🔭 astrophysics

Potential periodic signals in blazars: significance, forecasting and deep learning

This study analyzes five blazars with reported quasi-periodic oscillations (QPOs) and one candidate, finding that while detrending enhances signal detection, most QPOs appear transient rather than persistent; furthermore, by combining traditional statistics with Transformer-based deep learning to forecast future light curves, the authors predict a weakening of signals in transient cases but a strengthening in the nascent candidate PKS 0139-09, offering testable insights into the physical origins of these phenomena.

Original authors: M. A. Hashad, A. Hammad, Amr A. EL-Zant

Published 2026-02-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: M. A. Hashad, A. Hammad, Amr A. EL-Zant

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, chaotic orchestra. Most of the time, the music is just random noise—stars flickering, gas swirling, and black holes eating matter in a messy, unpredictable way. But every now and then, a specific instrument starts playing a steady, rhythmic beat. In the world of astronomy, these "beats" are called Quasiperiodic Oscillations (QPOs).

This paper is a detective story about five of these cosmic drummers (blazars) and one new candidate that might be starting to play. The authors, a team of astronomers and data scientists, wanted to answer three big questions:

Is the beat real? (Or is it just a trick of the light?)
Will the band keep playing? (Is the rhythm permanent, or will it fade away?)
What will they play next? (Can we predict the future music?)

Here is the breakdown of their investigation, explained in everyday terms.

1. The Setup: Listening to the Cosmic Radio

Blazars are supermassive black holes at the centers of galaxies that shoot powerful jets of energy straight at us. They are like cosmic lighthouses that flicker wildly. The team used data from the Fermi-LAT, a space telescope that acts like a giant radio ear, listening to these flickers for over 16 years.

They focused on five blazars that previous studies claimed were playing a steady beat (a cycle of about 2 to 4 years). They also looked at a sixth one, PKS 0139-09, which seemed to be just starting to hum a tune.

2. The Problem: The "Static" in the Signal

Imagine trying to hear a steady drumbeat while someone is slowly turning up the volume on a radio, or while the room is filling with fog. That's what these astronomers faced. The blazars weren't just beating; they were also slowly getting brighter or dimmer over the years (trends), and they had "seasonal" wobbles.

The Analogy: Think of a child on a swing.

The QPO: The child swinging back and forth at a steady rhythm.
The Trend: The child slowly being pushed higher and higher, or the swing slowly slowing down.
The Noise: The wind blowing randomly.

If you just look at the child's path, it's hard to tell if the rhythm is steady because the "pushing higher" (the trend) is messing up the math.

The Solution: The team used a digital tool called STL Decomposition. Think of this as a high-tech "noise-canceling headphone" for data. It stripped away the long-term "pushing" (the trend) and the "seasonal wobbles" to leave just the pure, rhythmic swinging.

Result: Once they cleaned the data, the rhythmic beats became much clearer and louder. It turned out that the "messy" trends were hiding the true strength of the rhythm.

3. The Twist: The "One-Hit Wonder" vs. The "Steady Drummer"

After cleaning the data, they looked closely at the rhythm over time using a technique called Wavelet Analysis (which is like looking at a spectrogram of a song to see how the notes change over time).

They found a shocking difference between the stars:

The Steady Drummer (PG 1553+113): This blazar is the only one that kept the beat perfectly steady for the entire 16 years. It's like a drummer who never misses a beat, no matter how long the concert goes on.
The One-Hit Wonders (The others): Most of the other blazars were like bands that played a great song for a few years and then stopped.
- Some started strong and faded away quickly (like a firework that burns out).
- Others started fading right at the end of the observation period.
- This suggests that for most of these objects, the "beat" isn't a permanent feature of the black hole, but a temporary event.

4. The Crystal Ball: Predicting the Future

This is where the paper gets really cool. Instead of just guessing, they used two different "crystal balls" to predict what these blazars would do for the next four years.

Crystal Ball #1 (The Statistician): A traditional math model (STLForecaster) that looks at past patterns and assumes the future will look like the past.
Crystal Ball #2 (The AI): A Deep Learning Transformer. Think of this as a super-smart AI that has read every song ever written. It doesn't just look for simple patterns; it understands complex, messy relationships. It can "see" that a trend is about to saturate (stop growing) or that a sudden flare is coming, even if the past data didn't explicitly show it.

