The South Pole Telescope (SPT) group manages and analyzes data from the eponymous telescope at the south pole. The telescope uses an array of detectors to measure the strength of radiation from the cosmic microwave background (CMB). Studying the distribution of this energy across the sky can tell us about the varying density of the early universe.
As an intern, I worked to eliminate some of the errors which found their way into the data via detector hardware (rather than software or statistics). Here is an example of one such error, called a spectral line:
It may be hard to see at this resolution, but there are small "ripples" across this particular view which don't appear in other views.
We call this a "line" because it appears as a vertical line, a very thin spike, when we look at frequency (or Fourier) space. This means the telescope sees a certain spatial oscillation or wave at a frequency that dominates all others, so much so that it becomes visible to the naked eye. Here is the same image transformed to 1-dimensional frequency space:
This type of plot is called a Power Spectral Density (PSD). The spike represents the single spatial frequency that dominates. Notice that all other frequencies are at a pretty even level (ignoring the low frequencies to the left, where noise naturally increases).
To address this problem, I proposed to test cutting out parts of the data. Not enough to lose valuable information, but enough to remove the peskiest spectral lines. One piece of code I wrote performs 3 main tasks:
- Identifies the power (i.e. height, strength, or energy) of each line in a PSD
- Determines whether that line is larger than a certain threshold
- Deletes the entire map if it is above this threshold (the map only represents a single detector, the telescope has 1500 of them)
My task, then, was to determine a suitable threshold which would not unnecessarily increase noise.
I applied my cut to the same dataset using a few different signal-to-noise thresholds. Below is the ratio between PSDs with the cut applied and those without the cut applied:
We see a greater reduction (Ratio < 1) in SN at certain frequencies: around 1800Hz and 3100Hz. At these frequencies, the cuts are most effective. But at other frequencies, any amount of cutting creates an increase in overall noise: ~3700Hz is a glaring example. A frequency-dependent cut was the reasonable solution to this.
To talk more about my work at SPT, feel free to send an email!