Detection and excision of radio-frequency interference in radio arrays
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: I'm Vera.
Jocelyn: And I'm Jocelyn, here with Subrahmanyan.
Vera: We're looking at the paper: "Detection and excision of radio-frequency interference in radio arrays".
Jocelyn: Here's the gist of the paper. The paper develops a model to determine which data product—individual antenna streams or coherent array combinations—is most detectable for suppressing radio-frequency interference (RFI) in radio arrays The detection and excision of radio-frequency interference (RFI) is usually approached by asking how best to identify RFI within a given data product.
Vera: The comparison between antenna-based and array-based detection reduces to a single quantity related to the mean per-baseline RFI coherence.
Jocelyn: Cross-correlation detection is preferred only above a coherence threshold of (N − 1)−1/2, where N is the number of antennas.
Vera: Compact, filled, arrays favor the crosscorrelations for stationary terrestrial emitters, whereas extended sparse arrays do not.
Title and authors: Vera: So what is the actual theoretical framework they use to set up this detectability problem? They start by defining visibilities and beams from channelized voltage data.
Jocelyn: It defines them in terms of how you sample spatial coherence, which they call the visibility, written as V p q equals the average of v p times q-star.
Subrahmanyan: Then they introduce a specific formula for a deterministic emitter at a finite distance that gives you the visibility contribution, which involves the gain and an integral over solid angle with an exponential term.
Vera: The suppression they derive comes from averaging that formula over bandwidth, time, or baselines to get those coherence factors.
Jocelyn: They break down the per-baseline RFI coherence into three main pieces: bandwidth smearing, which is γ bw pq; fringe rate time-smearing, which is γ fr pq; and the intrinsic spatial coherence of the RFI itself.
Subrahmanyan: So they’re saying the suppression depends on how smeared that RFI signal gets across the channel, how quickly that fringe rate smears over time, and how coherent that interference is inherently.
Vera: They also point out a key limitation regarding fringe tracking: it provides no suppression for interference that is impulsive on timescales shorter than one second in any ground-based array.
Jocelyn: It means if the RFI is really sharp and fast, like a very sudden spike, fringe tracking won't help you much in an array setting.
The paper's summary: Vera: One of the practical improvements they suggest is focusing on how you handle the data processing, moving towards real-time excision using GPU-based methods for large arrays.
Jocelyn: They describe a real-time implementation for an array like the eleven hundred antenna Deep Synoptic Array, which searches for fast radio bursts in real time.
Subrahmanyan: The improvement there is that they can form power moments once to reduce every operation in their processing from a second pass over the voltages down to just a reduction over resident tensors.
Vera: They also suggest using AI systems for hierarchical mitigation, meaning you flag individual antenna streams first and then inspect the autocorrelation sum afterwards.
Jocelyn: They argue that this ordering matters because the sum is not robust to those per-element tests they design to find RFI on a single antenna element.
Subrahmanyan: That connects back to the theory; it suggests that you should condition those individual streams first before looking at the final sum, because of how sensitive the sum itself is.
The paper's improvements: Vera: So summarizing this paper "Detection and excision of radio-frequency interference in radio arrays," the main finding is that autocorrelation sum is generally more sensitive to RFI than other methods they tested.
Jocelyn: And they suggest a specific workflow where you flag individual antenna streams first, and then inspect the autocorrelation sum afterwards because it has the advantage of ease of computation and lets you match time or frequency resolutions to the interference source.
Subrahmanyan: From a theoretical view, it confirms that for bright stationary terrestrial emitters on compact arrays, cross-correlations are better than extended sparse ones; but for moving emitters like planes or satellites, the sum of autocorrelations is superior due to fringe rotation effects.
Vera: So the practical implication is that the best path depends entirely on what you’re dealing with: the array you have, the local RFI environment, and what science case you are trying to pursue.
Jocelyn: They wrap up by emphasizing that for various RFI signals, suppression is achieved in visibility data and tiedarray beam data.
Subrahmanyan: That's a solid summary of what the paper established about these detection schemes in radio arrays.
Conclusion: Vera: So we've talked about how autocorrelation sums are more sensitive for radio interference in radio arrays, which is what this paper, "Detection and excision of radio-frequency interference in radio arrays," really gets at.
Jocelyn: It boils down to this: you condition those individual antenna streams first, and then the autocorrelation sum gives you a better signal to look at because it’s easier to compute and matches the time or frequency resolutions we need for the source.
