Band-Sampled Data Frequency-Hough Analysis

• Band-Sampled Data Frequency-Hough is a method that transforms time-frequency peakmaps from short FFTs into linear accumulations in an intrinsic parameter space, effectively linearizing signal tracks.
• It leverages band-limited data products like SFDBs and BSDs with adaptive coherence times to enhance detection sensitivity while reducing computational overhead in gravitational-wave searches.
• The approach extends classical continuous-wave analysis to cover power-law inspiral tracks, enabling robust searches for planetary-mass binaries with effective candidate vetoes and follow-up strategies.

to=arxiv_search.search ￣奇米json {"query":"Frequency-Hough band-sampled data BSD generalized frequency-Hough SFDB LIGO O3a", "max_results": 10} to=arxiv_search 】【。】【”】【json {"query":"Frequency-Hough band-sampled data BSD generalized frequency-Hough SFDB LIGO O3a", "max_results": 10} “Band-Sampled Data Frequency-Hough” is an Editor’s term for Frequency-Hough (FH) and generalized Frequency-Hough (GFH) analyses implemented on band-limited gravitational-wave data products, especially Short FFT Data Bases (SFDBs) and Band-Sampled Data (BSD). In this family of methods, short coherent transforms are used to build time-frequency peakmaps, and the Hough transform converts signal tracks into straight-line accumulations in an intrinsic-parameter space. In the classical continuous-wave setting, the target is a nearly monochromatic signal with Doppler modulation and spin-down; in the generalized setting, the same strategy is extended to power-law tracks such as long-lived inspirals of planetary-mass ultra-compact binaries. The central technical advantage is that a difficult curve-detection problem in the data plane becomes a peak-detection problem in a Hough accumulator [(Astone et al., 2014); (Miller et al., 2024)].

1. Frequency-Hough as a transform on band-limited peakmaps

The classical FH formalism starts from a detector-frame frequency model in which the observed frequency is the intrinsic source frequency modulated by detector motion. For an isolated continuous-wave source, the intrinsic frequency is expanded as

$$f_s(t)=f_0 + \dot f_0 (t-t_0) + \frac{1}{2}\ddot f_0 (t-t_0)^2 + \cdots,$$

while the detector-frame frequency is approximated by

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$

After Doppler correction at a trial sky position, the residual track is treated as linear in time,

$f'_i \approx f_0 + \dot f_0 (t_i-t_0),$

so that each peak $(t_i,f'_i)$ defines a line in the $(f_0,\dot f_0)$ plane,

$f_0 = f'_i - \dot f_0 (t_i-t_0).$

The FH histogram counts how many peaks are compatible with each intrinsic-parameter bin, thereby converting a time-frequency pattern into a source-parameter excess (Astone et al., 2014).

This line-based mapping is the defining feature of FH. It differs from the standard Hough usage in that the transform is organized directly in the source-parameter plane, which makes frequency over-resolution inexpensive and avoids the non-linear sky mapping characteristic of standard sky-Hough constructions. In the all-sky continuous-wave formulation, one Hough map is built for each sky position after Doppler correction. In later BSD-based follow-up work, the same principle is applied to narrow candidate-centered bands extracted from O3 SFDB data, so that heterodyned and downsampled re-analyses can be performed without manipulating the full wideband strain stream [(Astone et al., 2014); (Cesare et al., 5 Dec 2025)].

2. Band-sampled data products, SFDBs, BSD, and coherence-time choices

Band-limited FH analyses rely on a hierarchy of data products. The foundational one is the SFDB, built from interlaced short FFTs. In the continuous-wave FH method, the SFDB is constructed from interlaced FFTs of duration $T_{\rm FFT}\equiv T_{\rm coh}$, and narrow sub-bands can then be processed independently. In the O3 inspiral GFH analysis, the pipeline uses SFDBs rather than BSDs; the SFDBs are collections of 50% interlaced, cleaned FFTs of 1024 s produced for the full run, and the pipeline inverts these to the time domain and recomputes shorter FFTs matched to the instantaneous signal drift. In the BSD-based machine-learning follow-up for all-sky FH candidates, 1 Hz bands are extracted from the O3 SFDB, and 1 Hz is identified as the narrowest bandwidth that avoids introducing artifacts with the BSD procedure (Miller et al., 2024, Cesare et al., 5 Dec 2025).

