1: \documentclass[12pt,preprint]{aastex}
2: %\usepackage{lscape,graphicx,fullpage,doublespace}
3:
4: %\usepackage{chicago,graphicx,fullpage,doublespace,aas_macros}
5:
6: \topmargin=+0.5in
7:
8: \newcommand{\eps}{{\varepsilon}}
9:
10: \def\ie {{\it i.e.,~}}
11: \def\vs {{\it vs.~}}
12: \def\etc {{\it etc.~}}
13: \def\etal {{\it et al.~}}
14: \def\citeN {\cite}
15: \newcommand{\V}[1]{{\bf{#1}}}
16: \keywords{outer solar system, dynamics, chaos}
17:
18: \begin{document}
19: %\begin{spacing}{1.5}
20:
21: \title{Surfing on the Edge:\\ Chaos \vs Near-Integrability in the System of Jovian Planets}
22: \author{Wayne B. Hayes}
23: %\maketitle
24: \affil{Computer Science Department University of California, Irvine\\Irvine, California 92697-3435}
25: \email{wayne@ics.uci.edu}
26:
27: \begin{abstract}
28: We demonstrate that the system of Jovian planets
29: (Sun+Jupiter+Saturn+Uranus+Neptune), integrated for 200 million
30: years as an isolated 5-body system using many sets of initial conditions
31: all within the uncertainty bounds of their currently known positions, can
32: display both chaos and near-integrability. The conclusion is consistent
33: across four different integrators, including several comparisons against
34: integrations utilizing quadruple precision. We demonstrate that
35: the Wisdom-Holman symplectic map using simple symplectic correctors
36: as implemented in {\it Mercury 6.2} \citep{ChambersMercury99}
37: gives a reliable characterization
38: of the existence of chaos for a particular initial condition only
39: with timesteps less than about 10 days, corresponding to about 400 steps
40: per orbit. We also integrate the canonical DE405 initial condition out
41: to 5 Gy, and show that it has a Lyapunov Time of 200--400 My, opening
42: the remote possibility of accurate prediction of the Jovian planetary
43: positions for 5 Gy.
44: \end{abstract}
45:
46: \section{Introduction}
47:
48: \begin{flushright}
49: {\it Both the man of science and the man of art live always at the edge\\
50: of mystery, surrounded by it. Both, as a measure of their creation, have\\
51: always had to do with the harmonization of what is new with what is familiar,\\
52: with the balance between novelty and synthesis, with the struggle to make\\
53: partial order in total chaos.... This cannot be an easy life.}\\
54: --- J. Robert Oppenheimer
55: \end{flushright}
56:
57: When one speaks of the stability of our Solar System, one must
58: carefully define the meaning of ``stable''. We say that the Solar
59: System is {\it practically stable} if, barring interlopers, the known
60: planets suffer no close encounters between themselves or the Sun, over
61: the main-sequence lifetime of the Sun. In a practically stable Solar
62: System, the orbital eccentricities, inclinations, and semi-major axes
63: of all the planets remain within some bounded region, not too far from
64: their present values. In this sense, work by many authors over the past
65: 15 years has all but proven that the Solar System is practically stable
66: \citep{Laskar94,Laskar96,Laskar97,ItoTanikawa02}. Good reviews exist
67: \citep{Lissauer99,LecarFranklinHolmanMurray01}, and we will not discuss
68: it further in this paper. A second, more formal definition
69: involves the question of whether the Solar System is {\it chaotic} or not.
70: In a chaotic system, nearby solutions tend to diverge from each other
71: exponentially with time, although in a weakly chaotic system such as
72: the Solar System, the exponential divergence can be preceded by an initial
73: period of polynomial divergence. Let $d(t)$ be the distance between two
74: solutions, with $d(0)$ being their initial separation.
75: Then $d(t)$ increases approximately as $d(0) e^{\lambda t}$ in a chaotic
76: system, where $\lambda$ is the {\it Lyapunov exponent}. The inverse of
77: the Lyapunov exponent, $1/\lambda$, is called the {\it Lyapunov time},
78: and measures how long it takes two nearby solutions to diverge by a
79: factor of $e$. A system that is not chaotic is called {\it integrable}
80: or {\it regular}, and has a Lyapunov exponent of zero. A practical
81: consequence of being chaotic is that small changes become exponentially
82: magnified, so that uncertainties in the current positions of the planets
83: are magnified exponentially with time. Even though the Solar System is
84: practically stable, a positive Lyapunov exponent means that uncertainties
85: in the current positions of the planets are magnified to the point that
86: we cannot predict the precise positions of the planets in their orbits
87: after a few (or at most a few tens of) Lyapunov times.
88:
89: KAM theory tells us that essentially all Hamiltonian systems which
90: are not integrable are chaotic. An initial condition (IC) not lying
91: precisely on a KAM torus will eventually admit chaos, but with a
92: time scale that depends critically on the IC. Symplectic integrators
93: \citep{ChannellScovel90,WisdomHolman91,SanzSernaActaNum92} have many
94: nice properties when used for long-term integrations of Hamiltonian
95: systems, such as conservation of phase-space volume, and bounded energy
96: error. However, the validity of symplectically-integrated numerical
97: solutions also depends critically upon the integration time step $h$,
98: with the longevity of the solution's validity scaling as $e^{a/h}$
99: for some constant $a$ \citep{BenettinGiorgilli94,Reich99}. For linear
100: problems, the dependence is even stronger and manifests itself as a
101: bifurcation in the Lyapunov exponent, going discontinuously from zero to
102: a non-zero value (\citeN{Lessnick96}, \citeN{NewmanLee05} --- but see
103: \citeN{RauchHolman99}). Since the solar system is not integrable, and
104: experiences unpredictable small perturbations, it cannot lie permanently
105: on a KAM torus, and is thus chaotic. The operative question is the
106: time scale of the chaos. To compute the time scale accurately, we
107: must guarantee that the measured time scale is not an artifact of the
108: integration method.
109:
110: What is the Lyapunov time of the Solar System? \citeN{SussmanWisdom88}
111: first demonstrated that the motion of Pluto is chaotic with a
112: Lyapunov time of about 20 million years, corroborated over a longer
113: integration later by \citeN{KinoshitaNakai96}. \citeN{Laskar89} performed
114: an averaged integration of the 8 major planets (excluding Pluto)
115: and found that the Lyapunov time was about 5 million years, with the
116: divergence dominated by that of the inner planets. \citeN{Laskar90}
117: believed that secular resonances are the cause of the
118: chaos in the inner Solar System (but see \citeN{MurrayHolman01}),
119: although he did not believe the system
120: of Jovian planets was affected by the chaos displayed by the inner planets.
121: \citeN{SussmanWisdom92} performed a full (non-averaged) integration of the
122: entire Solar System and confirmed Laskar's 5 million year Lyapunov time,
123: and further found that the system of Jovian planets by itself had a Lyapunov
124: time of between 7 and 20 million years, although their measurement of
125: the Lyapunov time displayed a disturbing dependence on the timestep
126: of the integration. This dependence was later discovered to be due
127: to symplectic integration schemes effectively integrating a slightly
128: different set of ICs; the effect can be
129: corrected \citep{SahaTremaine92,WisdomHolmanTouma96}, although it decreases with
130: decreasing timestep.
131:
132: Since there are no two-body resonances amongst the Jovian
133: planets, the cause of the chaos between them was not understood until
134: \citeN{MurrayHolman99} identified the cause as being the overlap
135: of three-body resonances.
136: \citeN{MurrayHolman99} also performed Lyapunov time measurements
137: on a large set of Outer Solar Systems differing only in the
138: initial semi-major axis of Uranus. They found that their three-body
139: resonance theory correctly predicted which regions of ICs
140: were chaotic, and which were not, at least over the
141: 200-million-year integration timespan they used. For the ``actual''
142: Solar System, they found that the Lyapunov time was about 10 million
143: years. \citeN{Guzzo05} went on to corroborate the three-body
144: resonance theory by performing a large suite of integrations,
145: numerically detecting a large web of three-body resonances in the outer Solar System.
146:
147: \citeN{MurrayHolman99} noted that the widths $\Delta a/a$ of the individual
148: resonant zones was of order $3\times 10^{-6}$, so that changes in the
149: ICs of that order can lead to regular motion. They
150: note, however, that ``the uncertainties in the ICs,
151: and those introduced by our numerical model, are comfortably smaller
152: than the width of the individual resonances, so [the outer] Solar
153: System is almost certainly chaotic.'' Given that \citeN{Guzzo05}
154: has also detected many three-body resonances consistent with
155: Murry + Holman's theory, it would seem at first glance that chaos
156: in the outer Solar System is a fact.
