Abstract
The problem of evaluating the accuracy of Poisson approximation to the distribution of a sum of independent integer-valued random variables has attracted a lot of attention in the past six decades. Among authors who contributed to the topic are Prokhorov, Kolmogorov, LeCam, Shorgin, Barbour, Hall, Deheuvels, Pfeifer, Roos, and many others. From a practical point of view, the problem has important applications in insurance, reliability theory, extreme value theory, etc. From a theoretical point of view, the topic provides insights into Kolmogorov’s problem concerning the accuracy of approximation of the distribution of a sum of independent random variables by infinitely divisible laws. The task of establishing an estimate of the accuracy of Poisson approximation with a correct (the best possible) constant at the leading term remained open for decades. We present a solution to that problem in the case where the accuracy of approximation is evaluated in terms of the point metric. We generalise and sharpen the corresponding inequalities established by preceding authors. A new result is established for the intensively studied topic of compound Poisson approximation to the distribution of a sum of integer-valued r.v.s.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The task of approximating the distribution of a sum of independent random variables lies at the heart of the probability theory. The central role is played by the normal approximation. However, in situations where one deals with rare events Poisson or compound Poisson approximation may be preferable (cf. [4, 5, 35, 36, 40]).
Interest to the topic of Poisson/compound Poisson approximation arises in connection with applications in extreme value theory, insurance, reliability theory, etc. (cf. [6, 7, 29, 33]). The theory of Poisson/compound Poisson approximation underpins the extreme value theory [29, 33].
Let \(X_1,X_2,\ldots \) be integer-valued non-negative random variables (r.v.s). Denote \(S_0=0,\)
For example, given r.v.s \(\xi _1,\xi _2,\ldots \), one may deal with exceedances of a chosen “threshold” x. Set Then
denotes the number of exceedances of threshold x. In particular,
where \(\xi _{k,n}\) denotes the k-th largest element among \(\xi _1,\ldots ,\xi _n\). Thus, results concerning the distribution of sample extremes can be derived from the corresponding results concerning \(\mathcal{L}(S_n)\).
In applications, indicators may be dependent. A well-known approach (Bernstein’s blocks method [14]) consists of grouping observations into blocks which can be considered almost independent. The number of r.v.s in a block is an integer-valued random variable; hence, the number of rare events is a sum of almost independent integer-valued r.v.s.
In (re)insurance applications, the sum of integer-valued r.v.s allows to account for the total loss from the claims \(\{Y_i\}\) exceeding threshold x [22]. More information concerning applications can be found in [6, 7, 22, 29, 35, 36].
The distribution of a sum \(S_n(x)\) of integer-valued non-negative random variables can often be approximated by a Poisson or compound Poisson law. In early 1950s, Kolmogorov has formulated the task of evaluating the accuracy of approximation of the distribution of a sum \(S_n\) of independent and identically distributed (i.i.d.) r.v.s by infinitely divisible distributions (Kolmogorov’s uniform approximation problem). The topic has attracted a lot of attention among researchers (see, e.g., [4, 19, 35, 36, 40, 49] and references therein).
From a theoretical point of view, the question concerning the accuracy of Poisson/compound Poisson approximation is a particular case of Kolmogorov’s problem. Besides, it was noticed that estimates of the accuracy of approximation to the Binomial distribution can provide important insights in other areas of probability theory [32, 38]. In a sense, the Binomial distribution plays the role of a “testing stone” [4].
Note that there is a strong connection between the topics of Poisson and compound Poisson approximation [35, 36, 38, 48]; the latter plays a special role in approximating \(\mathcal{L}(S_n)\) by infinitely divisible laws since the class of infinitely divisible distributions coincides with the class of weak limits of compound Poisson distributions [26, Theorem 26].
In a range of situations, both normal and (compound) Poisson approximations can be applicable (cf. [4, 40]). Due to the complex structure of the compound Poisson distribution, in applications one would prefer pure Poisson approximation where possible.
One can choose between possible types of approximation by comparing estimates of the accuracy of approximation. Obviously, one would make a choice according to the sharpest estimate, thus the need of sharp estimates of the accuracy of approximation with explicit constants.
1.1 Notation
Let \(\mathcal{S}\) denote the class of measurable functions taking values in [0; 1]. Then
denotes the total variation distance between the distributions of r.v.s X and Y.
We denote by
the uniform (Kolmogorov’s) distance between the distributions of random variables X and Y with distribution functions (d.f.s) \(F_X\) and \(F_Y\).
The Gini–Kantorovich distance between the distributions of r.v.s X and Y with finite first moments (known also as the Kantorovich–Wasserstein distance) is
where \(\mathcal{L} = \{g\!:|g(x)\!-\!g(y)|\le |x\!-\!y|\}\) is the set of Lipschitz functions. Note that
where the infimum is taken over all random pairs \((X',Y')\) such that \(\mathcal{L}(X')=\mathcal{L}(X)\) and \(\mathcal{L}(Y')=\mathcal{L}(Y)\).
Set \(\Vert f\Vert = \sup _k |f(k)|.\) We denote by \(\Vert \cdot \Vert _{_1}\) the \(L_1\)-norm of a function. Given a discrete r.v. Y, we denote
In the case of discrete distributions, it is natural to exploit the point metric
where \(\Vert \cdot \Vert \) denotes the sup-norm.
Initially, estimates of the accuracy of Poisson approximation to the Binomial distribution \(\textbf{B}(n,p)\) were established in terms of the uniform distance \(d_K\) and the total variation distance \(d_{_{TV}}\!\) [27, 40, 47]. Metrics \(d_{_{TV}}\!,\ d_K,\ d_\textrm{o}\) have obvious merits. For instance, they are shift- and scale invariant. The \(O(n^{-2/3})\) estimate of the accuracy of approximation in Kolmogorov’s problem holds in terms of \(d_K\) [1,2,3] (and hence in terms of \(d_\textrm{o}\)) but generally not in terms of \(d_{_{TV}}\!\) [49]. Bounds in terms of \(d_\textrm{o}\) can be sharper than those in terms of the uniform and the total variation distances.
