) the discriminant ∆( µ ) is an entire function of µ . In the standard theory the so called region of stability is defined as those µ for which | ∆( µ ) | ≤
2. Even though we are interested in the case of positive µ = d = k /p forthe Euler equation, in figure 1 we include negative µ to connect to the classical theoryof Hill’s equation.3.2. Purely imaginary c . As is evident in figure 1 right, the discriminant and thespectrum of the Hill’s equation is real even though Q ( η ) is not real for purely imaginary c . The potential Q ( η ) has the property that Q ( η ) = Q ( − η ) which is called space-time orPT-symmetry [Ben07], which is a special case of so-called pseudo-hermiticity [SGH92],which means that there is a unitary involution that intertwines the operator with itsadjoint [CGS04]. The special case of the PT-symmetric Hill’s equation has been studiedin [BDM99]. The main conclusion from PT-symmetry in our case is that for purelyimaginary c Hill’s discriminant is real even though the potential Q ( η ) is complex, whichwe will see in the next section. Note that as in [Shi04] this does not imply that the wholespectrum is real.When c is purely imaginary and sufficiently small there is some spectrum for µ > µ ) with increasing purely imaginary c is illustrated in figure 2. From these graphs we observe that for c = it there is somespectrum with µ > µ = 0 for 0 ≤ t ≤ / √ t > / √
2. In theboundary case (figure 2 middle) we have ∆ (cid:48) (0) = 0. The condition for existence of It is usually formulated for π -periodic function with average zero; in our case the average g isnon-zero and shifts the spectral parameter µ . HOLGER R. DULLIN AND ROBERT MARANGELL - - - μ - - - Δ - - - μ - - - Δ - - - μ - - - Δ Figure 2.
Hill discriminant ∆( µ ) for equation (8) with purely imaginary c from left to right c = i/ , i/ √ , i ; dashed lines at ± πθ . For small | c | there is spectrum with 0 < µ <
1. When | c | reaches 1 / √ (cid:48) (0) = 0. For | c | > / √ µ > Figure 3.
Contours of ∆( d , c ) on the plane ( d, (cid:61) ( c )) for (cid:60) ( c ) = 0. Thecontours are equidistant in values of ∆ from − .
2. Themaximum of ∆ = 2 occurs at c = i/ √ d = 0 and the curve extendsto d = 1. The value ∆ = − c = 0 and d = √ /
2. The green curve is an approximation obtained froma 3 × θ are constant oneach contour, c.f. equation (9).some spectrum with µ > (cid:48) (0) <
0, since asymptotically ∆( µ ) → + ∞ for µ → + ∞ .The limiting case t → c = 0 in the limit) is easilysolvable because then Q ≡
1. This limit corresponds to a case when the parameters ofEuler’s equation are on the boundary of the unstable disk, compare theorem 2.2.
Lemma 3.2.
In the case c = 0 the real spectrum µ = d of the Hill’s equation (8) withboundary conditions equation (5) satisfies ( θ + l ) + d = 1 for any integer l . ULER, HILL, EVANS 9
Figure 4.
Unit circles in the ( θ, d )-plane with centres at ( l,
0) for l = − , . . . ,
2. On these circles c = 0 is in the spectrum. A fundamentaldomain is indicated by dashed lines. The numerals indicate the numberof circles in which a given point lies, see definition 5.5. Proof.
When c = 0 we have Q = 1, and the Hill’s equation (8) simplifies to g (cid:48)(cid:48) + (1 − d ) g = 0with solutions g ( η ) = exp( ± i √ − d η ). The boundary condition equation (5) is g (2 π ) =exp(2 πiθ ) g (0) and gives θ + l = ±√ − d for integer l . Squaring both sides and re-arranging gives the result. (cid:3) The equation ( θ + l ) + d = 1 for integer l defines a sequences of circles of radius one inthe ( θ, d )-plane. The centres of the circles are at ( − l, l ’s differ by 2 touch in one point, see figure 4. Considering θ as an angle with period 1 this becomes a picture of a single circle with diameter twowrapped around a cylinder with circumference 1. A convenient fundamental domain is0 ≤ θ ≤ / d ≥ ± πiθ ) one can restrict to non-negative θ . Since the spectral parameter µ = d wecan restrict to d ≥ d instead of µ = d on the horizontal axis. The regions of stability where | ∆( d ; c ) | < θ . Note that his diagram directly relates the spectral parameter µ = d of theHill’s equation to the spectral parameter c of Euler’s equation for the case of purelyimaginary c . The intersection of ∆ = 2 with the boundary µ = 0 occurs at c = i/ √
2, ascan also be seen in figure 2 (middle), where ∆ (cid:48) ( µ ) = 0 at µ = 0. Along the axis c = 0 wehave shown in lemma 3.2 that ∆( d , c = 0) = 2 cos(2 π √ − d ) such that the minimumwhere ∆ = − d = √ / Remark 3.3.
The derivative of ∆ can be computed at µ = 0 because when c / ∈ [ − , c + sin η is a solution to equation (8). Reduction of order gives a secondsolution, and using these, and variation of parameters, the derivative of ∆ with respectto µ at µ = 0 can be computed as(10) ∆ (cid:48) (0) = 2 π c (1 + 2 c )( c − / . Figure 5.
Contours of (cid:60) (∆( µ )) on the complex µ -plane with zero-locusof (cid:61) (∆( µ )) in red for c = i . c = 0 . i . c =0 . i . − , . . . , .