157:
158: However, the conclusion that the isolated outer Solar System is chaotic
159: cannot be taken for granted. For example, it is known that symplectic
160: integration with too-large a timestep can inject chaos into an integrable
161: system \citep{HerbstAblowitz89,NewmanEtAlDDA2000}. Although most authors verify their
162: primary results by performing ``checking'' integrations with smaller
163: timesteps, the checking integrations are not always performed for the
164: full duration of the main integrations. This, combined with the fact
165: that longer symplectic integrations require {\em shorter} timesteps
166: \citep{BenettinGiorgilli94,Reich99} means that one cannot assume that a
167: timestep good enough, for example, for a 100-million-year integration
168: is also good enough for a 200-million-year integration. There is
169: currently no known method for analytically choosing a short-enough
170: timestep {\it a priori}, and so the {\em only} method of verifying the
171: reliability of an integration is to re-perform the entire integration
172: with shorter-and-shorter timesteps until the results converge.
173: \citeN{NewmanEtAlDDA2000} used this method to demonstrate that, for a given
174: set of ICs, the \citeN{WisdomHolman91} symplectic mapping
175: with a 400 day timestep (about 11 steps per orbit, a commonly-quoted timestep)
176: admits chaos,
177: but that the results converge to regularity for any timestep less than
178: about 100 days. However, many authors who find chaos have also
179: performed reasonable convergence tests, demonstrating that the
180: chaos does not always disappear at convergence.
181:
182: There exists compelling evidence for the absense of chaos in the outer Solar System.
183: Laskar and others noted that when the entire solar system is integrated,
184: the inner Solar System manifests chaos on a 5-million-year timescale, but
185: the outer Solar System appears regular in these integrations.
186: Although Laskar's approximate theory can overlook some causes of chaos,
187: there also exist full-scale integrations that indicate the absense of chaos.
188: Grazier \etal (1999) and Newman \etal (2000) %\citep{GrazierNewmanKaulaHyman99,NewmanEtAlDDA2000}
189: utilized a St\"{o}rmer integrator whose per-step numerical errors were bounded by the
190: double-precision machine epsilon (about $2\times 10^{-16}$), as long as the orbital
191: eccentricity is less than 0.5 and 1000 steps or more are taken per orbit.
192: Furthermore, their integration method took care to ensure that the roundoff
193: error was unbiased.
194: Note that, except for the possibility of having the same property with a
195: larger timestep, this is practically as good an integration as is possible using
196: double-precision. Furthermore, such an integration is symplectic by default,
197: since it is essentially {\em exact} in double precision. Using this
198: method, they performed an integration of the Jovian planets lasting
199: over 800 million years, and found no chaos.
200: \citeN{VaradiRunnegarGhil03} performed a 207My integration of the entire
201: Solar System, including even the effects of the Moon, and placed a lower
202: bound of 30My on the Lyapunov time of the system of Jovian planets.
203:
204: We are thus left with the disturbing fact that, utilizing ``best
205: practices'' of numerical integration, some investigators integrate the
206: system of Jovian planets and find chaos, while others do not.
207:
208: In this paper, we demonstrate that this apparent dilemma has a simple
209: solution. Namely, that the boundary, in phase space, between chaos and
210: near-integrability is finer than previously recognized. In particular,
211: the current observational uncertainty in the positions of the outer
212: planets is a few parts in $10^7$ \citep{Standish98-DE405,MorrisonEvans98}.
213: Within that
214: observational uncertainty, we find that some ICs lead to
215: chaos while others do not. So, for example, drawing 7-digit ICs
216: from the same ephemeris at different times,
217: one finds some solutions that are chaotic, and some that are not.
218: Thus, different researchers who draw their initial coniditions from
219: the same ephemeris at different times can find vastly different
220: Lyapunov timescales.
221:
222: \section{Methods}
223:
224: With the exception of the two sets of initial conditions (ICs) we have
225: received from other authors \citep{MurrayHolman99,GrazierNewmanHymanSharp05} and the
226: set included in {\it Mercury 6.2} \citep{ChambersMercury99},
227: all ICs used in this paper are drawn at various epochs from
228: DE405 \citep{Standish98-DE405}, which is the latest planetary ephemeris
229: publicly available from JPL. It has stated uncertainty for the positions
230: and masses of the outer planets of a few parts in $10^7$.
231: To ensure that our integration agrees over the short term with DE405,
232: we verified in several cases that we can integrate between different
233: sets of DE405 ICs, separated by as much as 100 years, while maintaining
234: at least 7 digits of agreement with DE405.
235:
236: We integrate the system of Jovian planets using only Newtonian
237: gravity. The inner planets are accounted for by adding their masses
238: to the Sun and perturbing the Sun's position and linear momentum to
239: equal that of the Sun+Mercucy+Venus+Earth+Mars system. This ensures
240: that the resonances between the outer planets is shifted by an amount
241: that is second order in this mass ratio, roughly $3\times 10^{-11}$
242: \citep{MurrayHolman99}, which is far smaller than the uncertainty in the
243: outer-planet positions. We assume constant masses for all objects and
244: ignore many effects which are probably relevant over a 200My timescale
245: (see for example \citeN{Laskar99}).
246: We account for solar mass loss at a rate of ${\dot m}/m\approx 10^{-7}$
247: per million years \citep{Laskar99,Noerdlinger05}, but note that we
248: observe no noticable difference if we keep the solar mass constant.
249:
250: To reduce the possibility that our results are dependent on the
251: integration scheme,
252: we used three different numerical integration methods to verify many
253: of our results in this paper. First, we used the Mercury 6.2 package
254: \citep{ChambersMercury99}, with the Wisdom-Holman \citep{WisdomHolman91}
255: symplectic mapping option (called {\tt MVS} in the input files).
256: We used stepsizes varying from 2 days to
257: 400 days. Second, we used the {\tt NBI} package,
258: which contains a 14th-order Cowell-St\"{o}rmer method with modifications
259: by the UCLA research group led by William Newman
260: \citep{GrazierNewman95,GrazierNewman96,VaradiRunnegarGhil03}.\footnote{
261: {\tt NBI} is available at {\tt http://astrobiology.ucla.edu/\~{}varadi/NBI/NBI.html},
262: or by searching the web for ``{\tt NBI Varadi}''.}
263: {\tt NBI} has been shown to have relative truncation errors below the
264: double precision machine epsilon (about $2\times 10^{-16}$) when more than
265: 1000 steps per orbit are used and the orbital eccentricity is less than 0.5.
266: More precisely, if the largest component of the phase-space vector
267: of the solution at time $t$ has absolute value $M$, then the local errors
268: per step of each of the components are all less than $2\times 10^{-16}M$.
269: Note that this means that the component-by-component relative error
270: can be significantly greater than the machine precision for components
271: of the solution that are significantly less than $M$, but all components have
272: errors relative to $M$ which are smaller than the machine precision.
273: Furthermore, the authors of NBI have gone through great pains to ensure
274: that the roundoff error is unbiased.
275: We used a 4-day timestep for all {\tt NBI} integrations, which
276: gives more than 1000 timesteps per Jupiter orbit.
277: We have verified the
278: above ``exact to double precision'' property by comparison against
279: quadruple precision integrations described below.
280: Note that for the above-defined timestep, {\tt NBI} gives practically the best integration possible
281: using only double precision, and that such an integration is
282: symplectic by default since it essentially provides a solution which
283: is exact in double precision. In our 200 million year integrations,
284: {\tt NBI} always had relative energy errors and angular momentum errors
285: of less than $2\times 10^{-11}$, with an average of about
286: $2\times 10^{-12}$.
287:
288: Our third integrator was the {\it Taylor 1.4} package \citep{JorbaZou05}.
289: {\it Taylor 1.4} is a recent and impressive advance in integration
290: technology. It is a general-purpose, off-the-shelf integrator which
291: utilizes automatic differentiation to compute arbitrary order Taylor
292: series expansions of the right-hand-side of the ODE. {\it Taylor 1.4}
293: automatically adjusts the order and stepsize at each integration step
294: in an effort to minimize truncation error, and utilizes Horner's rule
295: in the evaluation of the Taylor series to minimize roundoff error.
296: As the authors note, integration accuracy is gained more efficiently
297: by increasing the order of the integration than by decreasing the
298: timestep, since the accuracy increases exponentially with the order
299: but only polynomially in the timestep. Although {\it Taylor 1.4}
300: allows the user to specify a constant order and timestep, we chose to
301: allow it to use variable order and timestep while providing it with
302: a requested relative error tolerance equal to 1/1000 of the machine
303: precision, in order to produce solutions which were exact to within
304: roundoff error. We found that {\it Taylor 1.4} typically used about 27th
305: order with about a 220 day timestep.
306: The fact that the solution is exact to machine precision over such a long
307: timestep guarantees that accumulated roundoff is by far the smallest in
308: the {\em Taylor 1.4} integrations. Furthermore,
309: Taylor integrators are extremely stable when applied to non-stiff problems,
310: with the radius of convergence increasing linearly with integration order,
311: and in our case the timestep is well within the radius of convergence \citep{BarrioBlesaLara05}.
