Strong approximation for additive functionals of geometrically ergodic Markov chains

Transcription

1 Strong approximation for additive functionals of geometrically ergodic Markov chains Florence Merlevède Joint work with E. Rio Université Paris-Est-Marne-La-Vallée (UPEM) Cincinnati Symposium on Probability Theory and Applications September 2014

2 Strong approximation in the iid setting (1) Assume that (X i ) i 1 is a sequence of iid centered real-valued random variables with a finite second moment σ 2 and define S n = X 1 + X X n The ASIP says that a sequence (Z i ) i 1 of iid centered Gaussian variables may be constructed is such a way that sup Sk σb k = o(b n ) almost surely, 1 k n where b n = (n log log n) 1/2 (Strassen (1964)). When (X i ) i 1 is assumed to be in addition in L p with p > 2, then we can obtain rates in the ASIP: b n = n 1/p (see Major (1976) for p ]2, 3] and Komlós, Major and Tusnády for p > 3).

3 Strong approximation in the iid setting (2) When (X i ) i 1 is assumed to have a finite moment generating function in a neighborhood of 0, then the famous Komlós-Major-Tusnády theorem (1975 and 1976) says that one can construct a standard Brownian motion (B t ) t 0 in such a way that ( ) P sup S k σb k x + c log n k n a exp( bx) (1) where a, b and c are positive constants depending only on the law of X 1. (1) implies in particular that sup S k σb k = O(log n) almost surely 1 k n It comes from the Erdös-Rényi law of large numbers (1970) that this result is unimprovable.

4 Strong approximation in the multivariate iid setting Einmahl (1989) proved that we can obtain the rate O((log n) 2 ) in the almost sure approximation of the partial sums of iid random vectors with finite moment generating function in a neighborhood of 0 by Gaussian partial sums. Zaitsev (1998) removed the extra logarithmic factor and obtained the KMT inequality in the case of iid random vectors. What about KMT type results in the dependent setting?

5 An extension for functions of iid (Berkes, Liu and Wu (2014)) Let (X k ) k Z be a stationary process defined as follows. Let (ε k ) k Z be a sequence of iid r.v. s and g : R Z R be a measurable function such that for all k Z, X k = g(ξ k ) with ξ k := (..., ε k 1, ε k ) is well defined, E(g(ξ k )) = 0 and g(ξ k ) p < for some p > 2. Let (ε k ) k Z an independent copy of (ε k ) k Z. For any integer k 0, let ξk = ( ξ 1, ε 0, ε ) 1,..., ε k 1, ε k and X k = g(ξk ). For k 0, let δ(k) as introduced by Wu (2005): δ(k) = X k X k p. Berkes, Liu and Wu (2014): The almost sure strong approximation holds with the rate o(n 1/p ) and σ 2 = k E(X 0 X k ) provided that δ(k) = O(k α ) with α > 2 if p ]2, 4] and α > f (p) if p > 4 with f (p) = 1 + p2 4 + (p 2) p p + 4 8p

6 What about strong approximation in the Markov setting? Let (ξ n ) be an irreducible and aperiodic Harris recurrent Markov chain on a countably generated measurable state space (E, B). Let P(x,.) be the transition probability. We assume that the chain is positive recurrent. Let π be its (unique) invariant probability measure. Then there exists some positive integer m, some measurable function h with values in [0, 1] with π(h) > 0, and some probability measure ν on E, such that We assume that m = 1 P m (x, A) h(x)ν(a). The Nummelin splitting technique (1984) allows to extend the Markov chain in such a way that the extended Markov chain has a recurrent atom. This allows regeneration.