2. In the middle figure there is a real positive value of µ in the spectrum, but not in the right figure.This shows that ∆ (cid:48) (0) is indeed zero at c = i/ √
2. Thus for some choice of c / ∈ [ − , (cid:48) (0) is real and negative, there will bepoints of spectrum of the Hill’s equation which are real and positive. These satisfy somequasi-periodic boundary conditions, and hence there is some spectrum of equation (8)on the real axis. For imaginary c = iβ, β ≥ (cid:48) (0) is negative for β ∈ (0 , / √ Complex c . For complex c the Hill’s equation (8) generally has a complex mon-odromy matrix, and hence also ∆( µ ) will in general be complex. Our boundary con-ditions do imply, however, that ∆( µ ) = 2 cos 2 πθ be real. This needs to be read as acomplex equation, in particular we require (cid:61) (∆( µ )) = 0. For complex c the spectrumwill in general be complex. Even though complex µ is irrelevant for the application tothe Euler equation we briefly digress into complex µ . In figure 5 we show a contourplot of (cid:60) (∆( µ )) on the complex µ -plane for fixed c . Overlaid are the curves where (cid:61) (∆( µ )) = 0in red. The reason to consider the curves of spectrum in the complex domain is thatwhat we are looking for the special case when these curves intersect the the axis of real µ for positive µ and with | ∆( µ ) | <
2. In figure 5 left such a case can be seen as theintersection of the red curve of spectrum with the positive real µ axis inside the shadedregion where | ∆( µ ) | < c for which this figure is drawn is a complexeigenvalue of the Euler equation with wave number k = ±√ µp .After this brief overview of the spectrum of the complex Hill’s equation we are now goingto compute the Hill discriminant as a Fredholm determinant and using this formulaconstruct an Evans function for the linearised Euler equation. ULER, HILL, EVANS 11 A Hill determinant
In order to prove facts about the spectrum of the linearised Euler equations via thediscriminant of Hill’s equation we are now going to use Hill’s determinant. This isslightly more general than in [MW66] because here Q ( η ) is a complex valued function,but this does not affect the computation of the Hill determinant. We need to first findthe complex Fourier series of the function Q ( η ) := sin ηc + sin η . The Fourier series is more conveniently expressed in terms of a parameter s related to c by(11) c = 12 (cid:18) s + 1 s (cid:19) . This is the well-known conformal Joukowski map that transforms the upper half-disk ofradius 1 into the lower half plane. The upper half of the unit circle s = e iφ is mappedonto the unit interval c = cos φ , and circles of radius r < r + 1 /r ) / r − /r ) /
2. The two solutions of thequadratic equation for s in terms of c are combined to produce a single-valued inversemap from s ( c ) such that for | c | → ∞ we have s → s ( c ) is not continuous along the branch cut where the argument of the squareroot is negative, which occurs along the real interval [ − , s ( c ) is holomorphic on C \ [ − , c to s implies | s | ≤ | s | = 1 is attained along the branch cut. The inversefunction s ( c ) satisfies s ( c ) = − s ( − c ) and s (¯ c ) = s ( c ), and so it is sufficient to considerthe positive quadrant of the s -plane. Substituting c = c ( s ) into Q ( η ) gives(12) Q ( η ) = 1 i ( s + s ) ( e iη − e − iη ) + 1 = 1 − − s s (cid:18) − sie iη + 1 − − s/ ( ie iη ) (cid:19) Formally expanding this as geometric series in s gives the Fourier series Q ( η ) = (cid:80) ∞ k = −∞ g k e ikη where g = 1 + κ, g k = κi k s | k | , κ = − s − s . Because | s | < | s | < κ is in the left half plane, (cid:60) ( κ ) <
0. Expandinga solution of Hill’s equation as a Fourier series the solvability condition leads to thenormalised Hill determinant. Following [MW66] in our particular case we find
Proposition 4.1.
The Hill-determinant of Hill’s equation (8) is (13) D (Λ) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ˜ g n − m Λ − n + δ n,m (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≈ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) − iκs Λ − κs Λ − iκs Λ − − iκs Λ − − κs Λ − iκs Λ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) where ˜ g k (cid:54) =0 = g k and ˜ g = 0 . The Hill-discriminant is (14) ∆( µ ) = 2 − D ( g − µ ) sin ( π √ g − µ ) . This is Hill’s fundamental result that relates the trace of the principal fundamentalsolution matrix of equation (8) – the Hill discriminant ∆( µ ) – to the Hill determinant D .In his original work [Hil78] Hill used these equations to determine the Floquet exponentand hence the stability of a particular solution. More generally Hill’s formula can beviewed as an equation for the spectral parameter µ = d for which quasiperiodic solutionswith Floquet exponent θ exist. Instead in our setting the spectral parameter d and theFloquet exponent θ are given, and we consider this as an equation for the parameter c (or, equivalently s ) for which such a solution exists. Here the Fourier coefficient g andthe higher order coefficients g k hidden in the definition of Hill’s determinant all dependon s . This is a change of interpretation by which the original spectral parameter λ andhence c of the linearised Euler equations becomes a parameter in Hill’s equation, whilethe spectral parameter µ = d in Hill’s equation is a parameter in the separated Euler’sequation.The three cases discussed in the previous section now appear like this: 1) For real c > s < µ . 2) For purely imaginary c we have purely imaginary s , so that all g k are real because i k s | k | gives an even power of i . Hence the Hill determinant is real andthe Hill discriminant is real for real µ . 3) For complex c also s is complex and the Hilldeterminant and discriminant are complex.The beauty of Hill’s formula is that it gives a very good approximation already for smalltruncation sizes. This was also important in Hill’s original application to the Moon,where he found good results already from 3 × D = 0. Now D factors because of discrete symmetry and expressing s in termsof κ in the 3 × D (1 + κ − d ) ≈ d ((2 + d ) κ − d )(3 κ − d κ − d )(1 + κ − d )( d − κ ) (1 − κ ) The vanishing of the last factor in the numerator is shown as a green line in figure 3. Thetwo endpoints of the green line correspond to the roots of this factor where ( d, κ ) = (1 , c = 0 and ( d, κ ) = (0 , − / √
3) and hence c = i/ √
2. It should be noted thateven though the result at the two endpoints is exact, for intermediate points along theboundary curve there are corrections from incorporating higher order approximations tothe determinant. As already noted, along the lower boundary of figure 3 where c = 0the exact answer ∆( d ) is obtained from Hill’s formula where D = 1 since s = ± i andhence κ = 0. As already noted for d = 0 we always have g = c + sin η as a solution thatis 2 π periodic, and hence ∆ = 2. For the Hill discriminant this leads to ∆(0) = 2, andhence D ( g ) = 0. This surprising identity appears because for d = 0 the three centralcolumns of the Hill determinant are linearly dependent. It should be noted that the 3 × ULER, HILL, EVANS 13 approximation to the Hill determinant produces the exact values at the edges d = 0 and c = 0 of figure 3.In section 5 we are going to show that when the circles in ( θ, d ) space from lemma 3.2where c = 0 are crossed in the correct direction then a pair of real eigenvalues is born.This is based on the following well-known formula, see, for example, [GK69]: Theorem 4.2.
The derivative of a Fredholm determinant f = det F is given by ˙ f = f tr( F − ˙ F ) . In general the computation of F − may of course be difficult, but in our case this turnsout to be simple at c = 0. Lemma 4.3.
The derivative of the Hill determinant D ( g − µ ) with respect to µ at c = 0 vanishes. The derivative of the Hill determinant D ( g − µ ) with respect to c at c = 0 vanishes. In both cases it is assumed that − µ (cid:54) = n for any integer n .Proof. From [MW66] we have that D ( g − d ) = det F is a Fredholm determinant.Indeed, there it is shown to be of Hill type, which means that the sum of the absolutevalues of the off-diagonal entries converges. A necessary condition for this is (cid:80) | g k | < ∞ which holds for | s | < s = ± i (so that κ = 0), and hence for all c ∈ C exceptfor c ∈ [ − , ∪ (0 , − µ (cid:54) = n ensures that the denominatorsin equation (13) are non-zero. So in particular we can apply theorem 4.2 at c = 0 aslong as d (cid:54) = 0 and d (cid:54) = 1. For this we need F − at c = 0. Now c = 0 implies s = ± i and κ = 0, hence all Fourier coefficients g k = 0 and F and F − are the identity. Thusthe trace of ˙ F needs to be computed. For either derivative the main observation is thatthe diagonal of F is constant, and hence the diagonal of ˙ F vanishes. Thus tr ˙ F = 0 andboth results follow. (cid:3) Note that this proof was easy because for c = 0 we have Q = ψ (cid:48)(cid:48)(cid:48) /ψ (cid:48) = ± const for ψ (cid:48) ( η ) = sin( η ) or cos( η ), and hence the Fourier series of Q is trivial. For a more generalperiodic function F will have additional entries, and hence F − will be more compli-cated, and even though the diagonal of ˙ F still vanishes the result does not immediatelyfollow.Even though d = 1 is excluded in the above proof, the Hill discriminant (and itsderivatives) are of course well defined in this case as well. This is so because the Hill-determinant D is multiplied by sin ( π Λ) in equation (14), and Λ = 0 for d = 1 and c = 0. 5. An Evans function
Hill’s determinant allows us to compute Hill’s discriminant, which is the trace of thefundamental solution matrix after one period. In our setting this trace is given by2 cos 2 πθ . Using Hill’s discriminant we can build a function that vanishes when c is inthe discrete spectrum of the linearised Euler operator. Figure 6.