Subrahmanyan: From a theoretical standpoint, it confirms that for bright stationary emitters on compact arrays, cross-correlations are better than extended sparse ones, but for moving emitters like planes or satellites, the sum of autocorrelations is superior because of those fringe rotation effects.
Vera: That’s what the data tells us then—the optimal path depends on the array you have and what kind of RFI you're hunting.
Jocelyn: And they show that real-time GPU processing can handle big arrays, like that eleven hundred antenna Deep Synoptic Array, by using power moments to cut down the operations from two passes to just one.
Subrahmanyan: It’s interesting that for moving RFI, the fringe tracking method actually works well in most large sparse arrays because you see significant suppression in those fringe-tracked visibilities.
Vera: So it’s about finding that balance between real-time speed and getting a clean estimate of the interference.
Jocelyn: That's what we were looking at, but this paper also flags a limitation: fringe tracking doesn't help with impulsive interference that happens on timescales shorter than one second in any ground array.
Subrahmanyan: Exactly, so if you’re dealing with something really sharp and fast, like a very sudden spike in the signal record, the fringe tracking won't give you much suppression.
Vera: So this whole study on "Detection and excision of radio-frequency interference in radio arrays" gives us a clear practical roadmap for how to clean up our data before we look at science.
Jocelyn: It really emphasizes that AI can use that hierarchical approach—flag the streams, then check the sum—because you know those per-element tests aren't robust enough on their own.
Subrahmanyan: For cosmology, this means if we’re looking for signals from moving sources like low and medium Earth orbit satellites, using the autocorrelation sum with those fringe tracking elements could be a way to isolate them from terrestrial noise.
Vera: It sounds like a solid piece of work that gives us concrete things to test on our next instrument designs.
Jocelyn: Yeah, it sets the stage for how we can get cleaner data streams for things like fast radio bursts or even just better pulsar timing in noisy environments.
Vikram Ravi
California Institute of Technology · Owens Valley Radio Observatory
astro-ph.IM
Submitted: 2026-10-31
Updated: 2026-10-31
Comments: 22 pages, 9 figures, 5 tables. To be submitted to the Open Journal of Astrophysics
Code: https://github.com/dsa110/dsa110-rt
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: The paper develops a model to determine which data product—individual antenna streams or coherent array combinations—is most detectable for suppressing radio-frequency interference (RFI) in radio…
Key concepts
- Per-baseline RFI coherence ($\gamma_{pq}$)
- This measures how much of the RFI power survives into a specific data measurement (visibility) between two antennas. It is a key factor in determining suppression; higher coherence means less suppression. This factor can be broken down into bandwidth smearing, fringe-rate time-smearing, and intrinsic spatial coherence.
- Autocorrelation Sum ($R_{auto}$)
- This estimator involves summing the autocorrelations across all antennas. The paper shows this sum is generally more sensitive to RFI than other detection schemes. It is computationally simple and offers good time and frequency resolution for matching RFI sources.
- Visibility ($V_{pq}$)
- Visibilities are samples of the spatial coherence function, calculated from channelized voltage data. They represent how the signal coherence changes across different baselines (distances between antennas). These are used in the theoretical framework to define detectability and suppression mechanisms.
- Fringe Tracking
- This technique is used to detect RFI emitters that are moving rapidly relative to the radio array, such as aircraft or satellites. It works by analyzing how fringe patterns change over time. This method provides significant suppression for moving interference in most large sparse arrays.