312: Finally, {\it Taylor 1.4} allows
313: the user to specify the machine arithmetic to use, including software
314: arithmetics. Out-of-the-box, {\it Taylor 1.4} supports the use of IEEE
315: 754 double precision (64 bit representation with a 53 bit mantissa),
316: Intel extended precision (80 bit representation with a 64-bit mantissa,
317: giving a machine precision of about $10^{-19}$, accessible as
318: {\tt long double} when using GCC on an Intel machine), the {\tt DoubleDouble}
319: datatype \citep{BriggsDoubleDouble} which provides software quadruple
320: precision in C++, and the GNU Multiple Precision Library, which
321: allows arbitrary precision floating point numbers in C++. Most of our
322: integations using {\it Taylor 1.4} used Intel extended precision, which
323: is almost as fast as double precision and gives about 19 decimal digits
324: of accuracy. Over our 200-million-year integrations using Intel extended
325: precision, Taylor typically had relative energy errors of less than
326: $8\times 10^{-14}$; the worst relative energy error observed in any of our
327: integrations was $2\times 10^{-13}$.
328: Integrations began with the Solar System's barycentre at the origin with zero velocity.
329: After 200 million years the barycentre drifted a maximum of
330: $3\times 10^{-10}$ AU,
331: while the $z$ component of the angular momentum was always conserved to
332: a relative accuracy better than $3\times 10^{-14}$.
333: We also performed a suite of quadruple precision integrations, in
334: which energy and angular momentum were each conserved to at least 26 significant
335: digits over 200 million years.
336: \clearpage
337: \begin{figure}[hbt]
338: \centering
339: \includegraphics[scale=0.5,angle=270]{f1.eps}
340: \caption{{\it Taylor 1.4} satisfies Brouwer's Law:\ when local tolerance is set
341: above machine precision of $10^{-19}$,
342: phase-space error grows as $t^2$ due to biased truncation errors (upper three curves).
343: But with tolerance set well below the machine precision, its result is exact (\ie correctly
344: rounded) to machine precision, so that phase-space error grows only as $t^{1.5}$ (lowest
345: curve).}
346: \label{fig:TaylorBrouwer}
347: \end{figure}
348: \clearpage
349: Now we analyze the error growth as a function of time for our
350: non-symplectic integrators, applied to a non-chaotic IC.
351: One consequence of having a solution whose numerical error is dominated
352: by unbiased roundoff (ie., exact to machine precision) is that when
353: integrating a non-chaotic system, the total phase-space error grows
354: polynomially as $t^{1.5}$. If the local error is biased (either
355: by biased roundoff, or due to truncation errors in the integration
356: scheme), then the total phase-space error grows as $t^2$. (This
357: is assuming that the integrator is not inherently symplectic, as is
358: the case with both NBI and {\it Taylor 1.4}.) This is
359: known as {\it Brouwer's Law} \citep{Brouwer37}. As noted above,
360: \citeN{GrazierNewmanHymanSharp05} have demonstrated that the error in
361: {\tt NBI}'s integration is dominated by unbiased roundoff when using
362: 1000 timesteps per orbit. We tested {\it Taylor 1.4} in Intel extended
363: precision for similar properties by integrating the system of Jovian planets
364: for 200 million years using various integration tolerances up to and beyond
365: the machine precision of $10^{-19}$, and compared these integrations to
366: a {\it Taylor 1.4} integration that used quadruple precision. In Figure
367: \ref{fig:TaylorBrouwer}, we plot the phase-space separation between
368: the quadruple precision integration and several integrations using
369: Intel extended precision. We see that when the relative integration
370: tolerance is set above the machine precision, the error grows as
371: $t^2$ and is therefore truncation dominated. But when the tolerance is
372: set to $10^{-22}$ (about a factor of 1000 below the machine precision), the
373: error grows as $t^{1.5}$, and is therefore dominated by unbiased roundoff.
374: This is consistent with {\it Taylor 1.4} producing results that are exact in
375: Intel extended precision when given a local relative error tolerance of
376: $10^{-22}$, just as {\tt NBI} produces exact results in double precision
377: when used with 1000 or more timesteps per orbit. We see that after 200
378: million years, the errors in the positions of the planets are of order
379: $10^{-5}$ AU (in the case of non-chaotic ICs). This translates into a phase
380: error, or equivalently an observational error, of substantially less
381: than one arc-minute. A comparison of our Figure \ref{fig:TaylorBrouwer}
382: with Figure 2 of \citeN{GrazierNewmanKaulaHyman99} at their 200 million
383: year mark also demonstrates that {\it Taylor 1.4} using Intel extended
384: precision provides about 3 extra digits of precision over {\tt NBI}
385: using double precision, as would be expected when comparing 19-digit
386: and 16-digit integrations. As we show later (Figure \ref{fig:final-ae-diffs}),
387: the differences in orbital elements are even smaller.
388:
389: For comparison, we briefly mention the run-times of the integration
390: algorithms. All timings are for a 2.8Ghz Pentium 4 processor. As we
391: shall see later, the largest timestep for which the {\it Mercury 6.2}
392: integrations agreed with the others was 8 days; a 16-day timestep was
393: almost as good. Thus we compare the Wisdom-Holman mapping with timesteps
394: of 16 and 8 days to NBI with a 4-day timestep, and {\it Taylor-1.4}.
395: In addition to the {\it Taylor-1.4} extended precision ({\tt long double})
396: timings used in this paper, we present its timings for standard IEEE
397: 754 {\tt double} precision, for comparison to the other {\tt double}
398: precision integrations. Table \ref{tab:runtime} presents the results.
399: We found that the total runtime was linearly proportional to the inverse
400: of the timestep, as would be expected. We note several observations.
401: First, {\it Taylor-1.4} is not competitive in terms of efficiency.
402: However, note that {\it Taylor-1.4} is a ``proof-of-concept'' software
403: package for general-purpose integration of {\em any} system of ODEs,
404: and it currently generates code that can be quite inefficient.
405: A carefully hand-coded Taylor series integrator for Solar
406: System integrations is far more efficient, and can be competitive with
407: the above codes (Carles Simo 2007, personal communication).
408: Second, Wisdom-Holman with an 8-day timestep is the fastest case among
409: the integrations we tested that showed complete convergence. Third,
410: Wisdom-Holman with a 4-day timestep (51 hours, not shown in the table)
411: is slower than NBI with a 4-day timestep. Finally, although we did {\em
412: not} test NBI for convergence at timesteps less stringent than 4 days,
413: it is possible that NBI maintains the lead in being more efficient than
414: Wisdom-Holman, if it also shows convergence at larger timesteps.
415: \clearpage
416: \begin{table}
417: \begin{tabular}{|c|ccccc|}
418: \hline
419: integrator & WH dt=16d & WH dt=8d & NBI & Taylor-1.4 {\tt double} & Taylor-1.4 {\tt long double} \\
420: \hline
421: time (hours) & 13 & 26 & 40 & 100 & 150 \\
422: \hline
423: \end{tabular}
424: \caption{Run times for integrating the 5-body system for 200 million years on a 2.8Ghz pentium 4,
425: for {\it Mercury 6.2} with timesteps of 16 and 8 days, for NBI, and for {\it Taylor-1.4} in
426: {\tt double} and {\tt long double} precision.
427: }
428: \label{tab:runtime}
429: \end{table}
430: \clearpage
431: We will not directly measure the Lyapunov time in this paper, since it
432: is difficult to create an objective measure of the Lyapunov time over a
433: finite time interval. Formally, the Lyapunov time is only defined over
434: an infinite time interval. In practical terms, the divergence is almost
435: always polynomial for some nontrivial duration before the exponential
436: divergence emerges, and it is difficult to pinpoint the change-over
437: objectively. However, when plotting the distance between two nearby
438: trajectories as a function of time, it is usually evident by visual inspection whether or
439: not exponential divergence has occured by the end of the simulation.
440: Thus we will plot the actual divergence between two numerical trajectories
441: initially differing by perturbing the position of Uranus by $10^{-14}$ AU
442: (about 1.5mm) in the $z$ direction. We will call these pairs
443: of trajectories ``siblings.'' In the cases where we see only polynomial
444: divergence between siblings over a 200 million year integration, we will
445: abuse terminology and call these systems ``regular'', ``near-integrable'',
446: or ``non-chaotic'', although formally all we have shown is that the
447: Lyapunov time is longer than can be detected in a 200-million-year
448: integration.
449: \clearpage
450: \begin{figure}[p]
451: \centering
452: \includegraphics[scale=0.5,angle=270]{f2.eps}
453: \caption{Murray + Holman (1999) in quadruple precision.}
454: \label{fig:MH-dd}
455: \end{figure}
456: \begin{figure}
457: \centering
458: \includegraphics[scale=0.5,angle=270]{f3.eps}
459: \caption{Murray + Holman (1999) in {\tt long double}.}
460: \label{fig:MH-ld}
461: \end{figure}
462: \clearpage
463: As a prelude to our main results, we first crudely reproduce the results
464: of \citeN{MurrayHolman99}. Murray + Holman performed a survey in which
465: the semi-major axis of Uranus $a_U$ was varied around its actual value,
466: plotting the measured Lyapunov time as a function of $a_U$. Broadly, they
467: found that below $a_U\approx 19.18$, the Lyapunov time was substantially
468: less than $10^7$ years, sometimes as small as $10^5$ years; in the range
469: $a_U\approx 19.20--19.22$, the Lyapunov time fluctuated between about
470: 1--10 million years, interspersed with cases in which the motion was
471: regular; near $a_U=19.24$ it was uniformly regular; near $a_U=19.26$
472: there was a tightly-packed region of both chaotic and regular orbits;
473: and near $a_U=19.28$ the motion was again uniformly regular. We have
474: reproduced this survey using quadruple precision integrations,
475: using a very course grid in $a_U$ due to computing constraints.