Complex colour wheel graphs of the Evans function E ( c ) for d = 0 . θ = 0 .
1, 0 .
22, 0 .
4, showing the positive octant of the complex c plane up to (1 + i )0 .
6. Shown are cases with a pair of imaginary roots,two pairs of imaginary roots, and a quadruplet of complex roots, respec-tively.
Theorem 5.1.
The function (16) E ( c ; θ, d ) = − πθ ) + 4 D ( g − d ) sin π (cid:112) g − d is an Evans function for the separated ODE equation (3) where D ( g − d ) is the Hilldeterminant depending on c through the Fourier coefficients g k of Q ( η ) = sin η/ ( c +sin η ) . Equating equation (9) and equation (14) we see that by construction we have that E ( c ) = 0 when c is in the discrete spectrum of the Euler equation. Unlike the Hilldiscriminant which foremost is a function of the spectral parameter µ of Hill’s equation,we define the Evans function as a function of the spectral parameter c of the linearisedEuler’s equation.For the Evans function µ = d and θ are parameters, and we are trying to find a (com-plex) c such that ∆( d ; c ) = 2 cos 2 πθ . In figure 8 we show a contourplot of (cid:60) (∆( d ; c ))on the complex c -plane for fixed d . Overlaid are the curves where (cid:61) (∆( µ ; c )) = 0 inred. Since the dependence of the Evans function on θ only appears in the single cosineterm, this way of presenting gives the location of complex roots of the Evans functionfor fixed d and varying θ .We note that we recover the circles from lemma 3.2 from E (0; θ, d ) = 0: When c = 0then s = ± i and hence g = 1 and ˜ g k = 0 which implies D (1 − d ) = 1, such that E (0; θ, d ) = − ( πθ ) + 4 sin ( π (cid:112) − d )and thus on the circles ( θ + l ) + d = 1 the Evans function vanishes. A contour-plotof E (0; θ, d ) is shown in figure 7 left. At the minimum ( ± ,
0) the value is −
4, while atthe maximum (0 , √ /
2) the value is 4. The critical point at the origin is degenerate. Inthe right part of figure 7 the eigenvalue is fixed to c = 0 . i , and the 0-contour showsfor which values of ( θ, d ) this eigenvalue occurs. Since the dependence of the Evansfunction on θ is simple the zero contour can be given explicitly by solving E = 0 for θ ULER, HILL, EVANS 15 - - θ d - - θ d Figure 7.
Contourplot of E ( c = 0; θ, d ) (left) and of E ( c = 0 . i ; θ, d )(right) as a function of the parameters θ and d . Contour lines haveinteger values, and in the green shading is the 0-level that shows wherethe corresponding eigenvalue c occurs (left c = 0, right c = 0 . i ).such that sin πθ = D ( g − d ) sin π (cid:112) g − d where the right hand side is considered as a function of d for constant c . The maximumof the 0-contour of E = 0 occurs where D ( g − d ) = 0. All 0-contours start at the originbecause D ( g ) = 0.One property of an Evans function is that it limits to a constant for large arguments: Lemma 5.2.
The function E ( c ; θ, d ) decays to a non-positive real constant for large | c | lim | c |→∞ E ( c ; θ, d ) = 2 cos(2 πθ ) − πd ) = − ( πθ ) − ( πd ) . Proof.
When | c | → ∞ then s → g k → D ( g − d ) = 1. Also g → ix ) = i sinh( x ) gives − πθ ) − ( π √ d ) and the result follows. (cid:3) Notice that the limiting values of the Evans function at infinity is non-zero unless θ = d = 0. Lemma 5.3. If c ∈ R , c (cid:54)∈ [ − , ∪ (0 , and if c ∈ i R , then E ( c ) is real for θ, d ∈ R .Proof. When c is real and c (cid:54)∈ [ − , ∪ (0 ,
1] then Q ( η ) is bounded, real, and smoothon [0 , π ], and so has a real Fourier series, and hence D ( g − d ) must be real. All ofthe other terms in the expression for E ( c ) are real if θ and d are. When c is purelyimaginary then by equation (11) s is also purely imaginary. This implies that κ < g k Hill’s matrix F isHermitian so that Hill’s determinant D ( g − d ) is real. The argument of sin in the lastterm may be either real of purely imaginary, but in either case the resulting value of E ( c ) is real. (cid:3) The first part of this lemma reflects the fact that if c is real and c (cid:54)∈ [ − , ∪ (0 ,
1] thenequation (8) with quasi-periodic boundary conditions represents a self adjoint operator,and so the spectrum must be real. The second part of this lemma is related to thefact that when c is purely imaginary then equation (8) is PT-symmetric as discussedabove.The Euler equation is an infinite dimensional Hamiltonian system, and hence the spec-trum must satisfy the symmetry of a Hamiltonian system. This leads to the follow-ing Lemma 5.4.
Let c be a root of the Evans function E ( c ) = 0 for fixed real ( θ, d ) . Thenalso E ( − c ) = 0 , E (¯ c ) = 0 , and E ( − ¯ c ) = 0 .Proof. Consider the potential Q in equation (8) and note that Q ( η, c ) = Q ( η, c ). Since µ is real equation (8) becomes its conjugate. Thus, if y , y is a basis of solutions for thesolution space to y (cid:48)(cid:48) + Qy = µy , then y , y are a basis for y (cid:48)(cid:48) + Qy = µy . Thus we havethat ∆( c ) = ∆( c ), that is, the Hill discriminant of the original equation (8), with c → c is the complex conjugate of the original Hill discriminant (again provided µ ∈ R ). Thismeans that E ( c ) = E ( c ) for real θ and d . In particular, if E ( c ) = 0 then also E ( c ) = 0.To consider what happens when we change the sign of c , note that for the potential Q in equation (8) we have that Q ( − η, − c ) = Q ( η, c ). This means that a fundamentalset of solutions (normalised to the identity) for y (cid:48)(cid:48) + Q ( η, − c ) y = µy is y ( − η, c ) and − y ( − η, c ), where y ( η, c ) and y ( η, c ) is a fundamental solution set to equation (8). Inparticular, we have that ∆( − c ) = y (2 π, c ) − y (cid:48) (2 π, c ) = y ( − π, c ) + y ( − π, c ). Nextwe observe that Floquet’s theorem says that for any t ∈ R Φ( η + 2 π ) = Φ( η )Φ(2 π ),where Φ( η ) is a principal fundamental solution matrix to equation (8). Letting η = − π gives the result. (cid:3) Alternatively, note that s ( − c ) = − s ( c ) and s (¯ c ) = s ( c ). The symmetry under s →− s can be seen directly in the Hill matrix. The factor κ ( s ) is unchanged, hence g is unchanged. In the Hill matrix every odd off-diagonal entry changes sign under s → − s . This can be compensated by conjugating by diag( . . . , +1 , − , +1 , − , +1 , . . . )which leaves Hill’s determinant unchanged. Conjugating c also conjugates s , κ and g . Further, since µ is real, in the Hill matrix, every odd off-diagonal entry willhave the opposite sign compared to the complex conjugated matrix entry, so againdiag( . . . , +1 , − , +1 , − , +1 , . . . ) provides a conjugacy to the original matrix such that D ( g − d ; s ) = D ( g − d ; s ). Together with the fact that θ ∈ R , this implies E ( c ) = E ( c )as claimed.We want to track the roots of E ( c ) as we move around in ( θ, d ) space. The circles where c = 0 shown in figure 4 divide the parameter plane ( θ, d ) into 3 types of regions: Definition 5.5.