In extreme value theory, one is interested in the distribution of the k-th largest sample element \(X_{n:k}\). Metric \(d_\textrm{o}\) appears more suitable than \(d_K\) and \(d_{_{TV}}\!\) if one evaluates probabilities like \(I\!\!P(S_n(x)\!<\!k)\) for “small” k, cf. (1). A similar observation can be made concerning the length of the longest head run (LLHR), the length of the longest match pattern (LLMP), etc. (see [35]). Metric \(d_\textrm{o}\) is shift and scale invariant. Another advantage of using \(d_\textrm{o}\) in comparison with \(d_{_{TV}}\!,\ d_K\) is the higher rate of approximation (cf. (4), (8)).
1.2 Independent Bernoulli r.v.s
Hereinafter, \(\{X_i\}\) are independent non-negative integer-valued random variables; multiplication is superior to the division. W.l.o.g. we may assume that \(I\!\!EX_i\!>\!0\ (\forall i)\).
Let \(\pi _\lambda \) denote a Poisson \(\mathbf{\Pi }(\lambda )\) r.v. Many authors worked on the problem of evaluating the accuracy of approximation
in the case where \(\{X_i\}\) are 0-1 r.v.s (see, e.g., [4, 7, 35, 36] and references therein). The problem goes back to Prokhorov [40]. It has attracted a lot of attention among specialists (see, e.g., [35, 36] and references therein).
Estimates of the accuracy of Poisson approximation to the distribution of a sum of independent 0-1 random variables in terms of the uniform distance \(d_K\) and the total variation distance \(d_{_{TV}}\!\) have been derived by Kolmogorov [27, Lemma 5], Tsaregradskii [47], LeCam [30], Kerstan [25], Romanowska [41], Shorgin [46], Barbour and Eagleson [12] and other authors (see, e.g., [20, 34, 39, 45]). Concerning estimates in terms of some other metrics, see [15, 21, 35,36,37] and references therein.
In the case of independent 0-1 r.v.s estimates with correct (the best possible) constants at the leading terms have been found by Roos [45]:
where \(\theta =\sum ^n_{i=1} p_i^2/\lambda ,\,p_i=I\!\!P(X_i\!=\!1)\ \ (i\!\ge \!1)\); constants at the leading terms in (2)–(4) cannot be improved. Note that \(3/4e\approx 0.276,\) \(1/2e \le 0.184,\ \frac{1}{2} (3/2e)^{3/2}\le 0.205\).
Estimate (3) has a sharper constant at the leading term than that in (2), estimate (4) has a sharper rate of decay if \(\lambda \!\equiv \!\lambda (n)\!\rightarrow \!\infty \) as \(n\!\rightarrow \!\infty .\)
Estimates of the accuracy of shifted (translated) Poisson approximation to the distribution of a sum of independent 0-1 r.v.s have been given by Barbour and Xia [9], Čekanavičius and Vaitkus [18], Barbour and Čekanavičius [11], Röllin [43], Novak [34, 37]. Kruopis [28] has evaluated the accuracy of shifted Poisson approximation to \(\mathcal{L}(S_n)\) in terms of \(d_{\textrm{o}}\!:\)
where \( Y = \lfloor \lambda _2\!+\!0.5\rfloor + \pi _{\lambda -\lfloor \lambda _2+0.5\rfloor }\), \(\lambda _2 \!=\! \sum _{j=1}^n p_j^2\). Note that \(I\!\!EY\!=\!\lambda ,\) \(|{\textrm{var}\,}Y\!-\!{\textrm{var}\,}S_n|\!\le \!1/2.\) In the case of the Binomial distribution \(\textbf{B}(n,p)\), the right-hand side (r.h.s) of (5) is of order \(\sqrt{p/n}\).
1.3 Integer-valued r.v.s
Let \(\{X_i\}_{i\ge 1}\) be independent non-negative integer-valued r.v.s. The problem of evaluating the accuracy of Poisson approximation to the distribution of a sum of independent non-negative integer-valued r.v.s has been considered, e.g. in [8, 11, 13, 21, 25, 33, 37]; an overview of the results on the topic can be found in [35, 36].
Röllin [43, formula (2.13)], assuming that third moments are finite, states that
where \(\sigma ^2=\hbox {var }S_n,\) \(\psi _i = \sigma ^2_i I\!\!EX_i(X_i\!-\!1) \!+\! |I\!\!EX_i\!-\!\sigma ^2_i| I\!\!E(X_i\!-\!1)(X_i\!-\!2) \!+\! I\!\!E|X_i(X_i\!-\!1)(X_i\!-\!2)|,\) \(d{\mathbb \,'} = \max _{i\le n} \Vert I\!\!P_{S_{n,i}+2}-2I\!\!P_{S_{n,i}+1}+I\!\!P_{S_{n,i}}\Vert _{_1}/2,\) \(S_{n,i}=S_n\!-\!X_i\), \([x]=\max \{k\!\in \!\textbf{Z}\!: k\!\le \!x\},\) \(\{x\}=x\!-\![x]\). In the case of the Binomial distribution \(\textbf{B}(n,p)\) bound (6) is of order 1/np.
Tsaregradskii [47] has shown that the rate of the accuracy of compound Poisson approximation to the Binomial distribution \(\textbf{B}(n,p)\) in terms of the uniform distance is O(1/np). Presman [38] has shown that
where C is an absolute constant and \(P_{n,p}\) is a particular compound Poisson distribution. Hence,
According to Čekanavičius [16], there exists an absolute constant C such that
In the case of i.i.d.r.v.s \(X,X_1,X_2,\ldots \) obeying \(I\!\!EX^4\!<\!\infty \) Čekanavičius [17] has shown that the accuracy of compound Poisson approximation to \(\mathcal{L}(S_n)\) is \(C_Xn^{-3/2}\), where \(C_X\) depends on \(\mathcal{L}(X)\).
The problem of establishing an estimate of the accuracy of Poisson approximation to the distributions of a sum of independent non-negative integer-valued r.v.s in terms of the point metric with a correct constant at the leading term remained open for a long while. In particular, an open question was whether \(\frac{1}{2} (3/2e)^{3/2}\) would remain the best possible constant at the leading term in the case of integer-valued r.v.s. We give below the affirmative answer to that question. We generalise and sharpen the corresponding results from [16, 21, 43, 45].