7 The Nummelin splitting technique (1) Let Q(x, ) be the sub-stochastic kernel defined by Q = P h ν The minorization condition allows to define an extended chain ( ξ n, U n ) in E [0, 1] as follows. At time 0, U 0 is independent of ξ 0 and has the uniform distribution over [0, 1]; for any n N, P( ξ n+1 A ξ n = x, U n = y) = 1 y h(x) ν(a) + 1 y>h(x) Q(x, A) 1 h(x) := P((x, y), A) and U n+1 is independent of ( ξ n+1, ξ n, U n ) and has the uniform distribution over [0, 1]. P = P λ (λ is the Lebesgue measure on [0, 1]) and ( ξ n, U n ) is an irreducible and aperiodic Harris recurrent chain, with unique invariant probability measure π λ. Moreover ( ξ n ) is an homogenous Markov chain with transition probability P(x,.).

8 Regeneration Define now the set C in E [0, 1] by C = {(x, y) E [0, 1] such that y h(x)}. For any (x, y) in C, P( ξ n+1 A ξ n = x, U n = y) = ν(a). Since π λ(c ) = π(h) > 0, the set C is an atom of the extended chain, and it can be proven that this atom is recurrent. Let T 0 = inf{n 1 : U n h( ξ n )} and T k = inf{n > T k 1 : U n h( ξ n )}, and the return times (τ k ) k>0 by τ k = T k T k 1. Note that T 0 is a.s. finite and the return times τ k are iid and integrable. Let S n (f ) = n k=1 f ( ξ k ). The random vectors (τ k, S Tk (f ) S Tk 1 (f )) k>0 are iid and their common law is the law of (τ 1, S T1 (f ) S T0 (f )) under P C.

9 Csáki and Csörgö (1995): If the r.v s S Tk ( f ) S Tk 1 ( f ) have a finite moment of order p for some p in ]2, 4] and if E(τ p/2 k ) <, then one can construct a standard Wiener process (W t ) t 0 such that S n (f ) nπ(f ) σ(f ) W n = O(a n ) a.s.. with a n = n 1/p (log n) 1/2 (log log n) α and σ 2 (f ) = lim n 1 n VarS n (f ). The above result holds for any bounded function f only if the return times have a finite moment of order p. The proof is based on the regeneration properties of the chain, on the Skorohod embedding and on an application of the results of KMT (1975) to the partial sums of the iid random variables S Tk+1 (f ) S Tk (f ), k > 0.

10 On the proof of Csáki and Csörgö For any i 1, let X i = T i l=t i 1 +1 f ( ξ l ). Since the (X i ) i>0 are iid, if E X 1 2+δ <, there exists a standard Brownian motion (W (t)) t>0 such that sup k X i σ(f )W (k) = o(n 1/(2+δ) ) k n i=1 a.s. Let ρ(n) = max{k : T k n}. If E τ 1 q < for some 1 q 2, then ρ(n) = n E(τ 1 ) + O(n1/q (log log n) α ) a.s. ρ(n) i=1 X i S n (f ) = o(n 1/(2+δ) ) a.s. W ( ) ( ρ(n)) W ( n = O n 1/(2q) (log n) 1/2 (log log n) α ) ) a.s. E(τ 1 ) With this method, no way to do better than O(n 1/(2q) (log n) 1/2 ) (1 q 2) even if f is bounded and τ 1 has exponential moment.

11 Link between the moments of return times and the coefficients of absolute regularity For positive measures µ and ν, let µ ν denote the total variation of µ ν Set β n = E Pn (x,.) π dπ(x). The coefficients β n are called absolute regularity (or β-mixing) coefficients of the chain. Bolthausen ( ): for any p > 1, E(τ p 1 ) = E C (T p 0 ) < if and only if n p 2 β n <. n>0 Hence, according to the strong approximation result of M.-Rio (2012), if f is bounded and E(τ p 1 ) for some p in ]2, 3[, then the strong approximation result holds with the rate o(n 1/p (log n) (p 2)/(2p) ).

