1:
2: \section{Numerical computation and results} \label{Result}
3: \reseteqnos
4:
5: \subsection{Evaluation of
6: Evans functions}\label{NumWron}
7:
8: To evaluate the Evans function $E_j(\lambda)$ in \eq{42Wdef}
9: for given $j$ and $\lambda$, we use a Runge-Kutta ODE package (rksuite) to
10: compute numerical approximations to the solutions
11: $\yf_1^{(0)}(r)$ and $\yf_2^{(0)}(r)$ of \eq{41yf} that are bounded at
12: the origin. To do this, we solve
13: two initial value problems for the system \eq{41yf}
14: on an interval $r_0\le r\le r_\infty$, where $r_0$ and $r_\infty$
15: are chosen so that $w(r)$ is small (usually $<10^{-4}$) on
16: $(0,r_0)$ and $(r_\infty,\infty)$. Then $\alpha(w(r))$ and $\beta(w(r))$
17: are close to their asymptotic values on
18: the intervals $(0,r_0)$ and $(r_\infty,\infty)$.
19: We make the approximations $\yf_i^{(0)}(r_0)\approx \hyf_i(r_0)$
20: ($i=1,2$) to determine the initial conditions.
21: We then evaluate the Wronskian of the computed approximations to
22: $\yf_1^{(0)}(r_\infty)$ and $\yf_2^{(0)}(r_\infty)$ with the
23: explicitly known approximations $\hyf_3(r_\infty)$ and
24: $\hyf_4(r_\infty)$ to the respective values $\yf_3^{(\infty)}(r_\infty)$
25: and $\yf_4^{(\infty)}(r_\infty)$ of the solutions to
26: the system \eq{41yf} that vanish at infinity.
27:
28: Our actual numerical implementation of this scheme makes use of a
29: change of variables and scaling to reduce the sometimes rapid variation of
30: solutions and improve the quality of the computation.
31: We represent a solution $\yf(r)$ of the system \eq{41yf} as
32: an $r$-dependent scaled linear combination
33: of the basis elements $\left\{\hyf_i:i=1,\dots,4\right\}$
34: for the asymptotic system \eq{yasymsys}, writing
35: \be
36: \yf(r)=\hY(r)\exp(\bmK r)\bmb(r)
37: \label{yYa}
38: \ee
39: where $\bmK:={\rm diag}\{-k_+,-k_-,k_+,k_-\}$,
40: Then we can anticipate that the vector $\bmb(r)$ is well-behaved
41: as $r$ becomes large, and it satisfies the system
42: \be
43: \bmb'(r)=\bmD(r)\bmb,
44: \label{asys}
45: \ee
46: whose coefficient matrix
47: $\bmD(r)=-\bmK+e^{-K r}{\hY}{}^{-1}(r)\tBf(r)\hY(r)e^{K r}$ remains
48: bounded as $r\to\infty$.
49:
50: The Evans functions $E_j(\lambda)$ are also approximated in a similar but
51: simpler way based on the expression in \eq{4Edef2},
52: by computing an approximation to $e^{-kr}\ysix^\lt(r)$
53: for $r_0\le r\le r_\infty$ and using the known asymptotic approximation
54: for $\zsix^\rt(r_\infty)$. The scaling factor $k$ is given by $k=k_++k_-$.
55: Numerical convergence studies indicate that we usually attain 8 to 10
56: digit accuracy with this approach.
57: This method of evaluating $E_j(\lambda)$
58: is very stable, since $\ysix^\lt$ is a solution of
59: the exterior system with maximal asymptotic growth rate.
60: It also avoids a potential difficulty
61: one can encounter with the Wronskian approach ---
62: maintaining the linear independence of
63: approximations to $\yf^\lt_1$ and $\yf^\lt_2$ is difficult
64: if the growth rates for these solutions differ greatly and if the
65: interval of computation is large.
66:
67: \subsection{Counting eigenvalues with the argument principle}\label{Roundup}
68:
69: Let us first recall what is known about the possible
70: location of unstable eigenvalues.
