Read the Gaussian lens chapter and the timing-statistics chapter first. Both inputs are observed CHIME data; the noise trials and injections are synthetic.
Get the files
Put these eight files in one empty folder:
- README.md, run instructions, results and failure cases
- propagation_fit.py, the one-burst and burst-plus-copy fits
- repeat_statistics.py, gaps, windows, folding and lens requirements
- reproduce.py, runs both and compares every number
- chime-frb20180725a-intensity.json, the frozen intensity cutout
- r3-catalog2-rows.csv, the 142 catalogue rows of FRB 20180916B
- expected-outputs.json, the numbers quoted here
- provenance.json, sources, data kinds and SHA-256 checksums
Then run:
python3 -m pip install numpy scipy
python3 reproduce.py
It exits with status 0 when every value matches the saved outputs to a relative 5×10−3 or an absolute 0.05. The burst fit uses Nelder-Mead, which stops at a tolerance, so the last digits can differ between platforms. We ran it with Python 3.9.6, numpy 2.0.2 and scipy 1.13.1. The data come from the CHIME/FRB Collaboration: the First CHIME/FRB Catalog (DOI 10.11570/21.0007) for the cutout and the Second CHIME/FRB Catalog (DOI 10.11570/25.0066) for the rows. Cite them with any copy; no broader licence is asserted.
Units and conventions
Step 1: fit one burst
The cutout has 64 channels from 400 to 800 MHz and 38 time bins of 0.983 ms. Ten channels are fully masked, leaving 54. The burst model is a Gaussian of width σ convolved with a one-sided exponential tail, shifted in each channel by a residual dispersion δDM. Each channel gets its own amplitude and baseline, solved linearly; noise per channel comes from the off-pulse bins 0 to 7 and 22 to 37.
A reduced χ2 of 1.28 means the model misses some structure. Keep that in mind when reading the significance below.
Step 2: add a lensed copy
A plasma-lens echo is the same burst, weaker, later and, when the dispersive term dominates, with a delay that grows as ν−2. The constrained model adds one copy with a single amplitude ratio r and delay shared by every channel:
A second version, achromatic, uses the same delay in every channel. The scan covers delays from 0.5 to 9.5 ms (chromatic; longer copies leave the window at 400 MHz) or 0.5 to 20 ms (achromatic), and ratios from 0.05 to 1.5. The statistic is the largest drop Δχ2 over the grid. Its null distribution comes from 300 synthetic data sets: the best H0 model plus white noise with each channel's off-pulse scatter, seed 20261114.
No copy is detected. The best grid point sits at the smallest ratio, 0.05, which is what an absent signal looks like. Rescaling the noise so the H0 fit has reduced χ2=1 gives false alarm probabilities of 0.68 and 0.86.
An earlier version gave each channel its own copy amplitude. It returned Δχ2=104 for an achromatic copy, an apparent detection. The 54 extra free amplitudes, fitted to a model that already misses structure, made that number meaningless. Sharing one ratio across the band is the physical constraint that turns the fit into a test.
Step 3: what the fit can see
Inject a chromatic copy into the best H0 model, add noise, refit H0 and rescan.
The first row is the main failure case. The copy is clearly detected at nearly the right delay, but the refit scattering tail triples and swallows most of it, so the ratio comes back three times too small. A scattering tail and a delayed copy are partly degenerate at 1 ms resolution. The second row is marginal, near the 95th percentile of the null. The third shows that a copy closer than one time bin is lost: the scan finds noise at a wrong delay.
Step 4: repeat statistics of FRB 20180916B
repeat_statistics.py reads the 142 catalogue rows, keeps 131 taken during nominal operation and merges sub-bursts into 104 events.
The fold is a consistency check with topocentric 400 MHz times, not a new period measurement. The published period used barycentric times at infinite frequency. Periodic activity from a fast radio burst source
Step 5: what a lens would need
For each delay, the script finds the gain-one image at u2=1/2, where the delay is close to tDM0α/2e. It fixes the lens width at a=3rF, solves for the column and computes free-free absorption with Eq. 26 of Cordes and colleagues. Lensing of fast radio bursts by plasma structures in host galaxies
Only the shortest gap could escape absorption, and it would need an extreme lens with the source 1.7×105 lens radii off axis. None of this changes the basic count: one lens gives at most three copies with fixed delays.
Failure cases and limits
- The null uses white noise. The reduced χ2 of 1.28 shows correlated residuals, so the false alarm probabilities are approximate.
- One copy, one ratio for all channels. A real lens gain changes with frequency; a copy that is bright in only part of the band would be missed.
- Delays below one 0.98 ms bin are not resolved, and a scattering tail can hide a copy a few ms late.
- The Weibull fit treats all gaps as one sample. The window-only trials are an idealized exposure; CHIME's real exposure varies from day to day.
- The lens requirement assumes a single Gaussian lens at 104 K and a=3rF. A hotter lens absorbs less: τff∝T4−1.32.
Lab tasks
Set the burst width in the plasma-lens echo Lab to about 0.6 ms, the width fitted here. Find a lens setting whose second copy is 4 ms late, and compare the time profile with the injection in Table 5.4.
The CHIME paper on the 16-day cycle is read figure by figure in the next guide: how to read the CHIME paper on the 16-day FRB cycle.