Notes

Things I learned doing research

Check what a number means before you use it

In the Gaia paper I took a published 2% calibration error as 2% of a star's total light, when it meant 2% of the light at each blue wavelength. For cool stars that turned a 2% error into a median 55% one, 27 times too big. The code ran fine the whole time. Now I read the source for the units and the definition of every input before running anything. The full story is in The 2% that was actually 55%.

Zero is a result, if you know what you could have found

My cross-band search went through 1,369 galaxies and found no source with the same period in X-rays and optical light. On its own that doesn't say how many there could be. So I put fake periodic signals into the real light curves and ran the whole search again, 1,800 times. That's what turned zero into a limit: at most about 3% of the Swift-BAT galaxies can have a shared period with a hard X-ray modulation of 30% or more.

A fit can look good and still be wrong

In the X-ray inference paper, shifting the detector's energy scale by up to 3% got past all three checks I tried, because the fit moves the source parameters to make up for the shift. The spectrum still fits and only the answer changes. So when a fit looks good, I still test it on simulations where I already know the right answer.