Theorem 1 presents an estimate of the accuracy of Poisson approximation to \(\mathcal{L}(S_n)\) in terms of the point metric with a correct constant at the leading term. Theorem 2 provides a first-order asymptotic expansion. Theorem 3 presents an estimate of the accuracy of shifted Poisson approximation. Theorem 4 provides an estimate of the accuracy of compound Poisson approximation in terms of the point metric. Theorems 1–3 only assume that the second moments are finite, Theorem 4 does not impose moment requirements, and the constants are explicit.
2 Results
Let \(\{X_i\}_{i\ge 1}\) be independent non-negative integer-valued r.v.s.
First, we present an estimate of the accuracy of pure Poisson approximation.
2.1 Poisson Approximation
If random variables \(X,\xi ,\eta \) have finite second moments, let
Denote \(X_0\!=\!0,\) \(\varepsilon ^*_\lambda = 1\!\wedge \!1/\sqrt{2\pi [\lambda ]}\). If \(i\!\in \!\{0,1,\ldots ,n\},\) let \(S_{n,i}=S_n\!-\!X_i\), \(\lambda _i\!=\!I\!\!ES_{n,i}\),
where \(X^*_i\) denotes a random variable with the distribution
(cf. [33, Ch. 4.4]). Note that \(U^2_*\!+\!(1\!-\!2u^*)U_*\!+\!1/4\!\ge \!0\). If \(u^*\!\le \!1/4,\) then \(U^2_*\!+\!(1\!-\!2u^*)U_*\!+\!1/4\!\ge \!U^2\), and \(r_{i,n}^* \le \min \!\big \{ \varepsilon _{i,n};8/\pi U\big \}\).
Set
Theorem 1
If \(X_1,\ldots ,X_n\) are independent non-negative integer-valued random variables with finite second moments, then
where \(c_\textrm{o}= \frac{1}{2} (3/2e)^{3/2}\).
If \(X,X_1,\ldots ,X_n\) are i.i.d.r.v.s, set \(\theta ^*\!=\!|\kappa _X|/I\!\!EX,\)
Then (10) becomes
According to [45], constant \(c_\textrm{o}\) in (10), (10\(^*\)) cannot be improved.
Note that the moment assumption in Theorem 1 can be relaxed at a cost of adding an extra term if one uses truncation at some levels \(\{K_i\}\) (i.e. switches from \(\{X_i\}\) to \(\{X_i'\},\) where ) since \(d_{_{TV}}\!((X_1,\ldots ,\!X_n);(X_1',\ldots ,\!X_n')) \le \sum ^n_{i=1} \!I\!\!P(X_i\!>\!K_i)\).
Example 1
Let \(\{X_i\}\) be independent Bernoulli \(\textbf{B}(p_i)\) random variables, where \(p_i\!\in \![0;1/2]\ (i\!\ge \!1).\) Then \(\lambda =\sum ^n_{i=1}p_i\), \(\kappa _{S_n} = \sum ^n_{i=1}p_i^2\), \(\varepsilon _1^\textrm{o}\!=\!0,\) and (10) yields
The constant at the leading term in (11) is sharper than that in (5).
In the case of the Binomial distribution \(\textbf{B}(n,p)\), one has \(\kappa _X\!=\!p^2\), \(u_1\!=\!p\),
where \(\tilde{\varepsilon }_n = \min \!\Big \{ \sqrt{2/\pi } \Big /\!\Big (1/4\!+\!(n\!-\!1)p\Big )^{1/2}; 1/\sqrt{2\pi [(n\!-\!1)p]}\,+2(1\!-\!e^{-np})p/(1\!-\!1/n) \Big \}.\) Then
The rate of the second-order term in (12) is sharper than that in (4) if \(p\!\equiv \!p(n)\!\rightarrow \!0\) as \(n\!\rightarrow \!\infty \). Note that \(\sup _k I\!\!P(S_n\!=\!k) = O(1/\sqrt{np}\,),\) cf. [31, 42].
Example 2
Let \(X,X_1,X_2,\ldots \) be independent geometric \(\mathbf{\Gamma _0}(p)\) r.v.s:
Then \(S_n\) is a Negative Binomial \(\mathbf{N\!B}(n,p)\) r.v.:
where \(\Gamma (y) = \int _0^\infty x^{y-1}e^{-x}dx.\) It is easy to see that \(I\!\!P(X_i^*=m) = (m\!+\!1)p^m(1\!-\!p)^2\). Hence,
Set \(r\!=\!p/(1\!-\!p).\) Note that \(\lambda \!\equiv \! I\!\!ES_n \!=\!nr\).
It is easy to check that \(\kappa _X\!=\!-r^2,\) \(\gamma _{X_1,X_1^*}\!=\!4r^2,\) \(u_1\!=\!p,\) \(\theta ^*\!=\!r,\)
where \(\varepsilon _{n,p}^\star = \sqrt{2/\pi }\!\Big /\!\Big (1/4\!+\!(n\!-\!1)p\Big )^{1/2}\! \wedge \Big ( 1\big /\sqrt{2\pi [(n\!-\!1)p]}\, + 2r/(1\!-\!1/n) \Big ).\) Theorem 1 yields
Estimate (13) has a correct constant at the leading term, \(\varepsilon _X^\textrm{o}=O(p\sqrt{p/n}+\!p/n)\).
Example 3
Let \(X,X_1,\ldots \) be i.i.d.r.v.s with the distribution
where \(p\!\in \![0;1/2].\) Then \(I\!\!EX = \textrm{var}X = p\), \(\mathcal{L}(X^*)=\textbf{B}(p)\).
Note that \(I\!\!EX\!=\!I\!\!EX^*=p,\)
Hence, \(\varepsilon _1^\textrm{o}\le 8p/\pi (1\!-\!p/2)(n\!-\!1),\) and Theorem 1 yields
While the rate of the accuracy of approximation in (12) is \(\sqrt{p/n},\) the rate is p/n in (14).
2.2 A First-Order Asymptotic Expansion
For any function f, denote \(\Delta f(\cdot ) = f(\cdot +\!1)-f(\cdot ).\) The following theorem provides a first-order asymptotic expansion in terms of \(d_{\textrm{o}}\).