476: (On a 2.8GHz Intel Pentium 4, a quadruple precision integration using
477: {\it Taylor 1.4} proceeds roughly at 100My per month of
478: CPU time.)
479: Figure \ref{fig:MH-dd} displays the distance between siblings,
480: for various values of $a_U$. Although we
481: were not able to perform the survey using a finer grid in $a_U$ due
482: to the computational expense, we see that in broad outline we obtain
483: results similar to Murray + Holman. In Figure \ref{fig:MH-ld},
484: we repeat the same survey using Intel extended precision. We see that
485: the results are qualitatively identical, demonstrating that 19 digits of
486: precision (which requires about 1/20 of the CPU time of quadruple precision)
487: gives qualitatively identical, and quantitatively very similar, results.
488: Henceforth in this paper, all {\it Taylor 1.4} integrations are
489: performed using Intel extended precision, with the relative local error
490: tolerance set to $10^{-22}$.
491:
492:
493: \section{Results}
494: \label{sec:results}
495:
496: \subsection{Corroborating previous results}
497:
498: As an early step, the author obtained from Murray + Holman
499: their initial conditions (ICs) used in \citeN{MurrayHolman99}, and
500: verified using accurate integrations that their system was chaotic.
501: The author then obtained from P. Sharp the ICs used by
502: \citeN{GrazierNewmanHymanSharp05}, and verified that their system was regular,
503: at least over a 200My timespan. The author then spent significant
504: time eliminating several possible reasons for the discrepancy, such as
505: incorrect ICs, incorrect deletion of the inner planets,
506: \etc The author also integrated from one set of ICs
507: to the other (having along the way to account for the fact that they were
508: in different co-ordinate systems), and finally found that the systems agreed
509: with each other to a few parts in $10^7$ when integrated to the same epoch.
510: After some time it became apparent that neither group of authors had made
511: any obvious errors. The conclusion seemed to be that both systems
512: nominally represented the outer Solar System to within observational
513: error, but one was chaotic and one was not.
514:
515:
516: \subsection{Ephemeris Initial Conditions Drawn at Different Times}
517:
518: It is reasonable to make the assumption that either our Solar System is
519: chaotic, or it is not. It cannot be both. This appears to be the assumption
520: that most practitioners make when measuring ``the'' Lyapunov time of
521: our Solar System. Although this is probably
522: a reasonable assumption, it does not follow that all initial conditions (ICs)
523: drawn from an ephemeris are equivalent. The most recent ephemeris published
524: by JPL is DE405 \citep{Standish98-DE405}. It is based upon hundreds of thousands
525: of observations, all of which have finite error. Thus, the ephemeris does
526: not represent the exact Solar System. For the outer planets, the best match
527: of DE405 to the observations yields residual errors of a few parts in $10^7$
528: \citep{Standish98-DE405,MorrisonEvans98}.
529: The masses of the planets are also known to only a few parts in $10^7$.
530: It is possible that within these error bounds there exist different solutions
531: with different Lyapunov times. In particular, it may be possible that some
532: solutions display chaos on a 200-million-year timescale, while others do not.
533: \clearpage
534: \begin{figure}[p]
535: \centering
536: \includegraphics[totalheight=7.5in]{f4.eps}
537: \caption{\small Distance in AU between ``sibling'' trajectories as a function
538: of time for 21 sets of initial conditions drawn
539: from DE405 at 30-day intervals. Each plot represents one initial
540: condition. Siblings are integrated with the Wisdom-Holman symplectic map
541: with timesteps: 50 (solid line), 100 (long dash), 200 (short dash),
542: and 400 days (dotted). For any given plot, note disagreement across timesteps,
543: and that the switch between chaos and regularity is not always monotonic with timestep.}
544: \label{fig:multiplot-big-steps}
545: \end{figure}
546: \clearpage
547: To test this hypothesis, we drew 21 sets of ICs from DE405,
548: at 30-day intervals starting at Julian date 2448235 (9.5 Dec, 1990).
549: We represent each of our 21 sets of ICs using a 3-digit numeral, 000,
550: 030, 060, $\ldots$, 570, 600, representing the
551: number of days after JD2448235 at which the intitial condition is drawn.
552: We chose a 2-year total interval across which to draw our ICs in order
553: to to ensure a reasonable sample of inner-planet positions before deleting
554: the inner planets (a Martian year is about 2 Earth years).
555: We drew the initial conditions by taking the output of the program
556: {\tt testeph.f}, which is included with the DE405 ephemeris. The output
557: of {\tt testeph.f} is rounded to 7 digits, since no more digits are
558: justifiable.\footnote{Although this is probably not the
559: best way to uniformly sample initial conditions from within the error ball
560: representing the error in the observations, it {\em does} represent a
561: reasonable way to reproduce how users of DE405, taking initial conditions
562: from DE405 to 7 digits, get their samples.} As
563: described above, we augment the mass of the Sun with that of the inner planets
564: and augment the Sun's position and linear momentum to match that of the inner
565: Solar System. For each of the 21 ICs, we generate a ``sibling''
566: IC which is offset by $10^{-14}$ AU (1.5mm). We then
567: integrate both of them for 200 million years, and measure the distance between
568: them at 1-million-year intervals. To ensure that our results are not integrator-
569: or timestep-dependent, we perform each integration in several ways:
570: first, with the Wisdom-Holman symplectic mapping included with {\it Mercury 6.2}
571: \citep{ChambersMercury99} with timesteps of 400, 200, 100, 50, 32, 16, 8, 4,
572: and 2 days; second, with {\tt NBI} with a timestep of 4 days; and third,
573: with {\it Taylor 1.4} in Intel extended precision with a relative error
574: tolerance of $10^{-22}$.
575: Results for the Wisdom-Holman integrations with larger timesteps are displayed in
576: Figure \ref{fig:multiplot-big-steps}. After staring at these figures for
577: awhile, several key observations become apparent. First, when comparing
578: across the 21 sets of initial conditions, there is
579: remarkable disagreement about whether or not the outer Solar System displays exponential
580: divergence. This will be discussed further below, using more accurate integrations.
581: Looking at an individual graph, but across timesteps, we note that there are very few cases
582: (e.g. $t=150,330,420$)
583: in which there is universal agreement between all four timesteps
584: that divergence is polynomial. There are also only a few cases that
585: universally agree that the divergence is exponential.
586: In most cases, there is substantial disagreement across timesteps whether a particular
587: initial condition admits chaos. This is consistent with the observation of
588: Newman \etal (2000). However, in contrast to Newman \etal (2000), we
589: note that the ``switch'' from chaos to non-chaos is not always monotonic
590: in timestep.
591: For example, for system {\tt 000}, timesteps of 400 and 50 days admit chaos, while
592: the ``in-between'' timesteps of 200 and 100 days do not. For system {\tt 180},
593: timesteps of 400 and 200 days display non-chaos, while for 100 and 50 day timesteps,
594: chaos is apparent; this is precisely opposite to what one would expect if
595: large timesteps were injecting chaos into the system. As we shall see later
596: in the paper (Table \ref{tab:chaosCounts}), there appears also to be no
597: observable correllation between
598: the timestep and the percentage of ICs that admit chaos. Thus we hypothesize
599: that the discrepancy across timesteps is due more to perturbations in
600: the ICs and our use of only low-order symplectic correctors
601: \citep{WisdomHolmanTouma96,ChambersMercury99},
602: than it is due to unreliable integration at large timesteps. Corroborating
603: this hypothesis would require us to re-perform these experiments using
604: higher-order symplectic correctors or ``warmup'' \citep{SahaTremaine92},
605: which is a possible direction for future research.
606: \clearpage
607: \begin{figure}[p]
608: \centering
609: \includegraphics[totalheight=7.5in]{f5.eps}
610: \caption{\small Similar to previous Figure, but more accurate integrations.
611: {\tt NBI}: solid lines; Wisdom-Holman: dashed lines with $dt=$8 days (long dash),
612: 4 days (short dash);
613: Dotted lines ({\it Taylor-1.4}) are shifted
614: below others because {\it Taylor-1.4} is more accurate.
615: For any given IC, there is good
616: agreement between the shapes of the curves, indicating agreement on
617: the existence (or lack) of chaos. (Divergence saturates at $\approx$ 100 AU.)
