Band-Sampled-Data Directed Pipeline

• The BSD Directed Pipeline is a semi-coherent continuous-wave search framework that uses band-organized Advanced LIGO O2 data with partial Doppler correction and FrequencyHough transform to target the Galactic Center.
• It efficiently organizes strain data into 10 Hz bands and applies long coherent FFTs with peakmap construction, significantly extending coherence times and reducing computational costs.
• The approach sets sensitive upper limits on gravitational-wave strain from isolated neutron stars and inspires analogous sampled-data pipelines in other fields like control and wireless sensing.

Searching arXiv for the exact topic and closely related papers. Band-Sampled-Data Directed Pipeline most specifically denotes the BSD-based semi-coherent continuous-wave search architecture used for the Galactic Center search in Advanced LIGO O2 data, where Band-Sampled-Data (BSD) files provide reduced, band-organized detector data and a directed search fixes the sky position while scanning broad frequency and first spin-down ranges (Piccinni et al., 2019). In that canonical usage, the pipeline combines BSD heterodyne-style Doppler correction, long coherent FFTs, peakmap construction, the FrequencyHough transform, inter-detector coincidence, and upper-limit setting for continuous gravitational waves from isolated asymmetric rotating neutron stars near Sgr A* (Piccinni et al., 2019). A broader use of the term is suggested by later sampled-data work in system identification, control, speech reconstruction, wideband conversion, and wireless sensing, where band-organized or nonuniform sampled data are routed through explicitly staged, direction-preserving processing chains; that broader reading is interpretive rather than standardized.

1. Defining meaning and scope

In its primary and explicit sense, the Band-Sampled-Data directed pipeline is the search framework introduced for a directed semi-coherent search for continuous gravitational waves from the Galactic Center using Advanced LIGO O2 data (Piccinni et al., 2019). The search is directed because the sky location is assumed known—taken to be the position of Sgr A*—while the intrinsic signal frequency and first spin-down are unknown and searched over broad ranges (Piccinni et al., 2019). It is therefore neither a fully targeted search, because the source rotational parameters are not known, nor an all-sky search, because only one sky position/bin is searched (Piccinni et al., 2019).

The target source model is the standard continuous-wave model of an isolated spinning neutron star that is non-axisymmetric about its rotation axis (Piccinni et al., 2019). The intrinsic strain amplitude is written as

$$h_0=\frac{4\pi^2G}{c^4}\frac{I_{zz}f_{GW}^2}{d}\epsilon,$$

with the usual dependence on moment of inertia, gravitational-wave frequency, source distance, and ellipticity (Piccinni et al., 2019). Operationally, the observed signal is treated as nearly monochromatic but modulated by intrinsic spin-down, detector motion relative to the source, and antenna-pattern amplitude modulation (Piccinni et al., 2019).

A broader Editor's term use can be inferred from later literature: namely, a pipeline that ingests band-limited, band-organized, or irregularly sampled data, applies stage-wise harmonization or correction, and then performs a directed estimation, reconstruction, or detection task. This broader interpretation is suggested by sampled-data continuous-time identification under band-limited intersample assumptions (Fang et al., 2024), sampled-data observer design with predictor-observer decomposition (Karafyllis et al., 2019), and multirate wideband DSP pipelines in hardware (Ming et al., 2023).

2. BSD data architecture and processing chain

The BSD framework is the reduced-data architecture on which the directed Galactic Center pipeline is built (Piccinni et al., 2019). Each BSD file contains the reprocessed strain time series $h(t)$, represented as a complex time series, and down-sampled to 10 Hz from the original 16 kHz data (Piccinni et al., 2019). BSD files are organized by 10 Hz frequency bands and by run sub-periods of about one month, so the data can be manipulated bandwise without repeatedly processing the full-rate detector stream (Piccinni et al., 2019).

