Interference Meets Inference: Bayesian Time-Series Modeling of Radio-Frequency Interference

arXiv:2609.26904 · astro-ph.IM · Submitted 2026-09-22 · Read on arXiv

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: "Interference Meets Inference".

Jocelyn: Here's the gist of the paper. Radio-frequency interference (RFI) detection is cast as a one-dimensional time-series anomaly-detection problem and a probabilistic mixture-model framework is developed for separating a smoothly varying astronomical background from anomalous contamination.

Vera: This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events.

Jocelyn: The framework is applied to observations from the Murchison Widefield Array collected in 2014, producing “soft” classification labels and seasonal RFI trends.

Vera: Modeling the Data Components The modeling framework expects input data in the form of a one-dimensional time series.

Title and authors: Vera: So, looking at the title of "Interference Meets Inference: Bayesian Time-Series Modeling of Radio-Frequency Interference," it really captures the core idea that we need inference to handle this kind of problem because standard methods aren't enough.

Jocelyn: It suggests a shift from just detecting spikes to building a model that describes the whole time series, separating what's natural variation from what's interference.

Subrahmanyan: From a theoretical side, modeling the data as three distinct components—smooth background, noise, and RFI contamination—is a solid way to handle this kind of superposition in astronomical observations.

Vera: Yeah, they explicitly model the smooth background component on longer timescales than the anomalous RFI events, which is how they try to tell them apart.

Jocelyn: And for the RFI part, they use a positively skewed Cauchy distribution because it accounts for that expected concentration of stronger signals while still allowing for those rare, very bright outliers.

Subrahmanyan: That choice of distribution helps capture both the typical and the extreme cases of interference in one mathematical form, which is important when you're trying to characterize the population parameters.

Vera: The core idea is defining a residual time series, which they call z sub n, by subtracting their smooth background model from the original observation at night n.

Jocelyn: And then they model this residual as a mixture: either it comes from a zero-centered normal distribution representing the clean data, or it comes from that skewed Cauchy distribution representing RFI contamination.

The paper's summary: Vera: To summarize what the paper does, they introduce a two-component mixture model for each residual time series at night n, where z sub n is either clean or contaminated by RFI.

Jocelyn: This mixture model lets them calculate the posterior probability of being in either class, which is that classification confidence we talked about earlier.

Subrahmanyan: The framework defines the likelihood across all nights and time steps as the product of these individual component probabilities, which is how they build their statistical inference.

Vera: Then they use Hamiltonian Monte Carlo with NUTS sampling to explore the joint posterior distribution, which gives them samples from that model.

Jocelyn: The results show that this probabilistic approach lets them fit parameters independently for every combination of pointing, frequency sub-band, and polarization used in their analysis.

Subrahmanyan: That means they can infer different contamination characteristics for each specific observational setup they test, which is a big step toward understanding how RFI varies across the sky.

Vera: They also show that this method performs better when compared directly against an established flagging approach, specifically the SkySubtracted Incoherent Noise Spectra or SSINS data products from MWA observations.

The paper's improvements: Jocelyn: One of the major things they suggest as an improvement is moving away from hard categorical flags and adopting this posterior probability estimation for every time sample.

Vera: That’s a big deal because it directly addresses that uncertainty, allowing us to quantify exactly how confident we are in any individual detection.

Subrahmanyan: This probabilistic formulation allows for the propagation of uncertainty into any downstream statistics you calculate from the data, which is vital for investigating those borderline cases where the signal isn't clearly clean or contaminated.

Jocelyn: They also point out that fitting this mixture model independently for fifty different combinations of pointing and frequency sub-bands provides a lot of parameter inference about the RFI population itself.

Vera: But they do have some limitations they mention, like the current model only uses a two-component mixture: clean versus RFI. They suggest that including multiple RFI components could give a more realistic picture of the data if there are different types of contamination present.

Jocelyn: And another limitation is that it treats each time sample independently and doesn't explicitly model how RFI events might persist over time, which is something they think we should address in future work.

