[ddahlen]

It's a bit of both, observing has uncertainty in a lot of places, if you are on the ground you get atmospheric effects, imprecision of timing, imprecision of optics, etc etc. You are also observing an object where you dont know how far away it is. That distance has to be solved by basically doing a sort of triangulation, which requires either the observer or the object to move enough. So if you observe over a short time (hours for example), you can see it is moving, but it is hard to tell distance.

Once you have an estimated orbit, if it has any interactions with planets (IE: flyby of Earth), small differences in positions during the close encounter make LARGE differences decades later. Add to this the effects of photons from the sun pushing on the smaller asteroids or dust, or out-gassing /dust from comets cause these objects to slightly drift from just the basic gravitational forces. Generally inner solar system asteroids (inside mars) are very chaotic over hundreds of years, though typically predictable less than a century.

Note that I am not an expert on impact calculations, I just know a bit about and and can do back of the envelope ones. There are a number of ways to get to the ~1%, the orbit fits have uncertainties on them and those can be propagated forward in time. However there are all sorts of complexities with doing that, and often the easiest method is to sample the uncertainty region a few hundred thousand times (Monte-carlo), and propagate those and see what hits.

[coderenegade]

Very cool. How are samples drawn from the uncertainty region? MCMC or does it simplify down? I'm guessing that this would drive the final percentage values that you guys determine, since the orbital dynamics would be deterministic.

[ddahlen]

I can tell you how I do it, but again I am not an impact study person. It helps to understand a bit of the background of how we fit orbits in general:

1) someone with a telescope sees something moving (typically these days these are bigger surveys)

2) These observations are submitted to the Minor Plant Center (MPC), the clearinghouse of all asteroid/comet observations.

3) Several groups pull observations from the MPC to fit orbits, including JPL Horizons (MPC also fits orbits)

4) You now have a pile of observations which you have to figure out which observation links to other observations, which is a complex math problem on its own. Solve that.

5) JPL Horizons for example then fits the orbits to the observations, and since the observations may be 100 years of data of wildly varying quality, from hand written notes in the 1920s through to modern data, this is very difficult. They publish a covariance matrix with the associated fit (IE: basically a gaussian error fit for the parameters).

6) I grab that covariance matrix and sample from it using some pretty vanilla statistics to build orbits.

7) Propagate and see what happens.

Here is an example of an observation from 1950: https://caltech-ipac.github.io/kete/tutorials/palomar.html The image was developed on a glass plate, this one was never even sent in to the MPC, the guy taking the observation just wrote down "Asteroid" on the cover slip for the image. It was not formally discovered until the 1980s. We now know its orbit very well, so this particular observation is not that interesting other than as a curiosity.

Here is an example of an orbit fit by JPL Horizons: https://ssd.jpl.nasa.gov/tools/sbdb_lookup.html#/?sstr=c%2F2...

Note the "condition code" on the right, which is a score of how good their orbit fit matches the data, 0 means we know the orbit with high precision. This one is an 8, meaning we have a fit, but its not that great. Most likely because we only have 31 days of observations.