0: A point in region 0 is outside all circlesI: A point in region I is contained inside exactly one circleII: A point in region II is contained inside exactly two circles
ULER, HILL, EVANS 17
Figure 8.
Contours of (cid:60) (∆( µ ; c )) on the complex c -plane with zero-locusof (cid:61) (∆( d ; c )) in red for d = 1 / d = 1 / πθ ) ranging from − , . . . , .
1. The red curve indicates the location ofthe spectrum of Euler’s equation for µ = d and θ determined by thecontourline intersected.In order to show how the spectrum changes when the parameters cross from one regionto another the derivatives of the Evans function are needed at the boundary circles where c = 0. Lemma 5.6.
The derivative of the Evans function at c = 0 satisfies ± i ∂E ( c = 0) ∂c = 2 π sin(2 π √ − d ) √ − d = 1 − d ∂E ( c = 0) ∂d . Proof.
Notice that c = 0 implies s = ± i . In lemma 4.3 we established that D = 1 at c = 0and that any derivative of D at c = 0 vanishes. Thus computing the c - or d -derivativeof the Evans function boils down to computing the derivative of 4 sin π (cid:112) g − d . Thisgives ∂E∂d (cid:12)(cid:12)(cid:12)(cid:12) c =0 = 4 π sin π √ − d cos π √ − d √ − d ( − d )using g = 1 at c = 0. Instead of the c -derivative we start by computing the s -derivate,using the same observation as before. The only difference is the final factor which comesfrom the chain rule, which now is dg ds (cid:12)(cid:12)(cid:12)(cid:12) ± i = − s (1 − s ) (cid:12)(cid:12)(cid:12)(cid:12) ± i = ∓ i . Using ∂c/∂s | ± i = 1 the result follows. (cid:3) The two different signs in the c -derivative are related to the fact that c = 0 is on thebranch cut. Considering the pre-images under the Joukowski map gives s = ± i . When approaching c = 0 from the inside of the unit disk in the s -plane we findlim (cid:15) → + c ( ± i (1 − (cid:15) )) = ∓ i(cid:15) − (cid:15)/ − (cid:15) so that s = ± i corresponds to approaching the origin c = 0 from below (for s → + i ) orfrom above (for s → − i ) the branch cut. Lemma 5.7.
The directional derivative in the normal direction (from inside to out) atthe circles ( θ − l ) + d = 1 , l ∈ Z , where c = 0 is ∇ ( θ + l,d ) E (0; θ + l, d ) = − π sin(2 π √ − d ) √ − d . This is positive for | d | < √ / and negative for | d | > √ / .Proof. By Lemma 3.2 the circles are given by ( θ + l ) + d = 1 and the normalisedgradient at the point ( θ + l, d ) is ( θ + l, d ). Using lemma 5.6 the directional derivativeof E ( c, θ, d ) along the locus where c = 0 thus is ∇ ( θ + l,d ) E (0 , θ + l, d ) = − π (cid:32) ( θ + l ) sin(2 π ( θ + l )) + d sin(2 π √ − d ) √ − d (cid:33) = − π (cid:32) ± (cid:112) − d sin( ± π (cid:112) − d ) + d sin(2 π √ − d ) √ − d (cid:33) and the result follows when putting this on the common denominator. The sin functionchanges sign when √ − d = 1 /
2. In the limit d → − π . (cid:3) The last lemma is describing the property of the contour-plot of E as a function of ( θ, d )shown in figure 7 left. On the upper arc of the circle θ + d = 1 where d > √ / E ( c = 0; θ, d ) is decreasing when crossing the circle towards its centre.On each of the lower arcs of the circles ( θ ± + d = 1 where d < √ / Proposition 5.8.
When crossing the circle ( θ + l ) + d = 1 , l ∈ Z , where c = 0 from outside to in (i.e towards the centre point) and d (cid:54) = 0 , ±√ / , a pair of purelyimaginary roots of E ( c ) is born out of the real axis at c = 0 .Proof. Consider the Evans function along a curve γ ( t ) in ( θ, d )-parameter space thatcrosses the circle ( θ + l ) + d = 1 for t = 0 in the normal direction with unit speed fromthe outside to the inside and define F ( c, t ) = E ( c ; γ ( t )). The function F ( c, t ) defines acurve c ( t ) through c = 0 at t = 0 if the implicit function theorem holds, i.e. when the c -derivative of F is non-zero at c = t = 0. According to lemma 5.6 this holds as long as ULER, HILL, EVANS 19 d (cid:54) = 0 and d (cid:54) = ±√ /
2. The derivative of the curve c ( t ) is given by the implicit functiontheorem as dcdt = − ∂F/∂t∂F/∂c (cid:12)(cid:12)(cid:12)(cid:12) , = − ∇ γ (cid:48) E∂E/∂c (cid:12)(cid:12)(cid:12)(cid:12) ,γ (0) = ∓ i using lemmas 5.6 and 5.7. As before the ± sign refers to the two sides of the branch cutat c = 0 corresponding to s = ± i . Thus two pure imaginary roots are created. (cid:3) The result fails at d = 0 and at d = ±√ / E as a function of ( θ, d ) shown in figure 7 right. In the right panel c = ± . i , and the contour-line E ( ± . i ; θ, d ) = 0 is “moved” relative to the nearbycircles in the direction of the inward normal of the circle. The result fails for technicalreasons at d = 1 because then lemma 4.3 does not hold.In some sense the last proposition is a bifurcation result where a pair of imaginaryroots bifurcates out of the origin. Note, however, that unlike in a bifurcation of rootsof a polynomial where the derivative of the polynomial vanishes, here the derivative isnon-zero.Having described how a pair of purely imaginary eigenvalues c (corresponding to a pairof real λ , hyperbolic and hence unstable) are created the last step is to make sure thatthey their imaginary part does not vanish. Pairs of purely imaginary eigenvalues cancollide and form a complex quadruplet. In principle a complex quadruplet of eigenvaluesmay collide on the real c -axis and form a pair of real eigenvalues (imaginary λ , ellipticand hence stable). The following lemma shows that this is impossible. Lemma 5.9.
