Abstract
We consider an asymptotically stable multidimensional random walk \(S(n)=(S_1(n),\ldots , S_d(n) )\). For every vector \(x=(x_1\ldots ,x_d)\) with \(x_1\ge 0\), let \(\tau _x:=\min \{n>0: x_{1}+S_1(n)\le 0\}\) be the first time the random walk \(x+S(n)\) leaves the upper half space. We obtain the asymptotics of \(p_n(x,y):= {\textbf{P}}(x+S(n) \in y+\Delta , \tau _x>n)\) as n tends to infinity, where \(\Delta \) is a fixed cube. From that, we obtain the local asymptotics for the Green function \(G(x,y):=\sum _n p_n(x,y)\), as \(|y |\) and/or \(|x |\) tend to infinity.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction, Main Results and Discussion
1.1 Notation and Assumptions
Consider a random walk \(\{S(n),n\ge 0 \}\) on \({\mathbb {R}^d}\), \(d\ge 1\), where
and \(\{X(n), n\ge 1\}\) are independent copies of a random vector \(X=(X_1,\ldots ,X_d)\). For \(x=(x_1, \ldots , x_d)\) in the (non-negative) half space, that is for \(x_1\ge 0\), let
be the first time the random walk exits the (positive) half space
When \(x=0\), we will omit the subscript and write
In this paper, we study the asymptotic, as \(n\rightarrow \infty \), behaviour of the probability
and the Green function
Here and throughout, we denote \(\Delta =[0,1)^d\) and for \(y=(y_1,\ldots ,y_d)\),
In this paper, we will mostly concentrate on the case when the random walk S(n) has infinite second moments. More precisely, we shall assume that S(n) is asymptotically stable of index \(\alpha <2\) when we study large deviations for local probabilities and asymptotics for the Green function. The asymptotics for the Green function of walks with finite variances have already been studied in the literature: (a) Uchiyama [20] has considered lattice walks in a half space; (b) Duraj et al [13] have derived asymptotics of Green functions for a wider class of convex cones. It is worth mentioning that the authors of [13] analyse first the case of a half space and use the estimates for the Green function for a half space in the subsequent analysis of convex cones. This fact underlines the importance of the case of half spaces.
We will say that S(n) belongs to the domain of attraction of a multivariate stable law, if
where \(\alpha \in (0,2]\) and \(\zeta _\alpha \) has a multivariate stable law of index \(\alpha \). Note that we assume that S(n) is already centred. This does not restrict generality, when \(\alpha \ne 1\), as one can subtract the mean for \(\alpha >1\) and the centring is not needed for \(\alpha <1\). This assumption excludes, however, some walks with \(\alpha =1\) from consideration.
Necessary and sufficient conditions for the convergence in (2) are given in [16]. When \(\alpha \in (0,2)\), the convergence will take place if \({\textbf{P}}(|X |>t)\) is regularly varying of index \(-\alpha \) and there exists a measure \(\sigma \) on the unit sphere \({\mathbb {S}}^{d-1} \) such that
for any measurable A on \({\mathbb {S}}^{d-1}\). We will write \(X\in {\mathcal {D}}(d, \alpha ,\sigma )\) when (2) holds and additionally \(\sigma ({\mathbb {S}}^{d-1}\cap {\mathbb {H}}^+)>0\) and \(\sigma ({\mathbb {S}}^{d-1}\cap {\mathbb {H}}^-)>0\) for \(\alpha \in (0,2)\), where \({\mathbb {H}}^- = {\mathbb {R}}^d\setminus {\mathbb {H}}^+\). Here, \(\sigma \) stands for the above measure on the unit sphere. Also, let \(g_{\alpha ,\sigma }\) be the density of \(\zeta _\alpha \). This assumption implies that the first coordinate \(X_1\) belongs to the one-dimensional domain of attraction.
Note that \(X\in {\mathcal {D}}(d, \alpha ,\sigma )\) implies that \(S_1(n)\) is oscillating, that is, \({\textbf{P}}(\tau <\infty )=1\) and \({\textbf{E}}[\tau ]=\infty \). For that recall that the random walk \(S_1(n)\) oscillates if and only if
Rogozin [17] investigated properties of \(\tau \) and demonstrated that the Spitzer condition
holds if and only if \(\tau \) belongs to the domain of attraction of a positive stable law with parameter \(\rho \). In particular, if \(X\in {\mathcal {D}}(d, \alpha ,\sigma )\), then (see, for instance, [23]) condition (3) holds with
It is well known that (3) is equivalent to the following Spitzer–Doney condition
The scaling sequence \(\{c_n\}\) can be defined as follows, see [16]. Denote \(\mathbb {Z}:=\left\{ 0,\pm 1,\pm 2,\ldots \right\} ,\) \(\mathbb {Z} _{+}:=\left\{ 1,2,\ldots \right\} \) and let \(\left\{ c_{n},n\ge 1\right\} \) be a sequence of positive numbers specified by the relation
where
It is known (see, for instance, [14, Ch. XVII, § 5]) that for every \( X\in \mathcal {D}(d,\alpha ,\sigma )\) the function \(\mu (u)\) is regularly varying with index \((-\alpha )\). This implies that \(\left\{ c_{n},n\ge 1\right\} \) is a regularly varying sequence with index \(\alpha ^{-1}\), i.e. there exists a function \(l_{1}(n),\) slowly varying at infinity, such that
Then, convergence (2) holds with this sequence \(\{c_n\}\).
