MS&E 321 Spring Stochastic Systems June 1, 2013 Prof. Peter W. Glynn Page 1 of 7

Transcription

1 MS&E 321 Spring Stochastic Systems June 1, 213 Prof. Peter W. Glynn Page 1 of 7 Section 9: Renewal Theory Contents 9.1 Renewal Equations Solving the Renewal Equation Asymptotic Behavior of the Solution of the Renewal Equation Renewal Equations Let X = (X(t) : t ) be a non-delayed regenerative process and suppose that f : S R +. Then, if a (t) = Ef(X(t)), a (t) can be expressed as follows: a (t) = Ef(X(t))I(τ 1 > t) + Ef(X(t))I(τ 1 ds). But Ef(X(t))I(τ 1 ds) = E[f(X(τ 1 + t s)) τ 1 = s]p(τ 1 ds) = Ef(X(t s))p(τ 1 ds) = a (t s)p(τ 1 ds). In other words, (a (t) : t ) satisfies the linear integral equation a(t) = b(t) + (F a)(t), for t, and F (t) = P(τ 1 t), b(t) Ef(X(t))I(τ 1 > t), (G h)(t) h(t s)g(ds) is the convolution operation (that is well-defined for any non-decreasing function G and non-negative h). 1

2 Definition Given a function b and a non-negative G, a linear integral equation of the form a = b + G a is called a renewal equation for a non-decreasing function G. Define the n-fold convolution of G via G (n) (t) = I(t ) for n = and G (n) (t) = G(t s)g (n 1) (ds) for n 1. Note that when G is the distribution function F of a positive rv τ 1, F (n) (t) = P(T (n) t), T (n) = τ 1 + τ τ n, and the τ i s are iid copies of τ 1. Proposition The function (a (t) : t ) = (Ef(X(t)) : t ) is the minimal non-negative solution of a = b + F a and is given by F (n) b. Here are some additional examples of renewal equations: Example Let a (t) = P(T (N(t) + 1) t > x) be the tail probability of the residual life process T (N(t)+1) t corresponding to a non-delayed renewal counting process N = (N(t) : t ). Note that a satisfies the renewal equation and a = b + F a, (9.1.1) b(t) = P(τ 1 > t + x) F (t) = P(τ 1 t). The function a is the minimal non-negative solution of (9.1.1) and is given by F (n) b. Example Let X = (X(t) : t ) be a non-delayed regenerative process and let a (t) = P(t T (N(t)) x) be the distribution of the current age process t T (N(t). Here, a satisfies a = b + F a, (9.1.2) b(t) = I(t x), 2

3 F (t) = P(τ 1 t). The function a is the minimal non-negative solution of (9.1.2) and is given by F (n) b. Example Let X = (X(t) : t ) be a non-delayed S-valued regenerative process and let T = inf{t : X(t) A} be the first hitting time if some subset A S. Put a (t) = P(T > t). Then, a satisfies a = b + G a, (9.1.3) b(t) = P(T τ 1 > t), G(t) = P(τ 1 t, T > τ 1 ). The function a is the minimal non-negative solution of (9.1.3) and is given by G (n) b. Note that G is not a probability distribution function here (since G( ) = P(T < τ 1 ) < 1 in general). Definition A renewal equation is said to be a proper renewal equation if G is a probability distribution function; otherwise, the renewal equation is said to be an improper renewal equation. In particular, if G() = and G( ) < 1, the improper renewal equation is said to be defective; if G() = and G( ) > 1, the improper renewal equation is said to be excessive. Improper renewal equations can frequently be transformed into proper renewal equations via use of the following trick. Note that if a is given by G (n) b, then ã(t) = ( G (n) b)(t), ã(t) = e γt a(t) b(t) = e γt b(t) G(dt) = e γt G(dt), and ã satisfies the renewal equation So if we can find γ such that ã = b + G ã. (9.1.4) G(dt) = e γ t G(dt) is a probability distribution function, (9.1.4) becomes a proper renewal equation. 3