71: According to Proposition \ref{Psomej} we may locate eigenvalues of $A$ by
72: locating zeros of the Evans functions $E_j(\lambda)$. For a given standing
73: wave $u_0=e^{i\omega t}e^{im\theta}w(r)$, Proposition \ref{L.eigbd}
74: guarantees that all such zeros lie in the strip where
75: \be
76: |\Re\lambda|\le \Lmax
77: := \max_{r\ge 0} (|\alpha(w(r))|+|\beta(w(r))|).
78: \ee
79: Theorem \ref{Plargej} guarantees that all unstable eigenvalues $\lambda$
80: will be zeros of $E_j$ with
81: \be
82: |j|\le j_{\rm max}:=|m|+
83: \max_{r\ge 0}\sqrt{r^2({\alpha (\ws(r))+
84: \left| {\beta (\ws(r))} \right|})}\ .
85: \ee
86: Furthermore, it follows from Proposition \ref{4Esym}
87: that $E_{-j}(\lambda)=0$ if and only if
88: $E_j(\bar\lambda)=0$, so that zeros of $E_{-j}$ in the strip
89: $0<\Re\lambda\le \Lmax$ are in one-to-one correspondence with zeros of $E_j$
90: in that strip.
91:
92: To search for zeros of the Evans functions in the strip
93: $|\Re\lambda|\le \Lmax$, we apply the argument principle to count the
94: number of zeros of the analytic functions $E_j$ inside a contour that
95: encloses a ``large" part of the strip. We do not have a rigorous bound
96: on $|\Im\lambda|$ for unstable eigenvalues, but in practice we find
97: that $|E_j(\lambda)|$ grows for large $|\lambda|$, so we take a large
98: contour determined so that an asymptotic trend appears established.
99: Figure~\figref{ContourH} shows a schematic of the contour $\gamma$
100: employed in our computations, lying in the cut plane $\Ccut$. The
101: contour is designed to enclose the region where $|\Re\lambda|\le\Lmax$
102: and $|\Im\lambda|\le K$, except for an $\epsilon$-neighborhood of the
103: essential spectrum $\{it\mid |t|\ge\sigma^2\}$. The parameter
104: $\epsilon$ is taken small ($\approx 10^{-6}$), and $K$ large.
105:
106: Using the computed approximation to $E_j(\lambda)$ for given $j$ and
107: $\lambda$ described above, we compute the winding number of the image
108: $E_j(\gamma)$ about 0. This is the change in the argument of the
109: complex number $E_j(\lambda)$ as $\lambda$ traverses the curve
110: $\gamma$, divided by $2\pi$. We use an adaptive stepping procedure to
111: traverse $\gamma$ in steps that result in a small relative change in
112: $E_j(\lambda)$, and accumulate the change in ${\rm arg} E_j(\lambda)$.
113: By the argument principle, this yields the number of zeros of $E_j$
114: enclosed by $\gamma$.
115:
116: Because of the symmetry $E_j(-\bar\lambda)=\overline{E_j(\lambda)}$ from
117: Proposition \ref{4Esym}(iii), in practice we need only to compute the
118: change in argument of $E_j(\lambda)$ along the portion $\gamma_r$ of the
119: contour $\gamma$ in the right-half plane $\Re\lambda\ge 0$. Doubling
120: yields the total change in argument along $\gamma$.
121:
122: By the same symmetry, the number of zeros enclosed by $\gamma$ in the
123: right half-plane is the same as the number in the left half-plane.
124: So, from the winding number above, we need to subtract the number of purely
125: imaginary zeros and divide by two to determine the number of {\em
126: unstable} eigenvalues of $A$ corresponding to index $j$ enclosed by
127: $\gamma$. To find the purely imaginary zeros we exploit the fact
128: established in Proposition~\ref{4Esym}
129: that $E_j(\lambda)$ is real if $\lambda$ is purely imaginary.
130: We plot the real function
131: $E_j(it)$ for $t\in[-\sigma^2+\epsilon,\sigma^2-\epsilon]$, refining and
132: checking the multiplicity of zeros with log-log plots as necessary.