In one-dimensional case, the study of asymptotics (1) was initiated in [21], where normal and small deviations of \(p_n(0,y)\) were considered. Asymptotics for \(p_n(x,y)\) with a general starting point x was studied then in [3] and [10]. Our assumption on \(X_{1}\) is the same as in these papers, and we use a similar approach for small and normal deviations. We used a different approach to study large deviations in the multidimensional case. Large deviations seem to be the most complicated part of the present paper.
As the first coordinate plays a distinctive role, we will adopt the following notation. For X(n), we will write \(X(n)=(X_{1}(n),X_{(2,d)}(n))\), where \(X_{1}(n)\) corresponds to the first coordinate and \(X_{(2,d)}(n)\) corresponds to the remaining coordinates. Similarly, we write \(S(n)=(S_{1}(n),S_{(2,d)}(n)), n=0,2,\ldots \).
The following conditional limit theorem will be crucial for the rest of this article. The weak convergence in this theorem can be proven similarly to [9]. Existence of the density is shown in the proof of Theorem 2, but can also be found similarly to Remark 2 in [21].
Theorem 1
If \(X\in \mathcal {D}(d,\alpha ,\sigma )\), then there exists a random vector \(M_{\alpha ,\sigma }\) on \({\mathbb {H}}^+\) with density \(p_{M_{\alpha ,\sigma }}(v)\) such that, for all \(u\in {\mathbb {R}}^d\),
Moreover, for every bounded and continuous function f,
uniformly in x with \(0\le x_1\le \delta _nc_n\), \(\delta _n\rightarrow 0\).
Our first result is an analogue of the classical local limit theorem, which is an extension of Theorems 3 and 5 in [21] when \(x=0\) and extends [10] and [3] for arbitrary starting point x to the case of half-planes.
Theorem 2
Suppose \(X\in \mathcal {D}(d,\alpha ,\sigma )\). If the distribution of X is non-lattice, then, for every \(r >0\), uniformly in x with \(0\le x_1\le \delta _nc_n\), \(\delta _n\rightarrow 0\),
If the distribution of X is lattice and if \({\mathbb {Z}}^d\) is the minimal lattice for X, then uniformly in \(x\in \mathbb {H}^+\cap {\mathbb {Z}}^d\) with \(0\le x_1\le \delta _nc_n\), \(\delta _n\rightarrow 0\),
If the ratio \(y/c_{n}\) varies with n in such a way that \(y_1/c_{n}\in (b_{1},b_{2})\) for some \(0<b_{1}<b_{2}<\infty \) and \(|y_{(2,d)} |=O(c_n)\), we can rewrite (9) as
However, if \(y_1/c_{n}\rightarrow 0\), then, in view of
relation (9) gives only
Our next theorem refines (11) in the mentioned domain of small deviations, i.e. when \(y_1/c_{n}\rightarrow 0.\) Let
Let \(\chi ^{+}:=S_{1}(\tau ^{+})\) (resp. \(\chi ^{-}:=-S_{1}(\tau )\)) be the first ascending(descending) ladder height and let \((\chi ^+_n)_{n=1}^\infty \) (resp. \((\chi _n^-)_{n=1}^\infty \)) be a sequence of i.i.d. copies of \(\chi ^+\) (resp. \(\chi ^-\)). Let
be the renewal function of the ascending (descending) ladder height process. Clearly, H is a left-continuous function.
Theorem 3
Suppose \(X\in \mathcal {D}(d,\alpha ,\sigma )\). If the distribution of X is lattice and if \(\mathbb {Z}^d\) is the minimal lattice, then
uniformly in \(x,y \in {\mathbb {H}}^+\cap {\mathbb {Z}}^d\)with \(x_1,y_{1}\in (0,\delta _{n}c_{n}]\) such that \(|x-y|\le A c_n \), where \(\delta _{n}\rightarrow 0\) as \(n\rightarrow \infty \) and A is a fixed constant.
If the distribution of X is non-lattice, then
uniformly in \(x_1,y_{1}\in (0,\delta _{n}c_{n}]\) such that \(|x-y|\le A c_n \), where \(\delta _{n}\rightarrow 0\) as \(n\rightarrow \infty \) and A is a fixed constant.
To obtain the asymptotics for the Green function of S(n) killed at leaving \(\mathbb {H}^+\), one has to estimate probabilities of local large deviations for S(n). For that, we shall assume that
where \(\phi (t):={\textbf{P}}(|X|>t)\) is regularly varying function of index \(-\alpha \).
The fact that global assumptions might in general give different asymptotics if the tails are too heavy is known in the multidimensional case since Williamson [22], who constructed a counterexample for the Green functions on the whole space, which has a different asymptotic behaviour without any local assumptions.
We now ready to formulate our bound for local large deviations.
Theorem 4
Let \(X\in {\mathcal {D}}(d, \alpha ,\sigma )\) with \(\alpha <2\) and (14) hold.
Then,
Having asymptotics for \(p_n(x,y)\) for all possible ranges (for small, normal and large deviations), one can easily derive asymptotics for the Green function. We start with the lattice case, where we combine Theorems 3 and 4 to obtain the asymptotics of the Green function near the boundary.
Theorem 5
Assume \(X\in {\mathcal {D}}(d,\alpha ,\sigma )\) with some \(\alpha <2\). Suppose that (14) holds.
-
1.
