Abstract
Branching random walks play a key role in modeling the evolutionary processes with birth and death of particles depending on the structure of a medium. The branching random walk on a multidimensional lattice with a finite number of branching sources of three types is investigated. It is assumed that the intensities of branching in the sources can be arbitrary. The principal attention is paid to the analysis of spectral characteristics of the operator describing evolution of the mean numbers of particles both at an arbitrary point and on the entire lattice. The obtained results provide an explicit conditions for the exponential growth of the numbers of particles without any assumptions on jumps variance of the underlying random walk.
Access provided by Autonomous University of Puebla. Download conference paper PDF
Similar content being viewed by others
Keywords
- Branching random walks
- Equations in Banach spaces
- Non-homogeneous environments
- Positive eigenvalues
- Population dynamics
1 Introduction: Model of BRW/r/k/m
We present results for continuous-time branching random walks (BRWs) on the lattice \(\mathbf {Z}^{d}\), \(d\in \mathbf {N}\), with a finite number of lattice sites in which the generation of particles can occur, which are called branching sources. By a BRW we mean a stochastic process that combines branching (birth or death) of particles at certain points on \(\mathbf {Z}^{d}\) with their random walk on \(\mathbf {Z}^{d}\). The goal of the paper is to study the distributions of the particle population \(\mu _{t}(y)\) at every point \(y\in \mathbf {Z}^{d}\) and \(\mu _{t}=\sum _{y\in \mathbf {Z}^{d}}\mu _{t}(y)\) over the lattice \(\mathbf {Z}^{d}\) for a BRW with branching sources of different type without any assumptions on the variance of jumps of the underlying random walk.
Suppose that there is a single particle at the moment \(t=0\) on the lattice situated at the point \(x \in \mathbf {Z}^{d}\). Each particle moves on the lattice \(\mathbf {Z}^{d}\) until it reaches a source where its behavior changes. There are three types of branching sources, depending on whether branching takes place or not and on whether random walk symmetry is violated or not. At sources of the first type, particles die or are born, and random walk symmetry is maintained, see, e.g., [1, 2, 11]. At sources of the second type, walk symmetry is violated through an increase in the degree of branching or walk dominance, see, e.g., [9]. Sources of the third type should be called “pseudo-sources,” because at these sources only the walk symmetry is violated, with no particle births or deaths ever occurring. BRWs with r sources of the first type, k of the second type, and m of the third type are denoted BRW/r/k/m and introduced in [12]. Particles exist on \(\mathbf {Z}^{d}\) independently of each other and of their antecedent history.
We define random walk by its generator
where \(\mathscr {A} = \left( a(x,y)\right) _{x,y \in \mathbf {Z}^{d}}\) satisfies the regularity property \(\sum _{y \in \mathbf {Z}^{d}}a(x,y)=0\) for all x, where \(a(x,y)\ge 0\) for \(x\ne y\), \(-\infty<a(x,x)<0\). From this it follows that A itself satisfies this regularity property [12, 13]. Additionally, we assume that the intensities a(x, y) are symmetric and spatially homogeneous, that is, \(a(x-y) : = a(x,y)=a(y,x)=a(0,y-x)\). Thus we can denote a(y, x), \(a(0,y-x)\), that is, \(a(x-y):= a(x,y)=a(y,x)=a(0,y-x)\). The matrix \(\mathscr {A}\) under consideration is irreducible, so for any \(z\in \mathbf {Z}^{d}\) there is such a set of vectors \(z_{1},\dots ,z_{k}\in \mathbf {Z}^{d}\) that \(z=\sum _{i=1}^{k}z_{i}\) and \(a(z_{i})\ne 0\) for \(i=1,\dots ,k\). It is fairly clear that the irreducibility property is inherited by the perturbed matrix A. This, however, does not hold true for the properties of spatial homogeneity and, most importantly, symmetry. We will, however, make use of the structure of A and the symmetry of the underlying matrix \(\mathscr {A}\) in order to overcome this complication.
According to the axiomatics outlined in [3, Ch. III, §2], the probabilities p(h, x, y) of a particle at \(x\notin \{v_{1},v_{2},\dots ,v_{k+r}\}\) to jump to a point y over a short period of time h can be presented as \(p(h,x,y)=a(x,y)h+o(h)\) for \(y\ne x\) and \(p(h,x,x)=1+a(x,x)h+o(h)\) for \(y=x\). From these equalities, see, for instance, [3, Ch. III], we obtain the Kolmogorov backward equations:
where \(\delta (\cdot )\) is the discrete Kronecker \(\delta \)-function on \(\mathbf {Z}^{d}\).
Infinitesimal generating functions \(f(u,v_{i})=\sum _{n=0}^\infty b_n(v_{i}) u^n\), \(0\le u \le 1\), govern branching process at each of the sources \(v_{1},v_{2},\dots ,v_{k+r}\). We denote source intensities \(\beta _{i}:=\beta _{i}^{(1)}=f^{(1)}(1,v_{i})= (-b_{1}(v_{i}))\left( \sum _{n\ne 1}n b_{n}(v_{i})/(-b_{1}(v_{i}))-1\right) \) where the sum is the average number of descendants a particle has at the source \(v_{i}\).