Theorem 2
Let \(X_1,\ldots ,X_n\) be independent non-negative integer-valued random variables with finite second moments. If where \(k\!\in \!I\!\!N,\) then
Let \(\pi ^\star _\lambda \) denote a random variable with the distribution
where \(\mathbb {Z}_+=\{0,1,2,\ldots \}\) is the set of non-negative integer numbers. Note that
(cf. [34, Remark 1]). Thus,
In particular,
According to Roos [44], \(\Big |\lambda \Vert \Delta ^2I\!\!P_{\pi _\lambda }\Vert - 1/\sqrt{2\pi \lambda }\,\Big | \le C/\lambda .\) Therefore,
If \(X_1,\ldots ,X_n\) are i.i.d. Bernoulli \(\textbf{B}(p)\) r.v.s, then for any \(k\!\in \!I\!\!N\)
where \(\varepsilon _X^\textrm{o}= 4p \left( \sqrt{2p/\pi (n\!-\!1)}\, + 4/\pi (n\!-\!1)\right) \); hence,
2.3 Shifted Poisson Approximation
Set \(\mu \!=\!\textrm{var}\,S_n\!+\!\{\kappa _{S_n}\}.\) The next theorem deals with shifted Poisson approximation
Note that \([\kappa _{S_n}] \!+\! I\!\!E\pi _{\mu } \!= I\!\!ES_n,\ \,|\textrm{var}\, \pi _{\mu }\!-\textrm{var}\,S_n|\!<\!1.\) W.l.o.g. we may assume that \(\mu \!>\!0\).
Given a r.v. Y, we denote \(\bar{Y} = Y\!-\!I\!\!EY.\) Let \(\sigma ^2\!=\!\textrm{var}\, S_n\), \(U_i \!=\! U\!-\!u_i,\)
Theorem 3
Let \(X_1,\ldots ,X_n\) be independent non-negative integer-valued random variables with finite second moments. Then
Example 1 (continued). Let \(\mathcal{L}(S_n)=\textbf{B}(n,p),\) where \(p\!\in \![0;1/2].\) Set \(q=1\!-\!p\). Clearly,
Therefore, \(\mu \!\le \!np,\) \(\hat{\varepsilon }_\mu \!\le \!2\mu ^{-1}\varepsilon _{0,n},\) and (19) yields \((n\!>\!1)\)
Using (12) as \(p\!<\!1/\sqrt{n}\) and (20) as \(1/\sqrt{n}\le \!p\!\le \!1/2,\) we derive
Indeed, (12) entails
if \(p\!<\!1/\sqrt{n}\). If \(p\!\ge \!1/\sqrt{n},\) then (20) yields
and (21) follows. A uniform in \(p\!\in \![0;1/2]\) estimate of the accuracy of Poisson approximation to the Binomial distribution \(\textbf{B}(n,p)\) in terms of the point metric seems to be new.
3 Compound Poisson Approximation
The topic of compound Poisson approximation to the distribution of a sum of random variables has attracted a lot of attention in the past decades (see, e.g., [19] and references therein).
The topic has applications in extreme value theory, insurance, reliability theory, patterns matching, etc. (cf. [6, 10, 22, 33]). In order to decide if a particular compound Poisson approximation to \(\mathcal{L}(S_n)\) is applicable, one would require an estimate of the accuracy of compound Poisson approximation indicating the distance between two distributions is “small”, hence the need of sharp bounds with explicit constants.
From a theoretical point of view, the interest to the topic arises in connection with Kolmogorov’s problem concerning the accuracy of approximation of the distribution of a sum of independent r.v.s by infinitely divisible laws (see [4, 19] and references therein). From a practical point of view, the problem has important applications in insurance, reliability theory, extreme value theory, etc., cf. [19], and references therein.
Estimates of the accuracy of compound Poisson approximation have been derived mainly in terms of the uniform distance and the total variation distance. Very few estimates of the accuracy of compound Poisson approximation are available in terms of the point metric. However, in situations where one needs to evaluate \(I\!\!P(S_n\!<\!k)\) for a “small” k the point metric may be advantages as estimates in terms of the point metric are expected to have better rate of approximation than estimates in terms of the uniform of the total variation distances, cf. (2)–(4).
By Khintchine’s formula (see [24, Ch. 2]), the distribution of any random variable X obeys
where \(\tau _p\) and \(X'\) are independent r.v.s, \(\mathcal{L}(X') = \mathcal{L}(X|X\!\ne \!0),\) \(\tau _p\) is a Bernoulli \(\textbf{B}(p)\) r.v., \(p=I\!\!P(X\!\ne \!0).\) Since \(X_1,\ldots ,X_n\) are independent r.v.s,
where \(\tau _1,X_1',\ldots ,\tau _n,X_n'\) are independent r.v.s, \(p_i=I\!\!P(X_i\!\ne \!0),\)
Assume that \(X',X_1',\ldots ,X_n'\) are i.i.d.r.v.s. Then
where r.v. \( \nu _n=\tau _1+\cdots +\tau _n\) is independent of \(\{X_i'\}\).
If \(\mathcal{L}(X')\) is degenerate (i.e. \(X'\!=\!c,\) where c is a constant), then \(S_n\!{\mathop {=}\limits ^{d}}\!c\nu _n,\) and the problem of compound Poisson approximation to \(\mathcal{L}(S_n)\) reduces to the problem of Poisson approximation to \(\mathcal{L}(\nu _n).\)
Assume that \(\mathcal{L}(X')\) is not degenerate. According to Kolmogrov [27], formula (30),
where \(\pi _\lambda \) denotes a Poisson \(\mathbf{\Pi }(\lambda )\) r.v.. Thus, an estimate of the accuracy of “accompanying” compound Poisson approximation (the terminology of Gnedenko and Kolmogorov [23]) in terms of the total variation distance follows from an estimate of the accuracy of pure Poisson approximation.
The following theorem presents an estimate of the accuracy of “accompanying” compound Poisson approximation to \(\mathcal{L}(S_n)\) in terms of the point metric.