If the distribution of X is lattice and if \({\mathbb {Z}}^d\) is minimal for X, then we have
$$\begin{aligned} G(x,y) \sim C \frac{H(y_1)V(x_1)}{|x-y|^d} \int _0^{\infty } g_{\alpha ,\sigma }\left( 0, \frac{y_{2,d}-x_{2,d}}{|x-y|} t\right) t^{d-1} \textrm{d}t \end{aligned}$$for \(x_1,y_1=o(|x-y|)\). In particular, in the isotopic case, that is when the limiting density \(\sigma \) is uniform on the unit sphere,
$$\begin{aligned} G(x,y) \sim C_{\alpha } \frac{H(y_1)V(x_1)}{|x-y|^d}, \quad |x-y|\rightarrow \infty , \end{aligned}$$for \(x_1,y_1=o(|x-y|)\).
-
2.
If X is non-lattice, then
$$\begin{aligned} G(x,y) \sim C \frac{\int _{y_1}^{y_1+1}H(u)\textrm{d}uV(x_1)}{|x-y|^d} \int _0^{\infty } g_{\alpha ,\sigma }\left( 0, \frac{y_{2,d}-x_{2,d}}{|x-y|} t \right) t^{d-1} \textrm{d}t \end{aligned}$$for \(x_1,y_1=o(|x-y|)\). In particular, in the isotopic case, that is when the limiting density \(\sigma \) is uniform on the unit sphere,
$$\begin{aligned} G(x,y) \sim C \frac{\int _{y_1}^{y_1+1}H(u)\textrm{d}uV(x_1)}{|x-y|^d}, \quad |x-y|\rightarrow \infty , \end{aligned}$$for \(x_1,y_1=o(|x-y|)\).
-
3.
In addition, there exists a constant C such that for all \(x,y\in \mathbb {H}^+\),
$$\begin{aligned} G(x,y) \le C \frac{H(y_1)V(x_1)}{|x-y|^d}. \end{aligned}$$(16)
Remark 1
Recall that \(H(x)\sim C x^{\alpha \rho } l(x), x\rightarrow \infty \) for some slowly varying function l, see [21, Lemma 13]. Similarly, \(V(x)\sim C x^{\alpha (1-\rho )}/l(x)\) as \(x\rightarrow \infty \).
Remark 2
For stable Lévy processes, an exact formula (which can be analysed asymptotically) for the Green function g(x, y) was obtained in [19]. We are not aware of any result of this kind for asymptotically stable random walks when \(\alpha <2\).
Remark 3
As we have already mentioned, walks with finite variances, which are a particular case of asymptotically stable walks with \(\alpha =2\), were considered in [13, Theorem 2] and [20] in the lattice case. In [13], estimates of the behaviour of the Green function relied on [5, 12] and were obtained in a more general situation of convex cones. Using our methods, we have obtained asymptotics for the Green function in a half space for all asymptotically stable walks with \(\alpha =2\). Specialising this result to the case of finite variances, we can obtain asymptotics under weaker moment conditions than in [13] and stronger than in [20]. The method we use is different from [20] who approached the problem using the potential kernel.
The only difference between \(\alpha <2\) and \(\alpha =2\) is the form of estimates for local large deviations. In the case \(\alpha =2\), one has to take care of the Gaussian component, which leads to very long calculations. For that reason, we have decided not to include walks with \(\alpha =2\) in the present paper and to consider this case in a separate paper.
Remark 4
It seems to be possible to extend the estimates for the Green function for asymptotically isotropic random walk in cones. This will be considered elsewhere and should also allow one to extend the results of [4, 5, 7, 8, 12, 13] to the stable case.
Since exit times from a half space can be considered as exit times for one-dimensional random walks, it is quite natural to use the methods, which are typical for walks on the real line. In the proofs of our results on the asymptotic behaviour of the probabilities, we follow this strategy and use a lot of methods from [10] and [21]. However, it is worth mentioning that additional dimensions cause many additional technical problems, because of the possibility that the random walk can have a large jump, which is close to the boundary hyperplane \(\{x:x_1=0\}\). While in the finite variance case one can try to control these jumps by assuming existence of additional moments, we cannot assume existence of some additional moments in the infinite variance case. For these reasons, our estimates for \(p_n(x,y)\) cannot be considered as a straight forward generalisation of one-dimensional results.
2 Preliminary Upper Bounds
In this section, we find bounds for \({\textbf{P}}(\tau _x>n)\). Since \(\tau _x\) is actually a stopping time for the one-dimensional walk \(S_1(k)\), we may apply Lemma 3 from [6], which gives us the following estimate.
Lemma 6
Assume that (4) holds. Then, there exists \(C_0\) such that for every \(x\in {\mathbb {H}}^+\) one has
The next lemma is an extension of [21, Lemma 20] to the case of half spaces. We will give a proof following a different approach, which relies on Lemma 6. This proof works in one-dimensional case as well, thus simplifying the corresponding arguments of [21].
For \(y=(y_1,\ldots , y_d),z=(z_1,\ldots , z_d)\in {\mathbb {R}}^d\), we will write \(y\le z\) if \(y_k\le z_k\) for all \(1\le k\le d\).
Lemma 7
Assume that \(X\in \mathcal {D}(d,\alpha ,\sigma )\). Then, there exists \(C>0\) such that for all \(x,y\in {\mathbb {H}}^+\) and all \(n\ge 1\) we have
Similar result holds for the stopping time \(\tau ^+\):
Proof
We prove the first statement only. The proof of the second estimate requires only notational changes.
