Interacting particle system

An interacting particle system is a stochastic process whose local states evolve through random transitions whose rates depend on nearby states.

At each site, a rate-one clock rings and the site copies the state of a randomly chosen neighbor. The model is a Markov process on Ising spin configurations, interpreted through its local update infinitesimal generator. All constant configurations are absorbing state. Tracing copying arrows backwards gives the voter-model duality with coalescing random walks.

Independent Poisson process copying arrows construct the voter model. A backward ancestral line from a site follows a continuous-time random walk; lines that meet then coincide. Thus a finite collection of Ising spins at time $t$ equals initial Ising spins at the positions of coalescing random walks. In particular $E_{\eta }{\xi }_t(x)={\sum }_z{p}_t(x,z)\eta (z)$, and the probability of disagreement at two sites is bounded by the probability their ancestral walks have not yet coalesced.

If the initial one-spin set $F$ is finite on the square lattice, voter-model duality gives $P({\xi }_t(x)=1)={\sum }_{z\in F}{p}_t(x,z)\to 0$. The transition probabilities tend to zero on this infinite lattice; their Fourier transform also proves the limit of a sequence by dominated convergence. A union bound on any finite set of observation sites then proves weak convergence of probability measures to the all-zero point mass in the product topology.

In a nearest-neighbor voter model on the square lattice, any two ancestral coalescing random walks meet almost surely. For an invariant probability measure of a Markov process, the disagreement probability is constant in time but bounded by the vanishing probability of noncoalescence. Hence every pair of sites agrees almost surely. Countability concentrates the measure on the two constant configurations; both are absorbing state, so every mixture of their Dirac measures is invariant. Translation invariance need not be assumed.

Particles on a countable lattice attempt jumps at the rates $p(x,y)$, and an attempt into an occupied site is suppressed. Thus each site contains at most one particle and the number of particles is conserved when finite. Here ${\eta }^{xy}$ interchanges the occupations at $x$ and $y$. Bounded nearest-neighbour rates give a well-defined Markov process. A symmetric jump kernel gives the symmetric simple exclusion process; an unequal left-right kernel gives the asymmetric simple exclusion process.

On $\mathbb{Z}$, put independent rate-$c$ Poisson processes on unoriented nearest-neighbour edges, and exchange their endpoint occupations at each mark. Its generator on cylinder functions is $Lf(\eta )=c{\sum }_x[f({\eta }^{x,x+1})-f(\eta )]$. Equal occupations do not change the state, and a particle next to a vacancy jumps across at rate $c$. The stirring representation of symmetric exclusion constructs it for arbitrary initial configurations. The simple occupation-product product self-duality of symmetric exclusion requires symmetry and is not the ordinary duality of an asymmetric process.

On a finite graph with symmetric edge-swap rates, the stirring representation of symmetric exclusion transports the occupations by a random permutation. An independent identically distributed Bernoulli distribution occupation field is invariant under every deterministic permutation, and hence under this random permutation when it is independent of the initial field. Therefore the displayed product measure is invariant, including densities zero and one. Equivalently, an edge swap preserves the number of particles and has the same reverse rate, giving detailed balance.

Give every site a distinct label and interchange labels at the edge marks defining the symmetric simple exclusion process. Each forward or backward label path makes only finitely many jumps in a bounded time interval, since its total rate is $2c$. Occupations are transported along these paths. The backward ancestors of a finite set have the law of the finite-particle symmetric exclusion process, by time reversal of the Poisson marks. This proves the product self-duality of symmetric exclusion.

For finite $A\subset \mathbb{Z}$ put $D(\eta ,A)={\prod }_{x\in A}\eta (x)$, with empty product one. Interchanging occupations gives $D({\eta }^{xy},A)=D(\eta ,A^{xy})$, where $A^{xy}$ interchanges membership of the two sites. The edge-swap generators therefore agree on $D$. The stirring representation of symmetric exclusion proves the displayed identity at all times by tracing the ancestors of $A$ backwards. A general asymmetric jump kernel does not satisfy this product self-duality.

If the probability of occupation at every site in a finite set depends only on its cardinality, the law is exchangeable. The product self-duality of symmetric exclusion and conservation of the finite dual particle count imply that all these occupation-product moments are unchanged by evolution. Inclusion-exclusion principle reconstructs every finite cylinder probability from them, proving invariance. This direct argument does not require a representation theorem for exchangeable laws.

In the contact process, infected vertices recover at rate $\mu$, while each oriented nearest-neighbour edge transmits infection at rate $\lambda$. The empty configuration is absorbing.

For the contact process with per-edge infection rate $\lambda$ and recovery rate one on a degree-$(d+1)$ regular tree, the infinitesimal generator gives $L{\nu }_{\rho }(A)=(1-\rho ){\nu }_{\rho }(A)(|A|/\rho -\lambda |{\partial }_{E}A|)$. The edge boundary of a finite forest in a regular tree implies $L{\nu }_{\rho }\le 0$ whenever $\rho \lambda (d-1)\ge 1$, including the absorbing empty state where the generator is zero. Thus the displayed functional is a bounded supermartingale, and the extinction probability from $A$ is at most ${\rho }^{|A|}$.