12 Main result: M. Rio (2014) Assume that β n = O(ρ n ) for some real ρ with 0 < ρ < 1, If f is bounded and such that π(f ) = 0 then there exists a standard Wiener process (W t ) t 0 and positive constants a, b and c depending on f and on the transition probability P(x, ) such that, for any positive real x and any integer n 2, ( ) P π sup Sk (f ) σ(f )W k c log n + x a exp( bx). k n where σ 2 (f ) = π(f 2 ) + 2 n>0 π(fp n f ) > 0. Therefore sup k n S k (g) σ(f )W k = O(log n) a.s.

13 Some comments on the condition The condition β n = O(ρ n ) for some real ρ with 0 < ρ < 1 is equivalent to say that the Markov chain is geometrically ergodic (see Nummelin and Tuominen (1982)). If the Markov chain is GE then there exists a positive real δ such that E ( e tτ 1 ) < and E π ( e tt 0 ) < for any t δ. Let µ be any law on E such that E Pn (x,.) π dµ(x) = O(r n ) for some r < 1. Then P µ (T 0 > n) decreases exponentially fast (see Nummelin and Tuominen (1982)). The result extends to the Markov chain (ξ n ) with transition probability P and initial law µ.

14 Some insights for the proof Let S n (f ) = n l=1 f ( ξ l ) and X i = T i l=t i 1 +1 f ( ξ l ). Recall that (X i, τ i ) i>0 are iid. Let α be the unique real such that Cov(X k ατ k, τ k ) = 0 The random vectors (X i ατ i, τ i ) i>0 of R 2 are then iid and their marginals are non correlated. By the multidimensional strong approximation theorem of Zaitsev (1998), there exist two independent standard Brownian motions (B t ) t and ( B t ) t such that and S Tn (f ) α(t n ne(τ 1 )) vb n = O(log n) a.s. (1) T n ne(τ 1 ) ṽ B n = O(log n) a.s. (2) where v 2 = Var(X 1 ατ 1 ) and ṽ 2 = Var(τ 1 ). We associate to T n a Poisson Process via (2).

15 Let λ = (E(τ 1)) 2. Via KMT, one can construct a Poisson process N Var(τ 1 ) (depending on B) with parameter λ in such a way that Therefore, via (2), and then, via (1), γn(n) ne(τ 1 ) ṽ B n = O(log n) a.s. T n γn(n) = O(log n) a.s. S γn(n) (f ) αγn(n) + αne(τ 1 ) vb n = O(log n) a.s. (3) The processes (B t ) t and (N t ) t appearing here are independent. Via (3), setting N 1 (k) = inf{t > 0 : N(t) k} := k l=1 E l, S n (f ) = vb N 1 (n/γ) + αn αe(τ 1)N 1 (n/γ) + O(log n) a.s. If v = 0, the proof is finished. Indeed, by KMT, there exists a Brownian motion W n (depending on N) such that αn αe(τ 1 )N 1 (n/γ) = W n + O(log n) a.s.

16 If v = 0 and α = 0, we have S n (f ) = vb N 1 (n/γ) + O(log n) a.s. Using Csörgö, Deheuvels and Horváth (1987) (B and N are independent), one can construct a Brownian motion W (depending on N) such that B N 1 (n/γ) W n = O(log n) a.s. ( ), which leads to the expected result when α = 0. However, in the case α = 0 and v = 0, we still have S n (f ) = vb N 1 (n/γ) + αn αe(τ 1)N 1 (n/γ) + O(log n) a.s. and then S n (f ) = W n + W n + O(log n) a.s. Since W and W are not independent, we cannot conclude. Can we construct W n independent of N (and then of W ) such that ( ) still holds?

17 The key lemma Let (B t ) t 0 be a standard Brownian motion on the line and {N(t) : t 0} be a Poisson process with parameter λ > 0, independent of (B t ) t 0. Then one can construct a standard Brownian process (W t ) t 0 independent of N( ) and such that, for any integer n 2 and any positive real x, ( P sup Bk 1 ) W N(k) C log n + x A exp( Bx), k n λ where A, B and C are positive constants depending only on λ. (W t ) t 0 may be constructed from the processes (B t ) t 0, N( ) and some auxiliary atomless random variable δ independent of the σ-field generated by the processes (B t ) t 0 and N( ).