618: }
619: \label{fig:multiplot-logy}
620: \end{figure}
621: \begin{figure}[p]
622: \centering
623: \includegraphics[totalheight=7.5in]{f6.eps}
624: \caption{Curves identical to the previous figure, but plotted on a log-log scale,
625: which depicts polynomial divergence as a straight line.}
626: \label{fig:multiplot-log}
627: \end{figure}
628: \clearpage
629: Figures \ref{fig:multiplot-logy} and \ref{fig:multiplot-log} plot the
630: divergence between sibling trajectories for all 21 ICs, as integrated
631: by more accurate integrations showing convergence: {\tt NBI},
632: the Wisdom-Holman mappings with timesteps of 8 and 4 days, and {\it
633: Taylor 1.4}. Wisdom-Holman with 2 day timesteps agreed with these curves,
634: but are omitted to reduce clutter; Wisdom-Holman with a timestep of
635: 16 days also showed good agreement in all but two cases.
636: Both figures are identical except that Figure
637: \ref{fig:multiplot-logy} uses a log-linear scale, while Figure \ref{fig:multiplot-log}
638: uses a log-log scale; different features are visible using the different scales.
639: After staring at these graphs for awhile, at least two observations present
640: themselves. First, if one looks at the system corresponding to
641: any single IC, there is usually good agreement between the
642: integrations as to the future divergence between the sibling trajectories
643: of that particular case. This demonstrates that convergence has occured
644: and makes it unlikely that the results are
645: integrator dependent. Second, looking across cases, it appears that the
646: future of the outer Solar System over the next 200 million years is quite
647: uncertain, varying from nearly integrable, to chaotic with a Lyapunov time
648: of order 10 million years or less. This is quite a startlingly diverse
649: array of possible outcomes, considering that the ICs
650: for these systems are all drawn from the same ephemeris, all less than
651: 2 years apart, and presumably differing from each other by only a few
652: parts in $10^7$. The author has, in fact, verified that several of
653: the ICs, when all integrated up to the same epoch in the
654: vicinity of 1991, agree with each other to a few parts in $10^7$.
655: \clearpage
656: \begin{figure}[hbt]
657: \centering
658: \includegraphics[scale=0.5,angle=270]{f7.eps}
659: \caption{
660: The final difference at the 200My mark between siblings,
661: in semi-major axis and eccentricity, for the 21 cases depicted in the
662: previous Figures.
663: The integrator was NBI. Eccentricity difference
664: is the Euclidean distance between siblings in the 4-dimensional space
665: consisting of the four orbital eccentricities. Semi-major axis difference
666: is the Euclidean distance between siblings in the 4-dimensional space
667: consisting of the four $\Delta a/a$ values. Some values have been moved
668: slightly for textual clarity, but in no case by more than the width or
669: height of a character.}
670: \label{fig:final-ae-diffs}
671: \end{figure}
672: \clearpage
673: Figures \ref{fig:multiplot-logy} and \ref{fig:multiplot-log} demonstrate
674: that in the chaotic cases, the siblings can have their respective planets on
675: opposite sides of the Solar System after 200My. However, Figure
676: \ref{fig:final-ae-diffs} demonstrates that changes in the orbital elements
677: are much less drastic, demonstrating that the Solar System is practically
678: stable over a 200My timescale even when it is chaotic.
679:
680:
681: \subsection{Explicitly Perturbed Initial Conditions}
682:
683: Following \citeN{MurrayHolman99}, we performed several surveys
684: in which we perturbed the semi-major axis $a_U$ of Uranus from its
685: current value, but kept all other initial conditions (ICs) constant.
686: We used all three previously mentioned integrators; Wisdom-Holman
687: with timesteps of 4 and 8 days; {\tt NBI}; and {\it Taylor-1.4}.
688: The IC was the default one from the file {\tt big.in}
689: included with {\it Mercury 6.2} \citep{ChambersMercury99}, which according
690: to the documentation is from JD2451000.5. The inner planets were
691: deleted, with their mass and momentum augmenting the Sun's as described
692: elsewhere in this paper.
693: We completed surveys in which $a_U$ was changed
694: in steps of $2\times 10^{-k}$ for $k=6,7,8$, which corresponds
695: to $\Delta a_U/a_U$ in steps of $10^{-(k+1)}$. We went 10 steps
696: in each direction for each value of $k$. For each step, we
697: generated a ``sibling'' IC by randomly perturbing the positions
698: of {\em all} planets by an amount bounded by $10^{-14}$ AU.\footnote{
699: Note that, for no good reason, this is different from the perturbations
700: used to generate siblings in the rest of the paper.}
701: We then integrated both for 200My, and plotted the distance between them
702: as a function of time. For $k=7,8$, there was no significant
703: difference between any of the integrations. That is, all siblings
704: at all steps had virtually identical divergences when changing
705: $a_U$ in 10 steps of $2\times 10^{-\{7,8\}}$ AU, for a total change of
706: $2\times 10^{-6}$. This corresponds to $\Delta a_U/a_U$ stepped by $10^{-\{8,9\}}$,
707: for a total change in $\Delta a_U/a_U$ of $10^{-7}$ in each direction.
708: However, for $k=6$, some of the steps showed chaos while others did not.
709: The change was not monotonic: over the 21 steps (10 in either direction plus
710: the ``baseline'' case), there were three ``switches'' between chaos and stability.
711: \clearpage
712: \begin{figure}[hbt]
713: \centering
714: \includegraphics[scale=1.3]{f8.eps}
715: \caption{Two systems, both using ICs within the error bounds of the best known positions
716: for the outer planets. Both have identical initial conditions except for the
717: semi-major axis $a_U$ of Uranus, which differs between the two systems by
718: $10^{-7}$ in $\Delta a_U/a_U$. One admits chaos, while the other does not.
719: The ``upper'' six curves (all starting with sibling distances near $10^{-5}$)
720: are all double-precision integrations, two using Wisdom-Holman with
721: timesteps of 4 and 8 days, and one using {\tt NBI}.
722: Of the six, the three plotted with points are the chaotic trajectory and
723: the three plotted with lines are the non-chaotic trajectory.
724: The ``lower'' two curves (starting with sibling distances near $10^{-8}$)
725: are integrated in extended precision with {\it Taylor 1.4}.
726: The chaotic one fits an exponential curve with a Lyapunov time of about 12 million
727: years, while the non-chaotic one has the two trajectories separating approximately
728: as $t^{1.5}$.}
729: \label{fig:surprise-all}
730: \end{figure}
731: \clearpage
732: Figure \ref{fig:surprise-all} plots two of these 21 systems.
733: The value of $\Delta a_U/a_U$ differs between the two systems by one part in $10^7$.
734: One of the systems appears chaotic, and the other does not,
735: over a 200My timespan. The non-chaotic one has a semi-major
736: axis of $a_U + 2\times 10^{-6}$, while the chaotic one has semi-major
737: axis $a_U+4\times 10^{-6}$. All other ICs in the two
738: systems are identical. To ensure that the result is not integrator
739: dependent, we have repeated the integrations with the Wisdom-Holman
740: mapping included in {\it Mercury 6.2} with timesteps of 8 and 4 days;
741: and with {\tt NBI} with a timestep of 4 days. As can be seen, all the
742: integrations agree quite well with one another. Note that the
743: {\it Taylor 1.4} integrations provide 3 extra digits of precision,
744: and so the curves for {\it Taylor 1.4} are displaced about 3 orders
745: of magnitude below the curves computed in double precision. Otherwise
746: the shapes of the curves are virtually identical. We note that the
747: chaotic one has a Lyapunov time of about 12 million years, while the
748: regular one has the sibling trajectories separating from each other
749: polynomially in time as $t^{1.5}$.
750:
751:
752: \subsection{Accurate integrations over the age of the Solar System}
753:
754: The author has reported related results for integrations lasting
755: $10^9$ years \citep{HayesNaturePhysics07}. The essential conclusion
756: is the same, in that even after $10^9$ years, there remain some
757: ICs (about 10\%) that show no evidence of chaos, although {\em some}
758: of the ICs appearing as regular over 200My develop exponential
759: divergence later.
760:
761: Figure \ref{fig:5Gy} displays the sibling divergence over $5\times 10^9$
762: years of the ``canonical'' IC used by DE405 (JED 2440400.5, June 28,
763: 1969). As we can see, this IC shows little evidence of chaos for about
764: the first 1.5Gy, and then develops slow exponential divergence with a
765: Lyapunov time between about 200 and 400 million years. The individual
766: planets each show similarly-shaped divergence curves (not shown), with the magnitude
767: of divergence increasing with orbital radius. After 5 Gy, the uncertainty
768: in Jupiter's position for this IC is less than 1 AU, while the uncertainty
769: in Neptune's position is about 9 AU. Thus, there is a non-negligible
770: chance that, if the Solar System lies close enough to the ``canonical''
771: IC of DE405, that we can know within about 10--15 degrees where each outer
772: planet will be in its orbit when the Sun ends its main-sequence lifetime
773: and becomes a red giant.