Conclusion: Vera: So, to wrap up the paper "Interference Meets Inference: Bayesian Time-Series Modeling of Radio-Frequency Interference," the main takeaway is that this probabilistic framework lets us perform RFI detection and characterization jointly.

Jocelyn: It moves beyond just flagging by giving us a measure of classification confidence for each sample, and it allows us to infer population parameters about the clean background and the contamination itself.

Subrahmanyan: This allows for broader statistical studies of the contamination, which is important for understanding how RFI evolves across different observational parameters like pointing or frequency.

Vera: So we get those soft classifications that show confidence, and we can use those inferred population parameters to study the RFI environment more broadly.

Jocelyn: It's a powerful tool for moving from simple detection to a full probabilistic understanding of the interference in radio astronomy data.

Subrahmanyan: Indeed, this paper provides a concrete methodology that applies to other telescopes beyond just MWA, offering a way to compare different contamination environments systematically.

Department of Physics, Brown University · Department of Physics and McGill Space Institute, McGill University

astro-ph.IM

Submitted: 2026-09-22

Updated: 2026-11-06

Comments: 21 pages, 12 figures, 2 tables. v2: replaced to match accepted ApJ version. Minor clarifying comments added, with no change to results

Code: https://github.com/mwilensky768/SSINS

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: Radio-frequency interference (RFI) detection is cast as a one-dimensional time-series anomaly-detection problem and a probabilistic mixture-model framework is developed for separating a smoothly…

Key concepts

Probabilistic Mixture Model
This is the core mathematical structure used to describe the data. It assumes that any observed signal at a specific time is a mixture of two underlying processes: a smooth, clean background and an anomalous RFI component. The model calculates the probability of observing any given data point belonging to either one or the other.
Posterior Responsibility Equation (Rt)
This equation quantifies the classification confidence for any single data sample. It tells you the probability that a specific measurement belongs to the RFI component, based on all the parameters learned from the entire dataset. This is a 'soft' label rather than a simple yes/no flag.
Skewed Cauchy Distribution
This distribution is used to model RFI contamination because it captures how RFI amplitudes are typically concentrated while also allowing for rare, very bright, and unusual events (heavy tails). It provides a flexible way to describe the non-Gaussian nature of the anomalous radio signals.
Uncertainty Propagation
This refers to tracking how errors or uncertainties in the model's parameters affect the final calculated results. By using this framework, researchers can propagate uncertainty from individual data points through to larger statistical summaries, which is crucial for understanding ambiguous detections.

Summary

Radio-frequency interference (RFI) detection is cast as a one-dimensional time-series anomaly-detection problem and a probabilistic mixture-model framework is developed for separating a smoothly varying astronomical background from anomalous contamination. p. 0

Where the paper says it

Wilensky , 2 , ∗ and Adrian Liu 2 1Department of Physics, Brown University, Providence, RI, USA 2Department of Physics and McGill Space Institute, McGill University, 3600 University Street, Montreal, QC H3A 2T8, Canada ABSTRACT Radio-frequency interference (RFI) remains a major challenge for modern radio astronomy experiments. In this work, we cast RFI detection as a one-dimensional time-series anomaly-detection problem and develop a probabilistic mixture-model framework for separating a smoothly varying astronomical background from anomalous contamination.

This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events. p. 0
Where the paper says it

The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags. This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events.

The framework is applied to observations from the Murchison Widefield Array collected in 2014, producing “soft” classification labels and seasonal RFI trends. p. 0
Where the paper says it

This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events. We apply the framework to observations from the Murchison Widefield Array collected in 2014, producing “soft” classification labels and seasonal RFI trends.

Modeling the Data Components

The modeling framework expects input data in the form of a one-dimensional time series. The input data is modeled as comprising three main components: a smoothly varying clean background component, noise, and an anomalous component corresponding to RFI. p. 1

Where the paper says it

We therefore introduce our modeling framework as a tool generally suitable for time-series anomaly detection. We model the input data as comprising three main components: a smoothly varying clean background component, noise, and an anomalous component corresponding to RFI.