If the particle is not in the branching source, then its random walk occurs in accordance with the above rules. Consider a combination of branching and walking processes observed when a particle is in one of the branching sources \(v_{1},v_{2},\dots ,v_{k+r}\). In this case, the following possible transitions, which can occur in a short period of time h, are as follows: the particle will either move to a point \(y\ne v_{i}\) with the probability of \( p(h,v_{i},y)=a(v_{i},y)h+o(h), \) or will remain at the source and produce \(n\ne 1\) descendants with the probability of \( p_{*}(h,v_{i},n)=b_{n}(v_{i})h+o(h) \) (we assume that the particle itself is included in these n descendants and we say that the particle dies if \(n=0\)), or no changes will occur to the particle at all, which has the probability of \( 1-\sum _{y\ne v_{i}}a(v_{i},y)h-\sum _{n\ne 1}b_{n}(v_{i})h+o(h). \) Thus, the time spent by the particle in the source \(v_{i}\) is exponentially distributed with the parameter \(-(a(v_{i},v_{i})+b_{1}(v_{i}))\). The evolution of each new particle obeys the same law and does not depend on the evolution of other particles.
Let us introduce the moments of the random variables \(\mu _{t}(y)\) and \(\mu _{t}\) as \(m_{n}(t, x, y)= \mathsf {E}_{x} \mu _{t}^{n}(y)\) and \(m_{n}(t, x)= \mathsf {E}_{x} \mu ^{n}_{t}\), respectively, where n is the order of the moment and \(\mathsf {E}_{x}\) is the mean on condition \(\mu _{0}(\cdot ) = \delta _{x}(\cdot )\).
In BRW/r/k/m more general multi-point perturbations of the self-adjoint operator \(\mathscr {A}\) generated of the symmetric random walk are used than in BRW/r/0/0 or in BRW/0/k/0, see, e.g., [13]. This follows from the statement, see [12], that the mean number of particles \(m_{1}(t)=m_1(t,\cdot ,y)\) at a point \(y\in \mathbf {Z}^{d}\) in BRW/r/k/m is governed by:
where
Here, \(\mathscr {A}:l^{p}(\mathbf {Z}^{d})\rightarrow l^{p}(\mathbf {Z}^{d})\), \(p\in [1,\infty ]\), is a symmetric operator, \(\varDelta _{x}=\delta _{x}\delta _{x}^{T}\), and \(\delta _{x}=\delta _{x}(\cdot )\) denotes a column-vector on the lattice taking the unit value at the point x and vanishing at other points, \(\beta _{s}\), \(\zeta _{i}\), \(\eta _{i}\), and \(\chi _{j}\) are some constants. The same equation is also valid for the mean number of particles (the mean population size) over the lattice \(m_{1}(t)=m_1(t,\cdot )\) with the initial condition \(m_{1}(0)=1\) in \(l^{\infty }(\mathbf {Z}^{d})\). Operator (3) can be written as
In each of the sets \(U=\{u_{i}\}_{i=1}^{k+m}\), and \(V=\{v_{j}\}_{j=1}^{k+r}\), the points are pairwise distinct, but U and V may have a nonempty intersection. The points from \(V\setminus U\) correspond to r sources of the first type; those from \(U\cap V\) to k sources of the second type; and those from \(U\setminus V\) to m sources of the third type.
Denote the largest positive eigenvalue of the operator Y by \(\lambda _{0}\). In contrast to [7] we consider BRW/r/k/m instead of BRW/r/0/0 and assume that in (4) the parameters \(\beta _{j}\) are real (\(\beta _{j}\in \mathbb {R}\)) instead of being positive (\(\beta _{j}>0\)). Under this assumption, we conclude that if \(\lambda _{0}\) exitsts then it is simple, strictly positive and guarantees an exponential growth of the first moments \(m_{1}\) of particle both at an arbitrary point y and on the entire lattice.
Theorem 1
Let for BRW/r/k/m under consideration the operator \(\mathscr {Y}\) have an isolated eigenvalue \(\lambda _{0}>0\), and let the remaining part of its spectrum be located on the halfline \(\{\lambda \in \mathbb {R}:~\lambda \leqslant \lambda _{0}-\epsilon \}\), where \(\epsilon >0\). If \(\beta _{i}^{(r)} = O(r! r^{r-1})\) for all \(i = 1, \ldots , N\) and \(r\in \mathbb {N}\), then the following statements hold in the sense of convergence in distribution
where \(\psi (y)\) is the eigenfunction corresponding to the eigenvalue \(\lambda _{0}\) and \(\xi \) is a nondegenerate random variable.
One approach to analysing Eqs. (2) and evolutionary equations for mean numbers of particles \(m_{1}(t,x,y)\) and \(m_{1}(t,x)\) is to treat them as differential equations in Banach spaces. To apply this approach to our case, we introduce the operators
on functions set u(x), \(x\in \mathbf {Z}^{d}\). We can represent the operator (4) in a more convenient form:
where \(\beta _{i}\in \mathbb {R}\), \(i=1,\ldots ,\beta _{k+r}\). All operators in (6) can be considered as linear continuous operators in any of the spaces \(l^{p}(\mathbf {Z}^{d})\), \(p\in [1,\infty ]\). Note that the operator \(\mathscr {A}\) is self-adjoint in \(l^{2}(\mathbf {Z}^{d})\) [12,13,14].
Now, treating for each \(t\ge 0\) and each \(y\in \mathbf {Z}^{d}\) the \(p(t,\cdot ,y)\) and \(m_{1}(t,\cdot ,y)\) as elements of \(l^{p}(\mathbf {Z}^{d})\) for some p, we can write (see, for example, [12]) the following differential equations in \(l^{p}(\mathbf {Z}^{d})\):
and for \(m_{1}(t,x)\) the following differential equation in \(l^{\infty }(\mathbf {Z}^{d})\):
Point out that for large t the asymptotic behaviour of the transition probabilities p(t, x, y), as well as of the mean particle numbers \(m_{1}(t,x,y)\) and \(m_{1}(t,x)\) is tightly connected with operators \(\mathscr {A}\) and \(\mathscr {Y}\) spectral properties.