For each BSD file covering one 10 Hz band and one sub-period, the single-detector pipeline follows a fixed sequence: partial Doppler correction at the assumed sky position using a modified BSD heterodyne in 1 Hz sub-bands; longer coherent FFTs and peakmap construction; FrequencyHough transform from the peakmap to a $(f,\dot f)$ map; summation of FrequencyHough maps from all BSD files spanning the same frequency/spin-down region; candidate selection by standard FrequencyHough ranking; inter-detector coincidence between Hanford and Livingston candidate sets; post-processing vetoes including a significance threshold and known-line exclusion; candidate inspection; and finally upper-limit setting by Monte Carlo injections and sensitivity-depth extrapolation (Piccinni et al., 2019).

The BSD-specific contribution is the reduced data format, the complex low-rate band-limited data products, BSD heterodyne-style corrections, the new partial Doppler correction in 1 Hz sub-bands (“multi-Doppler”), and efficient handling in 10 Hz chunks and monthly pieces (Piccinni et al., 2019). The FrequencyHough-specific contribution is the use of peakmaps, mapping of time-frequency peaks into the $(f,\dot f)$ plane, Hough-map construction and summation, candidate ranking by Hough number count and critical ratio, and coincidence logic in Hough-grid coordinates (Piccinni et al., 2019). The overall architecture is therefore a BSD front-end wrapped around a FrequencyHough semi-coherent back-end.

The multi-Doppler correction is implemented by multiplying each 1 Hz sub-band by

$\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$

where $p_{\vec n}$ is the detector position projected along the source direction and $f_i$ is the sub-band central frequency (Piccinni et al., 2019). The paper reports that, within a 1 Hz sub-band, the approximation is valid with a maximum error of 5% with respect to the source frequency, in the sense that the corrected signal remains in the expected frequency bin (Piccinni et al., 2019).

3. Search model, resolution, and computational strategy

The Galactic Center search fixes the sky position to Sgr A*, with RA(J2000) $=17^\mathrm{h}\,45^\mathrm{m}\,40.04^\mathrm{s}$, Dec(J2000) $=-29^\circ\,00'\,28.1''$, or ecliptic coordinates $(266.8517,-5.6077)^\circ$ (Piccinni et al., 2019). Only this sky position is used in the search grid, so the total number of templates is simply the product of the number of frequency bins and spin-down bins (Piccinni et al., 2019). The searched parameter space is

$h(t)$0

including both spin-down and a small spin-up range (Piccinni et al., 2019).

After sky-position Doppler correction, the intrinsic frequency evolution is treated as effectively linear over the observing span,

$h(t)$1

with higher derivatives not searched (Piccinni et al., 2019). The FrequencyHough grid resolutions are

$h(t)$2

with $h(t)$3 and $h(t)$4 in this search (Piccinni et al., 2019). Coincidence between Hanford and Livingston candidates is defined by the normalized distance

$h(t)$5

with threshold $h(t)$6 (Piccinni et al., 2019).

Even though only one sky bin is searched, the corresponding angular resolution is discussed through Doppler resolution. The number of frequency bins affected by Doppler modulation is reported as $h(t)$7 at the lowest frequencies and $h(t)$8 at the highest frequencies, corresponding at 8 kpc to a Galactic-Center-centered patch radius from about 150 pc at low frequency to 25 pc at high frequency (Piccinni et al., 2019).

The central computational advantage is that partial Doppler correction before peakmap generation allows longer coherence times at fixed cost (Piccinni et al., 2019). The coherence time scales as $h(t)$9, with reported values of 64208 s for 10–20 Hz and 10776 s for 700–710 Hz; crucially, the partial correction allows a coherence time 4 times longer than without correction (Piccinni et al., 2019). The effective Hough-grid frequency resolution ranges from $(f,\dot f)$0 to $(f,\dot f)$1, while the spin-down natural resolution ranges from $(f,\dot f)$2 to $(f,\dot f)$3 for Hanford and from $(f,\dot f)$4 to $(f,\dot f)$5 for Livingston (Piccinni et al., 2019). The total template counts are reported as $(f,\dot f)$6 for Livingston and $(f,\dot f)$7 for Hanford, yet the search required only 207 jobs per detector, about 30 min/job, and about 200 core hours total, excluding BSD production (Piccinni et al., 2019).