Terminology
Summary
The paper develops a model to determine which data product—individual antenna streams or coherent array combinations—is most detectable for suppressing radio-frequency interference (RFI) in radio arrays
The detection and excision of radio-frequency interference (RFI) is usually approached by asking how best to identify RFI within a given data product. The comparison between antenna-based and array-based detection reduces to a single quantity related to the mean per-baseline RFI coherence. p. 0 I develop a model for the suppression of RFI in visibilities and
tied-array beams by the antenna primary beam, near-field wavefront curvature, bandwidth smearing,
and fringe tracking. The comparison between antenna-based and array-based detection reduces to
a single quantity related to the mean per-baseline RFI coherence. The comparison between antenna-based and array-based detection reduces to
a single quantity related to the mean per-baseline RFI coherence. Cross-correlation detection is
preferred only above a coherence threshold of (N − 1)−1/2
, where N is the number of antennas. I
evaluate this model for four array configurations (the Very Large Array A and D configurations, and
the fixed MeerKAT and Deep Synoptic Array / DSA configurations) and five representative RFI
emitters spanning distances of 5 km to medium-Earth orbit. Compact, filled, arrays favor the crosscorrelations for stationary terrestrial emitters, whereas extended sparse arrays do not. Compact, filled, arrays favor the crosscorrelations for stationary terrestrial emitters, whereas extended sparse arrays do not. RFI emitters
with significant motion relative to radio arrays, in particular aircraft and low- and medium-Earth-orbit
satellites, favor the sum of antenna autocorrelations everywhere by one to seven orders of magnitude,
because their coherence is destroyed by rapid fringe rotation. For RFI below the single-antenna noise,
no phase-unknown cross-correlation statistic is more sensitive than the autocorrelation sum. Tiedarray beams are the least sensitive data product of all: the array factor supplies no suppression of
terrestrial RFI, so a beam sees an emitter as a single antenna does.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Theoretical Framework for RFI Detection
The theory establishes detectability in digital signal processing systems for radio arrays by defining visibilities and beams from channelized voltage data. The emitter is characterized by its bandwidth ∆νR, its coherence time tR — the timescale on which ˜a is stable, characterized by the symbol rate for a digital carrier or the pulse length for a radar — and its duty cycle δ within a data integration. p. 2 The sky term, sp is given by the far-field solid-angle
integral over the antenna voltage response (primary beam) Ap,
sp =
Z
dΩ Ap(ˆs, ν) E(ˆs, t, ν) e
−2πiν ˆs·rp/c
, (2)
where ˆs is a direction unit vector, and E is complex amplitude of the sky electric field, and the RFI term corresponds
to a single point emitter at finite distance Rp = |rR − rp| with complex envelope ˜a,
rp = Ap(ˆsR, ν)
e
2πiνRp/c
Rp
a˜(t, ν). (3)
The emitter is characterized by its bandwidth ∆νR, its coherence time tR — the timescale on which ˜a is stable,
characterized by the symbol rate for a digital carrier or the pulse length for a radar — and its duty cycle δ within a
data integration. (3)
The emitter is characterized by its bandwidth ∆νR, its coherence time tR — the timescale on which ˜a is stable,
characterized by the symbol rate for a digital carrier or the pulse length for a radar — and its duty cycle δ within a
data integration. Visibilities are samples of the spatial coherence function, Vpq = ⟨vpv
∗
q
⟩. Equation 4 holds only for far-field, spatially incoherent signals,
and is thus not applicable to the RFI signals of Eq. A deterministic emitter at finite distance contributes
V
R
pq = gpg
∗
qApA
∗
q
⟨|a˜|
2
⟩
RpRq
e
2πiν(Rp−Rq)/c
, (5)
of unit fractional coherence. That is, a single visibility is not intrinsically less sensitive to RFI than a single autocorrelation. All of the suppression derived in Section 2. All of the suppression derived in Section 2. 2 below arises from averaging Eq. 2 below arises from averaging Eq. 5 over bandwidth, time, or
baselines.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Suppression Mechanisms and Coherence Factors
The suppression of RFI is governed by several factors including the antenna primary beam, coherence factor, bandwidth smearing and near-field defocus. The per-baseline RFI coherence γpq is defined as the fraction of the RFI power surviving into the visibility. p. 3 8 becomes a lower bound and |Ap| acquires a distance dependence that differs between antennas. Coherence factor
I define the per-baseline RFI coherence
γpq ≡
|⟨rpr
∗
q
⟩|
P
(p)
R P
(q)
R
1/2
∈ [0, 1], (9)
as the fraction of the RFI power surviving into the visibility, with γpp ≡ 1 by construction. Coherence factor
I define the per-baseline RFI coherence
γpq ≡
|⟨rpr
∗
q
⟩|
P
(p)
R P
(q)
R
1/2
∈ [0, 1], (9)
as the fraction of the RFI power surviving into the visibility, with γpp ≡ 1 by construction. This can usually be