The components are modeled as follows:

  1. Smooth background: This is assumed to be smoothly varying on longer timescales with smaller amplitudes than anomalous events. p. 2

    Where the paper says it

    Smooth background In our model, we assume a smoothly varying background. Variations in the background are taken to occur on longer timescales and have smaller amplitudes than anomalous events.

    It is modeled using a Legendre polynomial basis where the background is modeled on a per-night basis. p. 2
    Where the paper says it

    We choose to model it using a Legendre polynomial basis. The background is modeled on a per-night basis: b n = β n 0L0 + β n 1L1 + β n 2L2 + .

  2. Noise estimation: The noise properties are estimated from theoretical principles based on how the data product is formed, using a modified form of an equation from M. p. 2

    Where the paper says it

    That is, we center each nightly ℓ = 0 coefficient prior mean on the median data value for that night, and give it flexibility by setting σ n β0 to five times the standard deviation of data on that night. Estimating the noise The noise properties of the data for our use case can be estimated from theoretical principles based on how the data product is formed.

    J. Wilensky et al. p. 0
    Where the paper says it

    Ducharme , 1 Andrei Li , 2 Michael J. Wilensky , 2 , ∗ and Adrian Liu 2 1Department of Physics, Brown University, Providence, RI, USA 2Department of Physics and McGill Space Institute, McGill University, 3600 University Street, Montreal, QC H3A 2T8, Canada ABSTRACT Radio-frequency interference (RFI) remains a major challenge for modern radio astronomy experiments.

    (2019).

  3. Anomalies due to RFI: RFI is modeled phenomenologically with a positively skewed Cauchy distribution, which captures the expected concentration of RFI amplitudes while allowing for a heavy tail of rare, exceptionally bright events. p. 3

    Where the paper says it

    We do not attempt to derive the resulting RFI distribution explicitly. Instead, we model the RFI component phenomenologically with a positively skewed Cauchy distribution, which captures the expected concentration of RFI amplitudes while allowing for a heavy tail of rare, exceptionally bright events.

The Probabilistic Mixture Model

The core probabilistic model defines the residual time series for night n as z n ≡ y n − b n. The clean residual is modeled using a zero-centered normal distribution, while RFI-contaminated residuals are modeled using a skewed Cauchy distribution. p. 3

Where the paper says it

Here, θ represents the set of all inferred model parameters to be explicitly defined at the end of this section. As discussed in the previous section, RFI-contaminated residuals are modeled using a skewed Cauchy distribution.

The two-component mixture model for each residual sample z n t is given by p(z n tθ) = wclean pclean(z n tθ) + wRFI pRFI(z n tθ). p. 3

Where the paper says it

This two-component mixture model of the residuals allows the class assignment to remain latent while the model parameters are jointly inferred. Specifically, for each residual sample z n t , p(z n t |θ) = wclean pclean(z n t |θ) + wRFI pRFI(z n t |θ), (10) where wclean and wRFI are the mixture weights satisfying wclean + wRFI = 1.

The likelihood across all N nights with Tn time steps each is L(θ) = Y N n=1 Y Tn t=1 p (z n t θ). p. 4
Where the paper says it

Note the difference in scale between both vertical axes. Assuming that the residual samples are conditionally independent given the model parameters5 , the likelihood across all N nights with Tn time steps each is L(θ) = Y N n=1 Y Tn t=1 p (z n t | θ), (12) where θ = {β n ℓ , µβℓ , σβℓ , µRFI, σRFI, wRFI} denotes the full set of model parameters.

Classification and Inference

The classification scheme adopts the posterior responsibility equation Rt = p(RFI z n t, wRFI) = wRFI pRFI(z n t) + wclean pclean(z n t). This Rt quantifies the probability that a given sample belongs to the RFI component under the inferred mixture model. p. 5

Where the paper says it

432): Rt = p(RFI | z n t , wRFI) = wRFI pRFI(z n t ) wclean pclean(z n t ) + wRFI pRFI(z n t ) . (17) The responsibility Rt ∈ [0, 1] quantifies the probability that a given sample belongs to the RFI component under the inferred mixture model.