4. Galactic Center O2 deployment and empirical performance

The search used Advanced LIGO O2 open data from Hanford and Livingston (Piccinni et al., 2019). O2 ran from 2016-11-30 to 2017-08-25; only science segments of the latest calibrated data were used, poor-quality periods were excluded, Livingston data before 2017-01-04 were discarded, and 35 days of Hanford data from mid-March to mid-April were excluded (Piccinni et al., 2019). Virgo was not used because of shorter observing time and lower sensitivity (Piccinni et al., 2019). The final search employed 1120 BSD files spanning 70 frequency bands of width 10 Hz across 10–710 Hz (Piccinni et al., 2019).

Peakmap construction uses the standard FrequencyHough threshold with noise-peak probability

$(f,\dot f)$8

(Piccinni et al., 2019). Candidate ranking retains approximately 1000 candidates per job, producing 203961 candidates for Livingston and 202556 candidates for Hanford (Piccinni et al., 2019). Coincidence with $(f,\dot f)$9 yields 237 coincident candidates; a second-stage veto uses the Critical Ratio

$(f,\dot f)$0

with a threshold chosen so that, on average, only one false candidate over the full parameter space is expected (Piccinni et al., 2019). The threshold range is $(f,\dot f)$1 for Hanford and $(f,\dot f)$2 for Livingston (Piccinni et al., 2019).

After the critical-ratio cut, only 9 coincident candidates survive (Piccinni et al., 2019). Of these, 4 are due to known instrumental lines, 1 is due to the hardware injection Pulsar_10, and the remaining 4 were judged non-astrophysical because they were associated with transient lines, especially in Livingston before 14 March 2017 (Piccinni et al., 2019). No astrophysical continuous-wave candidate was found (Piccinni et al., 2019).

Sensitivity is reported through upper limits and sensitivity depth. Using injection campaigns in 26 clean 1 Hz bands, the mean depths are about 44.30 $(f,\dot f)$3 for Hanford and 52.44 $(f,\dot f)$4 for Livingston (Piccinni et al., 2019). The detection efficiency is fit with

$(f,\dot f)$5

and the $(f,\dot f)$6 upper limit is defined by $(f,\dot f)$7 (Piccinni et al., 2019). The most stringent $(f,\dot f)$8 confidence upper limits are $(f,\dot f)$9 near 161 Hz for Livingston and $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$0 near 195 Hz for Hanford; these do not include calibration uncertainty (Piccinni et al., 2019). With $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$1 kpc and fiducial $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$2, the strongest ellipticity constraint is about $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$3 at the highest frequencies for Livingston (Piccinni et al., 2019). The paper characterizes the result as the most sensitive directed search for Galactic Center continuous waves to date and the first such search using O2 data (Piccinni et al., 2019).

5. Broader sampled-data formulations in estimation and control

Outside gravitational-wave astronomy, the phrase can be generalized only cautiously. A plausible broader reading is a staged architecture that starts from sampled data whose intersample, spectral, or timing structure matters, and then performs explicitly directed estimation or control.

In continuous-time system identification from sampled data, the decisive step is to specify both the intersample behavior and the past behavior of the continuous-time input (Fang et al., 2024). For band-limited input under the Nyquist frequency and periodically appended past behavior, the input can be reconstructed exactly from the samples by a finite Fourier series, and the kernel-based regularization estimator of the continuous-time impulse response acquires a closed form (Fang et al., 2024). The estimator is

$\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$4

and the paper reports from Monte Carlo experiments that the broader sampled-data KRM framework is more robust than SRIVC and PEM and more accurate when the sample size is small, although the experiments focus on the ZOH case rather than the band-limited case directly (Fang et al., 2024). This suggests a band-sampled-data pipeline in which sampled inputs are first lifted into a continuous-time model class before identification.