18 Construction of W (1/3) It will be constructed from B by writing B on the Haar basis. For j Z and k N, let and Y j,k = e j,k = 2 j/2( 1 ]k2 j,(k+ 1 2 )2j ] 1 ](k+ 1 2 )2j,(k+1)2 j ]), 0 e j,k (t)db(t) = 2 j/2( 2B (k+ 1 2 )2 j B k2 j B (k+1)2 j ). Then, since (e j,k ) j Z,k 0 is a total orthonormal system of l 2 (R), for any t R +, B t = j Z k 0 ( t To construct W, we modify the e j,k. 0 e j,k(t)dt ) Y j,k.

19 Construction of W (2/3) Let E j = {k N : N(k2 j ) < N((k )2j ) < N((k + 1)2 j )} For j Z and k E j, let f j,k = c 1/2 ( j,k bj,k 1 ]N(k2 j ),N((k+ 1 2 )2j )] a ) j,k1 ]N((k+ 1 2 )2 j ),N((k+1)2 j )], where a j,k = N((k )2j ) N(k2 j ), b j,k = N((k + 1)2 j ) N((k )2j ), and c j,k = a j,k b j,k (a j,k + b j,k ) (f j,k ) j Z,k Ej is an orthonormal system whose closure contains the vectors 1 ]0,N(t)] for t R + and then the vectors 1 ]0,l] for l N. Setting f j,k = 0 if k / E j, we define W l = j Z k 0 ( l 0 f j,k(t)dt ) Y j,k for any l N and W 0 = 0

20 Construction of W (3/3) W l = j Z k 0 ( l 0 f j,k(t)dt ) Y j,k for any l N and W 0 = 0 Conditionally to N, (f j,k ) j Z,k Ej is an orthonormal system and (Y j,k ) is a sequence of iid N (0, 1), independent of N. Hence, conditionally to N, (W l ) l 0 is a Gaussian sequence such that Cov(W l, W m ) = l m. Therefore this Gaussian sequence is independent of N By the Skorohod embedding theorem, we can extend it to a standard Wiener process (W t ) t still independent of N.

21 Take λ = 1. For any l N B l W N(l) = j 0 k 0 If l / ]k2 j, (k + 1)2 j [, Hence, setting l j = [l2 j ], ( l 0 e j,k(t)dt N(l) 0 ) f j,k (t)dt Y j,k. l 0 e j,k(t)dt = N(l) 0 f j,k (t)dt = 0 ( l N(l) ) B l W N(l) = e j,l j (t)dt f j,lj (t)dt Y j,lj. j Let t j = l l j 2 j 2 j. We then have B l W N(l) = U j,lj Y j,lj 1 tj ]0,1/2] + V j,lj Y j,lj 1 tj ]1/2,1[, j 0 j 0

22 The rest of the proof consists to show that ) P( Uj,l 2 j 1 tj ]0,1/2] c 1 log n + c 2 x j 0 and the same with V 2 j,l j 1 tj ]1/2,1[. c 3 e c 4x, Let Π j,k = N((k + 1)2 j ) N(k2 j ). We use in particular that the conditional law of Π j 1,2k given Π j,k is a B(Π j,k, 1/2) The property above allows to construct a sequence (ξ j,k ) j,k of iid N (0, 1) such that Π j 1,2k 1 2 Π j,k Π1/2 j,k ξ j,k This comes from the Tusnády s lemma: Setting Φ m the d.f. of B(m, 1/2) and Φ the d.f. of N (0, 1), Φ 1 m (Φ(ξ)) m ξ m

23 Thank you for your attention!