Let c be real and such that there exists a quasiperiodic eigenvalue of equa-tion (8) with a continuous eigenfunction. Then c = 0 .Proof. If c ∈ R and c ∈ [ − , ∪ (0 , Q ( η ) is singular and unbounded,and so there are no positive eigenvalues µ with continuous eigenfunctions here. When c ∈ R , c (cid:54)∈ [ − ,
1] equation (8) is a Hill’s equation with eigenvalue µ = 0. We have that g ( η ) = c + sin η is a non-vanishing periodic eigenfunction on [0 , π ]. Because it has nozeros, it is thus the eigenfunction corresponding to the largest periodic eigenvalue, andwe can conclude that all of the Hill’s spectrum will lie to the left of d = µ = 0. Thus wecan not have any bounded solutions for c ∈ R with c (cid:54)∈ [ − , , with a positive eigenvalue µ either. Thus the only value of c ∈ R for which there exists positive quasiperiodiceigenvalues of equation (8) is c = 0. (cid:3) We are now ready to state the main result of this section. We have already establishedthat there is an eigenvalue c = 0 on the circles in lemma 3.2. Whenever two circlesintersect the multiplicity doubles. In the following theorem we are counting all otherisolated eigenvalues. Theorem 5.10.
Consider the isolated eigenvalues c as given by the zeroes of the Evansfunction E ( c ; θ, d ) other than c = 0 . When ( θ, d ) is in region 0 there are no eigenvalues. θ d Figure 9.
Positive quadrant of the scaled unstable disk θ + d < ≤ θ < /
2. Curves of constant purely imaginary eigenvalues c/i =0 . , . , . . . , . c (or (cid:60) ( c ) = 0 . , . , . . . , .
8) in green, with varying imaginary part of c . Region I with a purely imaginary pair of eigenvalues is bounded by thered line. Below the red line in region II there are either two imaginarypairs of eigenvalues, indicated by intersecting blue contours, or there is acomplex quadruplet as indicated by the green contours. When ( θ, d ) is in region I there is a pair of purely imaginary eigenvalues. When ( θ, d ) isin region II there are 4 eigenvalues, either two purely imaginary pairs, a single pair withmultiplicity 2, or a complex quadruplet. On the boundary between region I and region IIthere is a pair of purely imaginary eigenvalues. For all cases half of the eigenvalues areunstable.Proof. We have been considering the quasiperiodic eigenvalue problem equation (8) withparameter c ∈ C and are looking for real positive eigenvalues µ = d with Floquetexponent 2 πθ . We have constructed an Evans function E ( c ; θ, d ) viewed as a function of c with parameters θ and d . By Li’s unstable disk theorem [Li00], see above theorem 2.2,there are no isolated eigenvalues in region 0. This implies that the Evans function E ( c )has no zeroes when the parameters ( θ, d ) are in region 0. Consider a path of parameters( θ, d ) from region 0 to region I that crosses the boundary circle away from d = √ / d = √ / d = 0 another pairof purely imaginary roots is created. Hence there are 4 isolated eigenvalues in regionII. Should the two pairs of eigenvalues coincide they are counted with multiplicity. Thisdoes in fact occur, and leads to a complex quadruplet of eigenvalues. Since we are ULER, HILL, EVANS 21 counting with multiplicity there are always 4 isolated eigenvalues in region II. Becauseof lemma 5.9 it is not possible that the quadruplet of eigenvalues bifurcates onto the realaxis and by Hamiltonian symmetry of the spectrum half of the eigenvalues correspondto unstable eigenvalues. (cid:3)
To conclude this section we present some numerical results that show how the eigenvalueschange in the fundamental domain in ( θ, d )-space. Throughout we have considered( θ, d ) as continuous parameters, such that the eigenvalues c become functions of theseparameters. For the case of purely imaginary c curves of constant c are plotted in the( θ, d )-plane, see figure 9. The blue curves are given in parametric form by ( θ ( d ) , d ) where θ ( d ) is determined by E ( c ; θ, d ) = 0, or equivalently ∆( d ; c ) = 2 cos(2 πθ ) for constantpurely imaginary value of c . For the green curves consider c = c r + ic i for constant realpart c r and the imaginary part c i as a parameter. For such given c find d such that (cid:61) ∆( d ; c r + ic i ) = 0. Then determine θ from ∆( d ; c ) = 2 cos(2 πθ ) as before and thusobtain the green curve ( θ ( d ( c i )) , d ( c i )) in parametric form. The extremal values of thecurve parameter c i are determined by ∆ = ±
2. These occur at the origin in ( θ, d ) andat θ = 1 /
2, respectively. 6.
Back to Euler’s equations
So far we have constructed an Evans function E ( c ) for given parameters ( θ, d ), deter-mined by the wave number k and the overall integer parameter vector p . the originalEuler equations these are related to an integer mode number vector a = k q + l p = d p ⊥ + θ p . When a is inside the unstable disk corresponding perturbations have eigen-values λ = − ikc with non-zero real part. To find the unstable spectrum of the Eulerequations linearised about cos( p x + p y ) the eigenvalues corresponding to all parameters( θ, d ) inside the unstable disk need to be combined into a single Evans function. In theprevious section it was convenient to define the Evans function E ( c ) to be a functionof c = iλ/k . Now putting them together we return to the original temporal spectralparameter λ . Theorem 6.1.