774: Note that the levelling-off that starts at about the 4 Gy mark is {\em
775: not} saturation (which occurs closer to 100 AU separation, while the separation
776: at 5 Gy here is less than 10);
777: the outer Solar System instead seems to be entering again into
778: a period of polynomial (non-chaotic) divergence.
779:
780: \clearpage
781: \begin{figure}[hbt]
782: \centering
783: \includegraphics[scale=1,angle=0]{f9.eps}
784: \caption{Sibling divergence over 5 Gy of the canonical IC of DE405, integrated using Taylor-1.4.}
785: \label{fig:5Gy}
786: \end{figure}
787:
788: \subsection{Percentage of initial conditions displaying chaos}
789:
790: Observing the distance between sibling trajectories in Figures \ref{fig:multiplot-logy} and
791: \ref{fig:multiplot-log} at the 200My mark, we can reasonably simplify the distinction between
792: chaotic and regular trajectories. By choosing a cutoff distance $C$
793: and restricting our view to the various double-precision integrations,
794: we can claim that siblings differing by less than $C$ after 200My
795: are regular (in the sense of having no observable Lyapunov exponent),
796: while those differing by more than $C$ are chaotic with a measurable
797: positive Lyapunov exponent. Table \ref{tab:chaosCounts} lists the number
798: of systems that are chaotic by the above definition, as a function of
799: timestep and cutoff. In addition to the 21-sample group of
800: ICs displayed in Figures \ref{fig:multiplot-logy} and \ref{fig:multiplot-log},
801: we also drew a second sample group of 10 samples, spaced at 10-year intervals
802: from 1900 to 1990.
803: Since the {\tt testeph.f} program included in DE405
804: provides only 7 digits of precision (corresponding to the accuracy
805: to which the positions are known), taking ICs from DE405 at different times
806: effectively takes ICs from different exact orbits, differing from
807: each other by as much as one part in $10^7$.
808: (As noted in a previous footnote, this method of sampling may not ideally represent
809: an unbiased sample from the observational error ball;
810: instead, it is an unbiased sample from the set of
811: 7-digit-rounded initial conditions drawn from DE405.)
812: We make several observations.
813: First, the fraction of sampled systems that are chaotic by this
814: simple definition is roughly about ($70\pm 10)\%$, and is relatively
815: independent of both the timestep and the two sample groups (30-day \vs 10-year samples),
816: although it of course increases as we decrease the cutoff.
817: Recall that for a given initial condition, different stepsizes can give
818: different results, so that {\em which} systems are chaotic
819: changes as the timestep changes. However, here we measure only the
820: {\em number} of chaotic systems as a function of timestep.
821: If chaos
822: were being ``injected'' into the system by the integrator, we would
823: expect that the number of systems displaying chaos should increase
824: with increasing timestep. However, this is not observed. A
825: possible interpretation of this is that
826: even a 400 day timestep reliably determines whether {\em some} system is chaotic,
827: but the system is slightly different for each timestep due to the
828: IC perturbation introduced by the symplectic
829: integration \citep{SahaTremaine92,WisdomHolmanTouma96}.
830: Verifying this would require us to implement ``warmup'' \citep{SahaTremaine92} or
831: higher-order symplectic
832: correctors than are included with {\it Mercury 6.2} \citep{ChambersMercury99},
833: which we have not done. However, the fact that the fraction
834: of systems displaying chaos is independent of timestep argues against the
835: ``chaos is injected by the integrator'' hypothesis, at least for the timesteps
836: used in this paper.
837:
838: \clearpage
839: \begin{table}
840: \begin{tabular}{c|cccc|cccc}
841: {} & {} & {C=1} & {} & {} & {} &{C=0.1} & {} & {} \\
842:
843: dt & /10 & /21 & total & \% & /10 & /21 & total & \% \\
844: \hline
845: 400 & 7 & 14 & 21 & 67.7 & 9 & 16 & 25 & 80.6 \\
846: 200 & 7 & 11 & 18 & 58.1 & 8 & 15 & 23 & 74.2 \\
847: 100 & 6 & 11 & 17 & 54.8 & 8 & 13 & 21 & 67.7 \\
848: 050 & 6 & 14 & 20 & 64.5 & 8 & 15 & 23 & 74.2 \\
849: 032 & 7 & 10 & 17 & 54.8 & 9 & 15 & 24 & 77.4 \\
850: 016 & 7 & 14 & 21 & 67.7 & 8 & 16 & 24 & 77.4 \\
851: 008 & 8 & 13 & 21 & 67.7 & 8 & 14 & 22 & 71.0 \\
852: 004 & 8 & 13 & 21 & 67.7 & 8 & 18 & 26 & 83.9 \\
853: 002 & 8 & 13 & 21 & 67.7 & 8 & 15 & 23 & 74.2 \\
854: \end{tabular}
855: \caption{The percentage of initial conditions drawn from DE405 that
856: admit chaos. Column labels:
857: $C$ is the cutoff described in the text;
858: $dt$ is the timestep; /10 means
859: ``out of the 10 initial conditions drawn at 10-year intervals from 1900
860: to 1990''; /21 means ``out of the 21 initial conditions drawn at 30-day
861: intervals starting at 1990''; {\it total} is the sum of the previous
862: two columns;
863: \% is the percentage of systems out of the 31 that display chaos according
864: to the cutoff.
865: }
866: \label{tab:chaosCounts}
867: \end{table}
868: \clearpage
869: \subsection{Chaos in the inner solar system is robust}
870:
871: \citeN{VaradiRunnegarGhil03} performed
872: a 207My integration of the entire Solar System, including some
873: non-Newtonian effects and a highly-tuned
874: approximation to the effects of the Moon, and placed a lower bound of
875: 30My on the Lyapunov time of the outer Solar System.
876: However, they still saw chaos in the inner solar system.
877: To test the robustness of chaos in the inner solar system, we performed
878: several integrations of 8 planets (Mercury through Neptune) using
879: the Wisdom-Holman mapping with timesteps of 8, 4, 2, 1, and 0.5 days.
880: We treated the Earth-Moon system as a single body. To ensure that
881: chaos in the outer solar system did not ``infect'' the inner solar system,
882: we used only those DE405 ICs from Figures \ref{fig:multiplot-logy} and
883: \ref{fig:multiplot-log} for which the outer solar system was
884: regular over 200My. We then
885: integrated the system until chaos appeared. In all
886: cases, even though the outer solar system was regular, the inner
887: solar system displayed chaos over a short timescale such that information
888: about the inner planetary positions was lost within about 20--50My.
889: Thus, unlike the outer Solar System, we observe that chaos in the
890: inner solar system is robust.
891:
892: \section{Discussion}
893:
894: We conclude that perturbations in the initial conditions (ICs) of the
895: outer planets as small as one part in $10^7$ can change the behaviour
896: from regular to chaotic, and back, when measured over a timespan
897: of 200 million years. We believe this is the first demonstration
898: of the ``switch'' from chaos to near-integrability with such a
899: small perturbation of the ICs. Since our knowledge of the orbital
900: positions of the outer planets is comparable to one part in $10^7$,
901: it follows that, even if our simplistic physical model accounting only
902: for Newtonian gravity were the correct model, it would be impossible at
903: present to determine the Lyapunov time of the system of Jovian planets.
904: Furthermore, it implies that an IC with 7 digits of precision (which
905: is all an ephemeris can justifiably provide) can randomly lie on a
906: chaotic or non-chaotic trajectory. Since our results converge in the
907: limit of small timestep for the Wisdom-Holman mapping, and the converged
908: results also agree with two very different high-accuracy integrations,
909: and finally since the high-accuracy integrations in turn agree very well
910: with quadruple-precision integrations, we believe that the results in
911: this paper are substantially free of significant numerical artifacts.
912:
913: \citeN{Guzzo05} corroborates the existence of a large web of 3-body
914: resonances in the outer Solar System, and finds that their placement is
915: consistent with Murray and Holman's (1999) theory. Guzzo used his own
916: fourth-order symplectic integrator \citep{Guzzo01}, and performed what
917: appear to be reasonable convergence tests to verify the robustness of
918: his main results. Thus, Murray and Holman's theory appears to explain
919: the existence and placement of 3-planet resonances. Furthermore, chaotic
920: regions in Hamiltonian systems are usually densely packed with both chaotic
921: and regular orbits \citep{LichtenbergLieberman92}.
922: This paper corroborates the observation of densely
923: packed regular and chaotic orbits, at a scale previously unexplored for
924: the system of Jovian planets.
925:
926: As discussed in the text relating to Figure \ref{fig:surprise-all}, we
927: performed surveys across the semi-major axis $a_U$ of Uranus in steps of
928: $2\times 10^{-k}$ AU for $k=6,7,8$. We found that, around the current
929: best-estimate value of $a_U$, perturbations smaller than $2\times 10^{-6}$
930: AU had no effect on the existence of chaos. However, this does not imply
931: that perturbations this small cannot have an effect; it simply means
932: that the ``border'' between chaos and regularity is not within $2\times
933: 10^{-6}$ AU of the current best estimate of $a_U$. However, it is clear
934: that the border between chaos and regularity {\em is} between the two
935: systems depicted in Figure \ref{fig:surprise-all}, which differ from each
936: other in $a_U$ by $2\times 10^{-6}$ AU. Although a survey across $a_U$
937: for values between those two systems may not be very relevant from a
938: physical standpoint, it might be very interesting from a dynamical systems
939: and chaos perspective to probe the structure of the border between chaos
940: and regularity. In particular, it may be interesting to see if the
941: border itself has some sort of fractal structure \citep{MandelbrotBook82}.