Context	Band product	Stated role
All-sky continuous-wave FH	SFDB	Interlaced short-FFT input for peakmap construction and per-band FH processing
Inspiral GFH in O3a	SFDBs, not BSDs	1024 s FFT database inverted to $h(t)$, then shorter FFTs recomputed
BSD follow-up of FH candidates	1 Hz BSD bands from O3 SFDB	Candidate-centered heterodyne and refined Hough-map construction

The coherence time $T_{\rm FFT}$ is not fixed globally. It is chosen so that the signal remains confined within one Fourier bin over each FFT, with the exact constraint depending on the problem class. In classical all-sky FH, $T_{\rm FFT}$ is band-dependent so that Doppler frequency drift during a single FFT does not exceed one bin; typical choices include $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$0 s for 10–128 Hz and $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$1 s for 128–2048 Hz. A later O3 FH implementation uses $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$2 and $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$3 s in the 10–128, 128–512, 512–1024, and 1024–2048 Hz bands respectively. In the inspiral GFH search, much shorter transforms are required, with $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$4 s across configurations, because the intrinsic chirp must also remain within half a bin and the 0PN track must stay within one bin of the 3.5PN track over $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$5 [(Astone et al., 2014); (Giovanni et al., 2024); (Miller et al., 2024)].

Band sampling introduces edge-management requirements. In BSD processing, the sub-band must encompass the full possible frequency excursion due to Doppler and spin-down, with margins large enough that peak shifts do not push signal power outside the processed band. The recommended margin is

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$6

plus a safety buffer of a few FH bins, and overlapping sub-bands are recommended to avoid edge losses. This makes BSD suitable for narrowband follow-up, but only when the band padding is coordinated with the expected Doppler and secular drift (Astone et al., 2014).

3. Peakmap construction, thresholding, and accumulator statistics

The basic observable entering FH is a peakmap derived from short FFTs. In the continuous-wave formulation, each periodogram is equalized by an autoregressive estimate of the average spectrum,

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$7

and a “peak” is a frequency bin that is above threshold and a local maximum. The use of local maxima rather than arbitrary threshold crossings reduces spectral leakage and computational load. In Gaussian noise, the probability that a bin is a local maximum above threshold $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$8 is

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$9

and the recommended threshold in the 2014 FH method is $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$0, yielding $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$1. In the presence of a weak signal with spectral amplitude parameter $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$2, the peak-selection probability becomes

$f'_i \approx f_0 + \dot f_0 (t_i-t_0),$3

The unweighted Hough count is

$f'_i \approx f_0 + \dot f_0 (t_i-t_0),$4

with mean and variance $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$5 and $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$6, and the standard significance variable is

$f'_i \approx f_0 + \dot f_0 (t_i-t_0),$7

Adaptive FH replaces the unit increments by segment weights $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$8, whereas several later implementations retain binary or non-adaptive voting for robustness or for image-like downstream processing [(Astone et al., 2014); (Cesare et al., 5 Dec 2025)].

The inspiral GFH pipeline keeps the same semi-coherent logic but modifies the ranking details. Its equalized power per time-frequency bin is compared to a fixed peak threshold $f'_i \approx f_0 + \dot f_0 (t_i-t_0),$9, and all selected peaks are assigned unit weight. The number of counts $(t_i,f'_i)$0 in a $(t_i,f'_i)$1 bin is compared, within a local square of the Hough map, to the neighborhood mean and standard deviation to form

$(t_i,f'_i)$2

The square-wise comparison is introduced because the nonlinear $(t_i,f'_i)$3 transformation skews the count distribution toward low $(t_i,f'_i)$4, i.e. high frequency. The resulting configuration-dependent thresholds lie in the interval $(t_i,f'_i)$5 and are set to achieve a $(t_i,f'_i)$6 overall false-alarm probability after trials (Miller et al., 2024).

This statistical layer is one of the main points of continuity across FH variants. The precise threshold and weighting scheme can change from pipeline to pipeline. For example, the Doppler-correlation veto paper describes an O3 FH implementation in which local maxima above $(t_i,f'_i)$7 are selected, corresponding to a peak false-alarm probability of approximately $(t_i,f'_i)$8. What is invariant is the architecture: equalization, local-maxima peak selection, Hough accumulation, and a significance ranking based on the departure from local or global noise expectations (Giovanni et al., 2024).

4. Generalized Frequency-Hough for power-law inspiral tracks

GFH extends FH beyond linear-in-time frequency evolution. For a generic power law

$(t_i,f'_i)$9

integration gives

$(f_0,\dot f_0)$0

Defining

$(f_0,\dot f_0)$1

one obtains a linear relation in the transformed variable,

$(f_0,\dot f_0)$2

For circular inspirals with negligible eccentricity and quasi-Newtonian evolution far from merger, the leading-order quadrupole law is

$(f_0,\dot f_0)$3

Specializing the generic GFH mapping to this case gives

$(f_0,\dot f_0)$4

or equivalently

$(f_0,\dot f_0)$5

The slope is therefore set by the chirp mass through $(f_0,\dot f_0)$6, while the intercept encodes the initial frequency or, equivalently, the coalescence time (Miller et al., 2024).