An Evans function for the 2-dimensional Euler equations on the toruswith coordinates ( x, y ) linearised about the steady state cos( xp + yp ) for integers p , p with gcd( p , p ) = 1 is given by E p ( λ ) = p − (cid:89) k =1 E ( iλ/k ; k p · q /p , k/p ) where p = (cid:112) p + p .Proof. The wave number k is an integer giving the periodicity of the perturbation in the ξ -direction. It determines the parameters( θ ( k ) , d ( k )) = ( k q · p /p mod 1 , k/p ) . As noted before θ is only defined mod 1. Now we are going to choose a fundamentaldomain so that θ ∈ ( − / , / By Li’s unstable disk theorem [Li00], see theorem 2.2,there are no unstable eigenvalues when | k | > p . When | k | = p we are on the boundaryof the unstable disk when θ ( k ) = 0, and outside otherwise. When we are on the boundaryof the unstable disk according to lemma 3.2 we have c = 0, and hence no eigenvalue λ with non-zero real part. To capture the point spectrum we can thus restrict to | k | < p .Note that for large | k | the corresponding point ( θ ( k ) , d ( k )) may be outside (or on theboundary of) the unstable disk.When k = 0 equation (3) becomes the trivial equation f (cid:48)(cid:48) = 0 which has no periodicsolutions other than the constant solution. So we can restrict to non-zero waver numbers k with | k | < p . Since E ( c ; θ, d ) in equation (16) is even in θ and in d we can restrict topositive k and instead square the Evans function. This shows that in E p every eigenvaluehas at least multiplicity 2. So we end up with a product of E ( c ; θ ( k ) , d ( k )) over k from1 to p − θ, d ) are functions of k . Finally also c is a function of k , since theEvans function of the full Euler equations instead of the separated ODE is a function of λ = − ikc .Since the point spectrum has only finitely many elements there is no convergence problemwith the product. Nevertheless, we note that using lemma 5.2 it is possible to normaliseeach Evans function such that its absolute value approaches 1 at infinity. (cid:3) Evans functions are only defined up to a non-zero multiplicative constant, so more, orfewer, terms could have been included in the product in Theorem 6.1. An obvious choicefor more terms would have been to include all k ∈ Z , while an obvious choice for fewerterms would be to remove any k such that ( θ ( k ) , d ( k )) lie outside the unstable disk (seethe leftmost panel of figure 10). Each of these choices would have produced an Evansfunction with the same roots as the one in Theorem 6.1. However, each of these choiceshas its own computational disadvantages. For example, choosing all k ∈ Z requires oneto normalise the Evans function for each class, in both λ , and k , while choosing only theterms inside the unstable disk is cumbersome (and different) for different choices of p ,and so becomes convoluted to compute in practice. We have chosen this particular setof values of k because it produces an Evans function that is readily computable, whilestill showcasing the importance of the unstable disk (and the lack of contribution fromterms outside it).In [LLS04, Theorem 3] it was proved that the number of non-imaginary isolated eigen-values does not exceed twice the number of integer lattice points inside the unstabledisk, not counting any lattice points on the line through p (in particular not countingthe origin). Numerical experiments in [Li00] and also in [DMW16] indicate that thisupper bound is sharp.For example for p = (4 ,
5) there are 128 lattice points in the open unstable disk of radius √
41, not counting the origin. In this case let k = 1 , . . . ,
40, and 11 times ( θ ( k ) , d ( k )) Even though the Evans function is even in θ , the set of points ( θ ( k ) , d ( k )) does not possess thatsymmetry. Thus for counting lattice points the domain is θ ∈ ( − / , / θ ∈ [0 , /
2] is enough.
ULER, HILL, EVANS 23 lies in region I, 26 times in region II, once (for k = 9) on the inner boundary betweenregion I and II, once (for k = 39) on the boundary of the unstable disk, and once (for k = 40) in region 0 (outside the unstable disk). On the inner boundary a pair is bornat the origin, but at that moment it is not yet an unstable eigenvalue. Thus there are(11 + 1) · · k = 39 and k = 40 on and outside the unstable disk produce an Evans function factorthat does not have any eigenvalues with non-zero imaginary part, in fact it is non-zeroeverywhere in C \ i [ − , θ, d ) inside the unstable disk, the other half are obtained for negative k . They sit in the opposite half of the circle, relative to the line through the latticepoint p . See [DMW16] for a few more examples of this type. Counting the number oflattice points inside a circle around the origin is a hard problem. But we don’t have toactually count this number, we just have to show that the Evans function E p ( λ ) hasthis number of zeroes. More precisely we have to establish that for each factor with 2or 4 roots according to theorem 5.10 there are 2 or 4 lattice points inside the unstabledisk. Combining all of the results about the Evans function we can now prove our maintheorem. Theorem 6.2.
The number of eigenvalues λ with non-zero real part for the 2-dimensionalEuler equations on the torus linearised about the steady state with vorticity cos( p x + p y ) for integers p , p with gcd( p , p ) = 1 is equal to twice the number of non-zero integerlattice points inside the unstable disk of radius (cid:112) p + p .Proof. By construction lattice points ( θ ( k ) , d ( k )) in the product that defines E p areinside the fundamental domain θ ∈ ( − / , / a = d p ⊥ + θ p which may or may not be inside the unstable disk defined by a < p .Now consider the number of lattice points a = k q + l p for given k and arbitrary l thatare inside the unstable disk. We are going to show that when ( θ ( k ) , d ( l )) is in regionI that number is 1, while when ( θ ( k ) , d ( l )) is in region II that number is 2. The mainobservation is that figure 10 and the regions it defines are translation invariant by ± q direction. The unstable disk is the disk in the centre. Any point in region I inthe unstable disk shifted by ± θ ( k ) , d ( k )) in region I that point is the only lattice point in the unstabledisk. For region II the situation is more interesting. Consider a lattice point ( θ ( k ) , d ( k ))(corresponding to a ) inside region II in the unstable disk, say with θ <
0. Shifting thislattice point by 1 to the right (corresponding to a + p ) it is again in region II, and stillinside the unstable disk. Similarly for θ > a − p ). Thus for every lattice point ( θ ( k ) , d ( k )) inside region II in the unstable diskthere are two lattice points inside the unstable disk.Now according to theorem 5.10 for ( θ, d ) in region I the Evans function has two roots,while for ( θ, d ) in region II the Evans function has four roots. Combining this with theobservation about lattice points just made we see that each lattice point accounts fortwo roots. The exceptional case is when ( θ ( k ) , d ( k )) is on the boundary between region I and regionII. As shown in theorem 5.10 this factor has two eigenvalues with non-zero real part.For this point ( θ ± , d ) (corresponding to a ± p ) is either outside or on the boundaryof the unstable disk, and hence again the number of the number of eigenvalues (2) istwice the number of lattice points (1). Any points ( θ ( k ) , d ( k )) in region 0 are outside theunstable disk and the corresponding factor has no eigenvalues. Any points ( θ ( k ) , d ( k ))in the boundary between region 0 and region 1 are not lattice points inside the unstabledisk and only have eigenvalue λ = 0, and hence have no eigenvalue with non-zero realpart. This exhausts all possible cases.Finally, for every point ( θ, d ) in the fundamental domain inside the unstable disk thereis another point ( − θ, − d ) inside the unstable disk. The Evans function E ( c ; θ, d ) is evenin θ and d , and this is why only positive k are included, but each factor is squared.This doubles the number of lattice points, but also doubles the number of roots. Thusaltogether the number of eigenvalues as given by roots of E p ( λ ) is equal to twice thenumber of non-zero lattice points inside the unstable disk. (cid:3) Summary and Discussion
We have constructed an Evans function for the linearised operator about shear flowsolutions to the Euler equations of the form ψ = cos( p x + p y ) for relatively prime p and p , and have proven the sharpness of the upper bound of the number of points inthe point spectrum found in [LLS04], that is the number of discrete eigenvalues of theoperator is exactly twice the number of non-zero integer lattice points inside the so-calledunstable disk. To compute the Evans function we separated variables in the linearisationand transformed the dispersion relation of a class into an equation of Hill’s type. Wewere then able to use properties of the Hill determinant to determine that points ofspectrum for this class coincided with integer lattice points inside the unstable disk.Finally, we multiplied each of the relevant Evans functions for each class intersectingthe unstable disk, resulting in the function given in the Evans function for the full 2+1dimensional linear operator given in theorem 6.1.We note that in theorems 6.1 and 6.2, we explicitly require gcd( p , p ) = 1. The reasonis that otherwise we cannot conclude that there is a uni-modular transformation to ournew coordinates in section 2. In particular, we are not allowed to treat p = (0 , p can be recovered.Lastly, we note that while consider the original shear flow and its perturbations as be-ing on the torus, we could have considered them on R , and then the perturbationswe have considered are those co-periodic in both x and y . Relaxing the conditions ofco-periodicity in x for example, we would get an additional parameter correspondingto the Floquet exponent of the (now only quasi-periodic) boundary conditions of equa-tion (3), which would then filter through the calculations, effectively shifting the pointsof spectrum in the complex plane. Relaxing the co-periodicity condition in y woulddo the same, only now the shifting parameter is the (now continuous) parameter k . Itremains to be seen which curves and/or surfaces would appear as the points of spectrum ULER, HILL, EVANS 25 are plotted in terms of these additional parameters, as well as their relation to points inand on an “unstable disk”, which to the best of our knowledge, has yet to be defined forsuch boundary conditions. 8.