factorized as
γpq ≃ γ
bw
pq γ
fr
pq γ
sp
pq, (10)
where γ
bw is bandwidth (delay) smearing across the channel, γ
fr is fringe-rate time-smearing across the integration,
and γ
sp is the intrinsic spatial coherence of the RFI, reduced by multipath and by multiple independent emitters. 9 km for 1 s and
290 m for 10 s. Fringe tracking provides no suppression of interference that is impulsive on sub-second timescales, in
any ground-based array.Where the paper says it
Where the paper says it
Where the paper says it
Comparative Analysis of Estimators
The comparison between estimators is reduced to the per-baseline coherence γpq and the number of antennas N. The estimators of interest have significances relative to R1, such as Rauto = √N R1 for the sum of the N autocorrelations. p. 4 Every statistic scans the available time–frequency cells; those that
additionally scan baselines, uv cells or trial emitter positions incur a false-alarm-rate penalty. The estimators of interest, with their significances relative to R1, are
Rauto =
√
N R1 (19)
for the sum of the N autocorrelations, in which the emitter appears at unit coherence and linearly in power;
Rpq =
√
2 γpq R1 (20 The
advantage of a cross-correlation statistic over the autocorrelation sum is the RFI coherence multiplied the square root
of the number of elements. The crossover follows immediately:
⟨γd⟩ > (N − 1)−1/2
, (26)
which is 0. That cancellation presumes the emitter lies in the antenna far field, R ≫ 2D2
a/λ (the Fraunhofer distance), which is
smaller than the corresponding array condition by (Da/D)
2
. A usual configuration for RFI viewed by a small-element
array is to lie simultaneously in the near field of every baseline and the far field of every antenna. 13 for 64 and 0. 025 for 1650.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Practical Implementation and Real-Time Excision
The paper describes a real-time, GPUbased implementation of autocorrelation-based excision for the 110-antenna Deep Synoptic Array, and outlines the design implications for upcoming arrays. The DSA-110 searches for fast radio bursts (FRBs) in real time. p. 1 Then, in Section 4, I discuss how this can be achieved in practise, and describe a
real-time, automated implementation for the 110-antenna Deep Synoptic Array (DSA-110). The results are discussed
in Section 5, with a focus on practical considerations for upcoming radio arrays, and the conclusions are enumerated
in Section 6.Where the paper says it
Conclusion on Optimal Data Product
The autocorrelation sum is more sensitive to RFI than other commonly used RFI detection schemes. The autocorrelation sum, prefaced by excision in individual-antenna data streams to remove antennaspecific RFI, has the advantages of ease of computation, and access to time- and frequency-resolutions that can be matched to RFI sources. p. 1 I begin this paper with a summary of how radio arrays detect RFI, and the common signal transformations that are
used (Section 2). I then present an assessment of common RFI excision strategies for radio arrays, with a focus on
comparing the detection of RFI in individual-antenna data streams and their incoherent combination, and in coherent
data combinations (Section 3). These strategies, whether automated and/or real-time or neither, are typically limited by the available
compute and data throughput. Here I address the parallel question of whether RFI is more easily detectable in the
data streams of individual antennas, including autocorrelation data, or in coherent combinations of data in radio arrays
(e.Where the paper says it
Where the paper says it
Final Summary of Findings
RFI signals of various types are suppressed in visibility and tied-array beam data. In many cases, the autocorrelation sum is more sensitive to RFI than other commonly used RFI detection schemes. p. 0 RFI emitters
with significant motion relative to radio arrays, in particular aircraft and low- and medium-Earth-orbit
satellites, favor the sum of antenna autocorrelations everywhere by one to seven orders of magnitude,
because their coherence is destroyed by rapid fringe rotation. For RFI below the single-antenna noise,
no phase-unknown cross-correlation statistic is more sensitive than the autocorrelation sum. I
evaluate this model for four array configurations (the Very Large Array A and D configurations, and
the fixed MeerKAT and Deep Synoptic Array / DSA configurations) and five representative RFI
emitters spanning distances of 5 km to medium-Earth orbit. Compact, filled, arrays favor the crosscorrelations for stationary terrestrial emitters, whereas extended sparse arrays do not. These strategies, whether automated and/or real-time or neither, are typically limited by the available
compute and data throughput. Here I address the parallel question of whether RFI is more easily detectable in the
data streams of individual antennas, including autocorrelation data, or in coherent combinations of data in radio arrays
(e.Where the paper says it
Where the paper says it
Where the paper says it
Acknowledgments