Put \(n_1=[n/4], n_2=[3n/4]-n_1, n_3=n-[3n/4]\). We split the probability of interest into three parts,
Now, we will make use of the time inversion. Let
and
Let \(1_d=(1,\ldots ,1)\). Then,
Then, using the concentration function inequalities, we can continue as follows:
As a result, we obtain the bound
Applying Lemma 6, we obtain
Here, recall that one can deduce by Rogozin’s result that (3) holds if and only if there exists a function l(n), slowly varying at infinity, such that, as \( n\rightarrow \infty \),
Then, we obtain that \({\textbf{P}}(\tau _0>n) {\textbf{P}}(\tau ^+_0>n)\sim \frac{C}{n}\) and arrive at the conclusion. \(\square \)
Lemma 8
Assume that the random walk S(n) is asymptotically stable. Then, there exists \(C>0\) such that for \(x,y\in {\mathbb {H}}^+\) and all \(n\ge 1\)
Proof
For \(n\ge 2\),
Applying now a concentration function inequality, we obtain
since
\(\square \)
Before proving the next lemma, we state the following result, which is a minor modification of [21, Lemma 13].
Lemma 9
Suppose that \(X\in {\mathcal {D}}(d, \alpha ,\sigma )\) Then, as \(u\rightarrow \infty \),
if \(\alpha \rho <1\), and
if \(\alpha \rho =1,\) where
In addition, there exists a constant \(C>0\) such that, in both cases,
Lemma 10
There exists a constant \(C\in \left( 0,\infty \right) \) such that, for all \(z\in [0,\infty )\times {\mathbb {R}}^{d-1},\)
In particular,
Proof
For all \(z\in [0,\infty )\times {\mathbb {R}}^{d-1},\) and all \(\varepsilon >0\), we have
Applying (25) gives
Recalling that H(x) is regularly varying with index \(\alpha \rho \) by Lemma 9 and taking into account (23), we get
Consequently,
This inequality shows that there exists a constant \(C\in (0,\infty )\) such that
as desired. \(\square \)
Corollary 11
Let \(X\in \mathcal {D}(d,\alpha ,\sigma ).\) There exists a constant \(C\in \left( 0,\infty \right) \) such that, for all \(n\ge 1\) and \(x,y\in {\mathbb {H}}^+\),
Proof
The desired estimates follow from (23) and Lemmas 7 and 8. \(\square \)
3 Baxter–Spitzer Identity
We will need the following multidimensional extension of one-dimensional Baxter–Spitzer identity, see [18, Lemma 3.2].
Lemma 12
For \(t\in {\mathbb {R}}^d\) and \(|s |<1\), the following identity
We will now follow closely [21]. Put
and \(b_n(y):= p_n(0,y)\). Lemma 15 of [21] extends as follows.
Lemma 13
The sequence of functions \(\{\, B_n(y), n\ge 1\,\}\) satisfies the recurrence equations
and
The proof is analogous to the proof of Lemma 15 of [21].
To deal with random walks started at an arbitrary point, we will prove Lemma 14 that extends (17) in [10]. Put
We will slightly abuse the notation and write
Lemma 14
For \(x\in \overline{{\mathbb {H}}^+},y \in {\mathbb {H}}^+\), we have
and for \(x\in \overline{{\mathbb {H}}^+} \cap {\mathbb {Z}},y \in {\mathbb {H}}^+\cap {\mathbb {Z}}\)
Proof
Decomposing the trajectory of the walk at the minimum of the first coordinate and using the duality lemma for random walks, we get
Integrating (28), we obtain the second equality (29). \(\square \)
4 Probabilities of Normal Deviations: Proof of Theorem 2 for \(x=0\).
Let \(H_{y_1}^+ = \{(z_1,z^{(2,d)}): 0<z_1<y_1 )\}\). It follows from (26) that
where, for any fix \(\varepsilon \in (0,1/2)\) and with a slight abuse of notation,
and
In what follows, we will show that the main contribution is due to \(R_{\varepsilon }^{(2)}(y)\) and other terms are negligible.
Fix some \(t\in {\mathbb {Z}}^d\). If \(z\in (t+\Delta )\), then
Consequently,
for all \(z\in t+\Delta \). Applying this estimate, we conclude that, for every \(A\subset {\mathbb {R}}^d\),
Combining this estimate with Corollary 11 with \(x=0\), we obtain
In order to bound \({R}^{(0)}_\varepsilon (y)\), we apply (31) with \(A=[y_1,y_1+1)\times {\mathbb {R}}^{d-1}\):
Noting that \({\textbf{P}}(S_1(k)\in (0,2))\rightarrow 0\) as \(k\rightarrow \infty \), we get
Taking into account (23), we conclude that
Using the Stone theorem, we obtain
Further, by (19),
As a result, we obtain
Applying (31) with \(A=[0,y_1)\times {\mathbb {R}}^{d-1}\), we get
From this estimate and (23), we deduce
Thus, in the non-lattice case we combine the Stone local limit theorem with the first equality in (8) and obtain, uniformly in \(y\in {\mathbb {H}}^+\),
According to (19),
Hence,
Since \(c_{k}\) and \(\textbf{P}(\tau >k)\) are regularly varying and \( g_{\alpha ,\sigma }(x)\) is uniformly continuous in \({\mathbb {R}}^d\), we let, for brevity, \(v=x/c_{n}\) and continue the previous estimates for \( R_{\varepsilon }^{(2)}(y)\) with
where, for \(0\le w_{1}\le w_{2}\le 1\),
Observe that, by boundedness of \(g_{\alpha ,\sigma }\left( y\right) \),
Further, it follows from (24) that \(\int \phi (u){\textbf{P}}(M_{\alpha ,\sigma }\in \textrm{d}u)\le C\int \phi (u)\textrm{d}u\) for every non-negative integrable function \(\phi \). Therefore,
As a result, we have
Combining (32)–(36) with representation (30) leads to
Since \(\varepsilon >0\) is arbitrary, it follows that, as \(n\rightarrow \infty \),
uniformly in \(y\in {\mathbb {H}}^+\). Recalling (8), we deduce by integration of ( 38) and evident transformations that
for all \(u\in {\mathbb {H}}^+, r>0\). This means, in particular, that the distribution of \(M_{\alpha ,\sigma }\) is absolutely continuous. Furthermore, it is not difficult to see that \(z\mapsto f(0,1;z)\) is a continuous mapping. Hence, in view of (39), we may consider f(0, 1; z) as a continuous version of the density of the distribution of \(M_{\alpha ,\sigma }\) and let \( p_{M_{\alpha ,\sigma }}(z):=\) f(0, 1; z). This and (38) imply the statement of Theorem 2 for \(\Delta =[0,1)^d\). To establish the desired result for arbitrary \(r\Delta , r>0\), it suffices to consider the random walk S(n)/r and to observe that
5 Probabilities of Normal Deviations When Random Walks Start at an Arbitrary Starting Point
Proof of Theorem 1
Due to the shift invariance in any direction orthogonal to \((1,0,\ldots ,0)\), we may consider, without loss of generality, only the case when the random walk starts at \(x=(x_1,0,\ldots ,0)\) with some \(x_1>0\).
