Strauss Model: Gibbs Point Process

• The Strauss model is a spatial Gibbs point process defined by pairwise interaction counts using activity, repulsion, and range parameters.
• It interpolates between complete spatial randomness and inhibited configurations, with a challenging intractable normalizing constant.
• Acceptance–Rejection Stitching is used to enhance simulation efficiency by recursively subdividing the domain to manage cross-boundary interactions.

Searching arXiv for recent and foundational papers on the Strauss model/process. The Strauss model, in the sense of the Strauss point process, is a Gibbs point process on a bounded observation window $S\subset \mathbb R^n$ that modifies a Poisson reference law by penalizing close pairs of points. In the formulation studied in "Generating from the Strauss Process using stitching" (Huber, 2020), a realization $X=\{x_1,\dots,x_k\}$ is weighted according to the number of pairs within distance $r$, so that pairwise interactions enter multiplicatively through a repulsion parameter $\gamma\in[0,1]$ and an activity parameter $\lambda>0$. The model is canonical in spatial statistics because it interpolates between complete spatial randomness and inhibited configurations, while remaining analytically nontrivial due to its intractable normalizing constant. The same surname also appears in distinct literatures—notably Strauss-type semilinear wave equations (Metcalfe et al., 2016, Metcalfe et al., 2017, Dai et al., 2020, Lai et al., 2019, Sobajima et al., 2024), Strauss’s transitive network model (Escribano et al., 2022), and the Schmidt–Strauss simultaneous logit model (Honoré et al., 2022)—but these are separate objects. In the strict spatial-statistical usage, the Strauss model refers to the interacting point process of (Huber, 2020).

1. Formal definition

Let $S\subset\mathbb R^n$ be bounded, with Lebesgue measure $|S|$, and let $\mu$ denote the unit-rate Poisson reference measure on $S$. Under $\mu$, the probability of observing exactly $X=\{x_1,\dots,x_k\}$0 points is

$X=\{x_1,\dots,x_k\}$1

and, conditional on $X=\{x_1,\dots,x_k\}$2, the locations are i.i.d. $X=\{x_1,\dots,x_k\}$3 (Huber, 2020).

For a finite configuration $X=\{x_1,\dots,x_k\}$4, define

$X=\{x_1,\dots,x_k\}$5

The Strauss density relative to $X=\{x_1,\dots,x_k\}$6 is then

$X=\{x_1,\dots,x_k\}$7

with $X=\{x_1,\dots,x_k\}$8 the interaction radius, $X=\{x_1,\dots,x_k\}$9 the activity, and $r$0 the repulsion parameter (Huber, 2020). The hard-core case discussed in the experiments of (Huber, 2020) is $r$1, in which any pair within distance $r$2 is forbidden.

The defining statistic $r$3 is purely pairwise. Each close pair contributes one factor of $r$4, so when $r$5 the model suppresses local crowding. This gives the Strauss process its standard interpretation as a finite-range repulsive Gibbs process. A plausible implication is that $r$6 controls the geometric scale of inhibition, while $r$7 chiefly controls the baseline density before interaction effects are imposed.

2. Reference measure, parameters, and interpretation

The model is specified relative to a Poisson point process rather than as a normalized density on Euclidean coordinates. This is important because the Poisson law supplies both the random cardinality and the conditional i.i.d. locations, while the Strauss factor reweights those configurations according to local pair counts (Huber, 2020).

The parameter roles are given explicitly in (Huber, 2020):

• $r$8 is the activity.
• $r$9 controls the interaction range.
• $\gamma\in[0,1]$0 penalizes each close pair.
• $\gamma\in[0,1]$1 is the observation window.

When $\gamma\in[0,1]$2, the pair penalty disappears, and the process reduces to the Poisson reference tilted only by the activity term. When $\gamma\in[0,1]$3, configurations with many pairs at distance at most $\gamma\in[0,1]$4 are downweighted. In the hard-core limit $\gamma\in[0,1]$5, the support is restricted to configurations with $\gamma\in[0,1]$6 (Huber, 2020).