Recall (24), where \(p_i=I\!\!P(X_i\!\ne \!0).\) Denote \(\bar{p}= \sum _{i=1}^np_i/n.\)
Theorem 4
If \(X,X_1,X_2,\ldots \) are independent integer-valued random variables, \(X',X_1',X_2',\ldots \) are i.i.d.r.v.s, where \(\mathcal{L}(X')=\mathcal{L}(X|X\!\ne \!0),\) then
If \(X,X_1,\ldots \) are i.i.d.r.v.s, then \(\bar{p}\!=\!p,\) where \(p\!=\!I\!\!P(X\!\ne \!0),\) and (27) becomes
Bound (27) appears suitable for evaluating probabilities \(I\!\!P(S_n\!<\!k)\) when k is “small”. A feature of the bound is that Theorem 4 does not impose moment requirements.
The rate of the accuracy of approximation in (27\(^*\)) is at least \(n^{-1/2},\) the rate is \(o(n^{-1/2})\) if \(p\!\equiv \!p(n)\!\rightarrow \!0\) as \(n\!\rightarrow \!\infty \).
If X is a Bernoulli r.v., then \(X'\!\equiv \!1,\) \(\sum _{i=1}^{\pi _{np}} X'_i = \pi _{np}\) is a Poisson r.v., and (27) coincides with (12). Thus, the constant at the leading term in (27) cannot in general be improved.
The advantage of employing the “accompanying” compound Poisson distribution is the simplicity of the approximating distribution. An open question is if the accuracy of compound Poisson approximation to \(\mathcal{L}(S_n)\) can be improved using more complex approximating laws? Relation (29) suggests the following hypothesis. Let \(X,X_1,X_2,\ldots \) be i.i.d.r.v.s. Denote \(P_{_X}=\mathcal{L}(X),\) and let \(\mathbb{C}\mathbb{P}\) denote the class of (shifted) compound Poisson r.v.s. An open question is if there exists an absolute constant \(C_\textrm{o}\) such that
Example 4
Suppose that \(X,X_1,\ldots \) are i.i.d. random variables,
Then \(I\!\!P(X'\!=\!1)=I\!\!P(X'\!=\!1)\!=\!1/2.\) Hence, \(\sum _{i=1}^{\pi _{np}} X'_i {\mathop {=}\limits ^{d}}\pi '\!+\!2\pi ''\), where \(\pi ',\,\pi ''\) are independent Poisson \(\mathbf{\Pi }(np/2)\) r.v.s. Note that \(\pi '\!+\!2\pi ''\) is a compound Poisson r.v. Theorem 4 yields
If \(p\!\equiv \!p(n)\!\rightarrow \!0\) as \(n\!\rightarrow \!\infty \), which is typically the case when one deals with rare events, then the r.h.s. of (28) is \(o(n^{-1/2})\).
We now present a uniform in \(p\!\in \![0;1/2]\) estimate of the accuracy of compound Poisson approximation to the Binomial distribution in terms of the point metric.
Let \(P_{n,p}\) denote the compound Poisson distribution from (7).
Proposition 5
There exists an absolute constant C such that
Bound (29) provides a better rate of approximation than (21). However, (21) offers a simpler approximating distribution, and the constants are explicit.
4 Proofs
First, we present the proof of Theorem 2, then the proof of Theorem 1.
We start with a particular lemma, which is needed in the proof of Theorem 2.
For any function \(h\!\in \!\mathcal{S}\), we denote by \(g\equiv g_h\) the solution of the Stein equation
Note that it is possible to write down the Stein equation in a different manner (see, e.g., [33, Remark 4.1]). The way shown in (30) is in line with the general approach to the characterisation of discrete distributions (cf. [33, Ch. 12]).
Lemma 6
Let for a particular \(k\!\in \!\mathbb {Z}_+\). If \(g_h\) is given by (30), then
Proof of Lemma 6
It is known that the solution of Eq. (30) is
(see, e.g., [7, 33]). The value of g(0) is irrelevant (we can set \(g(0)\!=\!0\)).
If where \(k\!\in \!\mathbb {Z}_+,\) then
Denote
It is known (see, e.g., [33, p. 82]) that function G is decreasing; function \(G_*\) is increasing. Therefore, (33) yields
and (31) follows (Röllin [43] mentions without proof that \(\Vert g_h\Vert \!\le \!1\)). \(\square \)
Proof of Theorem 2
We will use Stein’s method. The details of the method have been presented in many publications (see, e.g., [7, 33]).
According to (30),
Below we evaluate \(|I\!\!Eg(S_n\!+\!1)-\lambda ^{-1}I\!\!ES_ng(S_n)|\).
Let where \(k\!\in \!\mathbb {Z}_+.\) Then
If \(\Vert \Delta g\Vert \!=\!0,\) then g is a constant, and \(\lambda I\!\!Eg(S_n\!+\!1)-I\!\!ES_ng(S_n)=0.\) Therefore, without loss of generality we may assume that \(\Vert \Delta g\Vert \!>\!0.\)
Recall that \(S_{n,i}=S_n\!-\!X_i\). Set
It is known that
for any function f such that \(I\!\!E|X_i f(X_i)|<\infty \) (cf. [33, Ch. 4]). Therefore,
Given a function \(f\!:\mathbb {Z}_+\!\rightarrow \!I\!\!R,\) we denote \((\ell \!\ge \!0,m\!\ge \!0,k\!\ge \!0)\)
According to Proposition 4 in [35, 36],
Clearly,
It is known (see Barbour and Eagleson [12] or [7, Remark 1.1.2]) that
Thus, \(|c_1(g_i)| \le 2\Vert \Delta g_i\Vert \le 2(1\!-\!e^{-\lambda }).\) Using (51), we get
Recall that
Taking into account Proposition 6 and (42), we derive
Note that
Taking into account (40), \(\Vert \Delta ^2 g_i\Vert \le 2(1\!-\!e^{-\lambda }) \varepsilon _{i,n}\). By (51), (56) and (31),
Therefore,
Combining these estimates, we get
Here random variables \(X_i,X_i^*\) can be defined on a common probability space in such a way that \(I\!\!E|X_i\!-\!X_i^*| = d_{_G}(X_i,X_i^*).\) Notice that
(cf. (34)). Taking into account (39), (44), (45), we derive
By (51), \( I\!\!E\Delta g(Y) = -\sum _k g(k) \Delta I\!\!P_Y(k).\) Using (42) and (31), we get
This and (46) entail
Now we replace \(I\!\!E\Delta g(S_{n,i}\!+\!1)\) in (46) with \(I\!\!E\Delta g(S_n\!+\!1).\)
If \(m\!>\!k,\) then
for any function f. In particular,
According to [35], Lemma 5, for any bounded function f
An application of (50) with \(f=\Delta g_i\) yields
where \(\Vert \Delta g\Vert \!\le \!1\!-\!e^{-\lambda }\) by (40). Note that (43) and (49) yield
Therefore,
We have shown that
It remains to evaluate \(I\!\!E\Delta g(S_n\!+\!1)-I\!\!E\Delta g(\pi _\lambda \!+\!1)\).
