⚛️ general relativity

On correlated noise in the LIGO-Virgo network: a test of the stochastic gravitational-wave background hypothesis

By analyzing nine LIGO-Virgo candidate events and finding that their inter-detector cross-correlations are statistically indistinguishable from noise in most cases, this paper argues that the cross-correlation statistic is not a robust standalone validation tool and suggests the strain data may contain a continuous correlated physical component, such as a stochastic gravitational-wave background.

Original authors: S. C. Ulhoa, B. C. C. Carneiro, F. L. Carneiro, L. V. A. Cunha

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: S. C. Ulhoa, B. C. C. Carneiro, F. L. Carneiro, L. V. A. Cunha

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 you are a detective trying to catch a ghost. The ghost is a gravitational wave—a ripple in the fabric of space-time caused by cosmic collisions, like two black holes smashing together. To catch it, scientists use giant, ultra-sensitive ears called detectors (LIGO in the US and Virgo in Italy).

The standard rule for catching a ghost is simple: If the ghost is real, it should hit the ears at slightly different times. Since the ghost travels at the speed of light, it takes a tiny fraction of a second to cross the distance between detectors. If the ears in Washington (H1) and Louisiana (L1) hear a "thump" exactly 6.9 milliseconds apart, the scientists say, "Aha! That's a real ghost!"

But what if the background noise in the room also has a pattern that makes the ears thump at the same weird time, even when no ghost is there? That's the big question this paper asks.

The "Sticky Bead" vs. The "Echo"

For decades, scientists argued about whether gravitational waves were real or just mathematical tricks. A famous argument called the "sticky bead argument" proved they carry energy. Now, with the first detection in 2015, everyone thought the case was closed. But a few skeptics (Creswell et al.) noticed something odd: the "thump" delay between the two US detectors seemed to show up in the noise too, not just during the big events.

The official response from the LIGO team was, "That's just a glitch in our math! If you clean the data properly (using a special filter called 'whitening'), the ghost delay disappears from the noise, and only the real events remain."

The New Investigation: Nine Ghosts and Three Ears

The authors of this paper decided to play detective again, but with a twist. Instead of just looking at the two US detectors, they looked at nine different candidate events and checked all three pairs of detectors:

Hanford (H1) × Livingston (L1)
Hanford (H1) × Virgo (V1)
Livingston (L1) × Virgo (V1)

They used a very simple, "no-nonsense" approach. They didn't use the fancy, complex math filters that the official team uses. They just took the raw sound data, filtered out the very low and very high pitches (keeping the 35 to 350 Hz range), and asked: "Does the noise sound like the event?"

The Big Discovery: The Noise is Copycatting the Ghosts

Here is the plot twist: In 26 out of 27 cases, the noise looked exactly like the event.

When the authors checked the time delay between the ears, the "ghost" events didn't stand out. The background noise had its own little "thumps" that happened at the same time delays as the supposed ghosts.

The Virgo Factor: This is the most important part. The Virgo detector is in Italy, built by a totally different team, with different hardware, different power lines, and different earthquakes. If the "copycat" noise was just a shared glitch between the two US detectors, Virgo shouldn't see it. But it did. The noise in the US-Italy pairs looked just as "correlated" as the US-US pairs.
The One Exception: There was one event, GW190412, where the US detectors (H1×L1) showed a strong signal that didn't look like noise. But even this one didn't show up in the US-Italy pairs. It was like a loud shout in one room that the people in the next room didn't hear.

What Does This Mean?

The paper suggests that the "inter-detector time delay" (the time difference between ears) is not a reliable way to prove a gravitational wave is real on its own. If the noise itself has a pattern that mimics the ghost's travel time, you can't use that pattern to prove you found a ghost.

The authors propose a new idea: Maybe the "noise" isn't just random static. They suggest the data might contain a faint, continuous "hum" of gravitational waves coming from everywhere at once—a stochastic gravitational-wave background. Think of it like a constant, low-level ocean wave that is always there. Sometimes, a big splash (a black hole merger) happens on top of it, but the ocean wave itself is always rippling, creating those weird correlations in the noise.

How Sure Are They?

The authors are not saying they have proven a new type of wave exists. They are saying:

The current method of checking time delays is flawed because it can't tell the difference between a real event and this weird noise pattern.
The idea that this is a "stochastic background" is a plausible candidate that fits the data, but it is still just a hypothesis.
They rule out the idea that this is just a shared computer glitch between the US detectors, because the Italian detector (Virgo) sees the same thing.

The Future Test: The Japanese Ear

To solve this mystery, the paper suggests we need a fourth ear: KAGRA, a detector in Japan. It's underground, uses freezing cold mirrors, and has no shared equipment with the US or Italy.

If the "hum" is real, KAGRA should hear it too, with the correct time delays.
If KAGRA hears nothing, then the "hum" must be something specific to the US and Italy detectors, and the mystery will need a different explanation.

For now, the paper concludes that the "ghost hunting" rules need a rewrite. The time delay trick isn't enough to separate the signal from the noise, and the universe might be whispering a continuous song that we haven't fully understood yet.

Technical Summary: On Correlated Noise in the LIGO-Virgo Network

Problem Statement
The validation of gravitational-wave (GW) detections by the LIGO-Virgo-KAGRA (LVK) collaboration relies on two primary pillars: the matched-filter signal-to-noise ratio (SNR) and the inter-detector time delay. The latter is intended to serve as a model-independent confirmation that a signal propagating at the speed of light arrives at spatially separated detectors with a lag consistent with their baseline geometry.

However, a 2017 study by Creswell et al. [1] challenged the robustness of the time-delay criterion. Analyzing GW150914, GW151226, and GW170104, they demonstrated that the cross-correlation function $C(\tau)$ between the Hanford (H1) and Livingston (L1) detectors exhibited a peak at $\tau \approx 7$ ms not only in the event window but also in surrounding noise windows and even in residuals after subtracting the best-fit waveform. The LVK response attributed these correlations to spectral leakage (due to a lack of Tukey windowing) and the use of suboptimal templates. The authors of the present paper argue that this response reinforces a deeper concern: if a physically real time delay cannot be distinguished from the noise background without invoking the theoretical waveform used to generate the signal, the criterion lacks standalone validity. Furthermore, the original Creswell analysis was limited to the H1–L1 pair, which shares design heritage and infrastructure, leaving open the possibility that the correlations are instrumental artifacts rather than physical phenomena.

Methodology
The authors extend the Creswell et al. methodology to a larger dataset comprising nine candidate events from the O2 and O3 observing runs, including those with Virgo (V1) participation. The study analyzes all three possible detector pairs: H1×L1, H1×V1, and L1×V1.

The analysis employs a strictly model-independent pipeline:

Data Source: 256-second stretches of public strain data centered on the event GPS time, retrieved from the Gravitational-Wave Open Science Center.
Pre-processing: The only filtering applied is a fourth-order Butterworth band-pass filter over [35, 350] Hz (applied forward and backward to avoid phase distortion). Crucially, the authors exclude the standard LVK steps of Tukey windowing and power-spectral-density (PSD) whitening. This ensures the test is unbiased by theoretical waveform assumptions or noise models.
Statistical Framework: For each detector pair, the normalized cross-correlation function $C(\tau)$ is computed in the event window ( $w=0.2$ s). This is compared against an empirical ensemble of $N=200$ off-source noise windows drawn from the same 256-second segment (excluding a guard zone around the event).
Metric: An empirical $p$ -value ( $p_{emp}$ ) is calculated to determine if the peak correlation in the event window is statistically distinguishable from the noise ensemble.

Key Results
The analysis yields the following statistical outcomes across 27 valid detector-pair measurements:

Dominant Negative Result: In 26 of the 27 measurements, the peak cross-correlation in the event window is statistically indistinguishable from the noise ensemble ( $p_{emp} \geq 0.12$ ). The noise windows frequently produce peaks of equal or greater amplitude than the event windows.
Lag Inconsistency: The time lags ( $\tau_{ev}$ ) at which the event peaks occur often do not align with the physically expected light-travel times. In several cases (e.g., GW170818), the peak lag falls outside the physically allowed range for the baseline.
Virgo Independence: No detector pair involving the Virgo detector (H1×V1 or L1×V1) yields a $p_{emp} < 0.20$ for any event. This suggests that the correlated structure is not an artifact of shared LIGO infrastructure.
The Exception (GW190412): The only statistically significant outlier is the H1×L1 pair for GW190412, which shows a peak at $\tau = 3.9$ ms with $p_{emp} = 0.005$ . However, this excess is not reproduced in the H1×V1 or L1×V1 pairs for the same event.
Effect of Whitening: When the authors applied standard PSD whitening to the data (specifically for GW170814 and GW190412), the apparent excess for GW190412 disappeared entirely, and all pairs returned to the noise distribution range. This implies that the correlation structure extends into the surrounding data and is suppressed by the whitening process itself.

Significance and Claims
The paper concludes that the inter-detector time delay, when operationalized as the peak of the normalized cross-correlation function, does not provide a robust, standalone separation between event and noise under the tested conditions.

The authors interpret the persistence of comparable correlation structures in both event and noise windows—across independent detectors (LIGO and Virgo)—as evidence that the strain data may contain a continuous correlated physical component. They propose that a stochastic gravitational-wave background (SGWB) is a plausible candidate for this component.

The paper argues that:

The standard validation criterion (time delay) is compromised if the noise background itself possesses a correlated time structure.
The lack of correlation in LIGO-Virgo pairs rules out shared instrumental artifacts as the sole explanation for the observed correlations in LIGO-LIGO pairs.
The behavior of the data under whitening (where the only significant excess vanishes) supports the hypothesis that the correlation is a pervasive feature of the data rather than a transient signal confined to the event window.

The authors suggest that future observations involving the KAGRA detector, which operates under entirely different environmental and instrumental conditions, will provide a decisive test. If the correlated structure is physical (e.g., an SGWB), it should appear in KAGRA data with appropriate time delays; if not, the origin of the correlations in the Western network must be sought in a mechanism common only to those three detectors.