133:
134: The computation of the winding number can be considerably slowed by the
135: presence of zeros of $E_j(\lambda)$ very near the contour --- the
136: adaptive stepping procedure takes small steps to track the image curve
137: near such a zero. The reason we do not take our contour to be a
138: simple rectangle with one side along $\Re\lambda=\epsilon$ is to
139: avoid passing near numerous zeros on the imaginary axis between
140: $-i\sigma^2$ and $i\sigma^2$.
141:
142: Nevertheless, it sometimes happens that the values of $E_j(\lambda)$ become
143: small as $\lambda$ traverses the part of the contour $\gamma_r$
144: in the right half-plane near the essential spectrum on the
145: imaginary axis. This behavior
146: suggests that the contour passes near a zero of $E_j$. To verify that
147: these zeros do not lie in the right half-plane outside $\gamma$, we use
148: low-degree least-square polynomial fits to extrapolate an approximation to
149: $E_j(\lambda)$ near $\gamma_r$, and a root finder to locate the zeros.
150: In all cases the zeros appear to lie in the left half-plane.
151: In view of the symmetry $\overline{E_j(\lambda)}=E_j(-\overline{\lambda})$,
152: we interpret these findings to indicate
153: not that $E_j$ has zeros in the cut plane $\Ccut$ outside $\gamma$,
154: but rather that the analytic continuation
155: of $E_j(\lambda)$ across the imaginary axis has zeros near the axis.
156: Such points are known as {\it resonance poles}, see \cite{PW92}.
157:
158:
159:
160: \subsection{Spinless ground-state waves}\label{StabTh}
161:
162: The stability of ground state (nodeless) standing-wave solutions
163: to \eq{1nls} with
164: nonlinearities that include those we consider here is analyzed in
165: \cite{GSS2} and \cite{W86}.
166: These works establish that a family of ground state standing waves ($m=n=0$),
167: parametrized by standing wave frequency $\omega$, is
168: orbitally stable (stable modulo spatial translations and phase shifts
169: under $H^1$ perturbations of initial data) if
170: $dN/d\omega>0$
171: and unstable (to radial perturbations) if
172: $dN/d\omega<0$,
173: where $N=\|u_0\|^2$.
174: For general nonlinearities,
175: the stability of the ground state in the marginal case
176: $dN/d\omega = 0$
177: is open, but for pure-power nonlinearities
178: $g(u)=\gamma |u|^{p-1}u$ (for which $dN/d\omega = 0$
179: is equivalent in $d$ spatial dimensions to
180: $p=1+4/d$) the ground state is unstable in the sense
181: that there exist arbitrarily small $L^2$ perturbations of initial data for
182: which the corresponding solution of (NLS) blows up in $H^1$ norm in finite
183: time. (See \ccite{W83}.)
184:
185: The construction and computation of the Evans functions for these
186: spinless waves differs from the description we have given above,
187: since the wave amplitude $w(r)$ approaches a nonzero value as $r$
188: approaches zero. Nevertheless, through diagonalization of the coupling
189: matrix in \eq{ysys}, explicit (formal) asymptotic solutions may be written
190: and Evans functions computed. For brevity we omit the details,
191: since our primary concern is the spectral stability of standing waves
192: with spin.
193:
194: We performed computations for spinless waves for the cubic and
195: cubic-quintic nonlinearities, both to validate the numerical code and
196: to study the mechanism of transition to instability as the quantity
197: $dN/d\omega$ changes sign.
198:
199: The cubic nonlinear Schr\"odinger equation in two spatial dimensions
200: is a marginal case from the point of view of theory ($dN/d\omega=0$).
201: Recall that by scaling we may without loss of generality
202: take $g(u)=|u|^2u$ and consider only the fixed standing wave
203: frequency $\omega=1$.
204: The bounds from Theorem~\ref{Plargej} and Proposition~\ref{L.eigbd} yield that
205: $|j|\le\jmax=1.85$ and $|\Re\lambda|\le14.6$ for unstable eigenvalues.
206: In fact we find no eigenvalues with positive real part,
207: indicating that the mechanism for instability in this case is not
208: exponential linear instability.
209: The eigenvalue $\lambda=0$ has high multiplicity, however --- it
210: is found to be a zero of $E_j$ of order 4 for $j=0$ and of order 2 for $j=1$.