The properties of p(t, x, y) can be expressed in terms of the Green’s function which can be defined [11, § 2.2] as the Laplace transform of the transition probability p(t, x, y) or through the resolvent form:
where \(x,y\in \mathbf {Z}^{d}\), \(\lambda \ge 0\), and \(\phi (\theta )\) is the transition intensity a(z) Fourier transform:
The meaning of the function \(G_{0}(x,y)\) is as follows: it represents the mean amount of time spent by a particle at at \(y\in \mathbf {Z}^{d}\) as \(t\rightarrow \infty \) provided that at the initial moment \(t=0\) the particle was at \(x\in \mathbf {Z}^{d}\). The asymptotic behaviour of the mean numbers of particles \(m_{1}(t,x,y)\) and \(m_{1}(t,x)\) as \(t\rightarrow \infty \) can be described in terms of the function \(G_\lambda (x,y)\), see, e.g., [11]. Lastly, BRW asymptotic behaviour depends strongly on whether \(G_{0}: = G_{0}(0,0)\) is finite, it was shown in [10].
The approach presented in this section is based on representing the BRW evolution equations as differential equations in Banach spaces. It can also be applied to a wide range of problems, including the description of the evolution of higher-order moments of particle numbers (see, e.g., [11, 12]).
2 Key Equations and Auxiliary Results
We start off with a crucial remark. Since the operator \(\mathscr {Y}\) is in general not self-adjoint, the vast analytical apparatus, developed in [13] and relying heavily on the self-adjointness of the operators involved, is not applicable here directly. Due to the structure of \(\mathscr {Y}\), however, this difficulty can be obviated, to a certain extent, with relative ease. Indeed, consider the following differential equation in a Banach space
with \(\mathscr {Y}=\mathscr {A}+\sum _{i=1}^{k+m}\zeta _{i}\varDelta _{u_{i}}\mathscr {A}+ \sum _{j=1}^{k+r}\beta _{j}\varDelta _{v_{j}}\). Let us now introduce the operator
which is correctly defined for \(\zeta _{i} > -1\), and rewrite the equation using this notation:
which is equivalent to
By applying D to both parts of the equation above, we obtain
Since the operators D and \(\varDelta _{v_{j}}\) commute, the expression above is equivalent to
where \(g := Df\). We have thus rewritten the original equation in such a way that the previously non-self-adjoint operator \(\mathscr {Y}\) is replaced with the self-adjoint operator
and a one-to-one correspondence between the solutions f to the starting equation and the solutions g to the new equation can be established through the formula \(g = Df\). Therefore, when it comes to analysing Cauchy problems, the operator \(\mathscr {Y}\) can, for all intents and purposes, be considered self-adjoint.
We introduce the Laplace generating functions of the random variables \(\mu _{t}(y)\) and \(\mu _{t}\) for \(z \geqslant 0\):
where \(\mathsf {E}_{x}\) is the mean on condition \(\mu _{0}(\cdot ) = \delta _{x}(\cdot )\).
Theorem 2
Let the operator A have the form (1). The functions F(z; t, x) and F(z; t, x, y) are continuously differentiable with respect to t uniformly with respect to \(x,y\in {\mathbf{Z}}^{d}\) for all \(0 \leqslant z \leqslant \infty \). They satisfy the inequalities \(0 \leqslant F(z; t, x), F(z; t, x, y) \leqslant 1\) and are the solutions to the following Cauchy problems in \(l^{\infty }\left( {\mathbf{Z}}^{d} \right) \)
with the initial condition \(F(z; 0, \cdot ) = e^{-z}\) and
with the initial condition \(F(z;0, \cdot , y) = e^{-z \delta _{y}(\cdot )}\).
Theorem 2 allows us to advance from analysing the BRW at hand to considering the corresponding Cauchy problem in a Banach space instead. Note that, contrary to the single branching source case examined in [11], there is not one but several terms \(\varDelta _{v_{j}}f_{j}(F)\) in the right-hand side of Eqs. (8) and (9), \(j=1,2,\ldots ,N\).
Theorem 3
The moments \(m_{n}(t, \cdot , y)\in l^{2}\left( {\mathbf{Z}}^{d}\right) \) and \(m_{n}(t, \cdot )\in l^{\infty }\left( {\mathbf{Z}}^{d}\right) \) satisfy the following differential equations in the corresponding Banach spaces for all natural \(n \geqslant 1\):
the initial values being \(m_{n}(0, \cdot , y) = \delta _{y}(\cdot )\) and \(m_{n}(0, \cdot ) \equiv 1\) respectively. Here \(\mathscr {Y}m_{n}\) stands for \(\mathscr {Y}m_{n}(t, \cdot , y)\) or \(\mathscr {Y}m_{n} (t, \cdot )\) respectively, and
Theorem 3 will later be used in the proof of Theorem 8 to help determine the asymptotic behaviour of the moments as \(t\rightarrow \infty \).
Theorem 4
The moments \(m_{1}(t, x,\cdot )\in l^{2}\left( {\mathbf{Z}}^{d}\right) \) satisfy the following Cauchy problem in \(l^{2}\left( {\mathbf{Z}}^{d}\right) \):
This theorem allows us to obtain different differential equations by making use of the BRW symmetry.
