⚛️ general relativity

Sampling biases in sprinkled causal sets

This paper identifies a geometric sampling bias in finite-volume Poisson sprinkling of causal sets that distorts link statistics in dimensions $d \geq 2$ due to boundary dominance on non-compact hyperbolic spaces, and proposes specific estimators and boundary-free alternatives to resolve this issue.

Original authors: Marián Boguñá, Dmitri Krioukov

Published 2026-10-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Marián Boguñá, Dmitri Krioukov

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 trying to understand the rules of a vast, infinite universe by looking at only a small, finite patch of it. This is a common challenge in physics, where scientists often simulate a limited region of space to learn about the whole. In the flat, familiar geometry of our everyday world, this approach usually works well: as the simulated patch grows larger, the influence of its edges fades away, and the results settle into a stable picture of the infinite bulk. However, the universe of causality—the study of which events can influence others—has a hidden geometry that defies this intuition. When researchers try to map the connections between points in spacetime using a method called "sprinkling," they discover that simply making the simulation bigger does not always fix the problem. In fact, for certain dimensions of our universe, the edge of the simulation never truly disappears, leaving a permanent mark on the data that cannot be erased by adding more points.

This is the core puzzle tackled by Marián Boguñá and Dmitri Krioukov in their recent work on causal sets. A causal set is a way of describing spacetime not as a smooth fabric, but as a collection of discrete points, where the only thing that matters is which points can reach which others. To generate these sets mathematically, physicists use a "Poisson sprinkling," which is essentially a random scattering of points across spacetime, much like raindrops falling on a roof. This method is prized because it respects the fundamental symmetry of Einstein's relativity: no matter how fast you are moving, the pattern of sprinkled points looks the same. However, because the universe is infinite, computer simulations must stop somewhere, cutting off the sprinkling at a specific time or distance. The researchers set out to understand what happens to the statistics of the connections, or "links," between these points when such a cutoff is applied. They found that the way the simulation is cut off changes the answer in a fundamental way, and that this change does not vanish even as the simulation grows to encompass the entire universe.

The heart of the issue lies in the shape of the space where these connections live. When you pick a single event in spacetime and look at all the other events that are a fixed amount of time away from it, those events do not form a flat circle or sphere. Instead, they form a shape that curves away from you, resembling a hyperbolic surface. In dimensions higher than one, this surface has a strange property: it is "boundary dominated." This means that as the surface gets larger, the amount of space near its edge grows just as fast as the space in the middle. In a normal, flat space, the edge becomes a tiny fraction of the total area as you zoom out, but in this curved geometry, the edge remains a significant portion of the whole. Because of this, the way a simulation is bounded—whether by a flat wall of time or by a curved limit—determines which connections are counted.

The researchers demonstrated that if you simply cut off the simulation at a fixed time, like drawing a horizontal line across a graph, you inadvertently bias the results. This method creates a situation where the number of connections you see depends heavily on how far you are from the center of your simulation. In dimensions two and three, this bias persists forever. No matter how many points you sprinkle or how large you make the time window, the distribution of link lengths you measure will never match the true, infinite distribution. The data will always look slightly different, skewed by the fact that your "window" into the infinite universe was shaped like a cone rather than a balanced box. The authors showed that this is not a mistake in the calculation or a lack of data; it is a geometric inevitability. The infinite nature of the rapidity space—the mathematical space describing how fast things are moving relative to each other—means that the boundary always contributes a finite, unremovable fraction of the total volume.

To solve this, the team proposed a new way of looking at the data. Instead of trying to force the infinite universe into a simple time box, they suggested using a "product window." Imagine defining your simulation not just by a time limit, but by two independent limits: one for how much time can pass, and a separate one for how far away in rapidity you are willing to look. By keeping these two limits separate and carefully scaling them as the simulation grows, the researchers found a way to isolate the true signal. They proved that if you use this specific method, the bias disappears, and the results converge to the correct, intrinsic profile of the causal set. They tested this idea with computer simulations in one, two, and three spatial dimensions. The results were clear: the standard method of cutting off time produced the wrong distribution in higher dimensions, while their new product-window method successfully recovered the correct pattern.

The findings have a profound implication for how we understand the universe through simulation. It turns out that in the study of causal sets, the question "how big is big enough?" is not enough. The answer depends entirely on the shape of the boundary you choose. Simply adding more points to a simulation does not guarantee that you are seeing the true nature of the infinite spacetime. The researchers showed that the boundary is not just a technical detail to be ignored; it is an active participant in the physics. For dimensions two and higher, the edge of the simulation leaves a permanent fingerprint on the data. This means that to truly understand the structure of spacetime, scientists must be extremely careful about how they define the limits of their models. The paper concludes that while we cannot simulate the infinite universe directly, we can still extract the correct physics if we acknowledge the geometry of our limitations and choose our boundaries with precision. The universe, it seems, is not just about what is inside the box, but about how the box itself is shaped.

Technical Summary: Sampling Biases in Sprinkled Causal Sets

Problem Statement
The paper addresses a fundamental sampling bias arising in numerical simulations of causal sets generated by Poisson sprinkling in Minkowski spacetime. While Poisson sprinkling provides a Lorentz-invariant method to discretize continuum spacetime, numerical implementations are necessarily restricted to finite regions. The authors demonstrate that for observables involving causal-set links, this finite-volume restriction generates a bias that does not vanish simply by increasing the number of sprinkled points or the simulation volume.

The root of the problem is geometric. In $(d+1)$ -dimensional Minkowski spacetime, the set of events at a fixed proper time $\tau$ from a given event forms a hyperboloid with the geometry of $d$ -dimensional hyperbolic space, $H_d$ . For dimensions $d \geq 2$ , $H_d$ is non-amenable, meaning the volume of a large region is dominated by its boundary. Consequently, the fraction of the rapidity space volume close to the boundary of a finite cutoff does not vanish in the infinite-volume limit. This leads to different limiting normalized statistics depending on the specific finite-volume regularization (exhaustion) used, distinguishing this phenomenon from standard finite-size corrections.

Methodology
The authors analyze the link statistics of a Poisson-sprinkled causal set by conditioning on a root event $e_0$ at the origin. They utilize Milne coordinates $(\tau, \chi, \Omega_{d-1})$ , where $\tau$ is proper time and $\chi$ is rapidity. The probability of a link existing between $e_0$ and an event at proper time $\tau$ is determined by the emptiness of the Alexandrov interval, governed by the volume $V_{d+1}(\tau) \propto \tau^{d+1}$ .

The study compares two distinct approaches to handling the infinite rapidity volume:

Finite-Time Cone Regularization: A standard regularization where the simulation is restricted to a future light cone truncated by a constant time plane $t < T$ . This induces a proper-time-dependent rapidity cutoff, $\chi_c(\tau) \approx \ln(2T/\tau)$ .
Event-Centered Product Window: A proposed estimator where independent cutoffs are applied to proper time ( $\tau < \tau_c$ ) and rapidity ( $\chi < \chi_c$ ) relative to a timelike reference direction.

The authors derive the expected link-intensity measures and normalized proper-time distributions for both cases. They further propose a "controlled double-scaling limit" where the rapidity cutoff $\chi_c$ and proper-time cutoff $\tau_c$ scale with the density $\rho$ and simulation size $T$ in a specific manner to recover the intrinsic intensity profile.

Key Contributions and Results

Non-Commutativity of Limits: The paper establishes that the infrared (infinite-volume) limit and the continuum (infinite-density) limit do not commute for link statistics in $d \geq 2$ . The order in which the spatial volume and density limits are taken affects the resulting normalized distribution.
Dimensional Dependence:
- For $d \geq 2$ : The rapidity space $H_d$ is non-amenable. A finite-time cone regularization yields a limiting normalized proper-time distribution $p_T(\hat{\tau}) \propto \hat{\tau} e^{-\hat{\tau}^{d+1}}$ (Eq. 19), which differs significantly from the intrinsic intensity profile $q(\hat{\tau}) \propto \hat{\tau}^d e^{-\hat{\tau}^{d+1}}$ (Eq. 11). The bias persists even as the volume goes to infinity because the boundary contribution remains extensive.
- For $d = 1$ : The rapidity space is $H_1 \cong \mathbb{R}$ , which is amenable. The boundary dependence is only logarithmic and vanishes in the infinite-volume limit, allowing the distribution to converge to the intrinsic profile, albeit slowly.
Distinction Between Intensity and Probability: The authors clarify that for an infinite sprinkling, there is no uniform probability measure on the set of links (which is infinite). Equation (11) represents the normalized intensity profile per unit hyperbolic volume, not the probability distribution of a "uniformly chosen link." The latter is ill-defined without specifying the exhaustion method.
Numerical Verification: Simulations in $d=1, 2, 3$ $d = 1, 2, 3$ dimensions confirm the theoretical predictions.
- Simulations in a finite constant-time cone match the derived finite-cone distribution (Eq. 17) and deviate from the intrinsic profile (Eq. 11) for $d \geq 2$ .
- Simulations using the event-centered product window successfully recover the intrinsic intensity profile (Eq. 11) when the cutoffs are sufficiently separated and the double-scaling condition is met.
Controlled Estimator: The paper proposes an event-centered product-window estimator with a double-scaling limit (Eqs. 27–34). By ensuring the rapidity cutoff grows while the proper-time window remains large relative to the characteristic link scale, this method provides a controlled way to estimate the intrinsic link-intensity profile from finite simulations.

Significance
The paper argues that the sampling bias described is intrinsic to the observable (link statistics) in non-amenable geometries rather than a conventional finite-size correction. Simply increasing the number of sprinkled points or the simulation box size is insufficient to recover the bulk properties of the infinite spacetime if the regularization method (e.g., a finite time cone) induces a proper-time-dependent rapidity cutoff.

The authors conclude that for numerical studies of causal sets in dimensions $d \geq 2$ , the finite-volume regularization must be explicitly stated as part of the estimator. They propose the event-centered product window as a robust numerical strategy to isolate the intrinsic Lorentz-invariant intensity profile. Additionally, they suggest that compact hyperbolic quotients could serve as a boundary-free alternative for studying bulk behavior, analogous to periodic boundary conditions in Euclidean systems, though this requires a more complex construction involving non-trivial global topology.

The work highlights that in boundary-dominated geometries, the definition of the observable, the regularization scheme, and the order of limits are inextricably linked, and care must be taken to avoid misinterpreting boundary artifacts as physical properties of the causal set.

Technical Summary: Sampling Biases in Sprinkled Causal Sets

More like this