--- title: 'Sub-Poisson Distributions: Theory & Applications' url: https://www.emergentmind.com/topics/sub-poisson-distributions type: topic --- # Sub-Poisson Distributions: Theory & Applications Sub-Poisson distributions are not a single universally standardized class. Across probability, stochastic modeling, spatial statistics, and random-matrix theory, the term denotes objects that are compared to a Poisson benchmark and found to be more regular, less variable, or otherwise “below Poisson” in a precise technical sense. In one probabilistic usage, a non-negative random variable with mean $\mu$ is sub-Poissonian when its MGF is dominated by that of a Poisson law of mean $\mu$; in count-data modeling, sub-Poisson often means underdispersed, typically $\mathrm{Var}(X)<\E X$ or Fano factor $<1$; in point-process theory, it refers to directionally-convex domination by a Poisson point process with the same mean measure; and in spectral statistics it designates level-spacing laws of the form $a s e^{-b s^c}$ with $c<1$ [2103.17027] [1009.5696] [1705.09179]. ## 1. Terminological scope and Poisson benchmarks The cited literature uses “sub-Poisson” in several non-equivalent ways. What is common is the choice of Poisson behavior as a reference object: the comparison may be made at the level of MGFs, raw moments, variance-to-mean ratios, spatial ordering, or spectral spacing laws. The technical consequences therefore depend on the ambient category of random object. | Domain | Defining comparison | Poisson feature being benchmarked | |---|---|---| | Non-negative random variables | MGF domination | moments and tails | | Count distributions | $\mathrm{Var}(X)<\E X$ or $\mathrm{FF}<1$ | dispersion | | Point processes | $\Phi \le_{dcx} \Phi_{\mathrm{Poi}}$ | clustering and void structure | | Spectral statistics | $p(s)=a s e^{-b s^c}$ with $c<1$ | level-spacing profile | In the variance-based count-model literature, “sub-Poissonian” is often synonymous with underdispersion. In the gene-regulation setting, the defining condition is $\mathrm{FF}=\mathrm{Var}(p)/\E[p]<1$ [1110.2804]. In the point-process setting, by contrast, the notion is order-theoretic: a point process is sub-Poisson if it is directionally-convexly dominated by a Poisson point process with the same mean measure [1009.5696]. In random-matrix theory, the term is used for intermediate nearest-level-spacing statistics with weaker repulsion than Wigner-Dyson, and specifically for the parameter regime $c<1$ inside the family $p_{abc}(s)=a s e^{-b s^c}$ [1705.09179]. A further source of ambiguity is that Poisson also appears in adjacent but distinct phrases such as “compound Poisson” and “subexponential densities.” The class $\mathcal{S}_d$ of subexponential densities for compound Poisson sums concerns convolution asymptotics such as $(g*g)(x)\sim 2g(x)$, not sub-Poissonianity in any of the senses above [2001.11362]. Likewise, a study of the Poisson benchmark for the minimum of $P(X_\lambda\le \E[X_\lambda])$ explicitly notes that it does not define sub-Poissonianity via variance or moment comparison, although it uses Poisson concentration around the mean as a reference point [2210.16515]. ## 2. MGF domination and raw-moment theory A central probabilistic definition is the sub-Poissonian MGF bound for a non-negative random variable $X$ with mean $\mu$: \[ \E[e^{tX}] \le \exp(\mu(e^t-1)) \qquad \text{for all } t>0. \] This class includes the Poisson distribution, the Binomial distribution $\mathrm{Bin}(n,p)$, and sums of independent $[0,1]$-bounded random variables. It excludes distributions such as the geometric and negative binomial, whose MGFs exceed the Poisson bound [2103.17027]. For such $X$, the normalized raw moments admit the uniform bound \[ \E\!\left[\left(\frac{X}{\mu}\right)^k\right] \le \left(\frac{k/\mu}{\log(1+k/\mu)}\right)^k \le \left(1+\frac{k}{2\mu}\right)^k \le \exp\!