I thank Greg Hellbourg and Liam Connor for comments on this manuscript. This material is based in part upon work supported by the National Science Foundation under Grants No. 2537086 and 1836018. This research is enabled in part by development funded by Schmidt Sciences for the Deep Synoptic Array project (DSA). p. 0 Cross-correlation detection is
preferred only above a coherence threshold of (N − 1)−1/2
, where N is the number of antennas. I
evaluate this model for four array configurations (the Very Large Array A and D configurations, and
the fixed MeerKAT and Deep Synoptic Array / DSA configurations) and five representative RFI
emitters spanning distances of 5 km to medium-Earth orbit. Version November 3, 2026
Preprint typeset using LATEX style openjournal v. 09/06/15
DETECTION AND EXCISION OF RADIO-FREQUENCY INTERFERENCE IN RADIO ARRAYS
Vikram Ravi 1,2
1Cahill Center for Astronomy and Astrophysics, MC 249-17, California Institute of Technology, Pasadena, CA 91125, USA
2Owens Valley Radio Observatory, California Institute of Technology, Big Pine, CA 93513, USA
Version November 3, 2026
Abstract
The detection and excision of radio-frequency interference (RFI) is usually approached by asking how
best to identify RFI within a given data product. The statistical distribution of raw voltage data from individual antennas, immediately
after digital samplers (or analog to digital converters / ADCs), is typically determined by the system noise. In cases
Corresponding author: Vikram Ravi:
vikram@caltech. The statistical distribution of raw voltage data from individual antennas, immediately
after digital samplers (or analog to digital converters / ADCs), is typically determined by the system noise. In cases
Corresponding author: Vikram Ravi:
vikram@caltech.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Real-time RFI excision for large arrays like DSA-110 can be implemented on GPUs by forming power moments once to reduce every operation in Table 5 to a reduction over resident tensors rather than a second pass over the voltages.
-
AI systems can perform hierarchical RFI mitigation by first flagging individual antenna streams and then inspecting the autocorrelation sum, as
the ordering matters because the sum is not robust to the conditions per-element tests are designed to find.
-
The improved system can detect moving RFI emitters, such as aircraft and LEO/MEO satellites, using fringe-tracked visibilities in most large sparse arrays where
significant suppression is observed in fringe-tracked visibilities for most large sparse arrays.
-
AI systems can select the optimal data product for excision by ranking them based on their detection margin, as the paper suggests,
The broader point is one of framing. Most treatments of RFI mitigation ask how best to detect interference in a given data stream, whether visibilities, beamformed intensities or voltages.
Abstract
The detection and excision of radio-frequency interference (RFI) is usually approached by asking how best to identify RFI within a given data product. I address a prior question: which data product is a given RFI emitter is most detectable in? I develop a model for the suppression of RFI in visibilities and tied-array beams by the antenna primary beam, near-field wavefront curvature, bandwidth smearing, and fringe tracking. The comparison between antenna-based and array-based detection reduces to a single quantity related to the mean per-baseline RFI coherence. Cross-correlation detection is preferred only above a coherence threshold of (N-1)-1/2, where N is the number of antennas. I evaluate this model for four array configurations (the Very Large Array A and D configurations, and the fixed MeerKAT and Deep Synoptic Array / DSA configurations) and five representative RFI emitters spanning distances of 5,km to medium-Earth orbit. Compact, filled, arrays favor the cross-correlations for stationary terrestrial emitters, whereas extended sparse arrays do not. RFI emitters with significant motion relative to radio arrays, in particular aircraft and low- and medium-Earth-orbit satellites, favor the sum of antenna autocorrelations everywhere by one to seven orders of magnitude, because their coherence is destroyed by rapid fringe rotation. For RFI below the single-antenna noise, no phase-unknown cross-correlation statistic is more sensitive than the autocorrelation sum. Tied-array beams are the least sensitive data product of all: the array factor supplies no suppression of terrestrial RFI, so a beam sees an emitter as a single antenna does. I describe a real-time, GPU-based implementation of autocorrelation-based excision for the 110-antenna Deep Synoptic Array, and outline the design implications for upcoming arrays.
Sources
- Real-time RFI Mitigation Techniques in Radio Astronomy and Their Practical Limitations
- Radio Frequency Interference
- The DSA-2000 -- A Radio Survey Camera
- Can Dynamic Spectrum Sharing Protect Passive Radio Sciences?
- Nature and Evolution of UHF and L-band Radio Frequency Interference at the MeerKAT Radio Telescope
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer
- A design concept for the High Resolution Imager on the Habitable Worlds Observatory