The Winner: The AI (Transformer) was much better. It successfully predicted complex behaviors that the traditional math missed. For example, it correctly guessed that the steady growth of one blazar would suddenly flatten out, while the traditional model kept guessing it would keep growing forever.

5. The Forecast: What Will Happen Next?

Based on the AI's predictions, here is the "setlist" for the next few years:

PG 1553+113: The beat will continue. It's a reliable drummer. We expect it to keep playing its 2-year rhythm.
PKS 0139-09 (The New Kid): This is the exciting one. The AI predicts that the "nascent" (baby) rhythm we saw starting will get much stronger. It's like a drummer warming up; the beat is just getting started and will become loud and clear soon.
The Others (PKS 0454-234, PKS 2155-304, OJ 014, S5 1044+71): The beat is likely to fade or become very weak. The "temporary" nature of their rhythm means they might go silent or become very quiet in the coming years.

Why Does This Matter?

Why do we care if a black hole is beating a drum?

The "Binary" Theory: The most exciting explanation for these beats is that the black hole isn't alone. It might have a partner black hole orbiting it. As they dance around each other, they might be pushing the jet of energy in a wobbly, rhythmic way.
The Verdict: If the beat stops (transience), it might mean the "dance" is over or the setup is unstable. If the beat is steady (like PG 1553+113), it supports the idea of a stable binary system.

In a nutshell: The authors cleaned up the cosmic noise, found that most "rhythmic" black holes are actually just temporary performers, but one is a steady drummer and another is a rising star. They used AI to predict that the rising star will get louder, while the others will likely fade into the background. This helps us understand how black holes dance, and perhaps even find the elusive "binary black holes" that are the holy grail of modern astronomy.

1. Problem Statement

Blazars exhibit variable emission across diverse timescales, with some light curves (LCs) showing signs of Quasi-Periodic Oscillations (QPOs). These signals are significant because they may encode physical processes regarding jet emission mechanisms or indicate the presence of Supermassive Binary Black Holes (SMBHBs). However, confirming year-long QPOs is challenging due to:

Limited Observational Baselines: The total $\gamma$ -ray observation time (Fermi-LAT) is often comparable to the purported periodic timescales.
Non-Stationarity: Signals often exhibit transience (decaying or evolving amplitudes) and are superimposed on long-term trends (red noise) and seasonal components, which can distort statistical significance.
Methodological Limitations: Traditional statistical methods often assume linearity or stationarity, failing to capture the complex, non-linear, and multi-scale variability inherent in astrophysical data.

The study aims to re-analyze five blazars with previously reported high-significance QPOs and one new candidate (PKS 0139-09) using extended data up to February 2025. It seeks to determine the persistence of these signals, quantify the impact of detrending, and forecast future behavior using both statistical and deep learning approaches.

2. Methodology

Data Collection and Preprocessing

Sample: Six blazars (PG 1553+113, PKS 0454-234, PKS 2155-304, OJ 014, S5 1044+71, and the candidate PKS 0139-09).
Dataset: Fermi-LAT $\gamma$ -ray data (100 MeV–300 GeV) from August 2008 to February 2025, binned into monthly intervals.
Detrending: The authors applied STL (Seasonal-Trend Decomposition using LOESS) to decompose the light curves into three components:
1. Trend: Long-term systematic variation.
2. Seasonal: Cyclical variations (e.g., annual patterns).
3. Remainder: Residual noise and latent patterns.
  This was done to isolate the QPO signal from confounding long-term trends.

Periodicity Analysis

Two methods were applied to both the original and detrended light curves:

Lomb-Scargle Periodogram (LSP): Used to identify periodicity in unevenly sampled data. Significance was estimated using $10^6$ simulated light curves (Emmanoulopoulos et al. method) to account for red noise.
Weighted Wavelet Z-transform (WWZ): Used to detect time-dependent frequency evolution and transience, providing a 2D time-frequency representation.