Theorem 5
The moment \(m_{1}(t, x, y)\) satisfies both integral equations
The moment \(m_{1}(t, x)\) satisfies both integral equations
For \(k > 1\) the moments \(m_{k}(t, x, y)\) and \(m_{k}(t, x)\) satisfy the equations
This theorem allows us to make transition from differential equations to integral equations. It is later used to prove Theorem 8.
Theorems 3–5 are a generalization to the case BRW/r/k/m of Lemma 1.2.1, Theorem 1.3.1 and Theorem 1.4.1 from [11], proved there for BRW/1/0/0. The proofs of Theorems 3–5 differ only in technical details from the proofs of the above statements from [11] and are therefore omitted here.
3 Properties of the Operator \(\mathscr {Y}\)
We call a BRW supercritical if \(\mu _{t}(y)\) and \(\mu _{t}\) grow exponentially. As was mentioned in Introduction, one of the main results of this work is the numbers of particles limit behavior (5), from which it follows that the BRW with several branching sources with arbitrary intensities is supercritical if the operator \(\mathscr {Y}\) has a positive eigenvalue \(\lambda \). For this reason we devote this section to studying the spectral properties of the operator \(\mathscr {Y}\).
We mention a statement proved in [11, Lemma 3.1.1].
Lemma 1
The spectrum \(\sigma (\mathscr {A})\) of the operator \(\mathscr {A}\) is included in the half-line \((-\infty , 0]\). Also, since the operator \(\sum _{j=1}^{N} \beta _{j} \varDelta _{v_{j}}\) is compact, \(\sigma _{ess}(\mathscr {Y}) = \sigma \left( \mathscr {A} \right) \subset (-\infty , 0]\), where \(\sigma _{ess}(\mathscr {Y})\) denotes the essential spectrum [6] of the operator \(\mathscr {Y}\).
The following theorem provides a criterion of there being a positive eigenvalue in the spectrum of the operator \(\mathscr {Y}\).
Theorem 6
A number \(\lambda > 0\) is an eigenvalue and \(f \in l^{2}\left( {\mathbf{Z}}^{d} \right) \) is the corresponding eigenvector of the operator \(\mathscr {Y}\) if and only if the system of linear equations
for \(i = 1, \ldots , k+m\), and
for \(i = 1, \ldots , k+r\), with respect to the variables \(f(u_{j})\) and \(f(v_{j})\), where
has a non-trivial solution.
Proof
For \(\lambda > 0\) to be an eigenvalue of the operator \(\mathscr {Y}\) it is necessary and sufficient that there be a non-zero element \(f \in l^{2}\left( {\mathbf{Z}}^{d} \right) \) that satisfies the equation
Obviously, the solution sets of such an equation for the operators \(\mathscr {Y}\) and \(C^{-1} \mathscr {Y} C\) are the same for any operator C; let us set \(C := \left( I + \sum _{i=1}^{k+m}\zeta _{i}\varDelta _{u_{i}} \right) ^{\frac{1}{2}}\), which is correctly defined since \(\zeta _{i} > -1\) for all i. Thus the equation above can be rewritten as follows:
Since \((\varDelta _{v_{j}} f)(x) := f(x)\delta _{v_{j}}(x) = f(v_{j})\delta _{v_{j}}(x)\) and \(( \mathscr {A}\varDelta _{u_{i}} f)(x) := f(x)A\delta _{u_{j}}(x) \), the preceding expression can be rewritten as follows:
We apply Fourier transform to this equality and obtain
for \(\theta \in [-\pi , \pi ]^{d}\). The Fourier transform \(\widetilde{\mathscr {A}f}\) of \((\mathscr {A}f)(x)\) is of the form \(\phi \tilde{f}\), where \(\tilde{f}\) is the Fourier transform of f, and \(\phi (\theta )\) is defined by the equality (7), see [11, Lemma 3.1.1]. With this in mind, and making use of the fact that, by the definition of the Fourier transform,
we rewrite equality (17) as
or
where \(\theta \in [-\pi , \pi ]^{d}\). Since \(\lambda > 0\) and \(\phi (\theta ) \leqslant 0\), \( \int _{[-\pi , \pi ]^{d}} |\lambda - \phi (\theta )|^{-2} d\theta < \infty \), which allows us to apply the inverse Fourier transform to equality (18): as
we obtain
By choosing \(x = u_{i}\), where \(i = 1, \ldots , k+m\), or \(x = v_{i}\), where \(i = 1, \ldots , k+r\), we can rewrite (19) as follows:
We note that any solution of system (15) completely defines f(x) on the entirety of its domain by formula (19), which proves the theorem. \(\square \)
Let among \(k+r\) sources, in which branching occurs, in \(s \le k+r\) sources intensities are \(\beta _i > 0\), \(i=0, \ldots , s\), and in \(k+r-s\) sources intensities are \(\beta _i\le 0\), \(i=s+1, \ldots , k+r\). We represent the operator \(\mathscr {Y}\) defined by (6) as follows:
Define operator
then the eigenvector h corresponding to the eigenvalue \(\lambda \) of \(\mathscr {Y}\) satisfies the equation
Note that \(\langle \mathscr {A} x, x \rangle \le 0\). Besides, \(\beta _i<0\) for \(i=s+1, \ldots , k+r\), and therefore \(\langle \sum \limits _{i=s+1}^{k+r} \beta _i \varDelta _{v_i}x, x \rangle \le 0\). Hence, the operator \(\mathscr {B}\) is reversible. The problem of existence of positive eigenvalues of the operator \(\mathscr {Y}\) is converted to the question of the existence of nonzero solutions for the equation \(h = \mathscr {B}^{-1}\sum \limits _{j=1}^{s} \beta _j \delta _{v_j}\langle \delta _{v_j},h\rangle \), which, after introducing auxiliary variables \(q_{i}=\langle \delta _{v_i},h\rangle \) and scalar multiplication on the left of this equality by \(\delta _{v_{i}}\) reduces to a finite system of equations
Denote matrix \(B^{(\lambda )}\):
So the matrix representation (20) has the following form
and the problem on positive eigenvalues for \(\mathscr {Y}\) is reduced to the question of for which \(\lambda > 0\) the number 1 is the matrix \(B^{(\lambda )}\) eigenvalue.