211: (This remains true for the cubic for all spin/node-number combinations
212: considered.) This order is consistent with the enumeration of zero
213: modes in Appendix B. No other discrete eigenvalues are found.
214:
215: To study transition to instability, we perturb the cubic nonlinearity
216: and analyze the one-parameter family $g(u)=|u|^2u-\delta |u|^4u$.
217: Making use of the scaling transformation
218: $w(r)=\sqrt{\omega}\tilde{w}(\sqrt{\omega}\,r)$, we may without loss of
219: generality study eigenvalues for the single standing-wave frequency
220: $\omega=1$. For the wave to exist,
221: $\delta$ must satisfy $\delta<{3\over 16}$.
222: Numerical computation of $N=2\pi \int_0^\infty {w(r)^2r\,dr}$
223: for the wave profiles indicates that, by the stability criterion
224: for ground states, $u_0$ is stable ($dN/d\omega>0$)
225: for $0<\delta <\textstyle{3\over 16}$ and is unstable ($dN/d\omega<0$)
226: for $\delta<0$.
227:
228: Our numerical search for eigenvalues shows that the transition to
229: instability of the ground state as $\delta$ decreases through zero is
230: characterized by the collision of two purely imaginary conjugate eigenvalues
231: in the twist-0 ($j=0$) subspace. We observe the transition by plotting
232: $\Evans_0(i\tau)$ for $\tau\in (-\sigma^2,\sigma^2)$ for various values of
233: $\delta$. For $0<\delta<\textstyle{3\over 16}$, log-log plots
234: indicate that the function $\Evans_0(i\tau)$
235: has a zero of multiplicity two at $\tau=0$ and two other conjugate simple
236: zeros. As $\delta\to 0^+$, the simple zeros approach $\tau=0$, and at
237: $\delta=0$ there is a zero of multiplicity four at $\tau=0$.
238: For $\delta<0$, two real zeros of $\Evans_0(\lambda)$ emerge
239: with opposite sign on the real axis. This creates a positive eigenvalue
240: that renders the standing wave exponentially unstable.
241:
242: \subsection{Waves with nodes}
243: In all our computations for ``first excited states,''
244: waves whose profiles have a single node
245: where $w(r)=0$ for some $r>0$, we find many unstable eigenvalues.
246: This holds whether the waves have spin or not.
247: For example, for the cubic with $m=0$, $n=1$, we find 12
248: unstable eigenvalues, distributed as
249: $(2,1,1,1,1,1,0\ldots)$, which we abbreviate as
250: $(2,1(5\times),0\ldots)$, meaning
251: two in index $j=0$ and one each in index $j$ for $|j|=1,\ldots,5$.
252: For $m=1$, we find 24 unstable eigenvalues,
253: distributed as $(2(4\times),1(5\times),0\ldots)$.
254: For $m=2$, we find 34, distributed as $(2(6\times),1(6\times),0\ldots)$.
255: The existence of eigenfunctions with twist
256: $j=0$ leads us to expect that the higher-node-number states
257: have exponential instabilities to perturbations of initial data of the
258: form $e^{im\theta}v(r)$.
259: For the cubic-quintic $g(u)=|u|^2u-|u|^4u$ we only mention a few
260: cases with $n=1$ since now different values of $\omega$
261: are in principle different:
262: with $m=0$, $\omega=0.05$ we find 10 unstable
263: eigenvalues;
264: with $m=1$, $\omega=0.16$ we find 20;
265: and with $m=2$, $\omega=0.165$ we find 24.
266: It proved difficult to perform accurate computations for profiles
267: with more than one node ($n>1$), but we expect all such waves to be unstable.
268:
269:
270: \subsection{Spinning waves with no nodes}
271:
272: For the cubic nonlinearity (scaled so $\omega=1$) our numerical
273: results indicate that spinning standing waves with no nodes ($n=0$)
274: are unstable. With spin $m=1$, there are 6
275: unstable eigenvalues\footnote{$1.17\pm1.52i$, $1.97\pm1.69i$, $1.81\pm1.88i$},
276: one for each twist index with $|j|=1,2,3$. And
277: there are 8 eigenvalues in the ``gap'' $\{it\mid
278: -\sigma^2<t<\sigma^2\}$, distributed in $j$ as $(4,2,1,0,1,0\ldots)$. For
279: $m=2$, there are 6 unstable, one for each $j=1,\dots,6$, with 11 gap
280: eigenvalues distributed as $(4,2,1,1,1,0,0,1,1,0\ldots)$.