The framework is implemented using numpyro8 with a Hamiltonian Monte Carlo using the NUTS sampling strategy to produce samples from the joint posterior distribution. p. 8

Where the paper says it

Phan et al. We run a Hamiltonian Monte Carlo using the NUTS sampling strategy (M.

The mixture model is fit independently for each combination of pointing, frequency sub-band, and polarization, resulting in 50 independent fits in our implementation. p. 9
Where the paper says it

Compute requirements The mixture model is fit independently for each combination of pointing, frequency sub-band, and polarization, resulting in 50 independent fits in our implementation. As a representative example of the computational requirements, a single fit with 48 NUTS chains (6000 warmup and 3000 sampling iterations each) run on 12 CPU cores of a dual-socket AMD EPYC 9684X node reaches a peak memory usage of approximately 9 GB of RAM.

Validation and Findings

The analysis reveals that the XX polarization to be more heavily contaminated than the YY polarization. The TV7 frequency band (181–188 MHz) is more consistently contaminated than the other frequency sub-bands considered here. p. 7

Where the paper says it

1 – 174 TV6 174 – 181 TV7 181 – 188 TV8 188 – 195 TV9 195 – 197. Frequency sub-bands corresponding to Australian digital television allocations (Australian Communications and Media Authority 2005).

Pointing 1 is also found to be more highly contaminated than the other pointings. p. 8
Where the paper says it

In Figure 4, best-fit coefficients for a single pointing and night are compared to the posterior coefficient uncertainties for N = 40 terms. Coefficient value is found to decrease with term number, while parameter variance σ 2 p increases.

The principal contribution of this work is not simply an alternative flagging algorithm, but a probabilistic framework in which RFI detection and characterization are performed jointly. p. 0

Where the paper says it

In this work, we cast RFI detection as a one-dimensional time-series anomaly-detection problem and develop a probabilistic mixture-model framework for separating a smoothly varying astronomical background from anomalous contamination. The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags.

p. 0
Where the paper says it

In this work, we cast RFI detection as a one-dimensional time-series anomaly-detection problem and develop a probabilistic mixture-model framework for separating a smoothly varying astronomical background from anomalous contamination. The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags.

The resulting soft classifications quantify the confidence associated with individual detections, while the inferred population parameters enable broader statistical studies of the contamination itself.

Limitations and Future Work

A second limitation arises from the use of a twocomponent mixture model consisting of a clean component and a single RFI component. A richer mixture model containing multiple RFI components could therefore provide a more realistic representation of the data and potentially distinguish between qualitatively different classes of contamination. p. 8

Where the paper says it

In this limit, the clean and RFI components become difficult for the model to distinguish. The corresponding latent-state assignments therefore become weakly identified, producing a poorly constrained region of the posterior in which a chain may remain for an extended period and mix poorly with the others.

The present framework treats the latent state of each time sample independently and therefore does not explicitly model the temporal persistence of RFI events. p. 0
Where the paper says it

The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags. This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events.

A natural extension would be to introduce temporal dependence; for example, through a hidden Markov model, in which the probability of a sample being assigned to a clean or contaminated state depends on the state at neighboring time steps. p. 0
Where the paper says it

In this work, we cast RFI detection as a one-dimensional time-series anomaly-detection problem and develop a probabilistic mixture-model framework for separating a smoothly varying astronomical background from anomalous contamination. The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags.

Conclusion

In this work, we use interferometric observations from the Murchison Widefield Array as a case study and reformulate RFI identification and characterization as a general anomaly detection problem. The principal contribution of this work is not simply an alternative flagging algorithm, but a probabilistic framework in which RFI detection and characterization are performed jointly. The resulting soft classifications quantify the confidence associated with individual detections, while the inferred population parameters enable broader statistical studies of the contamination itself.