\left(\frac{k^2}{2\mu}\right), \] valid for all $k>0$ [2103.17027]. The paper presents this as a sharp and simple inequality, improving previous uniform bounds by a factor exponential in $k$. The asymptotic structure is two-regime. When $k^2/\mu \to 0$, the normalized raw moment approaches $1$, and the estimate is asymptotically tight: \[ \E\!\left[\left(\frac{X}{\mu}\right)^k\right]=1+\Theta\!\left(\frac{k^2}{\mu}\right) \] for Poisson $X$, and also for Binomial $X$ with suitable $n,p$ [2103.17027]. This is supported by the lower bound \[ \E[X^k] \ge \mu^k\left(1+\frac{k(k-1)}{2\mu}\right) \] for Poisson, and similarly for suitable binomial laws. In the opposite regime $x=k/\mu \to \infty$, the optimal moment growth satisfies \[ \E\!\left[\left(\frac{X}{\mu}\right)^k\right]^{1/k} \le \frac{x}{e\log x}\left(1+O\!\left(\frac{\log\log x}{\log x}\right)\right), \] with matching lower bounds from Poisson moments and Bell numbers up to constant-order corrections of the same logarithmic type [2103.17027]. This establishes asymptotic sharpness both for moderate $k$ and for large $k/\mu$. The significance of this formulation is that it yields explicit bounds depending on $k/\mu$ rather than only on universal constants. Earlier bounds of the form \[ c^k \le \E\!\left[\left(\frac{X}{\mu}\right)^k\right] \le C^k \] lose an exponential factor in $k$ relative to the refined estimate above [2103.17027]. This suggests that MGF-based sub-Poissonianity is especially effective when precise moderate-moment control is required. ## 3. Centered sub-Poisson concentration and variance proxies A more recent framework extends the MGF-domination idea to centered random variables and develops a nonasymptotic concentration theory. With \[ \phi(\lambda)=e^\lambda-1-\lambda, \] a random variable $X$ is upper sub-Poisson with variance proxy $\sigma^2$ if \[ \E e^{\lambda(X-\E X)} \le \exp(\sigma^2\phi(\lambda)) \qquad \forall \lambda\ge 0. \] Lower sub-Poisson means that $-X$ is upper sub-Poisson, and two-sided sub-Poisson means the same bound holds for all $\lambda\in\mathbb{R}$ [2508.12103]. Within this framework, the optimal sub-Poisson variance proxy is the smallest $\sigma^2$ for which the centered MGF inequality holds for all $\lambda$. The paper states that finiteness of this proxy is equivalent to sub-Poissonianity, that the proxy dominates the ordinary variance, and that it is zero if and only if $X$ is deterministic [2508.12103]. It also derives bounds connecting the proxy to sub-Gaussian and sub-exponential Orlicz norms. The principal concentration result is a Bennett-type inequality without boundedness assumptions. If $X$ is upper sub-Poisson with variance proxy $\sigma^2$, then for $t\ge 0$, \[ \Pr(X\ge t)\le \exp[-\sigma^2]\left(\frac{e\sigma^2}{\sigma^2+t}\right)^{\sigma^2+t}. \] A Bernstein-type corollary is \[ \Pr(X\ge t)\le \exp\left(-\frac{t^2/2}{\sigma^2+t/3}\right). \] The paper presents this as a tail theory tailored to Bernoulli and Poisson variables, and to signed versions of these variables, in their natural tail regime [2508.12103]. The same work proves that sub-Poissonianity is closed under independent sums and convex combinations, but not under all linear operations such as scalar multiplication [2508.12103]. This suggests a structural analogy with sub-Gaussian theory, but with $\phi(\lambda)$ replacing the quadratic log-MGF. ## 4. Underdispersion in discrete count models In several count-data literatures, a sub-Poisson distribution is an underdispersed count law, typically characterized by variance smaller than the mean. In stochastic auto-regulation models, the relevant quantity is the Fano factor \[ \mathrm{FF}=\frac{\mathrm{Var}(p)}{\E[p]}, \] with sub-Poisson behavior defined by $\mathrm{FF}<1$ [1110.2804]. The steady-state protein count distribution is represented as a Poisson mixture, \[ P(p)=\int_0^\infty \frac{e^{-\lambda}\lambda^p}{p!