In sampled-data observer design, the central architecture is a predictor-observer decomposition: a continuous-time observer is driven not by the unavailable continuous output $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$5, but by an auxiliary predictor state $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$6 that is reset at sample times and propagated between samples (Karafyllis et al., 2019). The general predictor is

$\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$7

with reset $\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$8, while the observer evolves as

$\exp\left(i\frac{2\pi}{c}p_{\vec n}f_i\right),$9

Under an IOS-type assumption on the underlying continuous-time observer and the sampling constraint

$p_{\vec n}$0

the sampled-data observer inherits exponential convergence in the noiseless case and robustness to measurement noise (Karafyllis et al., 2019). Here the “directed” character lies in the explicit information flow sample $p_{\vec n}$1 predictor $p_{\vec n}$2 observer $p_{\vec n}$3 state estimate.

A related directed sampled-data architecture appears in formation control with local measurements under directed graphs (Wang et al., 2019). The sampled-data controller is implemented with synchronous zero-order hold, requires only local-frame relative measurements, and admits local exponential stability for a static target when

$p_{\vec n}$4

and the interaction graph contains a directed spanning tree (Wang et al., 2019). In this case, directionality refers simultaneously to the graph topology and to the staged sample-and-hold control law.

6. Wideband signal-processing and communication instantiations

Several later hardware and sensing systems exhibit what can plausibly be called band-sampled-data directed pipelines, although not under a single standardized name.

Domain	Sampled-data organization	Directed stage
Wideband SRC	20 GSPS, 80 lanes, parallel then serial decimation	Parallel-to-serial multistage rate reduction
Real-time FFT metrology	40-lane ADC bus to 24 FFT lanes	Frame-directed demultiplexing and resequencing
Beyond-Nyquist reception	Bonded ADC channels with known phase/timing relation	Calibration-aware digital recombination
Orthogonal broad-band generation	$p_{\vec n}$5 low-rate branches with sinc-sequence weighting	Deterministic sample partition and analog summation
Wi-Fi sensing	Irregular CSI from 2.4/5 GHz and diverse packets	Sanitization plus time-aware attention

In wideband sample-rate conversion, a cascaded parallel-serial SRC structure converts a 20 GSPS input represented as 80 parallel lanes at 250 MSPS per lane into a lower-rate configurable stream by a parallel CIC and halfband front end followed by a serial CIC and halfband back end (Ming et al., 2023). In the worked design, the parallel front end decimates by $p_{\vec n}$6, the serial CIC is adjustable from 1 to 4000, and the total decimation range is 80 to 2,560,000 (Ming et al., 2023). The paper reports implementation on a Xilinx KU115 FPGA with total resources of 43,402 LUTs, 82,119 FFs, and 183 DSP48Es (Ming et al., 2023).

In real-time frequency measurement for time-stretched acquisition, a 40-lane high-rate ADC output is reorganized by a hierarchical parallel-to-serial stage into 24 FFT channels, each processing 440-sample frames zero-padded to 512 and refined by a simplified parabolic fit (Ming et al., 2023). The measured frequency is

$p_{\vec n}$7

and the paper reports a frequency precision better than 1 MHz while processing signals of bandwidth 4 GHz at a frame repetition rate of 22 MHz (Ming et al., 2023). Here the directed pipeline is explicitly frame-routed: incoming frames are distributed round-robin to FFT lanes, processed, then reorganized for host transfer (Ming et al., 2023).

In receiver bandwidth extension beyond Nyquist using channel bonding, two coherent 5 GSa/s ADC channels are digitally recombined to reconstruct 5 GHz instantaneous bandwidth using either a hybrid-coupler I/Q architecture or a time-interleaved architecture (Giehl et al., 2022). The paper reports up to 49 dB image rejection ratio, typically within 4 to 8 dB of theoretical front-end limits (Giehl et al., 2022). The key requirement is not mere oversampling but a known inter-channel phase or timing structure that makes digital recombination possible (Giehl et al., 2022).

In orthogonal sampling based broad-band signal generation, a broadband waveform of bandwidth $p_{\vec n}$8 is synthesized from $p_{\vec n}$9 low-bandwidth branches, each requiring only branch sampling rate $f_i$0 and branch analog bandwidth $f_i$1 (Hosni et al., 2023). The recombination uses orthogonal, time-shifted sinc-pulse sequences; the paper reports 60 GHz data generation from 20 GHz and 12 GHz electronics in simulation, and an ENOB improvement of about 2 bits at 1 ps DAC jitter (Hosni et al., 2023).

In Wi-Fi sensing from communication traffic, UniFi processes irregularly sampled CSI from diverse packets and multiple bands without packet injection (Dong et al., 14 Dec 2025). Its front end is a CSI sanitization pipeline with clustering, normalization, alignment, and burst filtering, and its back end is a time-aware attention model that learns directly from non-uniform sequences without resampling (Dong et al., 14 Dec 2025). On the dual-band CommCSI-HAR dataset, the paper reports $f_i$2 accuracy using 5 + 2.4 GHz, all packets, while fully preserving communication throughput (Dong et al., 14 Dec 2025).

7. Conceptual boundaries, misconceptions, and methodological cautions

A recurrent misconception is to equate directed with targeted. In the canonical BSD usage, the search is directed because the sky position is fixed, but it is explicitly not a targeted search because the source rotational parameters are unknown, and it is not an all-sky search because only one sky bin is used (Piccinni et al., 2019). Similar ambiguity appears in other fields: “directed” may refer to dataflow through a fixed stage sequence, to graph directionality, or to prior knowledge of geometry or timing rather than to a pre-specified source identity.

A second recurring issue is that sampled-data pipelines are only as valid as their intersample model. In continuous-time identification, the estimator becomes closed-form only after assuming ZOH or band-limited intersample behavior and an explicit past-input model; without that modeling step, the core covariance integrals have no tractable closed form (Fang et al., 2024). In observer design, robustness depends on a small-gain-type sampling bound, not on sampling alone (Karafyllis et al., 2019). In beyond-Nyquist reception, alias cancellation is only valid when the front end enforces a known quadrature or half-sample relation between channels (Giehl et al., 2022).

A third issue is the treatment of irregular or heterogeneous samples. UniFi provides a strong counterexample to the assumption that interpolation to a regular grid is always desirable: on irregular dual-band communication CSI, direct irregular-time modeling outperforms linear interpolation, and the framework eliminates intrusive packet injection entirely (Dong et al., 14 Dec 2025). Conversely, burst-induced redundancy still needs to be pruned, because dense packet bursts can overrepresent almost unchanged channel states (Dong et al., 14 Dec 2025).

A final caution comes from sampling on directed networks, where the sampling mechanism itself alters inferred structure. For BFS-type sampling on complete directed networks, the paper reports that at coverage below 40%, average degree, variance of out-degree, degree auto-correlation, and link reciprocity are overestimated by 30% or more, and values come within 10% of the complete-network values only when coverage exceeds 65% (Son et al., 2012). Although this use of “directed” is different, it illustrates a general principle relevant to all sampled-data pipelines: the acquisition rule is not a neutral front end, and inferred structure can be dominated by the sampling mechanism itself.

Taken together, these literatures suggest two levels of meaning. In the strict historical sense, the Band-Sampled-Data Directed Pipeline is the BSD-plus-FrequencyHough architecture for directed continuous-wave searches toward the Galactic Center (Piccinni et al., 2019). In the broader interpretive sense, it names a family of architectures in which band-organized or irregularly sampled data are passed through a deliberately staged, direction-preserving processing chain whose front-end sampling assumptions materially determine what can be inferred, reconstructed, or detected.