📊 statistics

Apparent criticality from common-drive aliasing in self-exciting networks

This paper demonstrates that self-exciting networks driven by a shared nonstationary background can appear critical due to statistical aliasing of the external drive into the interaction matrix, a mechanism that causes subcritical systems to be misdiagnosed as critical unless the common drive is explicitly corrected for.

Original authors: Mauricio Herrera-Marín

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mauricio Herrera-Marín

Original paper licensed under CC BY 4.0 (https://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 watching a massive, chaotic game of "telephone" in a crowded stadium. One person whispers a rumor, and suddenly, thousands of people are shouting it back and forth. You might look at the noise and think, "Wow, this rumor is so contagious! It's spreading on its own, like a virus that never dies out!" In the world of complex systems, scientists call this state "criticality"—a tipping point where a small spark can ignite an endless fire.

But what if the stadium isn't just echoing a rumor? What if, right before the game started, a giant, invisible speaker system blasted a siren that made everyone jump and shout at the exact same time?

This is the surprising discovery in Mauricio Herrera-Marín's new research. The paper argues that we often mistake a shared external siren for internal contagion. We think a network is super-powerful and self-sustaining, when in reality, it's just a sub-critical (weak) network that got a massive, temporary push from a common outside force.

The "Ghost" in the Machine

Let's break down the science with a simple analogy. Imagine a forest where trees can knock each other over.

The Real Tree Network (Subcritical): In a healthy, subcritical forest, if one tree falls, it might knock over one or two neighbors, but the chain reaction stops quickly. The forest is safe.
The Common Drive (The Wind): Now, imagine a massive, sudden gust of wind blows through the whole forest at once. Every tree sways and knocks into its neighbors simultaneously.
The Mistake: If you are a scientist watching the forest but you don't see the wind (because you forgot to measure it), you might look at the data and say, "Wow! The trees are knocking each other over so efficiently that the forest is in a 'critical' state! It looks like the trees are so connected that a single fall could destroy the whole forest!"

The paper shows that this "critical" state is an illusion. The trees aren't actually that connected; they were just all pushed by the same wind.

The Math Magic: The "Perron" Shift

The authors use some fancy math to prove this. They look at a "reproduction matrix," which is basically a scorecard of how much one part of a system triggers another.

The Real Score: The true score is low (subcritical).
The Fake Score: When you ignore the wind (the "common drive"), the math accidentally folds that wind into the scorecard. It adds a hidden "ghost" number to the matrix.
The Result: This ghost number pushes the score up. If the wind is strong enough, the fake score crosses the "danger line" of 1.0, making a safe forest look like a ticking time bomb.

The paper proves that this isn't just random noise. It happens because the "wind" pushes in the exact same direction as the forest's most sensitive path. They call this the Perron-mode shift. It's like pushing a swing: if you push it exactly when it's swinging toward you, it goes much higher. If you push it at the wrong time, it doesn't matter how hard you push. The "wind" in these networks pushes exactly when the network is most sensitive, making the whole system look critical.

Real-World Evidence: Earthquakes and Tweets

The authors didn't just dream this up; they tested it with real data.

1. The Earthquake Test
They looked at six different earthquake aftershock catalogs (like Maule, Iquique, and L'Aquila).

The Naive View: When they analyzed these quakes assuming the background activity was constant (ignoring the fact that big quakes naturally cause a wave of smaller ones that fades over time), three of the six catalogs looked supercritical. The math said the aftershocks were feeding each other endlessly.
The Corrected View: When they accounted for the "Omori field"—the natural, shared decay of earthquake productivity after a big event—the numbers dropped. Suddenly, all six catalogs were safely subcritical. The "endless chain reaction" was just the natural fading of the big quake's energy, not a magical self-sustaining network.
The Proof: They ran simulations and "bootstrap" tests (re-sampling the data thousands of times). The "supercritical" result was fragile; it disappeared as soon as they fixed the background. The "subcritical" result was rock-solid.

2. The Twitter Test (The Higgs Boson)
They also looked at Twitter activity around the announcement of the Higgs boson discovery in 2012.

The Setup: People were mentioning, replying to, and retweeting about the discovery.
The Naive View: A simple model that ignored the big news announcement gave a "criticality" score of 0.982. That's dangerously close to the critical limit of 1.0. It looked like the conversation was about to explode on its own.
The Corrected View: When they added a "flexible" model that recognized the CERN announcement as a massive, shared external event, the score plummeted to 0.156.
The Takeaway: The conversation wasn't a self-sustaining viral loop; it was just everyone reacting to the same big news at the same time. The "viral" nature was an illusion created by the shared siren.

What This Means for You

The paper is very clear about what it does not say. It doesn't say that all viral networks are fake. It doesn't say that earthquakes are never dangerous. It specifically rules out the idea that a high "reproduction number" (a score above 1) is always proof of a physical, self-sustaining explosion.

Instead, it offers a new rule of thumb: Before you declare a system critical, check for the wind.

If you see a network acting like it's on fire, ask: "Is there a shared external force pushing everyone at once?" If there is, and you ignore it, you might be looking at a "false-critical" wedge—a zone where a safe system looks like a disaster just because of a missing variable.

The authors show that in controlled simulations, you can take a perfectly safe, subcritical network and, simply by adding a stronger "common drive," make it look like it's crossing the critical threshold. It's a reminder that in the complex world of data, what looks like a runaway train might just be a train on a track that someone forgot to mention.

Technical Summary: Apparent Criticality from Common-Drive Aliasing in Self-Exciting Networks

Problem Statement
In the statistical physics of cascade systems, criticality is often diagnosed via a scalar control parameter: the effective branching or reproduction number. In self-exciting networks modeled by multivariate Hawkes processes, this parameter corresponds to the Perron–Frobenius radius ( $\rho$ ) of the reproduction matrix. If $\rho < 1$ , the system is subcritical (activity dies out); if $\rho > 1$ , it is supercritical (cascades propagate indefinitely).

The central problem addressed is that empirical cascades are rarely isolated. They are frequently driven by shared, nonstationary external fields (e.g., Omori-law aftershock productivity, regime changes in finance, or public announcements in social media). When such a common drive is omitted from the model and replaced by a constant baseline, the external field is not removed but instead aliased into the estimated interaction matrix. This paper investigates whether this omission can cause a physically subcritical network to appear critical, thereby leading to false diagnoses of phase transitions.

Methodology
The paper employs a combination of spectral perturbation theory, statistical physics analogies, and empirical auditing across synthetic and real-world datasets.

Theoretical Framework:
- Model: The authors consider a multivariate counting process where the conditional intensity depends on a background $\mu(t)$ and a history of past events. The true branching matrix is $B_0$ .
- Aliasing Mechanism: When the true background is a low-rank, nonstationary field $\mu(t) = Cf(t)$ (where $C$ is a loading matrix and $f(t)$ is a shared temporal signal) but the fitted model assumes a constant baseline, the estimated reproduction matrix becomes $\hat{B} = B_0 + \Delta$ .
- Spectral Shift: Using first-order perturbation theory, the paper derives that the shift in the apparent critical parameter is governed by the projection of the aliasing term $\Delta$ onto the dominant eigenmode (Perron mode) of the true matrix $B_0$ . Specifically, for a small perturbation, $\rho(\hat{B}) - \rho(B_0) \approx \frac{u_0^\top \Delta v_0}{u_0^\top v_0}$ , where $u_0$ and $v_0$ are the left and right Perron vectors.
- Rank Constraints: Theorem 1 establishes that the aliasing term $\Delta$ has a rank bounded by the rank of the common drive. Theorem 3 defines a "rank-one deflation interval," providing a certified range $[\rho_{lo}, \rho_{up}]$ for the true branching radius given an observed $\hat{B}$ and a fixed loading direction.
Controlled Simulations:
- The authors simulate subcritical branching networks ( $\rho(B_0) < 1$ ) driven by a shared nonstationary envelope.
- They compare constant-baseline fits against oracle-corrected fits to demonstrate the emergence of a "false-critical wedge": a region where increasing the amplitude of the common drive pushes the estimated $\rho(\hat{B})$ above 1, despite the underlying dynamics remaining subcritical.
Empirical Audits:
- Aftershock Catalogs: Six spatially resolved earthquake catalogs (Maule, Iquique, L'Aquila, Canterbury, Tocopilla, Illapel) are analyzed. The authors partition events into spatial zones and fit multivariate Hawkes networks. They compare naive constant-baseline fits against space-time declustering methods that explicitly account for the Omori-type shared decay.
- Higgs Twitter Stream: A dataset of mention, reply, and retweet events surrounding the 2012 Higgs boson announcement is analyzed. The announcement acts as a known global external field. The authors compare a constant-background fit against a flexible time-background regression with smooth time-basis knots.

Key Results

Theoretical: The paper proves that a low-dimensional external drive, when omitted, acts as a low-rank perturbation to the interaction matrix. The apparent criticality is determined not by the entrywise positivity of the bias, but by the alignment of this bias with the Perron mode of the network. Even if the bias contains negative entries, a positive projection onto the dominant mode is sufficient to inflate the spectral radius.
Simulations: Controlled experiments confirm the existence of a false-critical wedge. As the amplitude of the common drive increases, a fixed subcritical matrix is inferred as supercritical. Monte Carlo simulations show that naive estimates pile up near or above the critical threshold ( $\rho=1$ ), whereas oracle-corrected estimates remain centered on the true subcritical value.
Aftershock Analysis: In three of six catalogs (Iquique, L'Aquila, Tocopilla), the naive constant-background network radius exceeds 1, suggesting supercriticality. However, after applying space-time declustering to remove the shared Omori field, the corrected network radius falls below 1 in all cases. The naive supercriticality is shown to be sensitive to spatial partitioning and magnitude thresholds, whereas the subcritical certificate remains robust.
Higgs Twitter Analysis: The naive fit yields a near-critical radius ( $\rho_{naive} = 0.982$ ). The flexible common-drive fit reduces this significantly ( $\rho_{flex} = 0.156$ ). The difference matrix is essentially rank-one and exhibits a strong positive Perron projection (0.826), matching the observed spectral deflation. This confirms that the near-critical signal was manufactured by the shared burst of activity following the CERN announcement.

Significance and Claims
The paper claims to identify a general mechanism—common-drive aliasing—by which subcritical self-exciting networks can appear critical. The primary significance lies in reframing the interpretation of high branching ratios in complex systems:

Diagnostic Caution: A reproduction radius $\rho > 1$ is not sufficient evidence of a physical transition to supercritical branching if the system is observed under a shared nonstationary background.
Spectral Mechanism: The shift in the control parameter is a spectral phenomenon driven by the projection of the omitted field onto the network's dominant eigenmode, rather than a simple additive bias.
Methodological Workflow: The authors propose a diagnostic workflow: estimate the apparent operator, test for low-rank common-drive structure in the residuals, compute the Perron projection of the removed component, and report a deflation interval or structural correction rather than accepting the naive $\rho > 1$ as a critical transition.

The paper concludes that in domains with known external fields (like seismology), structural declustering can recover the true subcritical nature of the branching process. In domains with partial knowledge, the rank-one deflation interval serves to identify the "false-critical" ambiguity, preventing the misinterpretation of external forcing as endogenous criticality.

The "Ghost" in the Machine

The Math Magic: The "Perron" Shift

Real-World Evidence: Earthquakes and Tweets

What This Means for You

More like this