Recall that where \(k\!\in \!\mathbb {Z}_+.\) It is known that \(\Delta g(i)\!\le \!0\) if \(i\!\ne \!k,\) while \(0\!\le \!\Delta g(k)\!\le \!1\!-\!e^{-\lambda }\) (cf. [12] or the proof of Proposition 6). Note that \(\sum _{i\ne k} \Delta g(i) = -\Delta g(k)\). Therefore,
An application of (47) yields
Thus,
Note that
for any indicator function (cf. (4.31) in [33]). Since every function \(h\!\in \!\mathcal{S}\) can be represented as where \(\{c_k\}\) are constants,
Therefore,
This and (17) lead to (15). \(\square \)
Proof of Theorem 1
If \(\lambda \!=\!0,\) then \(S_n\!=\!\pi _\lambda \!=\!0,\) and (10) trivially holds. Therefore, w.l.o.g. we may assume that \(\lambda \!>\!0\).
Given a r.v. Y, denote
In particular, \(\Delta f(\cdot ,\cdot )\) means the increment of the first argument.
Clearly, for any bounded function h
Given an arbitrary \(k\!\in \!\mathbb {Z}_+\), let . We apply (15).
Note that
Therefore,
According to Lemma 3 in Roos [45],
The proof is complete. \(\square \)
The proof of Theorem 3 requires the following
Proposition 7
For any bounded function f and any integer-valued random variable Y
As a consequence,
If \(2u^*\!\le \!1,\) then (56) yields
If \(2u^*>1,\) then (56) entails
Proof of Proposition 7
Relation (53) follows from (51): \(I\!\!E\Delta f(Y) = -\sum _k f(k) \Delta I\!\!P_Y(k),\) hence
Similarly,
Bound (55) is an immediate consequence of (53) and (42). Relation (56) will follow from (54) and the following inequality:
We proceed with the proof of (57); the argument is similar to that behind (4.9) in [11].
Let \(I_x\) denote the distribution concentrated at point x, \(I\!\equiv \!I_\textrm{0},\) and let \(*\) denote the convolution of measures. Then
Let \(Q_1,\,Q_2\) be two measures. By the property of the total variation norm,
Set \(J\!=\!\{1,\ldots ,n\}.\) If \(J=C\cup D,\) we denote \(S'=\sum _{i\in C}X_i\), \(S''=\sum _{i\in D}X_i\). Since \(I\!\!P_{S_n}=I\!\!P_{S'}*I\!\!P_{S''}\), we have
We will exploit this bound and (59).
If \(U\!=\!0,\) then (57) trivially holds. Therefore, we may assume that \(U\!>\!0.\) Set J can be split into \(C\!\cup \!D\) so that sets C, D are non-empty,
where \(U_C=\sum _{i\in C}u_i,\ U_D:=\sum _{i\in D} u_i\).
Indeed, r.v.s \(X_1,\ldots ,X_n\) (and hence numbers \(u_1,\ldots ,u_n\)) can be rearranged without affecting \(S_n\). Therefore, we may assume that \(u_0\!:=\!0\!\le \!u_1\!\le \cdots \le \!u_n\!=\!u^*\). Denote
If \(\Sigma _{n-1}\!\le \!U/2,\) then \(\nu =n,\) hence \(u_n\!>\!U/2.\) One can choose \(C=\{1,\ldots ,n\!-\!1\},\,D\!=\!\{n\}.\) Then
If \(\Sigma _{n-1}\!>\!U/2,\) then \(\nu \!\le \!n\!-\!1\) and \(u^*\!\le \!U/2.\) One can choose \(C=\{1,\ldots ,\nu \!-\!1\},\,D=\{\nu ,\ldots ,n\}.\) Then
and (58) holds.
According to Mattner and Roos [31, Corollary 1.6],
Inequality (59) can be applied to \(S',\,S''\):
One can check that \((1/4\!+\!U_C) (1/4\!+\!U_D) \ge U^2/4\!+\!(1\!-\!2u^*)U/4\!+\!1/16\). Hence,
This leads to (57) and hence to (56). The proof is complete. \(\square \)
Proof of Theorem 3
Recall that \(\mathbb {Z}\) denotes the set of integer numbers. Set
Then
We need to evaluate \(|I\!\!Eh(S_n)-I\!\!Eh(\pi _{\mu }\!+\!a)|,\) where .
W.l.o.g. we may assume that \(\mu \!>\!0\): if \(\mu \!=0,\) then (60) yields \(\sigma ^2 = \{\lambda \}\!=\!0\); hence, every \(X_i\) is a constant, \(a=\lambda ,\) and (19) trivially holds.
In the case of shifted Poisson approximation, the basic equation is
(cf. (12.26) in [33]). The solution \(f \equiv f_{h}\) of Eq. (61) is
where \(\tilde{h}(m)=h(m\!+\!a)\ (m\!\ge \!0)\) and \(g\equiv g_{\tilde{h}}\) is given by (32) with \(\lambda \) replaced with \(\mu \).
According to (61),
Below we evaluate \(|\mu I\!\!Ef(S_n\!+\!1)-I\!\!E(S_n\!-\!a)f(S_n)|\).