The graphical representation of the contact process preserves coordinatewise order. Its law from the all-infected state decreases in stochastic domination with time, and its limit is invariant by the Feller semigroup property. The all-healthy state is absorbing. Any invariant probability law lies between these two laws. Duality of the contact process gives $\overline{\nu }(\eta (x)=1)=\theta (\lambda )$ when the recovery rate is one. Consequently the two extremal laws coincide exactly when the survival probability of the contact process is zero.

With recovery rate one, the number of infected sites in a nearest-neighbour contact process has death rate equal to its size and birth rate at most $2d\lambda$ times its size. It is dominated by a linear birth-death branching process with these per-particle rates. In particular, from one infection its expected size is at most $e^{(2d\lambda -1)t}$, so its survival probability is zero when $2d\lambda <1$. This proves the displayed lower bound on the critical rate.

Place recovery marks at each vertex according to independent rate-$\mu$ Poisson processes and infection arrows on each oriented nearest-neighbour edge according to independent rate-$\lambda$ Poisson processes. A vertex is infected at time $t$ exactly when a time-directed path from an initially infected vertex reaches it while following arrows and avoiding recovery marks.

For the one-dimensional contact process with recovery rate one and per-neighbour infection rate $\lambda$, use parity sites $(x,n)$ with $x+n$ even. Call a site good if $x$ has no recovery in $[(n-1)\delta ,(n+1)\delta ]$ and sends an arrow to each neighbour during $[n\delta ,(n+1)\delta ]$. The goodness indicators are independent: recovery intervals for successive parity sites at the same spatial point only touch at endpoints, and all outgoing-arrow intervals and clocks are disjoint. Their common probability is the displayed $q$. A path of good sites carries infection from each time layer to the next, because both its source and target avoid recovery during the transmitting interval. Taking $\delta ={\lambda }^{-1/2}$ makes $q\to 1$. The incoming-edge coupling of site and bond percolation and $p_c^{\mathrm{bond}}<1$ imply $p_c^{\mathrm{site}}<1$, so this proves a finite critical infection rate.

Map a site $x\in {\mathbb{Z}}^d$ to its height $h(x)={\sum }_i{x}_i$. Maintain one infected representative in height $n$ for every infected site of an auxiliary one-dimensional contact process. The representative's death makes that auxiliary site recover. Its $d$ outgoing arrows to each neighbouring height, each of rate $\lambda$, make an auxiliary infection at total rate $d\lambda$, using the arrow's target as the new representative. Arrows to an already infected auxiliary height do nothing there. Distinct active representatives use distinct death and outgoing-arrow clocks, so their conditional rates are exactly those of the one-dimensional process. Every auxiliary infection has a live representative upstairs, proving the survival comparison.

Reachability in the graphical representation gives

Time reversal of the Poisson graphical representation, together with reversal of every infection arrow, gives

The survival probability is the probability that the contact process begun from one infected vertex never reaches its absorbing empty state. Its critical infection rate is ${\lambda }_c(\mu )=\inf \{\lambda :\theta (\lambda ,\mu )>0\}$.

Local survival means repeated infection of a fixed vertex at arbitrarily large times in the contact process. It implies global survival, but infection can survive by moving away from every fixed vertex on an infinite graph. On a transitive graph the threshold is independent of the chosen vertex, and ${\lambda }_1\le {\lambda }_2$, with ${\lambda }_1$ the global survival threshold of the contact process.

Orient a degree-$(d+1)$ regular tree towards one end, and let height $g$ decrease along the unique parent edge and increase along the $d$ child edges. For $w_{\rho }(A)={\sum }_{x\in A}{\rho }^{g(x)}$, suppressing the occupied-target constraint only increases births, so its infinitesimal generator satisfies $L{w}_{\rho }\le [-1+\lambda ({\rho }^{-1}+d\rho )]w_{\rho }$. For $d>1$, choose $\rho =d^{-1/2}$. If $\lambda <1/(2\sqrt{d})$, the expected weight and the infection probability at each fixed site decay exponentially. The expected number of recoveries there equals its expected total occupation time, which is finite. Each infection episode ends almost surely at a recovery, so infinitely late infection is impossible. For $d=1$, take $\rho <1$ sufficiently close to one for each $\lambda <1/2$.

This is the critical infection rate for the contact process started from one infected site to avoid its absorbing empty state forever. It concerns infection somewhere in the graph at all times, rather than repeated infection of a fixed site. On a transitive graph it is independent of the starting site. Infection-arrow inclusion couples different rates monotonically.

Inside a rooted $d$-ary subtree of a degree-$(d+1)$ regular tree, pair each cell root with one distinguished child. Keep reinfection inside each pair, suppress infection back into its parent cell, and allow each child cell to be activated only once. Offspring counts of different cells are independent and identically distributed, so cell genealogies form a Galton-Watson process. A cell root has $d-1$ outgoing child-cell edges and its paired vertex has $d$. Starting at the root, the probabilities to send an arrow along one specified outgoing edge from these two vertices before internal extinction are respectivelyThey follow by first-step equations on the three nonempty pair states. The offspring mean is $M=(d-1)a+db$, andContinuity makes $M>1$ at some strictly smaller positive rate. The branching-process extinction criterion then gives positive survival probability for the suppressed process, hence for the original contact process.