The central structural difficulty is that the density is unnormalized. This places the Strauss model within the standard Gibbsian regime where simulation and inference depend on local factors or conditional intensities rather than closed-form partition functions. The 2020 stitching paper is concerned specifically with perfect simulation from this law (Huber, 2020).

3. Computational difficulty and failure of naive acceptance–rejection

A basic acceptance–rejection sampler for a penalty density $\gamma\in[0,1]$7 relative to $\gamma\in[0,1]$8 proceeds by drawing $\gamma\in[0,1]$9, then $\lambda>0$0, and accepting when $\lambda>0$1 (Huber, 2020). For the Strauss process,

$\lambda>0$2

Hence the acceptance probability is

$\lambda>0$3

According to (Huber, 2020), this probability decays roughly like $\lambda>0$4 once $\lambda>0$5 grows. The paper states that even moderate $\lambda>0$6, such as $\lambda>0$7 on the unit square, drives $\lambda>0$8 to machine-zero, so pure acceptance–rejection almost never accepts. This is the core computational pathology of the model: the same pairwise penalty that defines the process also makes direct perfect simulation effectively impossible in nontrivial regimes.

The paper further notes that several other perfect simulation methods had already been developed, but these also work poorly for reasonably large values of $\lambda>0$9 (Huber, 2020). The computational challenge is therefore not merely exact sampling in principle, but exact sampling at densities where local interactions are frequent enough to make rejection-based methods degenerate.

4. Acceptance–Rejection Stitching

The principal methodological contribution of (Huber, 2020) is Acceptance–Rejection Stitching, abbreviated ARS. The basic idea is to replace a single global rejection event by a recursive divide-and-conquer construction: split the observation window into subregions, sample within subregions recursively, and only then reject according to the cross-region interactions.

If $S\subset\mathbb R^n$0 is partitioned into $S\subset\mathbb R^n$1 and $S\subset\mathbb R^n$2, the algorithm recursively generates $S\subset\mathbb R^n$3 and $S\subset\mathbb R^n$4, then accepts the union according to the penalty contributed only by cross pairs. In the Strauss case,

$S\subset\mathbb R^n$5

and if $S\subset\mathbb R^n$6, $S\subset\mathbb R^n$7, the cross-penalty is

$S\subset\mathbb R^n$8

Because intra-region penalties have already been handled by the recursive calls, the top-level acceptance check depends only on pairs straddling the partition (Huber, 2020).

The recursive pseudocode in (Huber, 2020) is:

$\mu$6

A practical acceleration stated in the paper is to stop the recursion once $S\subset\mathbb R^n$9, for example $|S|$0, and use pure acceptance–rejection on that small cell (Huber, 2020). The rationale is that once the expected number of points in a cell is sufficiently small, exact rejection sampling again becomes fast.

This decomposition is especially effective when the split is chosen so that the number of cross-boundary interacting pairs is small. That is the geometric reason the method outperforms global rejection.

5. Correctness, runtime, and implementation

The paper proves a correctness proposition: under the condition

$|S|$1

$|S|$2 terminates almost surely and its output has unnormalized density $|S|$3 with respect to $|S|$4 (Huber, 2020). This is the exact-sampling guarantee: the method is not approximate and introduces no simulation bias.

A second proposition in (Huber, 2020) gives an expected-calls bound. If $|S|$5 denotes the supremal expected number of $|S|$6-draws when the initial acceptance chance is $|S|$7, then for any $|S|$8 there is a finite constant $|S|$9 such that

$\mu$0

The paper further states that $\mu$1 grows only mildly as $\mu$2, far more slowly than $\mu$3, so the exponential blow-up in $\mu$4 has a smaller exponent than for pure acceptance–rejection (Huber, 2020).

Several implementation details are given explicitly:

Component	Recommendation	Purpose
Data representation	Store each partial point set as an array of coordinates	Manage recursive samples
Cross-pair search	Use a uniform grid or kd-tree	Reduce worst-case $\mu$5 to near-linear
Partition rule	Split the longer side of the current rectangle in half	Keep regions as square as possible
Base case	Choose $\mu$6–$\mu$7	Make pure AR on small cells fast
Randomness	Use a single stream of uniform variates	Supply each recursive call with its own uniforms and Poisson count
Stack depth	$\mu$8	Area halves at each split

The cross-pair computation can be restricted to a boundary strip of width $\mu$9, since only points near the partition boundary can contribute interacting cross pairs (Huber, 2020). This is a crucial geometric optimization, and it explains why spatial indexing substantially improves runtime in practice.

6. Empirical behavior and relation to extensions

The experiments in (Huber, 2020) are conducted on $S$0 with $S$1 and varying $S$2. The reported comparison is qualitative but sharp:

• Basic AR: CPU grows like $S$3, with large $S$4.
• Partial Rejection Sampling: polynomial up to a critical $S$5, then $S$6 with $S$7 smaller but still large.
• AR-Stitch: looks polynomial over a wide range, switching to $S$8 only for very large $S$9, with $\mu$0 (Huber, 2020).

A concrete example is given at $\mu$1: pure AR is effectively impossible, PRS is already intractable, while ARS on a modern desktop returns a perfect draw in seconds (Huber, 2020). The log-timing plots described in the paper show ARS remaining nearly flat until approximately $\mu$2, whereas AR and PRS bend upward much earlier.

The 2023 spatio-temporal hybrid Strauss-hardcore model (Raeisi et al., 2023) shows how the classical Strauss construction extends to multiple interaction scales and a hard-core exclusion in $\mu$3. There, the density multiplies several Strauss-type interaction factors,

$\mu$4

and adds an explicit hard-core indicator (Raeisi et al., 2023). This confirms that the basic Strauss mechanism—pairwise interaction counted within finite-range neighborhoods—serves as a modular building block for more elaborate Gibbs models. The paper also formulates the corresponding Papangelou conditional intensity and uses logistic likelihood for inference and a birth-death Metropolis–Hastings algorithm for simulation (Raeisi et al., 2023). This suggests that the classical Strauss model occupies the simplest nontrivial point in a broader hierarchy of finite-range interaction processes.

A distinct but conceptually related extension appears in network theory. Strauss’s model of transitive networks is an exponential random graph with triangle interactions, admitting a free-energy density functional and a first-order phase transition above a critical triangle parameter (Escribano et al., 2022). Although this graph model is not a spatial point process, the commonality is that both are Gibbsian models with local interaction counts and nontrivial normalization.

7. Scope of the name and common ambiguities

The term “Strauss model” is not unique across mathematical sciences. In PDE, “Strauss conjecture” and “Strauss-type” equations refer to critical-power phenomena for semilinear wave equations, including asymptotically flat backgrounds (Metcalfe et al., 2016), coupled systems (Metcalfe et al., 2017), lifespan problems (Dai et al., 2020), scattering damping (Lai et al., 2019), and time-dependent propagation speeds (Sobajima et al., 2024). In network science, “Strauss’s model” denotes an exponential random graph with triangle interactions (Escribano et al., 2022). In econometrics, the Schmidt–Strauss model is a simultaneous logit model for bivariate binary outcomes (Honoré et al., 2022).

For spatial statistics, however, the established meaning is the repulsive Gibbs point process defined by a close-pair count and parameters $\mu$5 (Huber, 2020). Confusing this process with the PDE or graph-theoretic usages obscures both the object and the associated computational questions. The central issues for the point-process Strauss model are exact sampling, interaction-range geometry, and the behavior of algorithms under increasing activity. Within that domain, Acceptance–Rejection Stitching provides a perfect-simulation method that substantially enlarges the practically accessible parameter regime (Huber, 2020).