942: Such chaotic structure has already been observed in the circular
943: restricted three-body problem \citep{Murison89}.
944:
945: Newman \etal (2000) gave a compelling demonstration
946: that the Wisdom + Holman symplectic mapping with too-large a timestep
947: could introduce chaos into a near-integrable system
948: by first showing that they could reproduce the chaos with a 400-day timestep,
949: and then showing that the integration converged to being regular with
950: a timestep of about 50 days or less. Our results appear to hint
951: that an even smaller timestep may be required: our curves depicting
952: divergence-of-nearby-orbits did not fully converge until the timestep was
953: 8 days or less. On the other hand, the symplectic integrators produce
954: solutions that effectively integrate a system with slightly perturbed ICs,
955: although these perturbations decrease with decreasing timestep
956: \citep{SahaTremaine92,WisdomHolmanTouma96}.
957: Our Figure \ref{fig:multiplot-big-steps} demonstrates
958: that the behaviour does not always ``switch'' monotonically in timestep
959: from chaotic to non-chaotic as observed by Newman \etal
960: (2000); Table \ref{tab:chaosCounts} also hints that the
961: ``amount'' of chaos does not appear to increase with increasing timestep.
962: An alternate interpretation is that even a 400-day timestep accurately
963: integrates {\em an} orbit, but not {\it the} orbit that we chose.
964: In particular, it may accurately integrate an orbit whose IC is perturbed
965: slightly from the one we chose,
966: and an appropriate correction may allow us to recover the correct orbit using
967: an integration with much larger timestep \citep{SahaTremaine92,WisdomHolmanTouma96}.
968: However, the symplectic correctors in Mercury 6.2 are clearly not good
969: enough to perform this recovery at a 400-day timstep; better correctors
970: \citep{Wisdom06Correctors} or warmup \citep{SahaTremaine92} may be able to achieve this.
971:
972: The convergence of our results at timesteps of 8 days or less, as well as the
973: agreement with two different non-symplectic integrators (including the
974: one used by Newman \etal), indicate that the IC perturbations described
975: by \citeN{SahaTremaine92} and \citeN{WisdomHolmanTouma96} are negligible in our smaller timestep cases.
976: Thus, using ``warmup'' \citep{SahaTremaine92} or higher-order symplectic
977: correctors will not substantially alter
978: our conclusions, although they might allow the same conclusions to be drawn
979: using larger timesteps.
980:
981: With the exception of the results plotted in Figure \ref{fig:5Gy},
982: all of our simulations had a duration of 200 million years (My). All of
983: them display an initial period of polynomial divergence, before the
984: appearance of exponential divergence (if any). However, the duration
985: of initial polynomial divergence differs greatly across systems,
986: and has been observed by others \citep{LecarFranklinHolmanMurray01} to
987: last significantly longer that 200My; \citeN{GrazierNewmanHymanSharp05}
988: observed it to last the entire duration of their 800My simulation. It
989: would be interesting to create a table like Table \ref{tab:chaosCounts},
990: but including a ``simulation duration'' dimension as well. Certainly the
991: evidence hints that more systems make the ``switch'' from polynomial
992: to exponential divergence as the duration of the simulation increases.
993:
994: %Considering that approximately 70\% of our ICs chosen
995: %from DE405 admit chaos on a 200My timescale, it is disturbing that
996: %\citeN{GrazierNewmanKaulaHyman99} and \citeN{NewmanEtAlDDA2000} found
997: %non-chaos in all 16 sets of ICs chosen randomly
998: %from DE245. We have utilized their very own {\tt NBI} integrator
999: %with their recommended timestep in this study, so integrator issues are
1000: %unlikely to be the problem. The likelihood that either of our estimates
1001: %is accurate, while the other is by unlucky chance, is extremely small.
1002: %Although several possibilities for the wide discrepancy exist, a third-party
1003: %experiment may be in order to shed light on the discrepancy.
1004:
1005: Our physical model is very simplistic, accounting only for Newtonian
1006: gravity between the Sun and Jovian planets. Although we ignore many physical
1007: effects which are known to effect the detailed motion of the planets
1008: \citep{Laskar99,VaradiRunnegarGhil03}, it is unclear if such effects would
1009: substantially alter the chaotic nature of solutions. We at first believed
1010: that the largest such effect ignored was solar mass loss. Our first simulations
1011: did not account for solar mass loss, which amounts to about one part in $10^7$
1012: per million years \citep{Laskar99,Noerdlinger05}. Since we find that perturbations
1013: in position of that order can shift the system in-and-out of chaos, a naive
1014: analysis might lead one to suspect that solar mass loss might shift the
1015: planetary orbits in-and-out of resonance on a timesclase that is fast
1016: compared to the Lyapunov time, thus smoothing out the sibling divergence.
1017: We thus modified our model to include solar mass loss, but surprisingly it
1018: made absolutely no observable difference to any of the figures presented
1019: in this paper. To ensure that we did not make an error, we simulated systems
1020: with ever increasing mass loss until the Sun was losing 10\%
1021: of its mass per 100My. We noted that the planetary orbital semi-major axes
1022: expanded significantly, as would be expected, but that the sibling divergences
1023: did not change until mass loss was at a rate of about 1\% per 100My (1000 times
1024: greater than in reality). Thus, we conclude that solar mass loss also makes no
1025: difference to our results.
1026:
1027: %In any case, as noted by \citeN{Laskar99}, \citeN{VaradiRunnegarGhil03},
1028: %and others,
1029: %a model using only Newtonian gravity ignores many relevant effects
1030: %whose magnitude over a few Lyapunov times are comparable to the
1031: %current error in the positions of the outer planets.
1032: %This, combined with the large array of possible
1033: %outcomes depicted in Figures \ref{fig:multiplot-logy} and \ref{fig:multiplot-log},
1034: %may imply not only that it is not possible {\em at present}
1035: %to conclude whether or not the outer Solar System is chaotic, but
1036: %also that it is not possible {\em at all} to determine if the ``real''
1037: %system of Jovian planets admits chaos.
1038:
1039: %Finally, we recall again that a perturbation in the ICs
1040: %of just one part in $10^7$ can move a solution from a regular region into
1041: %a chaotic one. Thus, in order to clearly observe that a particular
1042: %IC is regular, one requires that the {\em global} error
1043: %in the solution is less than one part in $10^7$ over the {\em entire}
1044: %integration, lest the accumulation of numerical errors may cause the
1045: %numerical solution to drift between regular and chaotic regions,
1046: %which could spuriously add or delete chaos.
1047: %As Figure \ref{fig:TaylorBrouwer} shows, even {\it Taylor 1.4} using
1048: %Intel extended precision and a local tolerance of $10^{-22}$ does {\em not}
1049: %have global errors less than $10^{-7}$ AU after 200My. Since the integrator
1050: %is roundoff dominated at that tolerance, it is clear that no
1051: %integration using only double-precision can satisfy this property, either.
1052: %Thus even an integration that starts on a regular orbit can drift to
1053: %a chaotic one, and vice versa, due solely to roundoff error.
1054: %This seems at first to imply that even more accurate integrations are in order.
1055: %However, the modelling errors mentioned above have a comparable effect
1056: %on the evolution of the system.
1057: %Furthermore, unlike solar mass loss,
1058: %which is quite predictable, many modelling uncertainties mentioned
1059: %by \citeN{Laskar99} and others are inherently unpredictable.
1060: %Thus, even-higher-precision long-term integrations of a Newtonian-only model
1061: %would seem to be of dubious merit.
1062: %{\bf No, no, no. The convergence of our results clearly demonstrate
1063: %that even-more accurate results are NOT necessary; I think what's
1064: %happening is that the positional error along the orbit is bigger
1065: %than $10^{-7}$, but we must have the orbital elements correct to
1066: %higher precision than one part in $10^{-7}$, else we wouldn't see
1067: %convergence at all.}
1068:
1069: \section*{Acknowledgements}
1070: The author thanks Scott Tremaine, Norm Murray, Matt Holman, Philip Sharp,
1071: and Bill Newman for helpful discussions and comments on the manuscript;
1072: Norm Murray and Philip Sharp for sending me (and explaining) their ICs;
1073: and Ferenc Varadi for giving me the source code to his most recent version
1074: of {\tt NBI}.