Theorem 7
Let \(\lambda _{0}>0\) be the largest eigenvalue of the operator \(\mathscr {Y}\). Then \(\lambda _{0}\) is a simple eigenvalue of \(\mathscr {Y}\), and 1 is the largest eigenvalue of the matrix \(B^{(\lambda _{0})}\).
Proof
Denote \(\zeta := \max (0, \max \limits _{i}(\zeta _i)) \ge 0\) and note that the elements of the operator
are non-negative. It follows from Schur’s test [4] that in each of the spaces \(l^{p}(\mathbf {Z}^{d})\) for the operator norm \(\tilde{\mathscr {A}}\) there is an estimation
Operator \(\mathscr {B}\) can be represented as follows:
where the operator
is diagonal with all its diagonal elements no less than \(-\mathbf{a} (0, 0)(\zeta +1) + \lambda > 0\). Then
Then \(\mathscr {B}\) can be represented in the following form \(\mathscr {B} = \mathscr {F}_{\lambda }\left( I-\mathscr {F}_{\lambda }^{-1}\tilde{\mathscr {A}}\right) \) and therefore
Here by virtue of (23) and (24) the operator norm of \(\mathscr {F}_{\lambda }^{-1}\tilde{\mathscr {A}}\) is less than one:
and therefore the operator \(\left( I-\mathscr {F}_{\lambda }^{-1}\tilde{\mathscr {A}}\right) ^{-1}\) can be represented as a series:
Hence, by virtue of (25)
Note that the right-hand side (26) is the sum of the products of operators (infinite matrices) with non-negative elements. Therefore, each element of the operator (infinite matrix) \(\mathscr {B}^{-1}\) is non-negative.
Let us prove that each element of the operator \(\mathscr {B}^{-1}\) is strictly positive. For the proof, we use the fact that our random walk is irreducible. Note that, since the random walk under the action of the operator \(\mathscr {A}\) is irreducible, for any pair \(x, y \in \mathbf {Z}^{d}\) there exists \(n \ge 1\) and the set of points
such that \(a(u_{1}-u_{0})a(u_{2}-u_{1})\cdots a(u_{n}-u_{n-1}) > 0\), whence follows
Note that the elements of the infinite matrix \(\mathscr {B}^{-1}\) are indexed by pairs of points \(x, y\in \mathbf {Z}^{d}\). In addition, the element with indices (x, y) of the matrix \(\left( \mathscr {F}_{\lambda }^{-1}\tilde{\mathscr {A}}\right) ^{n}\) in (26) is a sum of the form
taken over all possible “chains” of n elements, satisfying (27), in which positive factors \(f_{u_{0},u_{1},\ldots ,u_{n}}\) are formed due to the presence of the diagonal matrix \(\mathscr {F}_{\lambda }^{-1}\) of \(\left( \mathscr {F}_{\lambda }^{-1}\tilde{\mathscr {A}}\right) ^{n}\). But by virtue of (28) at least one term in the sum (29) is strictly positive, while the rest are non-negative. Hence, the entire sum is also strictly positive, which implies that all elements of the operator (infinite matrix) \(\mathscr {B}^{-1}\) are strictly positive. Since the elements of \(\mathscr {B}^{-1}\) (see (21)) are positive, the matrix \(B^{(\lambda )}\) is positive.
The right side of (25) contains the operators \(\mathscr {F}_{\lambda }^{-1}\), whose elements, monotonically decreasing in \(\lambda > 0\), tend to zero as \(\lambda \rightarrow \infty \). Since in this case all multiplied and added operators (infinite matrices) are positive, then all their elements in this case will also decrease monotonically in \(\lambda > 0\) and tend to zero as \(\lambda \rightarrow \infty \).
We first show that if \(\lambda _{0}\) is the operator \(\mathscr {Y}\) largest eigenvalue, then the largest (absolute) eigenvalue of the matrix \(B^{(\lambda _{0})}\) is 1. Indeed, assume it is not the case.
It follows from (22) that \(\lambda _{0} > 0\) is an eigenvalue of \(\mathscr {Y}\) if and only if 1 is an eigenvalue of \(B^{(\lambda _{0})}\). All elements of \(B^{(\lambda _{0})}\) are strictly positive. Consequently, by the Perron-Frobenius theorem, see [5, Theorem 8.4.4], the matrix \(B^{(\lambda _{0})}\) has a strictly positive eigenvalue that is strictly greater (by absolute value) than any other of its eigenvalues.
