How much signal a segment boundary costs¶

A compact binary's buffer begins seconds before its coalescence. Data written in segments therefore has a choice to make: if the segment that claims an event is the one containing coa_time, then a buffer starting before that segment is cropped, and the earlier segments it belonged to have already been written.

WaveformBackend.pre_coalescence_duration exists so a caller can claim the event by where its waveform starts instead. This page records what the cropping costs when it does not.

The loss is not a single number¶

Every figure below is a fraction of the unweighted strain-squared energy, and every one of them moves by orders of magnitude with three things:

• the low-frequency cutoff,
• how far past the boundary coa_time lands,
• the waveform backend.

A percentage quoted without all three is not interpretable. The tables therefore name them, and so should anything that cites the tables.

A 30+25 solar-mass binary on LAL¶

IMRPhenomD, 1024 Hz, into a single detector (H1), coa_time 0.5 s past a segment boundary:

The dropped span is identical and the energy in it differs 35-fold. At 30 Hz those early samples lie below the cutoff and are near-silent; at 20 Hz they carry real signal, and they also enlarge the total the fraction is taken against.

So here the cutoff changes the content of the dropped span and not its geometry. That is a property of this chirp-time bin, not a rule. LAL rounds the buffer to a power of two, which happens to absorb the 20-to-30 Hz difference for 30+25 — and for 25+25, 30+30, 40+30 and 50+50 — but not for lighter systems:

The offset dominates¶

Same binary, LAL, H1. Energy lost at a 20 Hz cutoff against 30 Hz, by how far coa_time sits past the boundary:

The gap between cutoffs is widest in the middle, where the dropped span covers just the band between them.

A binary neutron star¶

1.4+1.35 solar masses, LAL, H1. Worse in absolute terms, and here the cutoff moves the geometry too, because the chirp time dominates the rounding:

Those last two rows are the same binary half a second apart. On the boundary the only thing retained is the near-silent post-merger tail; half a second later the merger itself is inside the segment. That pair is the reason an offset has to be given alongside any figure on this page.

The 256 s buffer is specific to 1.4+1.35. A 2.0+1.5 binary leads by about 115 s at the same cutoff.

The backend changes the geometry¶

Leads for the same 30+25 binary at 1024 Hz:

For ripple the cutoff moves the buffer itself, so none of the LAL figures above transfer to it. PyCBC does not report a length at all — its pre_coalescence_duration returns None, meaning unknown, and a caller must not read that as zero.

Other things that move it¶

All measured at 20 Hz, coa_time 0.5 s past the boundary, against the 32.3% baseline above.

Bigger than the cutoff-independent effects listed further down:

• Component spins are the largest single mover named here. Aligned spin1z = spin2z of 0.4 gives 35.7% and 0.8 gives 38.5%, against 32.3% for a non-spinning binary.

• The approximant matters as much, and this one has a trap in it:

  approximant	dropped h²
  IMRPhenomD	32.3%
  IMRPhenomXPHM	31.5%
  TaylorF2	26.4%

  Every table on this page uses IMRPhenomD. gwmock's default waveform model is IMRPhenomXPHM, so a default gwmock run does not produce the 32.3% quoted here — it produces 31.5%.

• GPS epoch moves it 30.9% to 33.3% across twelve epochs spanning a single day. The cause is the Greenwich mean sidereal time at the epoch, which sets where the detector is pointing when the signal arrives; it is not rotation during the buffer, which over 4 s is negligible.

  Sampling this one badly is easy, and the failure is worth recording. Four epochs taken on 1 January of different years give 32.32% to 32.34%, which looks like a flat knob — but sidereal time at a fixed calendar date barely changes from year to year, so those are nearly the same orientation measured four times. A quantity periodic in sidereal time has to be sampled across that period, not along a variable aliased with it.

Smaller:

• Right ascension, declination and polarization angle together span 29.7% to 33.4% over a 6×3×3 grid. Inclination is not part of that: on its own it moves the fraction by about 0.003 percentage points.
• Detector network: 32.9% for a three-detector ET triangle against 32.3% for H1.
• ringdown_fraction (a backend constructor argument, not a per-event parameter) changes the lead more than the fraction: 0.05 gives a 3.800 s lead and 32.37%, 0.1 gives 3.600 s and 32.34%, 0.2 gives 3.200 s and 32.29%.
• Distance cancels out of a fraction entirely.

Checked and genuinely negligible: sample rate, reference frequency, and the backend segment_duration pin.

The ripple leads quoted above are for its constructor defaults, including taper_fraction; changing the taper changes them.

These are a proxy, not an SNR loss¶

Every figure is unweighted h². A matched-filter SNR loss needs a detector PSD and frequency-domain weighting, neither of which is applied here. None of these numbers is anchored against an external SNR tool — they are this package's own measurement of its own output, and should be read as an indication of scale rather than as a detection-efficiency statement.