How the movies were made.
Carles Simo implemented the following
three main ideas, or methods, to find the
orbits depicted in these movies: (1) gradient descent, (2)
shooting (Poincare sections), and (3) the homotopy
(orbit following) method.
1. Gradient descent.
The gradient of the action functional
yields a flow in path space. This
flow is discretized, for example
using Fourier modes, to yield a
finite-dimensional optimization problem.
The descent method works well to get
the eight and Gerver's supereight. On
the more complicated choreographies it converges
too slowly to give good initial conditions to
check with an ODE solver. See step 2.
Slow convergence is in part because of
near collisions and high instability.
2. Shooting.
After time $T/N$ of a
period $T$ choreography, the $k$th mass is where the
$k-1$st mass started. In accordance with this, set up
a time $T/N$ slice (Poincare section) to
an initial guess orbit obtained via descent,
thus obtaining an $i \mapsto i + 1$ type
Poincare section. Look for fixed points to this map.
(We are ``shooting'' from the $t = 0$ to the $t = T/N$
slice.)
The shooting just described often fails,
due to dynamic instability. Multiple
shooting, where we further divide
$T/N$ into subintervals may help.
(3). (1) and (2) together
still fail to find many of the orbits depicted. However
if the Newtonian $1/r$ potential is
replaced by $1/r^2$ then gradient descent works
well enough to give initial conditions good
enough for step 2.
Better convergence occurs solutions cannot tend
to collision: any curve with collision has infinite
action for a $1/r^a$ potential, $a \ge 2$. Consider a family of
problems indexed by $a$ where the potential is $1/r^a$.
Implement descent for $a =2$. Use shooting to refine
the result.
Follow the orbit via implementing an implicit
function theorem as $a$ is varied, trying to bring
$a$ down from $2$ to $1$. This last step (3) is called the
homotopy method by some. It allows us to follow
periodic orbits for systems depending continuously
on some real parameter.
The basic set up for choreographies
and for (1) is described in
our paper (referenced above, w/ Gerver et al).
Which choreography classes are realized?
We have described the basic ideas Simo implements to
find the orbits. For any fixed $N$ there
are infinitely many topologically distinct choreography classes.
All are realized by solutions when the potential is
$1/r^2$. Probably some of these, maybe most,
disappear upon bringing $a$ down to the
the Newtonian case $a =1$. The subject is in
a state of flux (as of May 7, 2001). Simo has recently
found on the order of a million new 3-body choreographies
the methods outlined above, combined with other
ideas. Probably there are infinite families
of choreographies for all $N$.
Nothing is proved for the Newtonian case beyond
the existence fof the eight.