Let us denote the dominant eigenvalue of \(B^{(\lambda _{0})}\) by \(\gamma (\lambda _{0})\). Since we assumed that 1 is not the largest eigenvalue of \(B^{(\lambda _{0})}\), then \(\gamma (\lambda _{0})> 1\). Given that with respect to \(\lambda \) the functions \(I_{x_{i} - x_{j}}(\lambda )\) are continuous, then all elements of \(B^{(\lambda )}\) and all eigenvalues of \(B^{(\lambda )}\) are continuous functions with respect to \(\lambda \). All matrix \(B^{(\lambda )}\) eigenvalues tend to zero as \(\lambda \rightarrow \infty \), because for all i and j \(I_{x_{i} - x_{j}}(\lambda ) \rightarrow 0\) as \(\lambda \rightarrow \infty \). Hence there is such a \(\hat{\lambda } > \lambda _{0}\) that \(\gamma (\hat{\lambda }) = 1\). Then, as was shown earlier, this \(\hat{\lambda }\) has to be an eigenvalue of \(\mathscr {Y}\), which contradicts the initial assumption that \(\lambda _{0}\) is the largest eigenvalue of the operator \(\mathscr {Y}\).
We have just proved that 1 is the largest eigenvalue of \(B^{(\lambda _{0})}\); then we obtain from the Perron-Frobenius theorem the simplicity of this eigenvalue. Therefore, to complete the proof, we have to show the simplicity of the eigenvalue \(\lambda _{0}\) of \(\mathscr {Y}\).
Suppose it is not, and \(\lambda _{0}\) is not simple. In this case, there are at least two linearly independent eigenvectors \(f_{1}\) and \(f_{2}\) corresponding to the eigenvalue \(\lambda _{0}\). Therefore, we can again applying the equality (19), notice that the linear independence of the vectors \(f_{1}\) and \(f_{2}\) is equivalent to the linear independence of the vectors
From the definition of \(B^{(\lambda )}\) and Theorem 6 it follows that vectors \(\hat{f}_{1}\) and \(\hat{f}_{2}\) satisfy \(\left( B^{(\lambda _{0})} - I\right) f = 0\). It contradicts the simplicity of eigenvalue 1 of \(B^{(\lambda _{0})}\). This completes the proof. \(\square \)
Lemma 2
Let \(\mathscr {Y}\) be a self-adjoint continuous operator on a separable Hilbert space E, the spectrum of which is a disjoint union of two sets: fist one is a finite (counting multiplicity) set of isolated eigenvalues \(\lambda _{i} > 0\) and second one is the remaining part of the spectrum which is included in \([-s, 0]\), \(s> 0\). Then the solution m(t) of the Cauchy problem
satisfies the condition
where \(\lambda _{0} = \max _{i} \lambda _{i}\).
Proof
Let us denote by \(V_{\lambda _{i}}\) the finite-dimensional eigenspace of \(\mathscr {Y}\) that corresponds to the eigenvalue \(\lambda _{i}\).
We consider the projection \(P_{i}\) of \(\mathscr {Y}\) onto \(V_{\lambda _{i}}\), see [6]. Let
All spectral operators \(P_{i}\) and \(\left( I - \sum P_{i} \right) \) commute with \(\mathscr {Y}\), see [6]. Therefore
As \(x_{i}(t) \in V_{\lambda _{i}}\), we can see that \(\mathscr {Y} x_{i}(t) = \lambda _{i} x_{i}(t)\), from which it follows that \(x_{i}(t) = e^{\lambda _{i} t} x_{i}(0)\). Since the spectrum of the operator
is included into the spectrum of operator \(\mathscr {Y}\) and \(\mathscr {Y}_{0}\) has no isolated eigenvalues \(\lambda _{i}\), it is included into \([-s, 0]\). From this for all \(t\geqslant 0\) we obtain \(|v(t)| \leqslant |v(0)|\), see [11, Lemma 3.3.5]. Hence
and the proof is complete. \(\square \)
Remark 1
Let \(\lambda _{0}\) be the largest eigenvalue of \(\mathscr {Y}\). Denote \(P_{0} m(0 )=C(m_{0})\) in (30). Then \(C(m_0)\ne 0\) if and only if the orthogonal projection \(P_{0} m(0 )\) of the initial value \(m_0=m(0)\) onto the corresponding to the eigenvalue \(\lambda _{0}\) eigenspace is non-zero.
If the eigenvalue \(\lambda _{0}\) of \(\mathscr {Y}\) is simple and f is an eigenvector corresponding to \(\lambda _{0}\), the projection \(P_{0}\) is defined by the formula \(P_{0}x=\frac{(f,x)}{(f,f)}f\), where \((\cdot ,\cdot )\) denots scalar product in the Hilbert space E. In cases when this \(\lambda _{0}\) is not simple, describing the projection \(P_{0}\) is a significantly more difficult task.
We remind the reader that we proved the simplicity of the largest eigenvalue of \(\mathscr {Y}\) above allowing us to bypass this issue.
Theorem 8
Let defined by (6) operator \(\mathscr {Y}\) with the parameters \(\lbrace \zeta _{i} \rbrace _{i=1}^{k+r}\) and \(\lbrace \beta _{i} \rbrace _{i=1}^{k+m}\), has a finite number of positive eigenvalues (counting multiplicity). We denote by \(\lambda _{0}\) the largest of them, and the corresponding to \(\lambda _{0}\) normalized vector by f. Then for \(t \rightarrow \infty \) and all \(n \in {\mathbf{N}}\) the following statements hold:
where
and the functions \(C_{n}(x, y)\) and \(C_{n}(x) > 0\) for \(n \geqslant 2\) are defined as follows:
where \(D^{(j)}_{n}(x)\) are certain functions satisfying the estimate \(|D_{n}^{(j)} (x)|\leqslant \frac{2}{n\lambda _{0}}\) for \(n\geqslant n_{*}\) and some \(n_{*}\in {\mathbf{N}}\) and \(g^{(j)}_{n}\) are the functions defined in (12).