As we have already mentioned before, repeating the arguments from [9] one can easily show that \({\textbf{P}}(\frac{S(n)}{c_n}\in \cdot \mid \tau >n)\) and \({\textbf{P}}(\frac{S(n)}{c_n}\in \cdot \mid \tau ^+>n)\) converge weakly. Recall also that the limit of \({\textbf{P}}(\frac{S(n)}{c_n})\in \cdot \mid \tau >n)\) is denoted by \(M_{\alpha ,\sigma }\).
Fix an arbitrary Borel set \(A\subset \mathbb {H}^+\). According to Lemma 14,
Choose now a sequence \(\{N_n\}\) of integers satisfying
We start our analysis of the sum in (40) by noting that
Applying the second statement of Lemma 7 to the walk \(S_1(k)\) and recalling that the sequence \(\{c_k\}\) is regularly varying, we obtain
Taking into account (41), we conclude that
uniformly in \(x_1\le \delta _nc_n\). Consequently,
uniformly in \(x_1\le \delta _nc_n\).
Using once again Lemma 7, we obtain
Since \({\textbf{P}}(\tau >j)\) is also regularly varying, we conclude that
uniformly in \(x_1\le \delta _n c_n\).
Choose now a sequence \(\varepsilon _n\rightarrow 0\) so that \(c_{N_n}=o(\varepsilon _n c_n)\). By the convergence in the case of start at zero,
uniformly in \(k\le N_n\) and \(z\in B_{\varepsilon _n c_n}(0)\), where \(B_r(y)\) denotes the ball of radius r with centre at y. Therefore,
By the definition of V,
Using here (43), we obtain
uniformly in \(x_1\le \delta _nc_n\).
Furthermore,
Having all these estimates, one can easily see that it suffices to show that, uniformly in \(x_1\le \delta _nc_n\),
Indeed, applying (48) to (46) and (47) leads us to the equality
Plugging this and estimates (44), (45) into (40), we get
uniformly in \(x_1\le \delta _nc_n\). Recalling that \({\textbf{P}}(\tau _x>n)\sim V(x_1){\textbf{P}}(\tau >n)\), we have, uniform in \(x_1\le \delta _nc_n\), the convergence
To prove (48), we fix some \(R\ge 1\) and notice that
Similar to (42),
Furthermore, using the convergence of measures \({\textbf{P}}(\frac{S(n)}{c_n}\in \cdot \mid \tau ^+>n)\), we have
uniformly in \(R\le N_n\). Fix some \(\gamma >0\) and take R such that \(c_{R-1}<x_1/\gamma \) and \(c_R\ge x_1/\gamma \). Then,
and, by Lemma 9,
Combining these estimates, we conclude that
Since \(\gamma \) can be chosen arbitrary small, we get (48). Thus, the proof of the theorem is complete. \(\square \)
Proof
(First proof of Theorem 2) We give a proof in the non-lattice case only. The lattice case is even simpler.
As in the proof of Theorem 1, it suffices to consider the case \(x=(x_1,0,\ldots ,0)\) with \(x_1\in (0,\delta _nc_n]\). The case \(x_1=0\) is considered in Sect. 4. There, we have proven that, uniformly in \(y\in \mathbb {H}^+\),
To generalise this relation to the case of positive \(x_1\), we first notice that, by Lemma 14,
Fix some \(\gamma \in (0,1/2)\). The analysis of
is very similar to our proof of Theorem 1. If \(k\le (1-\gamma )n\), then we have the bound
which is an immediate consequence of (49). Using this uniform bound and the local limit theorem (49) directly, and repeating our arguments from the proof of Theorem 1, one can easily obtain
uniformly in \(x_1\le \delta _n c_n\).
For \(k>(1-\gamma )n\), the mentioned above bound for \(b_{n-k}\) is useless, and one needs an additional argument. We first notice that
Applying now the second statement of Lemma 7, we get
Consequently,
Noting now that the regular variation of \({\textbf{P}}(\tau >j)\) implies
we conclude that
for all \(x_1\le \delta _n c_n.\) Plugging this and (51) into (50) and letting \(\gamma \rightarrow 0\), we get the desired result. \(\square \)
Proof
(Second proof of Theorem 2) If the local assumption (14) holds, then we can use an approach similar to that of in [5], see Theorems 5 and 6 there. This approach allows one to avoid considering first the special case \(x=0\), as it is done in Sect. 4. Without loss of generality, we may assume that \(x=(x_1,0,\ldots ,0)\).