Acknowledgments

This work makes use of data obtained from Inyarrimanha Ilgari Bundara, the Murchison Radio-astronomy Observatory. This research was enabled in part by support provided by the Center for Computation and Visualization at Brown University (https://ccv. p. 0

Where the paper says it

Ducharme , 1 Andrei Li , 2 Michael J. Wilensky , 2 , ∗ and Adrian Liu 2 1Department of Physics, Brown University, Providence, RI, USA 2Department of Physics and McGill Space Institute, McGill University, 3600 University Street, Montreal, QC H3A 2T8, Canada ABSTRACT Radio-frequency interference (RFI) remains a major challenge for modern radio astronomy experiments.

brown.edu/), Calcul Québec (https://www. p. 0
Where the paper says it

These aim to identify contaminated Email: jade ducharme@brown. edu ∗ CITA National Fellow TSI Post-Doctoral Fellow samples, which are then excluded from subsequent processing and analysis.

calculquebec.ca/) and the Digital Research Alliance of Canada (https://www. p. 0
Where the paper says it

Ducharme , 1 Andrei Li , 2 Michael J. Wilensky , 2 , ∗ and Adrian Liu 2 1Department of Physics, Brown University, Providence, RI, USA 2Department of Physics and McGill Space Institute, McGill University, 3600 University Street, Montreal, QC H3A 2T8, Canada ABSTRACT Radio-frequency interference (RFI) remains a major challenge for modern radio astronomy experiments.

alliancecan.ca). This research made use of open-source scientific software, including NumPy, SciPy, Matplotlib, JAX, NumPyro, and ArviZ.

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

  1. Bayesian Time-Series Anomaly Detection Framework: The system can move beyond binary flagging to estimate posterior probability of belonging to either class, providing a measure of classification confidence rather than relying solely on binary flags.

  2. Uncertainty Propagation for Downstream Statistics: The framework enables the propagation of uncertainty into derived statistics, which is valuable for investigating ambiguous events.

  3. Joint Modeling with Smooth Background and Noise: The system can jointly model the three components—a smoothly varying clean background component, noise, and an anomalous component corresponding to RFI—to prevent RFI from biasing the model of the clean background.

  4. Probabilistic Classification Output: The system provides a posterior responsibility equation to quantify the probability that a given sample belongs to the RFI component (Rt), retaining this as a soft classification label that preserves information about classification confidence.

  5. Parameter Inference Across Multiple Dimensions: The system can perform inference independently for all 50 combinations of pointing, polarization, and frequency sub-band while jointly inferring parameters like the location and scale parameters of the skewed Cauchy distribution.

  6. Cross-Validation Using Established Pipelines: The system can compare its probabilistic classifications against established methods like the SkySubtracted Incoherent Noise Spectra (SSINS) pipeline to validate results on both clear outliers and samples near the boundary between clean and contaminated data.

  7. Systematic Identification of Temporal RFI Evolution: By comparing inferences from random splits versus chronological divisions, the system can reveal genuine temporal evolution in the RFI environment, suggesting a need for time-dependent occupancy parameters.

Abstract

Radio-frequency interference (RFI) remains a major challenge for modern radio astronomy experiments. In this work, we cast RFI detection as a one-dimensional time-series anomaly-detection problem and develop a probabilistic mixture-model framework for separating a smoothly varying astronomical background from anomalous contamination. The model jointly describes the clean and contaminated components and assigns each time sample a posterior probability of belonging to the RFI state, rather than relying solely on binary flags. This probabilistic formulation provides a measure of classification confidence, enables uncertainty propagation into derived downstream statistics, and offers additional information for investigating ambiguous events. We apply the framework to observations from the Murchison Widefield Array collected in 2014, producing "soft" classification labels and seasonal RFI trends. We perform a parallel analysis using the Sky-Subtracted Incoherent Noise Spectra software pipeline (SSINS), which produces "hard" classification labels. Overall, the mixture model provides similar classification labels but offers a more comprehensive framework by providing a complementary probabilistic description of RFI contamination and its uncertainty.

Sources

Related papers