Acknowledgements
The authors would like to thank G. Vasil for the very useful discussions which sparkedthis project. RM would like to acknowledge the support of Australian Research Councilunder grant DP200102130.
Appendix A. Relation to Jacobi Operators
In previous papers [Li00, LLS04, DMW16] the analysis of the spectrum of the 2-dimensionalEuler equations on the torus around a shear flow was done by analysing the operator inits representation in Fourier space, which after block-decomposition leads to so calledclasses which are represented by Jacobi operators. Let a j be the Fourier mode corre-sponding to the lattice point a + j p ∈ Z . Then the Jacobi operator of the class lead by a is given by the difference equation R ( j )( a j +1 − a j − ) = λa j , R ( j ) = (cid:107)| p || − − || a + j p || − , where λ is the spectral parameter. Since R ( j ) = P ( j ) /Q ( j ) is rational in j the recursionrelation can be written in polynomial form as P ( j )( a j +1 − a j − ) − λQ ( j ) a j = 0 . This can be considered as coming from a Fourier transform, and to identify this werelabel the coefficients such that P + ( j + 1) a j +1 + P − ( j − a j − − λQ ( j ) a k = 0 . Let’s start at the other end and consider a generalisation of a Hill’s equation in theform Au (cid:48)(cid:48) + Bu (cid:48) + Cu = 0where A, B, C are periodic functions. In the simple case at hand they are just linearcombinations of sin and cos. Now assuming a Fourier series solution u ( η ) = (cid:80) a j e ijη andinserting this into the ODE and collecting coefficients in terms of e ijη gives a recursionrelation for the coefficients a j . We obtain (cid:88) ( − Aj + iBj + C ) e ijη = 0Instead of derivatives we have multiplication by powers of ij . When multiplying theterm a j e ijη in the Fourier series by a term e ilη of a periodic coefficient the product is a j e i ( j + l ) η . To get the recursion relation for the a j all exponential terms with a shiftedexponent have to be relabelled to a j − l e ijη .Since the Fourier transform turns differentiation into multiplication we see that thenumber of Fourier modes in the coefficients A, B, C of the ODE determines the order ofthe DE, while the degree of the coefficients
P, Q in the DE determines the number ofderivatives in the ODE.
Note that the ODE can always be transformed into Hill’s form (different Q !) u (cid:48)(cid:48) + Qu = 0where AQ = C − B / (4 A ) − B (cid:48) / BA (cid:48) / (2 A ). However, this transformation changesthe boundary conditions, which is why we need to consider quasi-periodic boundaryconditions for a Hill’s equation in standard form. Appendix B. Counting Modes
We define the following map on classes. The class corresponding to the line y = p p x + kp is mapped to the point ( θ ( k ) , d ( k )) in the square S = [ − , ] × [ − ,
1] given by θ ( k ) = k p · q p mod 1 d ( k ) = kp . The θ coordinate is in the right range by construction, while the d coordinate can beseen to be in the right class by substituting the equation for the class y = p p x + kp into the equation for the unstable disk x + y ≤ p and observing that points inside the unstable disk on the class have x ∈ ( r , r ) the rootsof the corresponding quadratic. The discriminant of this quadratic is (cid:113) − k p = √ − d and so any (real) roots will have d ∈ [ − , θ ∈ [0 , θ ∈ [ − / , /
2] because of the lemma where the circles come from. The circles describethree regions in ( θ, d ) space which are described in figure 10. Now instead we considertwo circles in ( θ, d ) space, one centred at (0 ,
0) and one centred at (1 , θ + d = 1and ( θ ± + d = 1 in ( θ, d ) space.We have the following three lemmata Lemma B.1.
Suppose that θ ( k ) + d ( k ) > and that ( θ ( k ) − + d ( k ) > . Thenthere are no integer lattice points on the line y = p p x + kp inside the unstable disk. Lemma B.2.
Suppose that θ ( k ) + d ( k ) < and that ( θ ( k ) − + d ( k ) < . Thenthere are two integer lattice points on the line y = p p x + kp inside the unstable disk. Lemma B.3.
Suppose that either θ ( k ) + d ( k ) > and that ( θ ( k ) − + d ( k ) < or θ ( k ) + d ( k ) < and that ( θ ( k ) − + d ( k ) > . Then there is a single integer latticepoint on the line y = p p x + kp inside the unstable disk. ULER, HILL, EVANS 27 - - θ d - - θ d - - θ d Figure 10.
The regions in ( θ, d ) space (from left to right) lemmas B.1to B.3 respectively.
Proof.
All three lemmas can be proved at once, by computing the intersection of the lineof a class with the unstable disk in the coordinates set by p and q . That is we compute (cid:18) mk (cid:19) = (cid:18) q − q − p p (cid:19) (cid:18) x ± y ± (cid:19) where x ± and y ± are the intersection points of the class y = p p x + kp with the unstabledisk x + y = p . This gives (cid:18) m ± k (cid:19) = (cid:32) − k p · q p ± (cid:113) − k p k (cid:33) and so now the only thing left is to determine how many integers lie between m ± for afixed k , but this is exactly what is described in the statement of the lemmata. That is if θ ( k ) + d > θ ( k ) + 1) + d > m + and m − and hence no integer lattice points of the class y = p p x + kp inside the unstable disk, &tc.Expanding on this a bit, this is the number of integer points between a ± √ − b (inclu-sive). This can be thought of as the number of integers n such that ( n − ( a + √ − b ))( n − ( a − √ − b )) = n − an + ( a + b − ≤ a ∈ [ − , ] and b ∈ [ − , n = ± , a = b = 0. The lower bound is clearly 0, whichis achieved when b = ±
1, but more than that, if a + b > a + 1) + b > , (cid:3) References [Ach91] David J Acheson.
Elementary fluid dynamics . Clardendon, Oxford, 1991.[BDM99] Carl M Bender, Gerald V Dunne, and Peter N Meisinger. Complex periodic potentials withreal band spectra.
Physics Letters A , 252(5):272–276, 1999.[Ben07] Carl M Bender. Making sense of non-hermitian hamiltonians.