First, we show that
Let \(\{\tilde{X}_j\}\) denote independent copies of \(\{X_j\}\). Set \(\bar{X}_i = X_i\!-\!I\!\!EX_i,\)
Since \(r_i = f_i(X_i)-f_i(0)-X_i\Delta f_i(\tilde{X}_i),\)
Recall that \(\mu = \sigma ^2\!+\!b.\) Hence,
Note that \(|r_i| \le \delta _{X_i,0}^{\tilde{X}_i},\) \(c_1(f_i) \le 2\Vert \Delta f_i\Vert \le 4(1\!-\!e^{-\mu }) \varepsilon _{i,n}^\textrm{o}\) by (37), (41), (42), (31). Therefore,
According to (55), \(\Vert \Delta ^2 f_i\Vert \le 2\Vert \Delta g\Vert \varepsilon _{i,n},\) while (56) yields
In view of (40) and (31), \(\max \{ \Vert g\Vert ;\Vert \Delta g\Vert \} \le 1\!-\!e^{-\mu }\). Therefore,
Note that \((m\!-\!\ell )(m\!-\!\ell -\!1\!) - \ell (\ell \!+\!1) = m(m\!-\!1\!) - 2m\ell .\) Hence,
by (37), (40), (55), (31), (63). Thus,
i.e. (62) holds.
It is known (cf. [35, 36]) that
According to (51), (41), (42), (31),
Thus, (19) holds. The proof is complete. \(\square \)
Remark 1
Estimates of Theorems 1–3 work best if the maximum span of \(\mathcal{L}(X_1),\ldots ,\mathcal{L}(X_n)\) is 1. If the maximum span of \(\mathcal{L}(X_i)\) is \(>\!1\) for a particular i, then \(d_{_{TV}}\!(X_i;X_i\!+\!1)\) may be equal to 1, reducing \(U\!=\!\sum ^n_{i=1}u_i\) and hence increasing the bounds.
The following inequality is employed in the proof of Theorem 4.
Lemma 8
Let \(\tilde{\nu },\nu ,X,X_1,X_2,\ldots \) be independent r.v.s taking values in \(\mathbb {Z}_+\), \(X_i{\mathop {=}\limits ^{d}}X\ (\forall i).\) Denote \(S_0=0,\,S_k=X_1+\cdots +X_k\ (k\!\in \!I\!\!N).\) Then
Relation (64) is an analogue of (26) in terms of the point metric.
Proof of Lemma 8
W.l.o.g. we may assume that \(p\!:=\!I\!\!P(X\!\ne \!0)\!\ne \!0.\) For any \(m\in \mathbb {Z}_+\)
Therefore,
Denote
Estimate (64) will follow if we show that \(f(m)\!\le \!1/p\) for any \(m\!\in \!\mathbb {Z}_+\).
Clearly, \(f(0)\!=\!\sum _{k\ge 0}q^m \!=\!1/p,\) where \(q\!=\!1\!-\!p.\) Let \(m\!\in \!I\!\!N.\) By Khintchine’s formula (23),
where Binomial \(\textbf{B}(k,p)\) r.v. \(\nu _k\) is independent of \(\{X_i'\}\). Hence
where \(0^0\!:=\!1\); we have used the fact that \(\sum _{k\ge j} {k \atopwithdelims ()j} p^j q^{k-j} = 1/p\).
Denote by \(\eta (\cdot )\) the renewal process
Then \(\{S'_k\!\le \!m\} = \{k\!\le \!\eta (m)\},\) \(I\!\!E\eta (m) \!=\! \sum _{k\ge 1} I\!\!P(S'_k\!\le \!m),\)
Since \(X'\) takes values in \(I\!\!N,\) we have
Indeed, \(S_{\eta (m-1)}\!\le \!m\!-\!1,\ S_{\eta (m-1)+1}\!\ge \!m\) by (66). If \(S_{\eta (m-1)+1}\!>\!m,\) then \(\eta (m)=\eta (m\!-\!1).\) Therefore,
Thus, \(\sum _{j\ge 0}I\!\!P(S'_j\!=\!m)\!\le \!1\) for all \(m\!\in \!\mathbb {Z}_+,\) and (65) entails (64). \(\square \)
Proof of Theorem 4
According to (25),
where \(\nu _n=\tau _1+\cdots +\tau _n,\) \(\tau _1,X_1',\ldots ,\tau _n,X_n'\) are independent r.v.s, \(\mathcal{L}(X_i')=\mathcal{L}(X_i|X_i\!\ne \!0),\,\mathcal{L}(\tau _i)=\textbf{B}(p_i)\ (\forall i),\) \(p_i=I\!\!P(X_i\!\ne \!0),\)
Thus, \(d_{\textrm{o}}(S_n;S'_{\pi _{n\bar{p}}}) = d_\textrm{o}(S'_{\nu _n};S'_{\pi _{n\bar{p}}}).\) Since \(I\!\!P(X'\!\ne \!0)\!=\!1,\) inequality (64) yields
An application of (11) leads to (27). \(\square \)
Proof of Proposition 5
If \(p\!\le \!n^{-2/3},\) then (12) yields
If \(p\!>\!n^{-2/3}\), then we apply (8) to get
Combining these bounds, we derive (29). \(\square \)
Data Availability
Not applicable.