We first notice that if y is such that \(y_1\le 2\varepsilon c_n\), then, combining Lemmas 7 and 9,
Thus, uniformly in \(y\in \mathbb {H}^+\) with \(y\le 2\varepsilon c_n\),
Combining this bound with the fact that \(p_{M_{\alpha ,\sigma }}(z)\) goes to 0 as \(z\rightarrow \partial \mathbb {H}^+\), we conclude that
uniformly in x with \(x_1\le \delta _nc_n.\)
We next consider large values of y. More precisely, we assume that \(|y |>3A c_n\) with some \(A>1\). In this case, we have, by the Markov property at time \(m=[n/2]\),
where \(I(x)=\{z: |z-x |\le Ac_n\}\). If \(z\in I(x)\), then \(|y-z |>Ac_n\) for all sufficiently large values of n. Using (73), we have
Therefore,
Furthermore, using the standard concentration function estimate, we have
Combining these bounds, one gets easily
Applying now the integral limit theorem, we conclude that
uniformly in x with \(x_1\le \delta _nc_n.\)
Thus, it remains to consider y such that \(y_1>2\varepsilon c_n\) and \(|y |\le 3Ac_n\). To analyse this range of values of y, we set \(m=[(1-\gamma )n]\) with some \(\gamma <1/2\). Let \(B_{\varepsilon c_n}(y)\) denote the ball of radius \(\varepsilon c_n\) around y. Then, by the Markov property at time m, we have
Using the large deviations bound (74), one gets easily
Consequently,
For the integral over \(B_{\varepsilon c_n}(y)\), we have
By the strong Markov property,
Noting that \(|y-u |>\varepsilon c_n\) for all \(u\in \mathbb {H}^-\) and using once again (74), we obtain
Consequently,
By the Stone local limit theorem,
Recalling that \(g_{\alpha ,\sigma }\) is bounded, we then get
Applying now the integral limit theorem, we infer that
Since \(g_{\alpha ,\sigma }(y)\le C|y |^{-d-\alpha }\),
Finally, we notice that, uniformly in \(w\in \mathbb {H}^+\),
Combining (56)–(59), we conclude that
From this relation and from (55), we get
uniformly in x with \(x_1\le \delta _n c_n\). This completes the proof of the theorem. \(\square \)
6 Probabilities of Small Deviations When Random Walk Starts at the Origin
Proposition 15
Suppose \(X\in \mathcal {D}(d,\alpha ,\sigma )\) and the distribution of X is non-lattice. Then,
uniformly in \(y_{1}\in (0,\delta _{n}c_{n}]\), where \(\delta _{n}\rightarrow 0\) as \(n\rightarrow \infty \).
Proof
First, using once again (31), we get
When \(\alpha \in (1,2]\), using the Stone theorem, we proceed as follows:
Now, we will consider the case \(\alpha \le 1\). Fix \(\beta >0\) and notice that \(1/\alpha -1+\beta >0\) for any \(\alpha \le 1\). Since \(c_n\) is regularly varying of index \(1/\alpha \), by Potter’s bounds, there exists \(C>0\), such that for \(k\le n\),
Then, for the sequence \(\gamma _n=\delta _n^{\frac{1}{2(1/\alpha -1+\beta )}} \rightarrow 0\),
Therefore,
and, as a result, for any fixed \(\varepsilon >0\),
Since (61) holds for any fixed \(\varepsilon >0\), there exists a sequence \(\varepsilon _n\downarrow 0\), such that (61) is still true. Moreover, it will be true for any \(\varepsilon '_n\downarrow 0\) such that \(\varepsilon '_n\ge \varepsilon _n\). We will assume now that \(\varepsilon = \varepsilon _n\). It is clear and will be used in the subsequent proof that we can increase \(\varepsilon _n\) and (61) will hold as long as \(\varepsilon _n<1/2\).
Now, we represent
where
Let \(\varepsilon _n\) an arbitrary converging to zero sequence of positive numbers. First, by the Stone local limit theorem,
where \(\Delta _{1}(n-k,y)\rightarrow 0\) uniformly in z such that \(z_1\in (0,\delta _{n}c_{n})\) and \(k\in [1, \varepsilon _n n]\). Therefore,
where \(\Delta _{1}(n,y)\rightarrow 0\) uniformly in y with \(y_1\in (0,\delta _{n}c_{n}) \) and \(k\in [1, \varepsilon _n n]\). Now, we represent
For any fixed \(\varepsilon >0\), by Corollary 22 of [21],
uniformly in in y with \(y_1\le \delta _nc_n\). Hence, this bound holds for some sequence \(\varepsilon _n\downarrow 0\). Increasing the original sequence \(\varepsilon _n\) if needed, we obtain
uniformly in in y with \(y_1\le \delta _nc_n\).
