Tail behaviour of stationary distribution for Markov chains with asymptotically zero drift D. Denisov University of Manchester (jointly with D. Korshunov and V.Wachtel) April, 2014
Outline Statement of problem 1 Examples, main results and known results 2 General approach - random walk example 3 Harmonic functions and change of measure 4 Renewal Theorem 5 Further developments 6
Object of study One-dimensional homogenous Markov chain on R + . X n , n = 0 , 1 , 2 , . . . Let ξ ( x ) be a random variable corresponding to a jump at point x , i.e. P ( ξ ( x ) ∈ B ) = P ( X n +1 − X n ∈ B | X n = x ) . Let m k ( x ) := E [ ξ ( x ) k ] .
Object of study One-dimensional homogenous Markov chain on R + . X n , n = 0 , 1 , 2 , . . . Let ξ ( x ) be a random variable corresponding to a jump at point x , i.e. P ( ξ ( x ) ∈ B ) = P ( X n +1 − X n ∈ B | X n = x ) . Let m k ( x ) := E [ ξ ( x ) k ] . Main assumptions Small drift: m 1 ( x ) ∼ − µ x , x → ∞ ; Finite variance: m 2 ( x ) → b , x → ∞ .
Questions 1 If 2 xm 1 ( x ) + m 2 ( x ) ≤ − ε then X n is ergodic (Lamperti)
Questions 1 If 2 xm 1 ( x ) + m 2 ( x ) ≤ − ε then X n is ergodic (Lamperti) What can one say about the stationary distribution?
Questions 1 If 2 xm 1 ( x ) + m 2 ( x ) ≤ − ε then X n is ergodic (Lamperti) What can one say about the stationary distribution? 2 If 2 xm 1 ( x ) − m 2 ( x ) ≥ ε then X n is transient .
Questions 1 If 2 xm 1 ( x ) + m 2 ( x ) ≤ − ε then X n is ergodic (Lamperti) What can one say about the stationary distribution? 2 If 2 xm 1 ( x ) − m 2 ( x ) ≥ ε then X n is transient . What can one say about the renewal (Green) function ∞ � H ( x ) = P ( X n ≤ x ) , x → ∞ ? n =0
Continuous time - Bessel-like diffusions Let X t be the solution to SDE dX t = − µ ( X t ) dt + σ ( X t ) dW t , X 0 = x > 0 , X t where µ ( x ) → µ and σ ( x ) → σ. For Bessel processes µ ( x ) = const and σ ( x ) = const .
Continuous time - Bessel-like diffusions Let X t be the solution to SDE dX t = − µ ( X t ) dt + σ ( X t ) dW t , X 0 = x > 0 , X t where µ ( x ) → µ and σ ( x ) → σ. For Bessel processes µ ( x ) = const and σ ( x ) = const . We can use forward Kolmogorov equations to find exact stationary density � µ ( x ) � d 2 � � 0 = d + 1 σ 2 ( x ) p ( x ) p ( x ) dx 2 2 dx x to obtain � � x � 2 c 2 µ ( y ) p ( x ) = σ 2 ( x ) exp − y σ 2 ( y ) dy , c > 0 . 0
Continuous time - Bessel-like diffusions Let X t be the solution to SDE dX t = − µ ( X t ) dt + σ ( X t ) dW t , X 0 = x > 0 , X t where µ ( x ) → µ and σ ( x ) → σ. For Bessel processes µ ( x ) = const and σ ( x ) = const . We can use forward Kolmogorov equations to find exact stationary density � µ ( x ) � d 2 � � 0 = d + 1 σ 2 ( x ) p ( x ) p ( x ) dx 2 2 dx x to obtain � � x � 2 c 2 µ ( y ) p ( x ) = σ 2 ( x ) exp − y σ 2 ( y ) dy , c > 0 . 0 Then, � � � x 2 µ dy ∼ Cx − 2 µ/ b p ( x ) ≈ C exp − b y 1 and � ∞ p ( y ) dy ≈ Cx − 2 µ/ b +1 π ( x , + ∞ ) = x
Simple Markov chain Markov chain on Z P x ( X 1 = x + 1) = p + ( x ) P x ( X 1 = x − 1) = p − ( x ) .
Simple Markov chain Markov chain on Z P x ( X 1 = x + 1) = p + ( x ) P x ( X 1 = x − 1) = p − ( x ) . Then the stationary probabilities π ( x ) satisfy π ( x ) = π ( x − 1) p + ( x − 1) + π ( x + 1) p − ( x + 1) ,
Simple Markov chain Markov chain on Z P x ( X 1 = x + 1) = p + ( x ) P x ( X 1 = x − 1) = p − ( x ) . Then the stationary probabilities π ( x ) satisfy π ( x ) = π ( x − 1) p + ( x − 1) + π ( x + 1) p − ( x + 1) , with solution � x � x � � p + ( k − 1) π ( x ) = π (0) = π (0) exp (log p + ( k − 1) − log p − ( k )) , p − ( k ) k =1 k =1
Asymptotics for the tail of the stationary measure Theorem Suppose that, as x → ∞ , m 1 ( x ) ∼ − µ x , m 2 ( x ) → b and 2 µ > b . (1) Suppose some technical conditions and m 3 ( x ) → m 3 ∈ ( −∞ , ∞ ) as x → ∞ (2) and, for some A < ∞ , E { ξ 2 µ/ b +3+ δ ( x ); ξ ( x ) > Ax } O ( x 2 µ/ b ) . = (3) Then there exist a constant c > 0 such that R x 0 r ( y ) dy = cx − 2 µ/ b +1 ℓ ( x ) π ( x , ∞ ) ∼ cxe − as x → ∞ .