Reports on Progress in Physics ,70(6):947, 2007.[BHLL18] Blake Barker, Jeffrey Humpherys, Gregory Lyng, and Joshua Lytle. Evans function compu-tation for the stability of travelling waves.
Philosophical Transactions of the Royal SocietyA: Mathematical, Physical and Engineering Sciences , 376(2117):20170184, 2018.[BJK11] J. C. Bronski, M. A. Johnson, and T. Kapitula. An index theorem for the stability of periodictravelling waves of korteweg-de vries type. In
Proc. Roy , pages 1141–1173, Soc. Edinburgh,2011. Sect. A no. 6.[BJK14] J. C. Bronski, M. A. Johnson, and T. Kapitula. An instability index theory for quadraticpencils and applications.
Comm. Math , 327(2):521–550, 2014.[BNaKZ12] B. Barker, P. Noble, and L. Rodrigues adn K. Zumbrun. Stability of periodic Kuramoto-Sivashinsky waves.
Appl. Math. Letters , 25(5):824–829, 2012.[BR05] J. Bronski and Z. Rapti. Modulational instability for nolinear Schr¨odinger equations with aperiodic potential.
Dyn. Part. Diff. Eq. , 2(4):335–355, 2005.[CGS04] Emanuela Caliceti, Sandro Graffi, and Johannes Sj¨ostrand. Spectra of pt-symmetric opera-tors and perturbation theory.
Journal of Physics A: Mathematical and General , 38(1):185,2004.[CM20a] W.A. Clarke and R. Marangell. A New Evans Function for Quasi-Periodic Solution of theLinearised Sine-Gordon Equation.
Journal of Nonlinear Science , DOI: 10.1007/s00332-020-09655-4, 2020.[CM20b] W.A. Clarke and R. Marangell. Rigorous justification of the whitham modulation theory forequations of nls type. arXiv preprint arXiv:2011.09656 , 2020.[DK06] B. Deconinck and J.N. Kutz. Computing spectra of linear operators using the Floquet-Fourier-Hill method.
Journal of Computational Physics , 219:296–321, 2006.[DLM +
20] Holger R. Dullin, Yuri Latushkin, Robert Marangell, Shibi Vasudevan, and Joachim Wor-thington. Instability of unidirectional flows for the 2D α -Euler equations. Communnicationson Pure and Applied Analysis , 19:2051–2079, 2020.[DLMF]
NIST Digital Library of Mathematical Functions . http://dlmf.nist.gov/, Release 1.0.24 of2019-09-15. F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F.Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds.[DMW16] Holger R Dullin, Robert Marangell, and Joachim Worthington. Instability of equilibria forthe 2D Euler equations on the torus.
SIAM J. Appl. Math. , 76(4):1446–1470, 2016.[DMW19] Holger R Dullin, James D Meiss, and Joachim Worthington. Poisson structure of the three-dimensional Euler equations in Fourier space.
Journal of Physics A: Mathematical and The-oretical , 52(36):365501, 2019.[DW18] Holger R Dullin and Joachim Worthington. Stability results for idealised shear flows on arectangular periodic domain.
J. Math. Fluid Mech. , 20:473–484, 2018.[DW19] Holger R. Dullin and Joachim Worthington. Stability theory of the three-dimensional Eulerequations.
SIAM J. Appl. Math. , 79(5):2168–2191, 2019.[Gar93] R. A. Gardner. On the structure of the spectra of periodic travelling waves.
J. Math PuresAppl. , 72(5):415–439, 1993.[GH07] T. Gallay and M. Hˇarˇagu¸s. Stability of small periodic waves for the nonlinear Schr¨odingerequation.
J. Differential Equations , 234:544–581, 2007.[GK69] Izrail Cudikoviˇc Gohberg and M. G. Krein.
Introduction to the theory of linear non-selfadjoint operators , volume 18 of
Transl. Math. Monogr.
Amer. Math. Soc., 1969.[Hil78] G. Hill. Researches in the lunar theory.
Am. J. Math. , 1(5):129–245, 1878.[HLZ17] Jeffrey Humpherys, Gregory Lyng, and Kevin Zumbrun. Multidimensional stability of large-amplitude navier–stokes shocks.
Archive for Rational Mechanics and Analysis , 226(3):923–973, 2017.
ULER, HILL, EVANS 29 [HW79] G. H. Hardy and E. M. Wright.
An Introduction to the Theory of Numbers . Clarendon Press,Oxford, 1979.[JLM13] C. Jones, Y. Latushkin, and R. Marangell. The Morse and Maslov indices for matrix Hill’sequations.
Proceedings of Symposia in Pure Mathematics , 87:205–233, 2013.[JMMP13] C. K. R. T. Jones, R. Marangell, P. D. Miller, and R. G. Plaza. On the stability analysis ofperiodic sine-gordon traveling waves.
Physica D: Nonlinear Phenomena , 251:63–74, 2013.[JMMP14] C. K. R. T. Jones, R. Marangell, P. D. Miller, and R. G. Plaza. Spectral and modula-tional stability of periodic wavetrains for the nonlinear klein-gordon equation.
Journal ofDifferential Equations , 257(12):4632–4703, 2014.[KP13] Todd Kapitula and Keith Promislow.
Spectral and dynamical stability of nonlinear waves .Springer, 2013.[LBJM19] K. P. Leisman, J. C. Bronski, M. A. Johnson, and R. Marangell. Stability of traveling wavesolutions of nonlinear dispersive equations of nls type. arXiv preprint arXiv:1910.05392 ,2019.[Li00] Y. C. Li. On 2D Euler equations. I. On the energy-Casimir stabilities and the spectra forlinearized 2D Euler equations.
Journal of Mathematical Physics , 41(2):728 – 758, 2000.[LLS04] Y. Latushkin, Y. C. Li, and M. Stanislavova. The spectrum of a linearized 2D Euler operator.
Studies in Applied Mathematics , 112:259–270, 2004.[LV19] Y. Latushkin and S Vasudevan. Eigenvalues of the linearised 2D Euler equations via Birman-Schwinger and Lin’s operators.
J. Math. Fluid Mech. , 20(4):1667–1680, 2019.[MW66] W Magnus and S Winkler.
Hill’s Equation . John Wiley, New York, 1966.[SGH92] FG Scholtz, HB Geyer, and FJW Hahne. Quasi-hermitian operators in quantum mechanicsand the variational principle.
Annals of Physics , 213(1):74–101, 1992.[Shi04] Kwang C Shin. On the shape of spectra for non-self-adjoint periodic schr¨odinger operators.
Journal of Physics A: Mathematical and General , 37(34):8287, 2004.[TDK18] O. Trichtchenko, B. Deconinck, and R. Koll´ar. Stability of periodic traveling wave solutionsto the kawahara equation.
SIAM J. Appl. Dyn. Syst. , 17(4):2761–2783, 2018.
School of Mathematics and Statistics, The University of Sydney, Australia
Email address : [email protected] School of Mathematics and Statistics, The University of Sydney, Australia
Email address ::