1075:
1076: \bibliography{ms}
1077:
1078: %\end{spacing}
1079: \end{document}
1080: ------------------
1081:
1082: LONG INTRODUCTION
1083:
1084: Chaos, Stability, Resonances
1085:
1086: The stability of the Solar System has been a question of practical and
1087: theoretical interest since the invention of the theory of universal
1088: gravitation. Newton himself doubted the long-term stability of the
1089: Solar System, declaring of the apparent stable motion of the planets,
1090: ``{\it it is not to be conceived that mere mechanical causes could
1091: give birth to so many regular motions}'' \citep{NewtonPrincipia}.
1092: Laplace \citep{Laplace1798} created perturbation theory specifically in an
1093: attempt to answer the question, and proved analytically that, to first
1094: order, the Solar System is stable. However, Poincare proved that the
1095: general $n$-body problem is not analytically solvable, squelching hopes
1096: of unequivically proving the Solar System's stability. Further progress
1097: was severely hampered until the advent of numerical computing. In the
1098: late 1980s Laskar used computers to generate higher-order perturbation
1099: methods which demonstrated the existence of chaos in the Solar System.
1100:
1101: Since that time the practical question of the stability of the Solar
1102: System has been essetianly established using numerical methods: barring
1103: unforeseen intruders, the orbits of the planets will not substantially
1104: change over the next several billion years (Laskar, Wisdom+Holman,
1105: Pluto, etc.) However, in a more formal sense the Solar System may still
1106: be {\it chaotic}: although the orbits are relatively stable, the positions
1107: of the planets in the orbits, and minor fluctuations in the orientation
1108: of the orbits, cause the positions of the planets to have ``sensitive
1109: dependence on ICs'': two solutions with nearby initial
1110: conditions can diverge exponentially from each other with the passage of
1111: time, so that we cannot predict the precise positions of the planets,
1112: and in some cases their axis tilts and orientations, for more than a
1113: few tens of millions of years. This has the consequence, for example,
1114: of severely hampering our attempts at climate prediction over geological
1115: timescales \citep{Varadi}.
1116:
1117: In a chaotic system, nearby trajectories diverge from each other
1118: exponentially with time. If they are distance $d(0)$ apart at time $t=0$,
1119: then they tend to be about $d(t) = d(0)*e^{\lambda t}$ apart after time
1120: $t$, where $\lambda$ is called the {\it Lyapunov Exponent}. The
1121: {\it Lyapunov time} is $1/\lambda$, and measures the typical time it
1122: takes for nearby trajectories to diverge by a factor of $e$.
1123: If the distance between nearby trajectories increases
1124: only polynomially in time, then the Lyapunov exponent is formally
1125: zero, the Lyapunov time is formally infinite, and the system is termed {\it regular}.
1126: Strictly speakng, KAM theory tells us that every nonlinear dynamical
1127: system has a positive Lyapunov exponent. However,
1128: for some systems, the Lyapunov time can be a function of ICs.
1129: The Outer Solar System is such a system, as we shall see below.
1130: Furthermore, for some ICs, the
1131: Lyapunov time is so large that the system can appear to be regular
1132: over finite periods. For our purposes, we shall refer to such systems
1133: as ``regular'' although formally this is an abuse of the term.
1134: To make matters worse, some solutions can appear to be regular at first,
1135: and then suddenly develop exponential divergence.
1136:
1137: Sussman and Wisdom (1988) were the first to demonstrate a positive
1138: Lyapunov exponent in the Solar System, demonstrating that Pluto had
1139: a Lyapunov time of about 20 million years.
1140: Laskar \citep{Laskar89} first measured the Lyapunov time of the
1141: entire nine-planet Solar System to be about 5 million years,
1142: and found that the majority of the divergence was due to the inner planets.
1143: Laskar noted
1144: that the system of Jovian planets, consisting of the planets Jupiter, Saturn,
1145: Uranus and Neptune, appeared to be much more regular.
1146: The Lyapunov time of the isolated Jovian planets (ie., in a system
1147: where the inner planets have been carefully deleted) has been measured
1148: repeatedly to be positive, but with Lyapunov times ranging from 7
1149: million to 19 million years (Sussman + Wisdom 1992, etc).
1150: Extensive discussion can be found in the reviews by Lissauer (1999)
1151: and Lecar et. al (2001).
1152: The cause of chaos in the inner Solar System appears to be due to
1153: 2-body resonances (See Lecar et al and references therein). However,
1154: this explanation fails for the Jovian planets because there does not
1155: appear to be any 2-body resonances between any of the Jovian planets.
1156: Although there are several 2-body near-resonances, none of them appear
1157: adequate to explain the chaos seen in the isolated outer Solar System.
1158:
1159: To explain chaos in the outer Solar system, Murray + Holman (1999)
1160: proposed a theory based upon 3-body resonances. Guzzo (2001) went
1161: on to numerically detect the web of 3-body resonces in the Outer
1162: Solar System.
1163:
1164: Murray + Holman note that perturbs as small as 3e-6 are enough to
1165: destroy the resonces, and that error in ICs and their theory ensures
1166: the outer SS is almost certainly chaotic. In this paper we demonstrate
1167: that perturbs as small as 1e-7, which is within the observational errors
1168: of the positions of the planets, also can perturb us between chaos and
1169: regular motions.
1170:
1171:
1172: Integration Accuracy
1173:
1174: There is still some question as to whether the Outer Solar System is
1175: actually chaotic or not. In the course of studying the stability
1176: of planetesimal orbits in the Solar System, Grazier et. al (1999a)
1177: performed a suite of integrations providing compelling evidence of the
1178: absense of chaos in the Jovian planets. Their tests were performed
1179: using a highly accurate modified Cowell-Stormer integrator which was designed
1180: to provide a relative error which is effectively {\it exact} in double
1181: precision. More precisely, if the machine precision is ${\eps}$
1182: (typically $2^{-52}\approx 2\times 10^{-16}$ in IEEE 754 double
1183: precision), and the solution vector for the system is $\V y$, then
1184: when the stepsize is less than $1/1000$ of the orbital period and the
1185: orbital eccentricity is less than 0.5, then the local error $\delta$
1186: introduced at each step of the integration satisfies $\|\delta\|/\|\V y\|
1187: \le \eps$, where $\|\cdot\|$ is the max norm.\footnote{Note that this
1188: implies that components of $\V y$ which are not the maximim component
1189: will individually have relative errors larger than the machine precision.
1190: It is difficult to imagine how one could do better than this without
1191: explicitly extending the precision, however.} It is plausible to call
1192: such an integration ``exact to double precision'', and that is the term
1193: used by Grazier et al. Note that such an integrator is symplectic
1194: by default since, providing the exact solution, it obviously preserves
1195: symplectic structure to the extent possible in floating point. Using
1196: this integrator, they drew 16 ICs at random from
1197: the JPL ephemeris DE245 \citep{DE245} and performed integrations of
1198: various length up to $800\times 10^6$ years. In all 16 cases, the
1199: orbits appeared to be regular. Furthermore evidence of non-chaos
1200: was provided by Newman et al (American Astronomical Society, DDA Meeting
1201: \#31, \#04.01; Bulletin of the American Astronomical Society, Vol. 32,
1202: p.859). Using the well known Wisdom-Holman \citep{WisdomHolman} symplectic
1203: mapping with the then-common timestep of 400 days (1/12 of Jupiter's
1204: orbital period), they reproduced previous results demonstrating a
1205: Lyapunov time of about 10 million years. However, when the timestep
1206: was decreased to 200, then 100, then 50 days, the chaos disappeared
1207: and the orbits became regular. They concluded that a 50-day timestep
1208: was sufficient for a $100\times 10^6$ year integration.
1209:
1210: It is widely believed that a symplectic integrator exactly solves a nearby
1211: problem. However this is not the case. Formally, given a Hamiltonian
1212: $H$ symplectically integrated by algorithm $\V S$ with a timestep of $h$,
1213: then there exists a Hamiltonian $H'$, $|H'-H|\le c_1 e^{-c_2h}$, such that
1214: $\V S$ produces a solution which remains exponentially close in $h$ to
1215: an exact solution of $H'$ for a time which is exponentially long in $h$.
1216: \citep{BenettinGiorgilli94,Reich99}. Note that this means that if one
1217: wants to remain a fixed distance from $H'$, then one must decrease the
1218: timestep whenver the desired length of integration increases. Thus the
1219: expense of a {\em reliable} symplectic integration increases faster than
1220: linearly with the length of the integration; one cannot simply find a
1221: reliable timestep for a 100 million year integration and then expect
1222: the same timestep to be reliable for a 200 million year integration.
1223: Although there is no {\it a priori} method of knowing the constants $c_1,
1224: c_2$, and thus there is no way to {\it a priori} choose a timestep for
1225: a given integration, a reliable algorithm is to re-perform a given
1226: integration with a smaller-and-smaller timestep until the results
1227: converge.
1228:
1229: Computing the Lyapunov exponent using numerical integrations is not trivial.
1230: One fact that throws a wrench into the works is that inaccurate
1231: integrations can inject chaos into a system whose true behaviour
1232: is regular, and vice versa.
1233:
1234: In this paper, we aim to establsh why some researchers
1235: see chaos in the system of Jovian planets, while others do not.
1236:
1237: