Frequency-domain SETI Overview

• Frequency-domain SETI is the branch of radio technosignature research that analyzes power spectra and spectrograms using Fourier methods to identify signals consistent with engineered emissions.
• It employs a variety of techniques including FFT-based de-drifting, tree deDoppler, and deep-learning classifiers to analyze diverse observational regimes and signal morphologies.
• Advanced RFI rejection, multibeam coincidence matching, and sensitivity estimation are crucial for filtering noise and setting reliable upper limits on extraterrestrial transmitter power.

Frequency-domain SETI is the branch of radio technosignature research that searches channelized data products—power spectra, spectrograms, spectral-line image cubes, and related time–frequency representations—for signals whose morphology is more consistent with engineered emission than with known astrophysical or terrestrial backgrounds. Across Green Bank, the Murchison Widefield Array, FAST, SETI@home, and proposed ngVLA pipelines, the dominant targets are narrow-band and Doppler-drifting carriers, but the same general framework is also applied to pulsed signals, triplets, autocorrelation features, and arbitrary repeated waveforms (Siemion et al., 2013, Tingay et al., 2016, Zhang et al., 2020, Korpela et al., 17 Jun 2025, Ma et al., 2023).

1. Observational regimes and instrumental realizations

Frequency-domain SETI has been implemented across markedly different spectral bands, array architectures, and survey geometries. At the Green Bank Telescope, a targeted search of 86 Kepler Objects of Interest recorded baseband voltage data between 1.1 and 1.9 GHz and searched for narrow-band, drifting sinusoidal radio emission with spectral resolutions of 2.98 Hz, 1.49 Hz, and 0.75 Hz (Siemion et al., 2013). At low frequencies, the MWA generated spectral-line image cubes over 103–133 MHz and later 98–128 MHz, both at 10 kHz spectral resolution, using wide fields of view of about 400 deg$^2$ (Tingay et al., 2016, Tremblay et al., 2020). FAST has supported both commensal and targeted L-band SETI: the SERENDIP VI spectrometer operated over 1000–1500 MHz with 128 M channels and 3.725 Hz spectral resolution, while later targeted FAST campaigns searched 1.05–1.45 GHz toward 33 exoplanet systems and toward three nearby M dwarfs using multibeam coincidence matching (Zhang et al., 2020, Tao et al., 2022, Luan et al., 27 Feb 2025). SETI@home processed Arecibo data centered at 1.42 GHz by splitting each 2.5 MHz beam into 256 subbands and then searching multiple DFT lengths and drift hypotheses, while ngVLA studies describe a commensal 10–100 GHz program with full baseband capture, beam-forming, and Hz-scale channelization (Korpela et al., 17 Jun 2025, Anderson et al., 17 Jun 2025, Croft et al., 2018).

Facility or program	Band and representation	Search characteristics
Green Bank Telescope Kepler survey	1.1–1.9 GHz baseband, FFT channelization	tree deDoppler, on–off–on cadence, 25$\sigma$ threshold
MWA Galactic Centre and Vela surveys	103–133 MHz or 98–128 MHz image cubes	10 kHz channels, wide-field imaging, 5$\sigma$ narrow-band search
FAST SERENDIP VI and targeted FAST	1000–1500 MHz or 1.05–1.45 GHz	19-beam processing, Nebula, multibeam coincidence matching
SETI@home	1.42 GHz, 2.5 MHz beam split into 256 subbands	coherent de-drifting, 15 DFT sizes, multiple detectors
ngVLA concept	10–100 GHz full voltage capture	commensal processing, beam-forming, matched filters, tree-search de-chirpers

The diversity of these implementations defines the field. Some pipelines are pointed and target-centric, some are commensal, and some are effectively blind over very large stellar samples. The parameter space is correspondingly heterogeneous: low-frequency surveys emphasize sky coverage and underexplored bands, Green Bank and FAST emphasize high sensitivity in the microwave window, and ngVLA studies emphasize extension into the 10–100 GHz regime (Tingay et al., 2016, Croft et al., 2018).

2. Spectral representations, channelization, and drift compensation

The common computational substrate is Fourier analysis. In SETI@home, a sequence of length $N$ complex samples $x[n]$ is transformed through the standard $N$-point DFT,

$$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$

with power spectrum $\lvert X[k]\rvert^2$; successive DFTs of overlapping time segments form a two-dimensional array $P(k,t)$ (Korpela et al., 17 Jun 2025). The same paper describes DFT sizes $L_{\rm DFT}=2^3,\dots,2^{17}$, corresponding over each $\sigma$0 kHz sub-band to frequency resolutions from 1221 Hz down to 0.075 Hz, so that the search remains sensitive to both brief pulsed phenomena and maximally narrow carriers (Korpela et al., 17 Jun 2025).

Drift compensation is central because a nominally monochromatic extraterrestrial transmitter will not remain stationary in the observer frame. In the Green Bank Kepler survey, a modified “tree dedispersion” was repurposed as a “tree deDoppler” algorithm that sums power along linear tracks in the time–frequency plane. For each $\sigma$1, the search covers $\sigma$2 drift rates from $\sigma$3 to $\sigma$4, with

$\sigma$5

yielding, for example, $\sigma$6 Hz, $\sigma$7 Hz/s, and $\sigma$8 Hz/s (Siemion et al., 2013). SETI@home instead uses coherent integration: the raw time series is multiplied by a complex reference signal with instantaneous phase

$\sigma$9

so that a signal drifting at the trial rate is confined to a single DFT bin after de-drifting (Korpela et al., 17 Jun 2025). The front end performs this coherent search over 123 000 Doppler drift rates in the range $\sigma$0 (Korpela et al., 17 Jun 2025).

Not all modern pipelines explicitly de-Doppler before classification. The deep-learning Breakthrough Listen search keeps spectrograms intact and relies on a $\sigma$1-Convolutional Variational Autoencoder to recognize sloping tracks directly. Its 1.1–1.9 GHz scans are converted into $\sigma$2 inputs with $\sigma$3 Hz, and the maximum detectable drift rate within a single 4096-channel snippet is written as

$\sigma$4

By half-window overlap, the method achieves nearly 100 % coverage up to approximately 6 Hz/s and about 75 % coverage at the $\sigma$5 Hz/s extremes (Ma et al., 2023).

Image-domain implementations use different representations but the same basic logic. The MWA surveys imaged visibilities per 10 kHz channel and then searched extracted spectra or channel maps for isolated fine-channel excesses (Tingay et al., 2016, Tremblay et al., 2020). In this sense, frequency-domain SETI includes both beamformed and imaging spectrometers, provided that the detection logic is defined on discretized spectral bins or time–frequency cubes.

3. Detection statistics, candidate construction, and ranking

Classical frequency-domain SETI typically begins with thresholded excess power. The Green Bank Kepler search logged any deDoppler-summed spectral channel exceeding a 25$\sigma$6 threshold and reduced roughly $\sigma$7 detections by retaining only the highest-$\sigma$8 detection in each drift-and-frequency window before RFI rejection (Siemion et al., 2013). The MWA Galactic Centre and Vela surveys imposed a 5$\sigma$9 detection threshold in the channel map and corresponding single-pixel spectrum, with additional quality cuts requiring appearance in only one fine channel and exclusion of known astrophysical lines or catalogued RFI (Tingay et al., 2016, Tremblay et al., 2020). FAST SERENDIP VI normalizes spectra with a sliding window baseline estimator and records hits when $N$0, after which DBSCAN clustering with $N$1 and $N$2 is used to form candidate groups (Zhang et al., 2020).

SETI@home broadened the detection model beyond simple spikes. After de-drift and baseline removal, it applies five detectors: spike, Gaussian beam-crossing, pulsed-fold, triplet, and autocorrelation (Korpela et al., 17 Jun 2025). The spike detector uses a threshold $N$3 in a single bin, the Gaussian detector fits a 1D Gaussian plus constant background to power versus time, the triplet detector looks for three evenly spaced spikes above threshold, and the autocorrelation detector inverts the finest power spectrum to a pseudo-autocorrelation and searches delays up to $N$4 s (Korpela et al., 17 Jun 2025). In the back end, surviving detections are grouped into “multiplets” whose frequencies, drift rates, times, and sky positions are mutually consistent with a single celestial narrowband source; ranking combines power, density, and time factors, each expressed as a negative log-probability or normalized score (Anderson et al., 17 Jun 2025).

The deep-learning Green Bank search replaced much of the manual or rule-based candidate logic with a semi-unsupervised latent-space model. The $N$5-CVAE encoder has eight 2D convolution layers with $N$6 kernels and filter counts $N$7, a latent space $N$8, and a loss that adds reconstruction, $N$9-weighted KL divergence, and a cadence-clustering penalty with $x[n]$0 and $x[n]$1 (Ma et al., 2023). A Random Forest with 1 000 trees then assigns $x[n]$2, and a confidence cutoff of $x[n]$3 is used to keep the false-positive rate below 1 % (Ma et al., 2023). On held-out test benches of 24 000 simulated snippets, this $x[n]$4-CVAE+RF pipeline achieved ROC AUC $x[n]$5, precision $x[n]$6, recall $x[n]$7, and $x[n]$8; after searching about 115 million snippets across 1004 cadences, 8 “signals of interest” remained, but re-observations in May 2022 did not recover any persistent signals (Ma et al., 2023).

A recurrent pattern is that the candidate object is not a single spectral excess but a structured aggregate: a de-drifted detection, a cluster in $x[n]$9, or a multiplet persistent across epochs. This is significant because the practical burden in frequency-domain SETI is not only detecting power above noise, but discriminating between morphologies consistent with a sky-fixed transmitter and morphologies that are better explained by RFI or instrumental effects.

4. Radio-frequency interference rejection, multibeam logic, and verification

RFI rejection is the dominant methodological constraint in frequency-domain SETI. The Green Bank Kepler survey used an on–off–on cadence, requiring that a truly celestial signal appear in both on-source pointings but be absent in the intervening off-source; signals with $N$0 were further presumed local and filtered out (Siemion et al., 2013). The later Breakthrough Listen deep-learning search preserved the same logic in an extended ON–OFF–ON–OFF–ON–OFF pattern and trained its latent representation so that the three ON scans are pulled together and the OFF scans are pushed away for injected-ETI cadences, while pure-RFI cadences minimize all six cross-distances (Ma et al., 2023).

Multibeam discrimination is a second major line of defense. FAST commissioning data were filtered for zone RFI, drifting RFI, multi-beam RFI, and machine-learning RFI excision with $N$1 nearest hits in the $N$2 plane (Zhang et al., 2020). In this context, multi-beam RFI is defined by overlapping $N$3 boxes in different non-adjacent beams, and the later targeted FAST campaigns formalized an observing strategy named multi-beam coincidence matching (MBCM) for searches across 1.05–1.45 GHz in two orthogonal linear polarization directions separately (Tao et al., 2022). The 2025 FAST study likewise applied the multibeam coincidence matching blind search mode in each of the two orthogonal linear polarization directions (Luan et al., 27 Feb 2025). The underlying assumption is explicit in ngVLA design studies: a genuine astrophysical signal will appear only in the beam pointed at a star, whereas local RFI often leaks into multiple beams simultaneously (Croft et al., 2018).

Wide-field imaging surveys use both visibility-domain and image-domain defenses. In the MWA Vela survey, raw visibilities were flagged with AOFlagger on each 5 min snapshot, persistent narrow-band RFI in the 98–108 MHz FM radio region was blanked, and only about 64 % of the 30.72 MHz band remained for SETI (Tremblay et al., 2020). The earlier MWA Galactic Centre work similarly excluded the DC channel, coarse-band edges, and AO Flagger-flagged channels (Tingay et al., 2016). These low-frequency pipelines then reject any candidate coinciding with known astrophysical lines or catalogued RFI (Tremblay et al., 2020).

Verification often turns on secondary observables rather than on SNR alone. In the 2022 FAST exoplanet campaign, a signal at 1140.604 MHz from the observation toward Kepler-438 initially appeared roughly consistent with assumed ETI technosignatures, but evidence such as its polarization characteristics was almost able to eliminate the possibility of an extraterrestrial origin (Tao et al., 2022). In the 2025 FAST M-dwarf study, an unusual signal at 1312.50 MHz toward AD Leo was eliminated on the basis of polarization, frequency, and beam coverage characteristics (Luan et al., 27 Feb 2025). This repeated pattern addresses a common misconception: a visually narrow and drifting feature is not, by itself, an adequate technosignature claim. In the present literature, beam occupancy, cadence behavior, polarization, and repeatability are integral parts of the evidentiary standard.

5. Sensitivity, radiometer limits, and EIRP

The central quantitative summary in frequency-domain SETI is usually an upper limit on Equivalent Isotropic Radiated Power. In the MWA Galactic Centre study, the minimum detectable transmitter power is written as

$N$4

where $N$5 is the target distance, $N$6 is the flux-density limit in $N$7, and $N$8 is the channel bandwidth (Tingay et al., 2016). The same study gives the practical-unit approximation

$N$9

placing the best limits for the closest objects at order $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$0–$$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$1 W (Tingay et al., 2016). The Vela survey uses the same structure with a 5$$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$2 threshold and 10 kHz transmission bandwidth, obtaining $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$3 Jy for $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$4 Jy beam$$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$5 and, for a typical nearby star at $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$6 pc with $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$7 Jy, $$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$8 W (Tremblay et al., 2020).

For narrow-band microwave searches, the standard radiometer equation is explicit. The Green Bank Kepler survey writes

$$X[k] \;=\;\sum_{n=0}^{N-1}x[n]\;e^{-2\pi i\,k\,n/N}, \qquad x[n] \;=\;\frac{1}{N}\sum_{k=0}^{N-1}X[k]\;e^{+2\pi i\,k\,n/N},$$9

with $\lvert X[k]\rvert^2$0, $\lvert X[k]\rvert^2$1 K, $\lvert X[k]\rvert^2$2 K/Jy, $\lvert X[k]\rvert^2$3 s, and $\lvert X[k]\rvert^2$4 Hz, yielding $\lvert X[k]\rvert^2$5 Jy in a single 0.75 Hz channel (Siemion et al., 2013). For a nominal distance of $\lvert X[k]\rvert^2$6 kpc, the corresponding $\lvert X[k]\rvert^2$7 is about $\lvert X[k]\rvert^2$8 erg s$\lvert X[k]\rvert^2$9, approximately eight times the peak EIRP of the Arecibo Planetary Radar (Siemion et al., 2013). The ngVLA study uses a closely related form,

$P(k,t)$0

and, for $P(k,t)$1–20 Jy, $P(k,t)$2 Hz, $P(k,t)$3 s, $P(k,t)$4, and $P(k,t)$5, obtains $P(k,t)$6 Jy (Croft et al., 2018). The same study states that wide-angle airport-radar analogs with $P(k,t)$7 W become detectable at about 10 pc and that in the 30–50 GHz atmospheric windows the array could detect $P(k,t)$8 W at 10 pc and $P(k,t)$9 W at 100 pc for $L_{\rm DFT}=2^3,\dots,2^{17}$0 Hz and $L_{\rm DFT}=2^3,\dots,2^{17}$1 s (Croft et al., 2018).

FAST demonstrates how high collecting area and multibeam operation move these limits downward. The commissioning SERENDIP VI paper gives an EIRP estimate of about $L_{\rm DFT}=2^3,\dots,2^{17}$2 W at $L_{\rm DFT}=2^3,\dots,2^{17}$3 pc, $L_{\rm DFT}=2^3,\dots,2^{17}$4 Hz, and $L_{\rm DFT}=2^3,\dots,2^{17}$5 s (Zhang et al., 2020). The targeted FAST exoplanet campaign reports that the minimum equivalent isotropic radiated power it could detect reached $L_{\rm DFT}=2^3,\dots,2^{17}$6 W (Tao et al., 2022), while the three-source M-dwarf campaign reports a lower-limit detectable minimum equivalent isotropic radiant power of $L_{\rm DFT}=2^3,\dots,2^{17}$7 W (Luan et al., 27 Feb 2025). These numbers are explicitly described as being within the reach of current human technology in the 2025 FAST study (Luan et al., 27 Feb 2025).

A key interpretive point follows directly from these formulas: the reported upper limits are always conditional on bandwidth, dwell time, threshold, and assumed signal morphology. This suggests that comparisons across surveys are meaningful only when the channel width, drift treatment, and candidate logic are kept in view.

6. Search-space expansion, null results, and future directions

Frequency-domain SETI has expanded simultaneously in bandwidth, sky coverage, algorithmic complexity, and computational scale. The MWA Vela survey, combined with earlier Galactic Centre and Orion searches, reports examination of roughly 75 known exoplanets at low frequencies and EIRP upper limits toward over 10 million Gaia stellar sources in the Vela field, with a haystack fraction of approximately $L_{\rm DFT}=2^3,\dots,2^{17}$8, about 100 times the Galactic Centre value and 10 times the Orion value in the Wright et al. eight-dimensional metric (Tremblay et al., 2020). SETI@home moved the field in a different direction: instead of dedicated processing hardware, it distributed time-domain data to more than $L_{\rm DFT}=2^3,\dots,2^{17}$9 volunteered home computers, leveraging about $\sigma$00 floating-point operations per second and ultimately accumulating about $\sigma$01 front-end detections for later RFI excision and multiplet analysis (Korpela et al., 17 Jun 2025). The back end then reduced these detections to a few hundred manually examined high-ranking candidates, with final reobservations being carried out using FAST (Anderson et al., 17 Jun 2025).

Null results dominate the empirical record. The Green Bank Kepler survey found no signals of extraterrestrial origin (Siemion et al., 2013). The MWA Galactic Centre and Vela studies found no unknown signals (Tingay et al., 2016, Tremblay et al., 2020). FAST commissioning found no convincing extraterrestrial signals in the first $\sigma$02 hr of data (Zhang et al., 2020). The deep-learning search toward 820 nearby stars produced eight signals of interest but no claim of ETI, because re-observations did not recover persistent signals (Ma et al., 2023). The two targeted FAST papers each report initially interesting events that were subsequently eliminated by secondary evidence (Tao et al., 2022, Luan et al., 27 Feb 2025). This suggests that frequency-domain SETI has matured less as a sequence of detections than as a sequence of progressively stricter falsification procedures.

The future directions in the cited literature are correspondingly methodological. The FAST commissioning paper recommends increasing spectral resolution to about 1 Hz channels, implementing real-time RFI rejection with triggered 100 s raw-voltage dumps, streaming part of the raw band to SETI@home for 20$\sigma$03 greater sensitivity and coherent drift searches, and deploying a FAST Phase-Array Feed with about 100 beams (Zhang et al., 2020). Low-frequency work emphasizes that frequencies below 1 GHz remain largely unexplored and that future SKA-Low systems with bandwidth beyond 100 MHz, finer spectral bins, and lower $\sigma$04 could push $\sigma$05 into the $\sigma$06–$\sigma$07 W range for nearby stars (Tremblay et al., 2020). The ngVLA study frames 10–100 GHz as a critical complement to SKA, with commensal processing, beam-forming, up to $\sigma$08 GHz instantaneous bandwidth per receiver band, and machine-learning classifiers for spectrogram features (Croft et al., 2018).

A final misconception is that frequency-domain SETI is synonymous with a single narrow-band line search. The papers surveyed here show a broader reality: coherent de-drifting, tree deDoppler, wide-field image-cube searches, multibeam coincidence matching, latent-space cadence learning, multiplet ranking, and searches for pulses, triplets, Gaussian beam envelopes, and autocorrelation structure all fall within the same framework when the fundamental observable is spectral occupancy in discretized frequency space (Korpela et al., 17 Jun 2025, Anderson et al., 17 Jun 2025). Frequency-domain SETI is therefore best understood not as one algorithm, but as a family of inference strategies for distinguishing engineered radio structure from noise and interference over an expanding, explicitly quantified search space.