How far from the edge need a population be to survive? A probability model. Keywords: branching random walk, probability model, edge effect. Abstract Let N N be a natural number. We consider a population which lives on I N = { − N , − N + 1 , … , N − 1 , N } I_{N}=\{-N,-N+1,\dots,N-1,N\} . Each individual gives birth at rate λ \lambda on each of its neighboring sites and dies at rate 1. No births are allowed from the inside of I N I_{N} to the outside or vice-versa. The population on the whole line (i.e. N = + ∞ N=+\infty ) survives with positive probability if and only if λ > 1 / 2 \lambda>1/2 . On the other hand for any 1 / 2 < λ ≤ 2 / 2 1/2<\lambda\leq\sqrt{2}/2 there exists a natural number N c N_{c} such that the population survives on I N I_{N} for N ≥ N c N\geq N_{c} but dies out for N < N c N\sqrt{2}/2 the population on I_{N} survives for all N\geq 1 . On the other hand if \lambda\leq 1/2 the population dies out for all N . When 1/2<\lambda\leq\sqrt{2}/2 things get more interesting. Our main result is the following. Theorem 1. Let 1/2<\lambda\leq\sqrt{2}/2 then there exists a natural number N_{c} such that if N1/2 . Note that for \lambda\leq 1/2 the process restricted to I_{N} will die out for all N\geq 1 since (by our construction) the restricted process has less individuals than the unrestricted one. Let A_{t}(N) be the total number of individuals alive at time t for the brw restricted to I_{N} . Define the critical value by \lambda_{c}(N)=\inf\{\lambda>0:P_{\lambda}(A_{t}(N)\geq 1\mbox{ for all t})>0\}. Since P_{\lambda}(A_{t}(N)\geq 1\mbox{ for all t}) is an increasing function of \lambda , if \lambda>\lambda_{c}(N) the population restricted to I_{N} survives forever with positive probability while the population dies out if \lambda<\lambda_{c}(N) . Actually, the population also dies out when \lambda=\lambda_{c}(N) as will be explained below. Tom Liggett ( [ 4 ] ) considered branching random walks on finite homogeneous trees. In the infinite tree each site has d+1 neighbors. The finite tree T_{d,N} is obtained by retaining all those sites which can be reached from the origin (i.e. a fixed site on the tree) with a path of length less than or equal to N . In the particular case d=1 , T_{d,N} is exactly I_{N} and the root is 0 . The analysis in [ 4 ] (see page 319) is based on the representation of the branching random walk as a (non spatial) multi-type branching process. The type of the individual is defined as its distance from the root. Thus, possible types are 0,1,\dots,N . Let M_{N}(t) be the (N+1)\times(N+1) matrix whose (j,k) entry gives the mean number of type k individuals at time t produced by a single type j individual at time 0 . The generator of this matrix generator is denoted by A_{N} . That is, M_{N}(t)=\exp(A_{N}t). The entries of the matrix A_{N} are given by a(j,k)=\begin{cases}-1&\mbox{ if }j=k\\ \lambda&\mbox{ if }j=k+1\\ \lambda&\mbox{ if }j=k-1\geq 1\\ 2\lambda&\mbox{ if }j=0\mbox{ and }k=1\\ \end{cases} Let f_{N}(x)=\det(xI-A_{N}) be the characteristic polynomial of the matrix A_{N} . It is shown in [ 4 ] that for N\geq 2 , f_{N}(x)=(x+1)f_{N-1}(x)-\lambda^{2}f_{N-2}(x), with f_{0}(x)=x+1 and f_{1}(x)=(x+1)^{2}-2\lambda^{2} . Recall that a multi-type branching process survives if and only if the matrix A_{N} has a strictly positive eigenvalue, see for instance [ 1 ] (page 203). Note that we can write A_{N}=-I+\lambda C_{N} where the entries of the matrix C_{N} are constants. This shows that the eigenvalues of A_{N} can be written as -1+\lambda\sigma where \sigma is an eigenvalue of C_{N} . Hence, A_{N} has a positive eigenvalue if C_{N} has a positive eigenvalue \sigma and if \lambda>0 is such that -1+\lambda\sigma>0 . Therefore, the critical value \lambda_{c}(N) is the smallest value of \lambda such that A_{N} has zero as an eigenvalue. For instance, we see that f_{1}(0)=0 if \lambda=\sqrt{2}/2 . Hence, \lambda_{c}(1)=\sqrt{2}/2 . Similarly, \lambda_{c}(2)=\sqrt{3}/3 . Moreover, since \lambda_{c}(N) is the smallest value of \lambda so that A_{N} has zero as an eigenvalue we see that the process dies out for \lambda=\lambda_{c}(N) for all N\geq 1 . In [ 4 ] it is also proved that the critical value of the branching random walk on T_{d,N} has the following limit, \lim_{N\to\infty}N^{2}\left(2\sqrt{d}\lambda_{c}(N)-1\right)=\frac{\pi^{2}}{2}. We are interested in the case d=1 . Not surprisingly, \lambda_{c}(N) converges to 1/2 (i.e. the critical value of the branching random walk on the whole line \mathbb{Z} ). We also note that \lambda_{c}(N) is strictly larger than 1/2 for all N . See also [ 3 ] for an analogous (but less precise) result on {\mathbb{Z}}^{d} for all d\geq 1 . Let 1/2<\lambda\leq\sqrt{2}/2 . Define N_{c}=\min\{N\in\mathbb{N}:\lambda>\lambda_{c}(N)\}. Since \lambda_{c}(N) converges to 1/2 the set \{N\in\mathbb{N}:\lambda>\lambda_{c}(N)\} is not empty. By the well-ordering Principle it has a minimum N_{c} . Thus, the brw restricted to I_{N} with birth rate \lambda survives for all N\geq N_{c} . Since \lambda\leq\sqrt{2}/2=\lambda_{c}(1) we see that N_{c}\geq 2 . Therefore, N_{c}-1 is a natural number such that \lambda\leq\lambda_{c}(N_{c}-1) . Hence, the brw restricted to I_{N} with birth rate \lambda dies out for all N\leq N_{c}-1 . This completes the proof of our result.