The O3a inspiral search applies this mapping to planetary-mass primordial black-hole binaries with chirp masses roughly $(f_0,\dot f_0)$7. The parameter space is divided into $(f_0,\dot f_0)$8 configurations, each specified by

$(f_0,\dot f_0)$9

For each configuration, the pipeline computes short FFTs of duration $f_0 = f'_i - \dot f_0 (t_i-t_0).$0, builds a binary peakmap, transforms each peak $f_0 = f'_i - \dot f_0 (t_i-t_0).$1 to $f_0 = f'_i - \dot f_0 (t_i-t_0).$2, and accumulates votes along straight lines

$f_0 = f'_i - \dot f_0 (t_i-t_0).$3

The search deliberately avoids a sky grid: $f_0 = f'_i - \dot f_0 (t_i-t_0).$4 is chosen so that Doppler modulation stays within one frequency bin for any sky location over the full O3a span, and no barycentric demodulation is applied in the core search stage. This makes the method sensitive to sources anywhere on the sky, but it cannot localize them (Miller et al., 2024).

The same paper makes explicit the practical constraints required for the transformed linear model to remain faithful to the waveform. Intrinsic drift confinement requires

$f_0 = f'_i - \dot f_0 (t_i-t_0).$5

and PN consistency requires

$f_0 = f'_i - \dot f_0 (t_i-t_0).$6

These conditions are the technical basis for using a 0PN-inspired linearized accumulator on real inspiral tracks rather than only as a formal reparameterization (Miller et al., 2024).

5. Parameter-space grids, coincidences, vetoes, and follow-up

FH pipelines use discrete parameter grids whose resolution is set by coherence time and observing span. In the classical method, the natural frequency step is $f_0 = f'_i - \dot f_0 (t_i-t_0).$7, but FH can cheaply use frequency over-resolution

$f_0 = f'_i - \dot f_0 (t_i-t_0).$8

with $f_0 = f'_i - \dot f_0 (t_i-t_0).$9 recommended. The spin-down resolution is

$T_{\rm FFT}\equiv T_{\rm coh}$0

with coarse searches using $T_{\rm FFT}\equiv T_{\rm coh}$1 and refinement using $T_{\rm FFT}\equiv T_{\rm coh}$2, for example $T_{\rm FFT}\equiv T_{\rm coh}$3. In all-sky searches, a sky grid is also required; its density is controlled by the Doppler bandwidth

$T_{\rm FFT}\equiv T_{\rm coh}$4

and the high-frequency limit gives

$T_{\rm FFT}\equiv T_{\rm coh}$5

Reference epochs are placed at the mid-time of the observation interval to minimize stripe spread and parameter uncertainty (Astone et al., 2014).

Candidate handling is hierarchical. In the inspiral GFH search, the top $T_{\rm FFT}\equiv T_{\rm coh}$6 of candidates are retained within square tiles of the $T_{\rm FFT}\equiv T_{\rm coh}$7 plane rather than globally, to compensate for the skew introduced by the $T_{\rm FFT}\equiv T_{\rm coh}$8 transform. H1 and L1 are searched separately, and the resulting candidates are cross-matched using the metric

$T_{\rm FFT}\equiv T_{\rm coh}$9

This produced approximately $h(t)$0 million coincidences irrespective of $h(t)$1; after applying $h(t)$2, $h(t)$3 coincident candidates remained; after vetoing candidates within one frequency bin of known O3 lines, $h(t)$4 survived; and after heterodyne follow-up, all $h(t)$5 were vetoed on the first pass (Miller et al., 2024).

The heterodyne follow-up in that search is designed to demodulate the recovered inspiral phase,

$h(t)$6

If the recovered parameters are correct, the signal becomes nearly monochromatic near 0 Hz; the follow-up then doubles $h(t)$7 and applies the original classical FH to capture residual linear drift from parameter mismatch. The veto criterion is a reduction in $h(t)$8 relative to the original candidate (Miller et al., 2024).