281:
282: Next we consider the cubic-quintic nonlinearity
283: $g(u)=\gamma|u|^2u-\delta|u|^4u$, where the
284: constants $\gamma$ and $\delta$ are both positive.
285: As mentioned in section \ref{tax}, the scaling
286: $w(r)=\sqrt{\gamma\over\delta}\, \tilde w({\gamma\over{\sqrt\delta}} r)$
287: allows us to suppose without loss of generality that $\gamma=\delta=1$.
288: The allowable range for $\omega$, for which $F(s)>0$ for some $s>0$, is
289: $0<\omega<\omegas={3\over 16}$.
290:
291: {\it Spectrally stable waves.}
292: For the nodeless waves with spins $m=1,2,3,4,5$, we have
293: discovered ranges of standing-wave frequency $\omega$ for which all the
294: Evans functions $\Evans_j$, $j=0,1,\dots,\jmax$, appear to have no zeros with
295: positive real part. Our numerical computations indicate that, for a given
296: value of spin $m$, there is a transition frequency
297: $\mum$ such that $A$ possesses
298: unstable eigenvalues for $0<\omega <\mum$ but $A$ has no unstable
299: eigenvalues for $\mum<\omega<\omegas$.
300: In Table~\ref{T.transit} we list the transition frequency $\mum$,
301: the bound on instability index
302: from Theorem~\ref{Plargej}
303: and the computed distribution of all purely imaginary gap eigenvalues.
304: The figures in parentheses indicate the number of times a count is
305: repeated, so that ``$8(3\times),7(6\times)$'' in the last row of the
306: table means that there are 8 gap eigenvalues for each $j=0,1,2$,
307: and 7 eigenvalues for each $j=3,4,5,6,7,8$. Accordingly, the total number
308: of gap eigenvalues for $m=1,2,3,4,5$ respectively is
309: 22, 44, 78, 154, 268. %13, 24, 41, 80, 138.
310: The bound computed from Lemma \ref{L.eigbd} on the real part of any unstable
311: eigenvalue satisfies $\Lmax<1$ in all cases.
312: The value $\lambda=0$ is always found to be a double
313: zero of $E_j$ for $j=0$ and $1$, which corresponds to the four
314: zero modes enumerated in Appendix B.
315:
316: \begin{table}\label{T.transit}
317: \begin{center}
318: \caption{Eigenvalue distribution and bounds at the instability transition
319: for the cubic-quintic nonlinearity.
320: Waves with spin $m$ are spectrally stable for $\mum\le \omega<\omegas$.
321: Unstable eigenvalues emerge at $\lamcr$,
322: and can exist only for $|j|\le \jmax$ by Theorem \ref{Plargej}.}
323: \medskip
324: \begin{tabular}{|c|c|c|c|c|l|}
325: \hline $m$ &$\mum$&$\lamcr$ & $\jmax$ & \# of zeros in gap,
326: for $j=0,1,\ldots$ \\\hline
327: 1 & 0.1487 & 0.0478$i$ & 7.66 & $4,3,3,2,1,0\ldots$ \\
328: 2 & 0.1619 & 0.0271$i$ & 14.2 & $4,4,3(4\times),1(4\times),0\ldots$\\
329: 3 & 0.1700 & 0.0136$i$ & 23.4 & $4,4,5,5,4,3(4\times),2,1(5\times),0\ldots$\\
330: 4 & 0.1769 & 0.0063$i$ & 38.3 &
331: $6(5\times),5(4\times),3(7\times),2,1(7\times),0\ldots$\\
332: 5 & 0.1806 & 0.0033$i$ & 57.6 &
333: $8(3\times),7(6\times),6,5(4\times),3(10\times),
334: 2(3\times),1(10\times),0\ldots$ \\
335: \hline \end{tabular}\end{center}
336: \end{table}
337:
338: %Lmax values: .45 .45 .55 .58 .58
339:
340: For each $m=1,\ldots,5$ the mechanism for transition to instability
341: as $\omega$ decreases below $\mum$ is the same. A pair of
342: purely imaginary gap eigenvalues, always with index $j=2$,
343: collides when $\omega=\mum$ at $\lambda=\lamcr$ (as shown in Table 1)
344: and move away from the imaginary axis in opposite directions, creating
345: an unstable eigenvalue. Figure~\figref{TransitMech} shows how the
346: graph of the real-valued function $E_2(it)$ ($-\omega<t<\omega$)
347: changes as $\omega$ passes through the critical
348: frequency for the case of spin $m=1$.