Forecasting Models

To predict future light curves (4-year horizon) and test signal persistence, two models were employed:

STLForecaster (Statistical): A decomposition-based approach that extrapolates trend and seasonal components independently and combines them with stochastic residual estimates.
Transformer-based Deep Learning:
- Architecture: A custom Transformer encoder tailored for time series.
- Input Processing: The input sequence (96 past observations) was decomposed into trend and residual components, which were processed by parallel encoder stacks to learn long-term and short-term dependencies separately.
- Mechanism: Utilized Self-Attention to capture global dependencies without the vanishing gradient issues of RNNs. It incorporated positional encodings to preserve temporal order.
- Training: Trained end-to-end using PyTorch Lightning with Adam optimization, minimizing Mean Squared Error (MSE).

3. Key Contributions

Extended Analysis: Provided the most up-to-date analysis of QPOs in these blazars, extending data by several years beyond previous studies.
Impact of Detrending: Demonstrated that removing long-term trends via STL decomposition generally increases the statistical significance of QPO signals by reducing red noise contamination, though the effect varies by source.
Deep Learning Application: Successfully applied a Transformer architecture to astrophysical time series, showing superior performance over traditional statistical methods in capturing non-linear trends and complex variability.
Signal Persistence vs. Transience: Systematically categorized the sources based on the temporal stability of their QPOs, distinguishing between persistent signals, early transients, and late-time attenuations.
Testable Predictions: Generated specific forecasts for future flaring events and signal evolution, providing a roadmap for observational verification.

4. Key Results

Periodicity and Significance

PG 1553+113: Confirmed a persistent, strong QPO with a period of $\approx 2.1$ years. Detrending increased the significance from $3.6\sigma$ to $4.5\sigma$ .
PKS 0139-09 (New Candidate): Initially showed a weak signal ( $1.6\sigma$ ). However, the signal is nascent and strengthening. Forecasts predict the significance will rise to $3.2\sigma$ in the coming years.
Transient Sources (PKS 0454-234, PKS 2155-304): These sources showed clear transience. The QPO signals decayed significantly within the observed timespan (on scales of $\lesssim 4000$ days). Even after detrending, the significance remained lower than in previous shorter-term studies.
Late-Time Attenuation (OJ 014, S5 1044+71): These sources showed QPOs that were strong earlier but weakened near the end of the observational period.

Forecasting Performance

Accuracy: The Transformer model outperformed the STLForecaster on the test set, achieving lower Mean Absolute Percentage Error (MAPE) across all sources.
Behavioral Prediction:
- The Transformer successfully predicted the saturation of the upward trend in PG 1553+113, whereas the statistical model overestimated the continued linear increase.
- It captured unexpected flaring patterns in S5 1044+71 and PKS 0139-09 that simple inspection of past data would not have predicted.

Future Predictions (Next 4 Years)

PG 1553+113: Expected to maintain a strong, persistent QPO ( $>4.8\sigma$ after detrending). Next flare predicted for April 2027.
PKS 0139-09: Expected to show a significant strengthening of the QPO signal. Predicted twin peaks in May and December 2027.
OJ 014 & S5 1044+71: Predicted to show suppressed QPO signals due to late-time transience. Specific flaring events predicted for Feb 2027 (OJ 014) and Oct 2027 (S5 1044+71).
PKS 0454-234 & PKS 2155-304: Expected to remain in a quasi-stationary state with suppressed QPOs, as the transient behavior occurred early in the timeline.

5. Significance and Implications

Physical Origins: The findings challenge simple geometric models of SMBHBs (e.g., rigid jet precession on circular orbits), which typically predict stable, non-decaying signals. The observed transience and complex trends favor richer scenarios, such as:
- Ballistic motion of plasma blobs in curved jets.
- Tidal interactions between SMBH disks.
- Complex magnetohydrodynamic (MHD) processes or Lense-Thirring precession.
Methodological Advancement: The study validates the use of STL decomposition to isolate periodic signals from red noise and demonstrates that Transformers are a powerful tool for forecasting non-stationary astrophysical data, outperforming classical decomposition methods.
Observational Strategy: The specific predictions for flaring dates and signal strengthening/weakness provide a concrete framework for future multi-wavelength monitoring campaigns to confirm the existence of SMBHBs or validate jet physics models.

In conclusion, the paper establishes that while QPOs are real phenomena in blazars, their detectability is highly sensitive to long-term trends and intrinsic transience. Only PG 1553+113 and the nascent candidate PKS 0139-09 show evidence of persistent or strengthening signals, while others exhibit transient behavior that complicates the search for stable SMBHB signatures.