Fix a large positive number A and let
Note that \(c_{c^\leftarrow (y)}>y\) and \(c_{c^\leftarrow (y)-1}\le y\). Recalling the definitions of sequence \(c_n\) and of function \(\mu \) in (5) and (6) and using the fact that \(\mu (y)\) is regularly varying, we infer that \(c^\leftarrow (y)\sim 1/(y)\) as \(y\rightarrow \infty \). Then,
where
and
We can estimate \(R_{\varepsilon }^{(4,2)}(x)\) similarly to the above
Clearly, taking A sufficiently large we can make this term much smaller than \(H(y_1)\). By Theorem 1,
uniformly in \(k\le \varepsilon _n n\). Combining this with (23), we conclude that
uniformly in y with \(y_1\le \delta _n c_n\). Hence,
uniformly in y with \(y_1\le \delta _nc_n\). Combining (62), (63), (64) and (65), we obtain that
where \(\Delta _{2}(n,y)\rightarrow 0\) uniformly in y such that \(y_1\in (0,\delta _{n}c_{n})\). Using (65) and the Stone local limit theorem, one can easily conclude that
uniformly in y such that \(y_1\in (0,\delta _{n}c_{n})\).
Analysis of \(R^{(5)}_\varepsilon (y)\) is very similar to that of \(R^{(4)}_\varepsilon (y)\). First, we make use of the Stone theorem,
where \(\Delta _{3}(n,y)\rightarrow 0\) uniformly in y with \(y_1\in (0,\delta _{n}c_{n})\). Then, using the same arguments as above, we obtain
where \(\Delta _{4}(n,y)\rightarrow 0\) uniformly in y with \(y_1\in (0,\delta _{n}c_{n})\). Integrating by parts, we can complete the proof now. \(\square \)
7 Probabilities of Small Deviations When Random Walks Start at an Arbitrary Starting Point
Proof of Theorem 3 in the lattice case
For the lattice distribution, we have from (28)
Let \(N_n\) be the sequence of integers, which was constructed in the proof of Theorem 1. We shall use the representation
where
To estimate \(P_2(x,y)\), we shall proceed as in the analysis of normal deviations. Note that by Lemma 7,
Then,
Using now (43) and increasing, if needed \(N_n\), we conclude that
Now, we will consider the first term \(P_1(x,y)\). Let \(\varepsilon _n\downarrow 0 \) be the sequence, which we define in the proof of Theorem 1. We will need the following sets:
Applying now the asymptotics for small deviations of walks starting at zero, we get
Next, we note that
and
Taking into account (48), we obtain
where we also replaced \(\sum _{0}^{N_n}\) by \(\sum _{0}^{\infty }\) using the arguments in the proof of Proposition 11 in [10]. Analogous arguments give us
Then, the arguments at the end of the proof of Proposition 11 in [10] give
\(\square \)
Proof of Theorem 3 in the non-lattice case
The proof is very similar to the proof in the lattice case.
For the non-lattice distribution, we will make use of (29). We split the sum as follows:
where
There is virtually no difference in estimates for \(P_2(x,y)\). So repeating the same arguments we obtain
Now, we will consider the first term \(P_1(x,y)\). We will need the following sets:
Now, we have,
Now, note that by (48)
provided that \(\gamma _n\) and \(\varepsilon _n\) converge to 0 sufficiently slowly. As a result,
where we also replaced \(\sum _{0}^{\varepsilon _n n}\) by \(\sum _{0}^{\infty }\) using similar arguments. Analogous arguments give us
Now, note that integration by parts of the first integral gives
As a result,
\(\square \)
8 Probabilities of Large Deviations When Random Walk Starts at the Origin
We will need the following large deviations estimates.
Proposition 16
Let \(X\sim {\mathcal {D}}(d,\alpha ,\sigma )\) with some \(\alpha <2\).
Then, there exists constant \(C_H\) such that for \(|x |\ge c_n\) we have
If, in addition, (14) holds, then
This result is proven in [1, Theorem 2.6] in the lattice case. We omit the proof of non-lattice case, as it can be done very similarly to [1, Theorem 2.6].
Using Definition (5) of \(c_n\), we obtain from Corollary 11 the following upper bound. (Recall that \(g(r)=\frac{\phi (|r |)}{r^d}\).)
Lemma 17
For any \(A>1\), there exists \(c_A\) such that
for x with \(c_n\le |x |\le Ac_n\).
The main goal of this section is to obtain an upper bound for \(b_n(x)\) in the case \(|x |>Ac_n\). We now obtain a bound, which will be valid also for \(|x |>Ac_n\).
Lemma 18
Suppose that X is asymptotically stable with \(\alpha \in (0,2)\). If (14) holds, then there exists \(\gamma >0\) such that for all y with \(|y |>c_n\) we have
Proof
We will first introduce some constants and sequences that will be used throughout the proof. Set
Fix \(\delta \in (0,1)\) such that
and let \({\widetilde{A}}\) be such that
for y with \(|y |>1\) and \(k\ge 1\). Let A be such that
By Lemma 17, there exists \(c_A>1\) such that (76) holds for y with \(c_n\le |y |\le Ac_n\).
The proof will be done by induction. We will inductively construct an increasing sequence \(\gamma _n\) such that
for y with \(|y |>c_n\) and \(n\ge 1\). Then, we will show that \(\sup _n \gamma _n <\infty \). We put \(\gamma _1=c_A\) and then the base of induction \(n=1\) is immediate. Since \(\gamma _n\) will be increasing, it follows from the definition of \(A\) that (80) holds for y such that \(|c_n |<y\le A c_n\). Hence, we will consider only y with \(|y |>c_n\).