References
Arak, T.V.: On the convergence rate in Kolmogorov’s uniform limit theorem. I. Theory Probab. Appl. 26(2), 219–239 (1981)
Arak, T.V.: On the convergence rate in Kolmogorov’s uniform limit theorem. II. Theory Probab. Appl. 26(3), 437–451 (1981)
Arak, T.V.: An improvement of the lower bound for the rate of convergence in Kolmogorov’s uniform limit theorem. Theory Probab. Appl. 27, 767–772 (1982)
Arak, T.V., Zaitsev, A.Y.: Uniform limit theorems for sums of independent random variables. Proc. Steklov Inst. Math. 174, 3–214 (1986)
Arenbaev, N.K.: Asymptotic behavior of the multinomial distribution. Theory Probab. Appl. 21, 805–810 (1976)
Balakrishnan, N., Koutras, M.V.: Runs and Scans with Applications. Wiley, New York (2001)
Barbour, A.D., Holst, L., Janson, S.: Poisson Approximation. Clarendon Press, Oxford (1992)
Barbour, A.D.: Asymptotic expansions in the Poisson limit theorem. Ann. Probab. 15(2), 748–766 (1987)
Barbour, A.D., Xia, A.: Poisson perturbations. ESAIM Probab. Stat. 3, 131–150 (1999)
Barbour, A.D., Chryssaphinou, O.: Compound Poisson approximation: a user’s guide. Appl. Probab. 11(3), 964–1002 (2001)
Barbour, A.D., Čekanavičius, V.: Total variation asymptotics for sums of independent integer random variables. Ann. Probab. 30(2), 509–545 (2002)
Barbour, A.D., Eagleson, G.K.: Poisson approximation for some statistics based on exchangeable trials. Adv. Appl. Probab. 15(3), 585–600 (1983)
Barbour, A.D., Jensen, J.L.: Local and tail approximations near the Poisson limit. Scand. J. Stat. 16, 75–87 (1989)
Bernstein, S.N.: Sur l’extensiori du theoreme limite du calcul des probabilites aux sommes de quantites dependantes. Math. Ann. 97, 1–59 (1926)
Borisov, I.S., Vorozheikin, I.S.: Accuracy of approximation in the Poisson theorem in terms of \(\chi ^2\) distance. Sib. Math. J. 49(1), 5–17 (2008)
Čekanavičius, V.: Remarks on infinitely divisible approximations to the Binomial law. In: Grigelionis, B. et al. (eds) Probability Theory and Mathematical Statistics. Proceedings 7th Vilnius Conference, VSP/TEV, pp. 135–146 (1999)
Čekanavičius, V.: On compound Poisson approximations under moment restrictions. Theory Probab. Appl. 44(1), 18–28 (1999)
Çekanavičius, V., Vaitkus, P.: A centered Poisson approximation via Stein’s method. Lith. Math. J. 41(4), 319–329 (2001)
Čekanavičius, V., Novak, S.Y.: Compound Poisson approximation. Probab. Surv. 19, 1–80 (2022)
Deheuvels, P., Pfeifer, D.: A semigroup approach to Poisson approximation. Ann. Probab. 14(2), 663–676 (1986)
Franken, P.: Approximation der verteilungen von summen unabhängiger nichtnegativer ganzzahliger zufallsgrössen durch Poissonsche verteilunged. Math. Nachr. 27, 303–340 (1964)
Gerber, H.U.: An Introduction to Mathematical Risk Theory. Huebner Foundation, Philadelphia (1979)
Gnedenko, B.V., Kolmogorov, A.N.: Limit Distributions for Sums of Independent Random Variables. Addison-Wesley, Cambridge (1954)
Khintchine, A.Y.: Asymptotische Gesetze der Wahrscheinlichkeitsrechnung Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer, Berlin (1933)
Kerstan, J.: Verallgemeinerung eines satzes von Prochorow und Le Cam. Z. Wahrsch. Verw. Gebiete 2, 173–179 (1964)
Khintchine, A.Y.: Limit laws for sums of random variables. ONTI, Moscow (1938). (in Russian)
Kolmogorov, A.N.: Two uniform limit theorems for sums of independent random variables. Theory Probab. Appl. 1(4), 384–394 (1956)
Kruopis, J.: Precision of approximations of the generalized Binomial distribution by convolutions of Poisson measures. Lith. Math. J. 26(1), 37–49 (1986)
Leadbetter, M.R., Lindgren, G., Rootzen, H.: Extremes and Related Properties of Random Sequences and Processes. Springer, New York (1983)
LeCam, L.: An approximation theorem for the Poisson binomial distribution. Pacif. J. Math. 19(3), 1181–1197 (1960)
Mattner, L., Roos, B.: A shorter proof of Kanter’s Bessel function concentration bound. Probab. Theory Relat. Fields 139, 407–421 (2007)
Meshalkin, L.D.: On the approximation of polynomial distributions by infinitely-divisible laws. Theory Probab. Appl. 5(1), 114–124 (1960)
Novak, S.Y.: Extreme Value Methods with Applications to Finance. Chapman & Hall/CRC Press, London. ISBN 9781439835746 (2011)
Novak, S.Y.: On the accuracy of Poisson approximation. Extremes 22, 729–748 (2019)
Novak, S.Y.: Poisson approximation. Probab. Surv. 16, 228–276 (2019)
Novak S.Y.: Poisson approximation. Addendum. Probab. Surv. 18, 272–275 (2021)
Novak, S.Y.: Poisson approximation in terms of the Gini–Kantorovich distance. Extremes 24(1), 67–84 (2021)
Presman, E.L.: Approximation of binomial distributions by infinitely divisible ones. Theory Probab. Appl. 28, 393–403 (1983)
Presman, E.L.: On Poisson approximation in total variation for a sum of independent Bernoulli random variables. Theory Probab. Appl. 30(2), 391–396 (1985)
Prokhorov, Y.V.: Asymptotic behavior of the binomial distribution. Uspehi Matem. Nauk 8(3), 135–142 (1953)
Romanowska, M.: A note on the upper bound for the distribution in total variation between the binomial and the Poisson distribution. Stat. Neerlandica 31, 127–130 (1977)
Rogozin, B.A.: On evaluating the concentration functions. Theory Probab. Appl. 6(1), 94–97 (1961)
Röllin, A.: Approximation of sums of conditionally independent variables by the translated Poisson distribution. Bernoulli 11, 1115–1128 (2005)
Roos, B.: Asymptotic and sharp bounds in the Poisson approximation to the Poisson-Binomial distribution. Bernoulli 5(6), 1021–1034 (1999)
Roos, B.: Sharp constants in the Poisson approximation. Stat. Probab. Lett. 52, 155–168 (2001)
Shorgin, S.Y.: Approximation of a generalized binomial distribution. Theory Probab. Appl. 22(4), 846–850 (1977)
Tsaregradskii, I.P.: On uniform approximation of the binomial distribution with infinitely divisible laws. Theory Probab. Appl. 3(4), 470–474 (1958)
Zaitsev, A.Y.: On the accuracy of approximation of distributions of sums of independent random variables which are nonzero with a small probability by means of accompanying laws. Theory Probab. Appl. 28(4), 657–669 (1983)
Zaitsev, A.Y.: An example of a distribution whose set of n-fold convolutions is uniformly separated from the set of infinitely divisible laws in distance in variation. Theory Probab. Appl. 36(2), 419–425 (1991)
Acknowledgements
The author is grateful to the reviewer for many helpful remarks.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Competing interests
The author declares no competing interests.
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
Novak, S.Y. On Poisson Approximation. J Theor Probab 37, 2277–2303 (2024). https://doi.org/10.1007/s10959-023-01310-4
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-023-01310-4