Post-coincidence vetoes have also been developed for the standard all-sky FH pipeline. The Doppler correlation-driven veto exploits the approximate orbital relations

$h(t)$9

together with

$T_{\rm FFT}$0

A genuine continuous-wave signal and its nearby sky-grid “children” thus form near-straight lines in the $T_{\rm FFT}$1 and $T_{\rm FFT}$2 planes. The veto inserts two additional Hough steps after multi-detector coincidences, calibrates an effective orbital phase $T_{\rm FFT}$3, constrains the slopes to $T_{\rm FFT}$4 within $T_{\rm FFT}$5, and on average vetoes $T_{\rm FFT}$6 of candidates while preserving the dominant signal-consistent clusters (Giovanni et al., 2024).

An alternative follow-up path uses BSD explicitly. In the 2025 all-sky FH study, 1 Hz bands are extracted from the O3 SFDB, a heterodyne demodulation is applied using $T_{\rm FFT}$7, and a refined Hough map is built with

$T_{\rm FFT}$8

The resulting Hough maps are normalized per detector by the number of FFTs and passed to a neural-network classifier. This path is designed for the subthreshold regime below the standard follow-up threshold $T_{\rm FFT}$9, and the paper reports almost perfect confusion matrices at $T_{\rm FFT}$0, corresponding through the stated conversion factor $T_{\rm FFT}$1 to $T_{\rm FFT}$2 (Cesare et al., 5 Dec 2025).

6. Sensitivity, computational profile, limitations, and adjacent generalizations

The FH literature provides analytic sensitivity estimates that tie peak-selection probabilities and Hough thresholds to detectable strain or distance. In the classical continuous-wave method, the minimum detectable strain at confidence level $T_{\rm FFT}$3 is written as

$T_{\rm FFT}$4

with the associated analytic ROC framework derived from the Bernoulli/binomial model of peak counts. In the inspiral GFH search, the analogous figure of merit is a median $T_{\rm FFT}$5-CL distance reach $T_{\rm FFT}$6, whose closed-form expression makes explicit the dependence on $T_{\rm FFT}$7, $T_{\rm FFT}$8, the track frequency sequence $T_{\rm FFT}$9, the PSD $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$00, and the threshold terms involving $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$01, $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$02, $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$03, and $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$04. For a representative configuration at $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$05, O3a H1 injections yielded median reaches of $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$06 kpc empirically versus $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$07 kpc analytically, after accounting for duty cycle of approximately $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$08 and H1’s approximately $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$09 worse strain sensitivity than L1 [(Astone et al., 2014); (Miller et al., 2024)].

The computational profile depends strongly on whether a sky grid is present. Classical all-sky FH scales with the number of selected peaks, the spin-down bins they touch, and the number of sky points,

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$10

which is precisely why band-limited processing, peak reduction, and simple line drawing are important. The planetary-mass inspiral GFH avoids the sky grid entirely by constraining $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$11 so that Doppler stays below one bin, and the resulting search is much cheaper than matched filtering or all-sky continuous-wave FH. The reported total wall time is approximately $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$12–$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$13 days on $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$14 CPUs, with total core-hours of approximately $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$15, and each $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$16 point takes approximately $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$17 to process. By contrast, the paper explicitly notes that a dominant cost in all-sky CW FH searches, of order $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$18 core-hours in O3, is the sky grid (Miller et al., 2024).

Several limitations recur across implementations. FH sensitivity formulas assume Gaussian, stationary noise at the statistical level; spectral lines, non-stationarity, and duty-cycle gaps alter the peak-selection probabilities and count distributions. BSD processing requires sufficient margins and overlap to avoid band-edge losses after Doppler shifts. In the inspiral GFH setting, the method enforces that the 0PN track differ from the 3.5PN track by no more than one bin over $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$19, but higher PN terms and eccentricity are still neglected, and avoiding a sky grid prevents localization. In the continuous-wave setting, first spin-down only is typically retained when the minimum spin-down age is sufficiently large, while binary orbital modulation is outside the isolated-source model used by the Doppler-correlation veto [(Astone et al., 2014); (Miller et al., 2024); (Giovanni et al., 2024)].

A broader implication is that FH is best understood as a linearizing strategy rather than a single fixed astrophysical pipeline. A radio-burst analogue maps the dispersed law

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$20

to straight lines

$$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$21

in a $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$22 accumulator, so that points on one dispersed $$f(t)\approx f_s(t)\left(1+\frac{\mathbf v(t)\cdot \mathbf n}{c}\right).$$23 curve intersect at a single Hough peak. That method is not the gravitational-wave FH pipeline, but it demonstrates the same structural principle on band-sampled dynamic spectra: transform a curve family to straight lines, sparsify the data by thresholding, and detect the source as a compact excess in parameter space (Zuo et al., 2019).