Known results Menshikov and Popov (1995) investigated Markov Chains on Z + with bounded jumps and showed that c − x − 2 µ/ b − ε ≤ π ( { x } ) ≤ c + x − 2 µ/ b + ε .
Known results Menshikov and Popov (1995) investigated Markov Chains on Z + with bounded jumps and showed that c − x − 2 µ/ b − ε ≤ π ( { x } ) ≤ c + x − 2 µ/ b + ε . Korshunov (2011) obtained the following estimate π ( x , ∞ ) ≤ c ( ε ) x − 2 µ/ b +1+ ε .
General approach - random walk example Consider a classical example, of Lindley recursion W n +1 = ( W n + ξ n ) + , n = 0 , 2 , . . . , W 0 = 0 , assuming that E ξ = − a < 0.
General approach - random walk example Consider a classical example, of Lindley recursion W n +1 = ( W n + ξ n ) + , n = 0 , 2 , . . . , W 0 = 0 , assuming that E ξ = − a < 0. This is an ergodic Markov chain (Note drift is not small).
General approach - random walk example Consider a classical example, of Lindley recursion W n +1 = ( W n + ξ n ) + , n = 0 , 2 , . . . , W 0 = 0 , assuming that E ξ = − a < 0. This is an ergodic Markov chain (Note drift is not small). A classical approach consists of three key steps Step 1: Reverse time and consider a random walk S n = ξ 1 + · · · + ξ n , n = 1 , 2 . . . , S 0 = 0 . Then, d → W = sup S n . W n n ≥ 0
Random walks ctd. Step 2: Exponential change of measure. Assuming that there exists κ > 0 such that E [ e κ ξ ] = 1 , E [ ξ e κ ξ ] < ∞ 1
Random walks ctd. Step 2: Exponential change of measure. Assuming that there exists κ > 0 such that E [ e κ ξ ] = 1 , E [ ξ e κ ξ ] < ∞ 1 one can perform change of measure P ( � ξ n ∈ dx ) = e κ x P ( ξ n ∈ dx ) , n = 1 , 2 , . . .
Random walks ctd. Step 2: Exponential change of measure. Assuming that there exists κ > 0 such that E [ e κ ξ ] = 1 , E [ ξ e κ ξ ] < ∞ 1 one can perform change of measure P ( � ξ n ∈ dx ) = e κ x P ( ξ n ∈ dx ) , n = 1 , 2 , . . . ξ 1 > 0 Under new measure � S n = � ξ 1 + · · · + � ξ n and E � � S n → + ∞ , and S n → −∞ .
Random walks ctd. Step 3: Use renewal theorem for � S n .
Random walks ctd. Step 3: Use renewal theorem for � S n . This step uses ladder heights construction and represents P ( M ∈ dx ) = CH ( dx ) = Ce − κ x � H ( dx ) . Now one can apply standard renewal theorem to � H ( dy ) ∼ dy / c to obtain P ( M ∈ dx ) ∼ ce − κ x , x → ∞ .
Asymptotically homogeneous Markov chains One can repeat this programme for asymptotically homogenous Markov chains. Namely, assume ξ ( x ) d → ξ, x → ∞ , where E [ e κ ξ ] = 1 , and E [ ξ e κ ξ ] < ∞
Asymptotically homogeneous Markov chains One can repeat this programme for asymptotically homogenous Markov chains. Namely, assume ξ ( x ) d → ξ, x → ∞ , where E [ e κ ξ ] = 1 , and E [ ξ e κ ξ ] < ∞ Borovkov and Korshunov (1996) showed that if �� � � ∞ x E e κ ξ ( x ) < ∞ , e κ t | P ( ξ ( x ) < t ) − P ( ξ < t ) | dt sup dx 0 R then π ( x , ∞ ) ∼ Ce − κ x , x → ∞ , x → ∞ .
Problems in our case Problem 1 (easier) In our case drift E ξ ( x ) → 0 , x → ∞ . Hence, for 1 = E exp { κ ξ ( x ) } ≈ 1 + κ E ξ ( x ) , x → ∞ to hold we need
Problems in our case Problem 1 (easier) In our case drift E ξ ( x ) → 0 , x → ∞ . Hence, for 1 = E exp { κ ξ ( x ) } ≈ 1 + κ E ξ ( x ) , x → ∞ to hold we need κ = κ ( x ) → ∞ , x → ∞ .
Problems in our case Problem 1 (easier) In our case drift E ξ ( x ) → 0 , x → ∞ . Hence, for 1 = E exp { κ ξ ( x ) } ≈ 1 + κ E ξ ( x ) , x → ∞ to hold we need κ = κ ( x ) → ∞ , x → ∞ . Hence, exponential change of measure does not work.
Problems in our case Problem 2 Suppose we managed to make a change of measure. As a result a . s � → + ∞ X n and E � ξ ( x ) → 0 .
Problems in our case Problem 2 Suppose we managed to make a change of measure. As a result a . s � → + ∞ X n and E � ξ ( x ) → 0 . Then, there is no renewal theorem about ∞ � � H ( x ) = P ( X n ≤ x ) . n =0 Main reason for that � X n d → Gamma ( α, β ) n c which makes the problem difficult.
Harmonic function Step 1 Change of measure via a harmonic function.
Harmonic function Step 1 Change of measure via a harmonic function. Let B be a Borel set in R + with π ( B ) > 0, in our case B = [0 , x 0 ]. Let τ B := min { n ≥ 1 : X n ∈ B } . Note E x τ B < ∞ for every x . V ( x ) is a harmonic function for X n killed at the time of the first visit to B , if V ( x ) = E x { V ( X 1 ); τ B > 1 } = E x { V ( X 1 ); X 1 / ∈ B } If V is harmonic then V ( x ) = E x { V ( X n ); τ B > n } for every n . (4)
Recommend
More recommend