349: Figure~\figref{TransitWaves} shows the critical wave profiles for
350: $m=1,\ldots,5$.
351:
352: Figure~\figref{TransitCurv}
353: shows a log-log plot of $\mum$ versus $m$. (The
354: transition frequency is calculated for noninteger values of $m$
355: by solving the profile equation \eq{1weq} and computing the associated
356: Evans functions.) From this data it appears that
357: \be
358: \omegas-\mum\sim \frac{0.18}{m^2}
359: \ee
360: as $m$ becomes large, suggesting that
361: stable spinning waves may exist for any spin $m$.
362:
363: After this paper was submitted,
364: the existence of a stability transition frequency $\mum$ for $m=1$ and 2 was
365: also reported in a paper of Towers {\it et al.}\ \cite{TB2},
366: who solve numerically the algebraic eigenvalue problem arising from
367: discretization of a system of equations equivalent to \eq{ysys}.
368: These authors also report finding (at the limit of their numerical accuracy)
369: a weak instability for $j=\pm1$ (see Figure~5 in \cite{TB2})
370: which affects the location of $\mum$ for $m=1$
371: and $2$ and causes them to find all waves with $m=3$ unstable.
372: If our present numerical results are correct,
373: this weak instability is likely to be a numerical artifact,
374: since for $j=1$ the existence of a double eigenvalue at $\lambda=0$
375: is consistent with the analysis of zero modes in Appendix~\ref{ZeroModes},
376: and we find no other zeros of $E_1(\lambda)$ inside or near the
377: contour $\gamma$ except on the imaginary axis, where $E_j(\lambda)$ is real
378: and changes of sign clearly locate zeros.
379:
380: We examined a few cases involving nonlinearities other than
381: cubic and cubic-quintic to see whether spectrally stable localized standing
382: waves may exist.
383: For the focusing-defocusing nonlinearity $g(w)=w^3-w^7$,
384: the critical frequency at which $F$ has two
385: zero-height hilltops is $\omegas\approx 0.2722$.
386: Our computations indicate that the wave with
387: spin $m=1$ for $\omega=0.24$ is spectrally stable
388: ($\jmax=9$ and $|\Re\lambda|<0.91$ for unstable eigenvalues in this case).
389: For the focusing-defocusing nonlinearity $g(w)=w^5-w^7$,
390: the critical frequency $\omegas\approx 0.0878$,
391: and there is a spectrally stable wave
392: with spin $m=1$ for $\omega=0.07$
393: ($\jmax=11$ and $|\Re\lambda|<1.27$).
394:
395: For the saturable nonlinearity $g(w)=w^3/(1+w^2)$, which is always
396: focusing, we found no spectrally stable spinning waves.
397: Spinning solitary waves exist for $0<\omega<1$, growing in
398: amplitude without bound as $\omega$ increases. With spin $m=1$,
399: we find for small $\omega$ that there are three unstable eigenvalues,
400: as in the case of the focusing cubic above.
401: For $\omega\ge0.35$, however, we find only one unstable eigenvalue;
402: it has twist index $j=2$.
403: And as $\omega$ increases towards 1, the real part of the unstable
404: eigenvalue becomes small. E.g., for $\omega=0.5$ the unstable eigenvalue
405: is $\lambda=0.174+0.199i$, and for $\omega=0.9$ it is $0.0372+0.0547i$.
406: As $\omega$ increases, then, we see that the nonlinearity saturates
407: and the wave becomes less unstable.
408:
409: