--- title: 'Poisson RCM: A Continuum Connection Model' url: https://www.emergentmind.com/topics/poisson-random-connection-model-rcm type: topic --- # Poisson RCM: A Continuum Connection Model The Poisson random connection model (RCM) is a continuum random graph built on a Poisson point process, with edges inserted independently between pairs of points according to a symmetric connection function. In its standard unmarked form, the vertex set is a homogeneous Poisson point process \(\eta\) on \(\mathbb R^d\) of intensity \(\lambda>0\), and distinct vertices \(x,y\in\eta\) are joined with probability \(g(x,y)\), often in the translation-invariant form \(g(x,y)=\varphi(x-y)\), where \(\varphi:\mathbb R^d\to[0,1]\) is measurable, symmetric, and integrable. Marked versions replace \(\mathbb R^d\) by \(\mathbb R^d\times E\) or \(\mathbb R^d\times M\), allowing the edge law to depend on both spatial displacement and vertex marks. This framework encompasses homogeneous, weighted, enhanced, finite-range, long-range, and marked continuum percolation models, and supports a substantial theory of percolation, connectivity, critical behavior, limit theorems, and stochastic-geometry applications [2004.06313] [2305.07398] [2403.17762]. ## 1. Definition and model classes In the homogeneous Poisson RCM, \(\eta\) is a homogeneous Poisson point process of intensity \(\lambda>0\) on \(\mathbb R^d\), and each unordered pair \(\{x,y\}\subset\eta\) is connected independently with probability \(\varphi(y-x)\), where \(\varphi(x)=\varphi(-x)\) and \[ 0<\int_{\mathbb R^d}\varphi(x)\,dx<\infty. \] Equivalent formulations use a radial kernel \(g:[0,\infty)\to[0,1]\) and write the connection probability as \(g(|x-y|)\) in \(\mathbb R^2\), or \(g(d(x,y)/r)\) on a torus in finite-window connectivity problems [2004.06313] [1908.00346] [1510.05440]. The integrability condition is the basic bounded-degree hypothesis. In the marked setting, the underlying space is \(X=\mathbb R^d\times E\), with intensity measure \[ \mu=\lambda(\mathrm{Lebesgue}\otimes P), \] and a measurable symmetric adjacency kernel \[ \phi((t,a),(s,b))=\phi(t-s;a,b)=\phi(s-t;b,a). \] For \(P\otimes P\)-almost every pair of marks \((a,b)\), one assumes \[ D(a,b):=\int_{\mathbb R^d}\phi(x;a,b)\,dx<\infty, \] so the expected number of \(b\)-marked neighbors of an \(a\)-marked vertex is finite. The unmarked RCM is recovered by taking \(E=\{\ast\}\) [2305.07398]. Several important subclasses appear repeatedly. In the inhomogeneous weighted RCM on \(\mathbb R^2\), Poisson points carry i.i.d. weights with tail \[ \Pr(W>w)=w^{-\beta}1_{[1,\infty)}(w),\qquad \beta>0, \] and distinct weighted vertices at \(x,y\) are connected with probability \[ 1-\exp\!\left(-\frac{\eta W_xW_y}{|x-y|^\alpha}\right),\qquad \eta,\alpha>0. \] In enhanced RCMs, present edges are viewed as straight line segments, and continuous paths are allowed to switch at geometric intersections of segments. In finite-range models, \(\varphi(r)=0\) for \(r>R\). In long-range models, one studies kernels with asymptotic decay such as \(g(x)\asymp |x|^{-\alpha}\) or \(g(x)=1-\exp(-c|x|^{-\alpha})\) [1908.00346] [2508.11562] [1904.13288]. A further extension fixes deterministic endpoints \(y_1,\dots,y_m\in\mathbb R^d\) and studies subgraphs in the RCM on \(\eta\cup\{y_1,\dots,y_m\}\), with the same independent edge rule \(H(x,y)\in[0,1]\). This produces rooted or terminal-constrained subgraph counts and explicit cumulant formulae [2312.12745]. ## 2. Percolation thresholds, sharpness, and uniqueness The basic percolation observable is the probability that a typical vertex belongs to an infinite cluster. In the planar homogeneous RCM, \[ \theta(\lambda)=\Pr(\text{the component of the origin is infinite}), \qquad \lambda_c=\inf\{\lambda>0:\theta(\lambda)>0\}. \] Penrose (1991), as reported in later work, established that \[ 0<\int_0^\infty r\,g(r)\,dr<\infty \quad\Longleftrightarrow\quad 0<\lambda_c<\infty, \] so the model has a nontrivial phase transition under the standard integrability condition. For the enhanced planar model \(G_\lambda^e\), a nontrivial transition is proved under the stronger moment condition \[ 0<\int_0^\infty r^3\,g(r)\,dr<\infty, \] and both the ordinary and enhanced models satisfy absence of percolation at criticality: \[ \theta(\lambda_c)=0,\qquad \theta^e(\lambda_c^e)=0. \] For the weighted planar model, a nontrivial transition in the unenhanced graph occurs only if \(\beta>2\) and \(\alpha>4\), while the enhanced weighted model requires \(\beta>4\) and \(\alpha>8\), and again there is no percolation at criticality [1908.00346]. Marked RCMs support a parallel threshold theory. Writing \(\theta_\lambda(a)=P_\lambda(|C(0^a)|=\infty)\) and \(\chi_\lambda(a)=E_\lambda|C(0^a)|\), one defines the percolation threshold \[ \lambda_c:=\sup\{\lambda\ge0:\theta_\lambda(a)=0\ \text{for }P\text{-a.e. }a\} \] and the susceptibility threshold \[ \lambda_s:=\inf\{\lambda\ge0:\|\chi_\lambda\|_p=\infty\}. \] Under bounded-degree and irreducibility assumptions, these thresholds coincide: \[ \lambda_s=\lambda_c. \] This is the continuum marked analogue of sharpness [2305.07398]. A more refined susceptibility-type threshold was introduced through \[ t_T:=\sup\{t\ge0:\operatorname{ess\,sup}_{x\in\mathbb R^d}E_t|C(x)|<\infty\}, \] alongside the percolation threshold \[ t_c:=\sup\{t\ge0:P_t(|C(x)|<\infty)=1\ \text{for a.e. }x\}. \] Always \(t_T\le t_c\). If \(t0\) such that \[ \operatorname{ess\,sup}_{x\in\mathbb R^d}E_t e^{\theta |C(x)|}<\infty. \] Under an exponential decay assumption on the connection function, one also obtains exponential tails for cluster diameters. In the stationary marked case, under a uniform moment condition on the connection function, one has \[ 0d, \] and \(d_n\) is the smallest radius for which there are no isolated vertices, then almost surely \[ \lim_{n\to\infty}\frac{\alpha n d_n^d}{\log n}=1. \] Thus the isolated-node threshold occurs at the scale \[ r_n^\ast=\Bigl(\frac{\log n}{\alpha n}\Bigr)^{1/d}. \] For connectivity, define \[ \beta=\inf\{a>0:\,a\,g(\alpha/(a\theta))>1\}, \] where \(\theta\) is the volume of the unit ball in \(\mathbb R^d\). If \(\gamma>\beta\) and \[ r_n=\Bigl(\frac{\gamma\log n}{\alpha n}\Bigr)^{1/d}, \] then \[ P\bigl(G(P_n,r_n)\text{ is connected}\bigr)\to1. \] When \(g=1_{[0,1]}\), one has \(\beta=1\), recovering the standard random geometric graph threshold [1510.05440]. A related planar scaling theory considers a Poisson process of intensity \(\rho\) on the unit square \(A=[0,1]^2\), with connection radius \[ r_\rho=\sqrt{\frac{\log \rho+b}{C\rho}}, \qquad C=\int_{\mathbb R^2}g(\|\boldsymbol x\|)\,d\boldsymbol x. \] Under rotational invariance, monotonicity, and integrability of \(g\), together with the tail condition \[ g(x)=o\!\bigl(x^{-2}\log^{-2}x\bigr), \] the number \(N_0\) of isolated nodes satisfies \[ E[N_0]\to e^{-b}, \qquad N_0\Rightarrow \mathrm{Poisson}(e^{-b}). \] Boundary effects do not alter this limit: if \(W_E\) denotes the number of extra isolates created by boundary artifacts, then \(P(W_E=0)\to1\). Since asymptotic connectivity requires the absence of isolated nodes, a necessary condition for asymptotic almost sure connectivity is \(b\to+\infty\), equivalently \[ r_\rho=\sqrt{\frac{\log \rho+b(\rho)}{C\rho}} \quad\text{with }b(\rho)\to+\infty. \] This places the RCM on the same logarithmic-over-density scale as the unit-disk model [1012.5693]. ## 4. Stabilizing functionals, subgraph counts, and limit theorems A large part of the modern theory concerns functionals of \(G(\eta)\) in the thermodynamic regime. For a bounded window \(W\subset\mathbb R^d\), a graph functional \(f\), and the induced graph \(G(\eta)|_W\), the add-one cost at the origin is \[ D_o f(W)=f\bigl(G(\eta\cup\{o\})|_W\bigr)-f\bigl(G(\eta)|_W\bigr). \] The functional is weakly stabilizing if there exists a random variable \(\Delta\) such that for every sequence of cubes \(W_n\uparrow\mathbb R^d\), \[ D_o f(W_n)\to_p \Delta, \] and if, for some \(\delta>0\), \[ \sup_W E|D_o f(W)|^{2+\delta}<\infty. \] If \(f\) is translation-invariant and these conditions hold, then for cubes with \(|W_n|\to\infty\), \[ \frac{f(G(\eta)|_{W_n})-E[f(G(\eta)|_{W_n})]}{\sqrt{|W_n|}} \Rightarrow N(0,\sigma^2), \] with exact limiting-variance formula \[ \sigma^2=\lim_{n\to\infty}\frac{\mathrm{Var}[f(G(\eta)|_{W_n})]}{|W_n|} =\lambda\int_0^1 E\!\left[E[\Delta_t\mid \eta_t]^2\right]dt. \] This general theorem yields CLTs for induced subgraph counts, component counts, the number of connected components, Betti numbers of clique complexes, and, in the supercritical regime, the size of the largest component [2004.06313]. For a fixed connected graph \(A\) on \(k+1\) vertices, the induced-copy count \[ \xi_n^{(A)}=\text{number of induced copies of }A\text{ in }G(\eta\cap W_n) \] satisfies \[ \frac{\xi_n^{(A)}-E[\xi_n^{(A)}]}{\sqrt{|W_n|}}\Rightarrow N(0,\sigma_A^2), \] and the component count \[ \zeta_n^{(A)}=\text{number of components isomorphic to }A\text{ in }G(\eta\cap W_n) \] obeys the same normalization and limit law. For clique complexes, the \(k\)th Betti number \(\beta_k(W)\) also satisfies a CLT, and if \(\varphi(x)\in(0,1)\), then the limiting variance is strictly positive. In the supercritical regime \(\lambda>\lambda_c\), with radial \(\varphi\), \(\varphi(0)=1\), and sufficiently fast decay, the largest cluster size \(|C_{\max}(W_n)|\) satisfies \[ \frac{|C_{\max}(W_n)|-E|C_{\max}(W_n)|}{\sqrt{|W_n|}} \Rightarrow N(0,\sigma^2), \qquad \sigma^2>0 \] [2004.06313]. An earlier CLT for isolated vertices in shrinking-range RCMs required correction. For sequences \(\lambda_n/n^d\to\lambda>0\) and \(g_n(x)=g(n|x|)\), the normalized isolated-vertex count \(I_n(g)\) in a bounded Borel set \(K\) converges to \(N(0,1)\). The proof previously claimed by Roy and Sarkar was shown to contain errors; the corrected argument also extends to larger components when \(g\) has bounded support [1404.2096]. Beyond stabilizing methods, subgraph counts admit exact combinatorial cumulant expansions. For connected graphs with fixed endpoints, the moments and cumulants are indexed by connected non-flat partitions of \([n]\times[r]\), yielding explicit partition-sum formulae. Under translation-invariant integrable \(H\), the cumulants satisfy \[ 0<\kappa_n\bigl(N^G_{y_1,\dots,y_m}\bigr) \le n!^r\,r!^{\,n-1}(C\lambda)^{1+(r-1)n}, \] while the variance has matching lower order \(\Omega(\lambda^{2r-1})\). Consequently, the standardized count obeys a CLT, and the Kolmogorov distance satisfies the explicit bound \[ \sup_{x\in\mathbb R}\Bigl|\Pr(\widetilde N^G\le x)-\Phi(x)\Bigr| \le C_r\,\lambda^{-1/(4r-2)}. \] A second-moment bound gives \[ \Pr\bigl(N^G_{y_1,\dots,y_m}>0\bigr) \ge \frac{(E[N^G_{y_1,\dots,y_m}])^2}{E[(N^G_{y_1,\dots,y_m})^2]}, \] and this lower bound tends to \(1\) as \(\lambda\to\infty\) [2312.12745]. If the edge kernel is rescaled as \(H_\lambda(x,y)=c_\lambda H(x,y)\) with \(c_\lambda\approx \lambda^{-\alpha}\), subgraph counts exhibit a normal-to-Poisson phase transition. For a graph \(G\) with \(m\) fixed endpoints, the critical decay rate is \[ \alpha_m^\ast(G)= \max\Bigl\{ \frac{v(G)-m-1}{e(G)-a_m(G)}, \frac{v(G)-m}{e(G)} \Bigr\}. \] Then the standardized count is asymptotically normal for \(0<\alpha<\alpha_m^\ast(G)\), has a Poisson limit at \(\alpha=\alpha_m^\ast(G)\), and vanishes in probability for \(\alpha>\alpha_m^\ast(G)\) [2409.16222]. Related normal-approximation results based on connected partition diagrams and the Statulevičius condition also cover dense, dilute, and sparse tree-dominated regimes [2301.12145]. ## 5. Critical exponents, lace expansion, and sharpness theory High-dimensional and spread-out RCMs admit a mean-field theory based on the lace expansion. In the marked setting, one defines the two-point operator \(T_\lambda\) with kernel \[ T_\lambda(a,b)=\int_{\mathbb R^d}\tau_\lambda((0,a),(x,b))\,dx \] and the triangle diagram \[ \Delta(\lambda)=\|T_\lambda^3\|_\infty. \] Under bounded-degree and irreducibility assumptions, one has \(\lambda_c=\lambda_T=\lambda_O\), where \(\lambda_T\) is the susceptibility threshold and \(\lambda_O\) the operator-critical value. If the triangle condition \[ \Delta(\lambda_c)<\infty \] holds, then the critical exponents exist and take their mean-field values: \[ \gamma=1,\qquad \beta=1,\qquad \delta=2. \] More precisely, there are matching upper and lower bounds showing \[ \chi_\lambda(a)\asymp (\lambda_c-\lambda)^{-1}\quad (\lambda\uparrow\lambda_c), \] \[ \theta_\lambda(a)\asymp (\lambda-\lambda_c)\quad (\lambda\downarrow\lambda_c), \] and at criticality \[ P_{\lambda_c}(|C(0^a)|>n)\asymp n^{-1/2}. \] The proof uses differential inequalities for a magnetization function \(M(\lambda,h,a)\), together with analyticity of \((\lambda,h)\mapsto M(\lambda,h,a)\) [2305.07398]. A complementary high-dimensional theory for unmarked RCMs derives a continuum lace expansion with remainder term, proves convergence in sufficiently large dimension or sufficiently spread-out range, and establishes both an infra-red bound and the triangle condition. The results cover finite-variance, spread-out, and long-range connection functions. In particular, for sufficiently large \(d\) or large spread-out parameter \(L\), one obtains continuity of the percolation function at criticality and mean-field behavior. In the marked Euclidean setting, the lace-expansion analysis yields an infrared bound of the form \[ \widehat T(k)=O(|k|^{-2})\qquad (k\to0), \] which implies finiteness of the triangle diagram at criticality. Under the stated kernel hypotheses, the mean-field regime holds for all \(d>d_c=6\) [1908.11356] [2210.07727]. Sharpness below criticality has two distinct forms. For bounded-edge models, randomized-algorithm methods show that if \(\lambda<\lambda_c\), then \[ \Pr_\lambda^0(0\leftrightarrow S_n)\le c_1 e^{-c_2 n}, \] while for \(\lambda>\lambda_c\), \[ \Pr_\lambda^0(0\leftrightarrow\infty)\ge C(\lambda-\lambda_c). \] These are continuum analogues of sharp-threshold inequalities familiar from discrete percolation [1712.07016]. More recently, for non-increasing integrable connection functions with unbounded support, the subcritical cluster-size tail was shown to decay exponentially: \[ \psi_n(\lambda)=P_{\lambda,g}(|C_0|\ge n)\le C e^{-cn}, \qquad \lambda<\lambda_c. \] The proof constructs the subcritical RCM as site percolation on a very high-intensity RCM and uses an “asymptotic transitivity” argument to transfer sharpness methods beyond bounded range [2509.02310]. ## 6. Related constructions and applications The RCM functions as a unifying object across continuum percolation, stochastic geometry, Gibbsian point processes, and transport on random media. In enhanced planar models, ordinary graph edges are replaced by straight line segments, and continuous paths may switch at segment intersections. This changes the relevant moment assumptions for phase transitions but retains the critical-continuity phenomenon \(\theta^e(\lambda_c^e)=0\) [1908.00346]. In continuum statistical mechanics, the RCM provides a disagreement-percolation criterion for Gibbs uniqueness. Given a non-negative pair potential \(v\ge0\), set \[ g(x,y)=1-\exp(-v(x,y)). \] If the Poisson RCM with connection function \(g\) and intensity \(\lambda\alpha\) is subcritical, then the Gibbs measure with pair potential \(v\) and reference \(\lambda\alpha\) is unique. Equivalently, for every \(\lambda<\lambda_c\), where \(\lambda_c\) is the RCM critical intensity associated with \(1-e^{-v}\), there is a unique Gibbs distribution in the sense of Dobrushin–Lanford–Ruelle [2108.06303]. The geometry of the infinite cluster also governs dynamical behavior. For simple random walk on long-range Poisson RCMs, recurrence holds in \(d=2\) when \(g(x)\le M|x|^{-s}\) for large \(|x|\) with \(s\ge4\), for every intensity \(\lambda>0\). In contrast, transience holds on the unique infinite cluster for all sufficiently large \(\lambda\) when \(d\ge3\) and \(\int g<\infty\), and also for \(d=1,2\) under the lower bound \[ g(x)\ge 1-\exp(-c|x|^{-\alpha}), \qquad \alpha\in(d,2d), \] again for sufficiently large \(\lambda\) [1904.13288]. In supercritical marked finite-range RCMs, crossing statistics support stochastic-homogenization results. If \(\rho\) exceeds the supercritical parameter and the model has a unique infinite cluster, then there exist constants \(c_1,c_2>0\) such that for all \(\ell\ge1\), \[ P_{\rho,\nu}\otimes Q\{N_\ell\ge c_1\ell^{d-1}\} \ge 1-\exp(-c_2\ell^{d-1}), \] where \(N_\ell\) is the maximal number of vertex-disjoint left-to-right crossings of \(\Lambda_\ell=[-\ell,\ell]^d\) by the infinite cluster. Equipping each edge with conductance \(1\), one obtains a deterministic homogenized matrix \[ D(\rho)=\kappa(\rho)I, \] and the crossing bound implies \(\kappa(\rho)>0\), so \(D(\rho)\) is strictly positive-definite [2507.03965]. A recent supercritical result establishes a continuum analogue of Grimmett–Marstrand slab percolation. For \(d\ge3\) and a nonincreasing finite-range connection function \(\phi\), if \(\lambda>\lambda_c(\phi,\mathbb R^d)\), then the RCM remains supercritical in sufficiently thick slabs \[ S_M=\mathbb R^2\times[-M/2,M/2]^{d-2}, \] and \[ \lim_{M\to\infty}\lambda_c(\phi,S_M)=\lambda_c(\phi,\mathbb R^d). \] This identifies supercritical slab percolation as a stable feature of the full-space supercritical phase [2508.11562]. Taken together, these results show that the Poisson RCM is not merely a single continuum graph model, but a general analytic framework in which percolation, connectivity, subgraph statistics, topological observables, Gibbs uniqueness, random-walk behavior, and homogenized transport can all be formulated in a common language of Poisson input and independently randomized edges. Source: https://www.emergentmind.com/topics/poisson-random-connection-model-rcm