Assume now that we have already constructed the elements \(\gamma _k\) for \(k\le n-1\). We shall construct the next value \(\gamma _n\). It follows from (26) that
where
Using first the local large deviations bound (73) and then the regular variation of g, we get
Second, integrating by parts and using then Definition (79) of A and the induction assumption, we obtain
Third, using the induction assumption and (74),
We can estimate the integral as follows:
using Definition (78) of \({\widetilde{A}}\). Hence,
Combining (82), (83) and (84), we obtain that
Then, for
inequality (80) holds. Then, using Definition (77) of \(\delta \) it is not difficult to show that
Hence, the statement of the lemma holds with
\(\square \)
9 Probabilities of Large Deviations When Random Walks Start at an Arbitrary Starting Point
Proof of Theorem 4
Let
If \(|y-z |\ge \frac{1}{2}|y-x |\), then, by Lemma 18,
This implies that
By the same argument,
These estimates give the desired bound. \(\square \)
10 Asymptotics for the Green Function Near the Boundary
Proof of Theorem 5
We consider the lattice case only. Fix \(A>0\). Then,
Using Theorem 4, we obtain
For the second term, make use of Theorem 3
We will now analyse the series. Using the regular variation of \(c_n\), we can write it as
as \(|x-y|\rightarrow \infty \).
Here, we used the fact that since \(c_n\) is regularly varying of index \(1/\alpha \), for a fixed \(z>0\),
as \(|x-y|\rightarrow \infty \). Thus, in the general case,
Letting \(A\rightarrow \infty \) and substituting \(z^{1/\alpha }=t\), we arrive at the conclusion. Noting that in the isotropic case the ratio \(\frac{y_{2,d}-x_{2,d}}{\left|x-y \right|}\) belongs asymptotically to the unit sphere, we obtain the result in this case as well.
To obtain the upper bound (16) in the analysis of the second term, we make use of Lemma 7 instead of of Theorem 3. \(\square \)
Data Availability
Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
References
Berger, Q.: Strong renewal theorems and local large deviations for multivariate random walks and renewals. Electron. J. Probab. 24, 1–47 (2019)
Bingham, N.H., Goldie, C.M., Teugels, J.L.: Regular Variation, p. 494. Cambridge University Press, Cambridge (1987)
Caravenna, F., Chaumont, L.: An invariance principle for random walk bridges conditioned to stay positive. Electron. J. Probab. 18, 1–32 (2013)
Denisov, D., Wachtel, V.: Conditional limit theorems for ordered random walks. Electron. J. Probab. 15, 292–322 (2010)
Denisov, D., Wachtel, V.: Random walks in cones. Ann. Probab. 43, 992–1044 (2015)
Denisov, D., Wachtel, V.: Exact asymptotics for the moment of crossing a curved boundary by an asymptotically stable random walk. Theory Prob. Appl. 60(3), 481–500 (2016)
Denisov, D., Wachtel, V.: Alternative constructions of a harmonic function for a random walk in a cone. Electron. J. Probab. 24, 1–26 (2019)
Denisov, D., Wachtel, V.: Random walks in cones revisited. arXiv:2112.10244 (2021)
Doney, R.A.: Conditional limit theorems for asymptotically stable random walks. Z. Wahrsch. Verw. Gebiete 70, 351–360 (1985)
Doney, R.A.: Local behaviour of first passage probabilities. Probab. Theory Relat. Fields 152, 559–88 (2012)
Doney, R., Jones, E.: Large deviation results for random walks conditioned to stay positive. Electron. Commun. Probab. 17, 1–11 (2012)
Duraj, J., Wachtel, V.: Invariance principles for random walks in cones. Stoch. Process. Appl. 130, 3920–3942 (2020)
Duraj, J., Raschel, K., Tarrago, P., Wachtel, V.: Martin boundary of random walks in convex cones Ann. H. Lebesgue 5, 559–609 (2022)
Feller, W.: An Introduction to Probability Theory and Its Applications, vol. 2. Willey, New York (1971)
Nagaev, S.V.: Large deviations of sums of independent random variables. Ann. Probab. 7, 745–789 (1979)
Rvacheva, E.L.: On domains of attraction of multi-dimensional distributions. In: Selected Translations in Mathematical Statistics and Probability, vol. 2. American Mathematical Society, Providence, pp. 183–205 (1962)
Rogozin, B.A.: On the distribution of the first ladder moment and height and fluctuations of a random walk. Theory Probab. Appl. 16, 575–595 (1971)
Tamura, Y., Tanaka, H.: On a fluctuation identity for multidimensional Lévy processes. Tokyo J. Math. 25, 363–380 (2002)
Tamura, Y., Tanaka, H.: On a formula on the potential operators of absorbing Lévy processes in the half space. Stoch. Proc. Appl. 118(2), 199–212 (2008)
Uchiyama, K.: Green’s functions of random walks on the upper half plane. Tohoku Math. J. 66(2), 289–307 (2014)
Vatutin, V.A., Wachtel, V.: Local probabilities for random walks conditioned to stay positive. Probab. Theory Relat. Fields 143, 177–217 (2009)
Williamson, J.A.: Random walks and Riesz kernels. Pac. J. Math. 25(2), 393–415 (1968)
Zolotarev, V.M.: Mellin–Stiltjes transform in probability theory. Theory Probab. Appl. 2, 433–460 (1957)
Acknowledgements
D. Denisov was supported by Leverhulme Trust Research Project Grant RPG-2021-105. V. Wachtel was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)—Project ID 317210226—SFB 1283.
Author information
Authors and Affiliations
Contributions
All authors contributed equally to this work. All authors reviewed the manuscript.
Corresponding author
Ethics declarations
Conflict of interest
The authors have no competing interests to declare that are relevant to the content of this article.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Denisov, D., Wachtel, V. Green Function for an Asymptotically Stable Random Walk in a Half Space. J Theor Probab 37, 1745–1786 (2024). https://doi.org/10.1007/s10959-023-01283-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-023-01283-4