}\rho(\lambda)\,d\lambda. \] If $\rho(\lambda)$ is a positive, normalizable density, then the mixture is super-Poisson, with $\mathrm{FF}\ge 1$; sub-Poisson behavior therefore requires parameter regimes in which no such positive mixing density exists [1110.2804]. For the auto-repression model, the effective parameter \[ \beta = \frac{p_b r}{(1+r)^2} + \frac{c_b - c_f r}{1+r} \] controls this transition. The sub-Poisson regime occurs when $\beta<0$, whereas auto-activation remains in the $\beta>0$ regime and cannot produce sub-Poisson statistics [1110.2804]. The paper further states that the region of parameter space yielding sub-Poisson behavior under auto-repression is narrow. A different route to underdispersion is obtained by Markov chain-based generalizations of discrete parent laws, including Poisson-derived models. In that construction, the mean number of successes is \[ \mu = \frac{r_1 \E[N]}{r_1+r_2}, \] and dispersion is controlled by $r_1+r_2$: overdispersion when $r_1+r_2<1$, equidispersion when $r_1+r_2=1$, and underdispersion or sub-Poisson behavior when $r_1+r_2>1$ [2006.13766]. The mechanism is explicit: when $r_1+r_2>1$, successive successes are negatively correlated and fluctuations are suppressed. The mean-parametrized Conway–Maxwell–Poisson distribution provides an extreme underdispersion result. For fixed mean $\mu$, as the underdispersion parameter $\nu\to\infty$, the limiting law is a point mass at $\mu$ when $\mu$ is an integer, and a shifted Bernoulli on $\lfloor\mu\rfloor$ and $\lceil\mu\rceil$ with weights given by the fractional part of $\mu$ when $\mu$ is non-integer [2011.07503]. The paper states that this limiting law is the most underdispersed discrete distribution possible for a given mean and that this is currently the only known generalization of the Poisson distribution exhibiting arbitrary underdispersion for any mean. The discrete weak-stability literature imposes a further restriction. In the class of Poisson–delayed Sibuya distributions, which includes Poisson and Hermite laws, the paper states that only the Poisson law is sub-Poissonian in the sense of variance not exceeding the mean; there are no other nontrivial sub-Poissonian discrete stable laws in that family [2509.12070]. This corrects a common overgeneralization: Poisson-related families need not contain genuinely underdispersed non-Poisson members. ## 5. Sub-Poisson point processes and spatial connectivity For point processes, sub-Poissonianity is defined through the directionally-convex order. A function $f:\mathbb{R}^d\to\mathbb{R}$ is directionally convex (dcx) if for vectors $x,y,p,q$ with $p\le x,y\le q$ and $x+y=p+q$, \[ f(x)+f(y)\le f(p)+f(q). \] For point processes $\Phi_1,\Phi_2$, one writes $\Phi_1\le_{dcx}\Phi_2$ when for every finite collection of bounded Borel sets $B_1,\ldots,B_n$, \[ (\Phi_1(B_1),\ldots,\Phi_1(B_n)) \le_{dcx} (\Phi_2(B_1),\ldots,\Phi_2(B_n)). \] It is enough to verify this for mutually disjoint $B_i$ [1009.5696]. A point process $\Phi$ is sub-Poisson if \[ \Phi \le_{dcx} \Phi_{\mathrm{Poi}}, \] where $\Phi_{\mathrm{Poi}}$ is a Poisson point process with the same mean measure [1009.5696]. The paper also introduces weaker notions, including weakly sub-Poisson point processes defined by the joint-intensity inequalities \[ \rho^{(k)}(x_1,\ldots,x_k)\le \lambda^k \] for all $k\ge 1$, almost everywhere, in the stationary case of intensity $\lambda$ [1009.5696]. The geometric interpretation is that point processes smaller in dcx order are less variable and more regular than Poisson; they exhibit less clustering and fewer voids. The order is compatible with comparisons of Ripley’s $K$-function, correlation functions, and shot-noise fields [1009.5696]. These comparisons have concrete consequences for continuum percolation. For a homogeneous dcx-sub-Poisson point process $\Phi$ of intensity $\lambda$, the critical communication radius in the Boolean or Gilbert model satisfies \[ 0