4 Exercise a.) Prove that if a renewal equation is excessive, there always exists a unique γ such that G is a probability distribution function. b.) Prove that if a renewal equation is defective, there need not exist a γ such that G is a probability distribution function. 9.2 Solving the Renewal Equation so Hence, If F is the distribution of an exponential rv with parameter λ >, then F (n) (dt) = λ (λt)n 1 (n 1)! exp( λt)dt F (n) (dt) = δ (dt) + λdt. (F (n) b)(t) = b(t) + λ t b(s) ds. So, the renewal equation can be explicitly solved when F is exponential. It can also be solved in closed form in a limited number of other cases. Let ã(γ) = e γt a(t) dt, b(γ) = e γt b(t) dt, G(γ) = e γt G(dt), and note that ã, b, and G are the Laplace transforms of a, b, and G, respectively. If a satisfies a = b + G a, it follows that so that ã = b + Gã, ã = b 1 G. If b and G can be computed in closed form, then a can potentially be computed in closed form by calculating the inverse Laplace transform corresponding to b(1 G) 1. For example, this can generally be done when G has a rational Laplace transform (i.e. is the ratio of two polynomials in γ). 4

5 9.3 Asymptotic Behavior of the Solution of the Renewal Equation Typically, the solution to the renewal equation can not be computed in closed form. In such settings, one usually must be satisfied with computing the limiting behavior of when t tends to infinity. Definition The function (F (n) b)(t) U(t) = F (n) (t) is called the renewal function and the corresponding distribution/measure is called the renewal measure. If F has a density f, u(t) = f (n) (t) is called the renewal density. If F is the distribution function of an integer-valued rv τ 1, then ( p (k) n u n = n=1 k=1 p (k) n = P (τ τ k = n)) is called the renewal mass function. We focus first on the case F is the distribution function of an integer-valued rv τ 1. Note that u n = = E P (τ τ k = n) k=1 I(T (k) = n) k=1 = EI(regeneration at time n). It turns out that the asymptotic behavior of u n as n can be analyzed via consideration of an appropriate Markov chain X = (X n : n ). In particular, let X n = inf{t (k) n : T (k) n} be the residual life process corresponding to the sequence of event times T (1), T (2),.... Note that u n = P (X n = ), X = (X n : n ) is an Z + -valued Markov chain with transition probabilities given by P (i, i 1) = 1 for i 1, and P (, i) = P (τ 1 = i + 1). 5

6 Note that the state is always recurrent for X. Furthermore, X is positive recurrent if and only if <, in which case π is given by j>i π(i) = P (τ 1 < j) = P (τ 1 > i). Finally, X is aperiodic if and only if If X is aperiodic and <, it follows that as n. It follows that if gcd{k 1 : P (τ 1 = k) > } = 1. u n = P (X n = ) π() = 1 b k <, (9.3.1) k= then the Dominated Convergence Theorem implies that as n. (F (k) b)(n) = k= k= u n k b k 1 k= b k Theorem (Discrete Renewal Theorem) Suppose < and If (9.3.1) holds, then gcd{k 1 : P (τ 1 = k) > } = 1. k= as n. If = and (9.3.1) holds, then u n k b k 1 k= b k as n. u n k b k k= Let X = (X n : n ) be a S-valued non-delayed regenerative sequence and suppose f : S R is a bounded function (i.e. sup{ f(x) : x S}). As noted earlier, satisfies a renewal equation, namely a n = b n + [,n] a n = Ef(X n ) a (t s)p (τ 1 ds) = b n + 6 a n jp j, j=1

7 so a n = u n j b j, b j = Ef(X j )I(τ 1 > j). The discrete renewal theorem guarantees that if τ 1 is aperiodic and has finite mean, then j= a n j= b j as n. In other words, as n. Ef(X n ) j= Ef(X j)i(τ 1 > j) = E τ1 1 j= f(x j) Exercise Prove that if X = (X n : n ) is a delayed S-valued regenerative sequence with < and f : S R is a bounded function, then as n, provided that τ 1 is aperiodic. Ef(X n ) E τ 1 1 j= f(x j) All the above discussion for discrete-time renewal theory extends to the continuous-time setting. 7