Proof
For \(n \in {\mathbf{N}}\) we consider the functions \(\nu _{n} := m_{n} (t, x, y) e^{-n \lambda _{0} t}\). From Theorem 3 (see Eqs. (10) and (11) for \(m_{n}\)) we obtain the following equations for \(\nu _{n}\):
the initial values being \(\nu _{n}(0, \cdot , y) = \delta _{y}(\cdot ), n \in {\mathbf{N}}\).
Since \(\lambda _{0}\) is the largest eigenvalue of \(\mathscr {Y}\), the spectrum of \(\mathscr {Y}_{n} := \mathscr {Y} - n \lambda _{0} I\) for \(n \geqslant 2\) is included into \((-\infty , -(n-1)\lambda _{0}]\). As it was shown, for example, in [11, p. 58], that if the spectrum of a self-adjoint continuous operator \(\widetilde{\mathscr {Y}}\) on a Hilbert space is included into \((-\infty , -s], s > 0\), and also \(f(t) \rightarrow f_{*}\) as \(t \rightarrow \infty \), then the solution of the equation
satisfies \(\nu (t) \rightarrow -\widetilde{\mathscr {Y}}^{-1} f_{*}\) condition. For this reason for \(n \geqslant 2\) we obtain
Now we prove the existence of such a natural number \(n_{*}\) that for all \(n\geqslant n_{*}\) the estimates
hold. We evaluate the operator \(\mathscr {Y}_{n}^{-1}\) norm. For this, let us consider two vectors u and x such that \(u=\mathscr {Y}_{n}x= \mathscr {Y}x - n\lambda _{0} x\). Then \(\Vert u\Vert \geqslant n\lambda _{0} \Vert x\Vert - \Vert \mathscr {Y}x\Vert \geqslant (n\lambda _{0} -\Vert \mathscr {Y}\Vert )\Vert x\Vert \), hence \(\Vert \mathscr {Y}_{n}^{-1}u\Vert =\Vert x\Vert \leqslant \Vert u\Vert /\left( n\lambda _{0} -\Vert \mathscr {Y}\Vert \right) \), and for all \(n\geqslant n_{*}=2\lambda _{0}^{-1}\Vert \mathscr {Y}\Vert \) the estimate \(\Vert \mathscr {Y}_{n}^{-1}\Vert \leqslant \frac{2}{n\lambda _{0}}\) holds. From this we conclude that
Now we have to estimate the particle number moments asymptotic behaviour. It follows from (14) that as \(t \rightarrow \infty \) the following asymptotic equivalences hold:
The function \(m_{1}(t, x, 0)\) exhibits exponential growth as \(t \rightarrow \infty \) and \(m_{1}(t, x)\) will display the same behaviour.
We can now infer the asymptotic behaviour of the higher moments \(m_{n}(t, x)\) for \(n \geqslant 2\) from Eqs. (11) in a similar way to how it was done above for \(m_{n}(t, x, y)\).
We proceed to prove the equalities for \(C_{1} (x, y)\) and \( C_{1}(x)\). The eigenvalue \(\lambda _{0}\) is simple by Corollary 7 and it follows, according to Remark 1, that
But \(m_{1}(0, x, y) = \delta _{y}(x)\), hence
We also obtain from (32) that
which concludes the proof. \(\square \)
Corollary 1
\(C_{n}(x, y) = \psi ^{n}(y) C_{n}(x)\), where \( \psi (y) = \frac{\lambda _{0}f(y)}{\sum _{j=1}^{k+r} \beta _{j} f(v_{j})}\).
Proof
We prove the corollary by induction on n. For \(n=1\) the induction basis holds due to Theorem 8. Let us now deal with the induction step: according to Theorem 8,
therefore, it suffices to prove that for all j the equalities
hold. As a consequence of the definition and hypothesis of induction,
which proves the corollary. \(\square \)
4 Proof of Theorem 1
Let us introduce the function
The following auxiliary lemma is proved in [7, Lemma 9].
Lemma 3
There is such a constant \(C>0\) that \(f(n, r) < C \frac{n^{n}}{r^{ r -1}}\) for all \(n \geqslant r \geqslant 2\).
We now turn to proving Theorem 1.
Proof
Let us define the functions
as follows from Theorem 8 and \(G_{\lambda }(x,y)\) being positive, these definitions are sound. Corollary 1 yields
From the above equalities and the asymptotic equivalences (31) we have Theorem 1 statements in terms of convergence of moments of the random variables \(\xi (y)=\psi (y)\xi \) and \(\xi \).
The distributions of the limit random variables \(\xi (y)\) and \(\xi \) to be uniquely determined by their moments if, as was shown in [11], the Carleman condition
We establish below that the series for the m(n, x) diverges and that, therefore, said moments define the random variable \(\xi \) uniquely; the statement concerning \(\xi (y)\) and its moments can be proved in much the same manner.
