BASIC IDEAS BEHIND PRECOVERIES
'ephem0.cpp' now has a 'find_precovery_plates' function. As of this
writing (mid-August 2017), it is able to integrate an orbit backward in
time and tell you on which of about three million CSS images that object
might be found. Extending this to handle other multi-million-image lists
would be trivial, and I've hopes of doing so if I can persuade (for
example) PanSTARRS, WISE, Spacewatch, etc. to cough up their lists
of "what images were taken in what parts of the sky at what times."
There is a fast way of doing this, and a simpler but slower way of
doing this. We expect this to be a common operation, so fast would be
good. (It takes a few seconds, usually, to search all three million
fields on my somewhat elderly machine.)
The simpler, slower method is to say :
For each of the three million fields :
Numerically integrate our object of interest to the time that image
Is it in the field of view? (If so, add it to the output.)
The problem with this is that computing three million positions to
full, numerically-integrated precision takes a long time. I actually
did this, but only for testing purposes.
The faster, but more complicated, method computes a somewhat
approximate position for each of those three million times. Almost
always, we can then see immediately that there's no hope of finding
the object on that particular field (the sky is big and each field is
tiny). If it _does_ look possible, we do the numerical integration
to that instant and verify that the object was on that field.
The resulting algorithm looks like this :
(1) Sort the fields in order of the time midpoint of the exposure.
(2) Starting with the first field, compute geocentric positions of our
object of interest for a time "before" and "after" that of the
exposure. Make those times close enough that motion between them
is passably close to linear. (I set the time difference to depend
on the distance to the object: if it's very close, you need a small
(3) We can now compute approximate positions for our object of interest
between those start/end times, using linear interpolation. We
allow a wide margin of about .1 degree for the fact that the motion
does have some real curvature, plus enough extra margin to allow
for the maximum possible topocentric parallax over that interval.
The extra margin is (radius_earth / distance_to_object), in radians.
(4) If, after allowing for that extra margin, we still think the object
might have been on that plate, do the full "slower" method above:
find out _exactly_ where the object would have been at that time.
This time, we do a topocentric calculation that includes light-time
lag... no approximations this time.
(5) If this precise location would have been on that image, add that
field to the output.
(6) Move on to each field in turn, running it through (2) to (5).
The reason this is fast is that much of the time, we're doing a quick
linear interpolation and finding that it's nowhere near the field in
question. (That is to say, most images will be quite far away from
whatever object we're actually interested in.) We'll get to step (3)
and be done. Only in a percent or two of cases will we have to do that
full positional computation (step (4)) and maybe find out that yes,
we really did image the object (step (5)).
(1) is not theoretically necessary. But it saves computing time if
you integrate backward through time once, rather than hopping about at
random through the decades of CSS observations. It also makes it very
likely that the field you're looking at will be "close" in time to the
preceding one, meaning you can recycle the "before" and "after"
computations from the preceding field. (The sorted list of fields
is precomputed. It takes a minute or so to do it... not something
you'd want to have to wait for with each precovery search.)
The folks at CSS thought they might want to do certain types of
filtering. Currently implemented are :
(1) Ignore fields where the computed magnitude was less than
some specified limit (no point trying to recover the object from a
field if it would have been at mag 32 at the time).
(2) Ignore fields where we've already measured the astrometry.
(3) Show _only_ fields where we've measured the astrometry.
(4) Show only fields within a given date range, _or_ only those
_outside_ that date range.
The above works nicely if you have a single orbit with low
uncertainties. But suppose the orbit isn't well-determined for the
Fortunately, Find_Orb can (usually) determine ephemeris uncertainties.
Those uncertainties are (usually) elongated ellipses, to the point
where we can say that the object is on a particular line, spread
out in a Gaussian distribution along that "line of variation" with
a particular uncertainty.
So the above scheme is modified slightly: Find_Orb integrates
the nominal orbit for the object, _plus_ a one-sigma variant object.
As we go backward in time, we can compute a probability that the
object would appear on a particular image. We may have cases
where the nominal position isn't on the image, but can say that
if the object is (say) 0.2 to 1.3 sigmas off the nominal position,
it would be on the image. That would give us a 32.4% chance that
the object would be in the image. Given that, we might (or might
not) try to find it.
The problem becomes one of "does the uncertainty region intersect
the image rectangle" instead of "does the nominal position land in the
image rectangle." We can provide a little more information as well:
"the object will probably be found here on your image" (or, in
the above case, "is probably a little off the image on this side")
"and is probably within thus-and-such distance, at thus-and-such
position angle." As described above, we can also give a probability
that it's on the image at all.
This does slow things down a bit, because we're now integrating
two orbits instead of one. Fortunately, much of the work (computing
perturber positions, for example) only need be done once, so it
doesn't actually take twice as long as the single-orbit case. And
in any case, we really need to know if we're apt to find an object
or if we'd just be wasting time looking for it. So the extra time
Also, the above description of the uncertainty area being nearly
linear is really only true once the orbit can be fitted with least
squares. At the early stages, when the observed arc is small,
almost _any_ orbit linking the first observation to the last observation
will have reasonable residuals and has to be considered. (Within
some limits.) So for _really_ short, just-posted-on-NEOCP orbits,
this scheme doesn't work. I've ideas for handling that situation
as well, but thus far, I don't have actual working code.
PLANNED IMPROVEMENTS :
Finding out which of several million tracklets (ITF or short-arc
astrometry) would fit your observations is only slightly different
from telling you which images to look at. It has the advantage of
being somewhat automatable, and it doesn't require access to images.
Ideally, Find_Orb would say : "Your object _might_ be linked to
objects A, B, and/or C in the ITF." It would then try computing
an orbit with A's observations added in, then with B's, and then
with C's, to see if the residuals looked good.
Also, note that if you're a survey, for "ITF", you might say
"our file of observations that weren't quite good enough to submit
to MPC." You might, for example, require that an object be
detected on three out of four images before it could be submitted
to MPC. But you might also keep a file of two-out-of-four detections,
and only use them when you could confirm them by linking them to