Since \(\beta ^{(r)}_{j} = O(r! r^{r-1})\), there is such a constant D that for all \(r \geqslant 2\) and \(j = 1, \ldots , k+r\) the inequality \(\beta ^{(r)}_{j} < D r! r^{r-1}\) holds. Without loss of generality we assume that for all n
Let \(\gamma := 2 N C D E\frac{\lambda _{0}\beta _{2}}{2} C_{1}^{2}(v_{1})\), where C is defined in Lemma 3, and the constant E is such that \(C_{n}(v_{1}) \leqslant \gamma ^{n-1} n! n^{n}\) for \(n \leqslant \max \{n_{*}, 2 \}\), where \(n_{*}\) is defined in Theorem 8.
From this point on, the proof follows to the scheme of proof of [7, Th. 1] and is included only for readability.
Let us show by induction that
The induction basis for \(n = 1\) is valid due to the C choice. In order to prove the step of induction, we will show that
It follows from \(C_{n+1}(v_{1})\) formula and the estimate for \(D_{n}^{(j)}(x)\) from Theorem 8 that
By the induction hypothesis
which, added to the fact that \(\beta _{j}^{(r)} < D r! r^{r-1}\) and \(\gamma ^{n+1-r} \leqslant \gamma ^{n-1}\), yields
We infer from Lemma 3 that
Hence, by referring to the \(\gamma \) definition we obtain
which completes the proof of the step of induction.
Since \(n! \leqslant \left( \frac{n+1}{2} \right) ^{n}\), \(C_{n}(x) \leqslant \frac{\gamma ^{n}}{2^{n}} (n+1)^{2n}\). Thus,
from which we obtain that
The condition (33) is satisfied, and the corresponding Stieltjes moment problem for the moments m(n, x) has a unique solution [8, Th. 1.11]. Hence, statements (5) are valid in terms of convergence in distribution and Theorem 1 is proved. \(\square \)
References
Albeverio, S., Bogachev, L.V., Yarovaya, E.B.: Asymptotics of branching symmetric random walk on the lattice with a single source. C. R. Acad. Sci. Paris Sér. I Math. 326(8), 975–980 (1998). https://doi.org/10.1016/S0764-4442(98)80125-0
Bogachev, L.V., Yarovaya, E.B.: A limit theorem for a supercritical branching random walk on \(\mathbf{Z}^{d}\) with a single source. Russian Math. Surv. 53(5), 1086–1088 (1998). https://doi.org/10.1070/rm1998v053n05ABEH000077
Gikhman, I.I., Skorokhod, A.V.: The Theory of Stochastic Processes. II. Classics in Mathematics. Springer, Berlin (2004).Translated from the Russian by S. Kotz, Reprint of the 1975 edition
Halmos, P.R.: A Hilbert Space Problem Book. Graduate Texts in Mathematics, vol. 19, 2nd edn. Springer, New York (1982). https://doi.org/10.1007/978-1-4684-9330-6Encyclopedia of Mathematics and its Applications, 17
Horn, R.A., Johnson, C.R.: Matrix Analysis. Cambridge University Press, Cambridge (1985). https://doi.org/10.1017/CBO9780511810817
Kato, T.: Perturbation theory for linear operators. Die Grundlehren der mathematischen Wissenschaften, Band 132, Springer, New York (1966). https://doi.org/10.1007/978-3-662-12678-3
Khristolyubov, I.I., Yarovaya, E.B.: A limit theorem for a supercritical branching walk with sources of varying intensity. Teor. Veroyatn. Primen. 64(3), 456–480 (2019). https://doi.org/10.1137/S0040585X97T989556
Shohat, J.A., Tamarkin, J.D.: The Problem of Moments. American Mathematical Society Mathematical Surveys, vol. I. American Mathematical Society, New York (1943)
Vatutin, V.A., Topchiĭ, V.A., Yarovaya, E.B.: Catalytic branching random walks and queueing systems with a random number of independent servers. Teor. Ĭmovīr. Mat. Stat. 69, 1–15 (2003)
Yarovaya, E.: Spectral asymptotics of a supercritical branching random walk. Theory Probab. Appl. 62(3), 413–431 (2018). https://doi.org/10.1137/S0040585X97T98871X
Yarovaya, E.B.: Branching random walks in a heterogeneous environment. Center of Applied Investigations of the Faculty of Mechanics and Mathematics of the Moscow State University, Moscow (2007). (in Russian)
Yarovaya, E.B.: Spectral properties of evolutionary operators in branching random walk models. Math. Notes 92(1–2), 115–131 (2012). Translation of Mat. Zametki 92 (2012), no. 1, 123–140. https://doi.org/10.1134/S0001434612070139
Yarovaya, E.B.: Branching random walks with several sources. Math. Popul. Stud. 20(1), 14–26 (2013). https://doi.org/10.1080/08898480.2013.748571
Yarovaya, E.: Positive discrete spectrum of the evolutionary operator of supercritical branching walks with heavy tails. Methodol. Comput. Appl. Probab. 19(4), 1151–1167 (2016). https://doi.org/10.1007/s11009-016-9492-9
Acknowledgment
The research was supported by the Russian Foundation for the Basic Research (RFBR), project No. 20-01-00487.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2021 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Yarovaya, E., Balashova, D., Khristolyubov, I. (2021). Branching Walks with a Finite Set of Branching Sources and Pseudo-sources. In: Shiryaev, A.N., Samouylov, K.E., Kozyrev, D.V. (eds) Recent Developments in Stochastic Methods and Applications. ICSM-5 2020. Springer Proceedings in Mathematics & Statistics, vol 371. Springer, Cham. https://doi.org/10.1007/978-3-030-83266-7_11
Download citation
DOI: https://doi.org/10.1007/978-3-030-83266-7_11
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-83265-0
Online ISBN: 978-3-030-83266-7
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)