Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

Moments and tails of hitting times of Bessel processes

and convolutions of elementary mixtures of exponential


distributions

Rafal Lochowski
Warsaw School of Economics
joint work with Witold Bednorz (University of Warsaw)

GPSD Mannheim 2021

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 1 / 24
Fast exit of a Brownian motion from a ball

The most interesting result stemming from the presented results seem to
be the following estimates for the quick exit of d-dimensional Brownian
motion from a ball centered at the origin, with radius R.
If T ≤ R 2 /(2d), d = 2, 3, . . . and τ R is the exit time of d-dimensional
standard Brownian motion from the ball with radius R centered at the
origin then
1 c2
   
P τ R ≤ T ≤ exp − ,
32 T

c2
  1  
R
P τ ≤ T ≥ exp −4 .
3 T

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 2 / 24
Bessel processes

Xt , t ≥ 0, is a δ-dimensional Bessel process, starting from x0 ≥ 0, if it is a


non-negative process, whose square Zt = Xt2 satisfies the following sde
Z tp
Zt = x02 + 2 Zs dβs + δ · t,
0

where β is a standard Brownian motion.


In my talk I will present universal estimates of ordinary and central
moments as well as right and left tails of hitting times of Bessel processes:

τ c := inf t ≥ 0 : Zt = c 2 = inf {t ≥ 0 : Xt = c} ,

(1)

when c > x0 .

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 3 / 24
Purpose: universal estimates of moments and tails of
hitting times of Bessel processes
The estimates will be universal in the sense that I will present explicit
functions F , G , H, K of parameters δ > 0, c > 0, x0 ∈ [0, c), p ≥ 2, t > 0,
for which there exist universal constants C1 , C2 , C3 , C4 > 0, independent
from δ, c, x0 , p, t and such that

C1−1 F (δ, c, x0 , p) ≤ (E (τ c )p )1/p ≤ C1 F (δ, c, x0 , p);

C2−1 G (δ, c, x0 , p) ≤ (E |τ c − Eτ c |p )1/p ≤ C2 G (δ, c, x0 , p);


P (τ c ≥ C3 t) ≤ C3 H(δ, c, x0 , t), P τ c ≥ C3−1 t ≥ C3−1 H(δ, c, x0 , t);


ln P (τ c ≤ C4 t) ≥ C4 K (δ, c, x0 , t), ln P τ c ≤ C4−1 t ≤ C4−1 K (δ, c, x0 , t).




The results which will be presented were obtained in collaboration with


Witold Bednorz, they are published in a working paper available at arXiv
(arXiv:2102.07015).

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 4 / 24
Some history
Times τ c were topic of research of many authors:
Paul Lévy calculated in [Lév53] the cumulative probability function of
τ c in the case x0 = 0, δ - natural, as a series with osscillating
coefficients; but he did not find values of these coefficients.
The coefficients were calculated by Zbigniew Ciesielski and S. James
Taylor in [CT62].
John T. Kent generalized these results for the case of arbitrary
x0 > 0, δ > 0 in [Ken80], and found the following formula for arbitrary
ν = δ/2 − 1 > −1, x0 ∈ [0, c):
 
c ν−2 Jν (jν,n x0 /c) jν,n −jν,n
2 t/ 2c 2
( )  dt.
X
P (τ c ∈ dt) =  ν e
x0 Jν+1 (jν,n )
n≥1

Grzegorz Serafin proved in [Ser17] some estimates of the density of


τ c , but these estimates depend on δ.

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 5 / 24
Some history

The literature on τ c in the case when the starting point x0 of the Bessel
process belongs to the interval (c, +∞) seems to be much larger.
Two papers whose authors are Yuji Hamana and Hiroyuki Matsumoto:
[HM12, HM13]; few papers whose authors are Tomasz Byczkowski,
Tomasz Grzywny, Jacek Malecki, Michal Ryznar, Andrzej Stós, Bartosz
Trojan: [BR06, TBS07, TBR13].

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 6 / 24
Times τ c , x0 ∈ [0, c), - first estimates of tails from above
First, easy to obtain result on τ c , x0 ∈ [0, c) is the following estimate .
For any η ∈ (0, +∞)

c 2 − x02 η2
 
P (τ c ≥ (1 + η) Eτ c ) ≤ exp −δ (2)
c 2 8(1 + η)

and for any η ∈ (0, 1)

c 2 − x02 η2
 
c c
P (τ ≤ (1 − η) Eτ ) ≤ exp −δ , (3)
c 2 8(1 − η)

where
c 2 − x02
Eτ c = .
δ
A method to obtain these estimates - investigation of the
process Yt = exp (λ (Zt − δ · t)) stopped at τ c (Itô’s formula, Chebyshev’s
inequality, optimization with repect to λ ∈ R).

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 7 / 24
Times τ c , x0 ∈ [0, c), - first estimates of tails from above

Estimates (2) and (3) give quantitative description of strong concentration


of τ c around its mean and they were the reason why we decided to look
closer if they are optimal ones.
In particular, (3) seems to be interesting since it gives strong estimate on
probability of fast exit of a standard, multidimensional Brownian motion
from a ball. Estimates of this type (but worse) are used in the Large
Deviation Theory, see for example ([DZ98, Lemma 5.2.1]).

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 8 / 24
Probability of fast exit of a standard, multidimensional
Brownian motion from a ball
Let W be a standard d-dimensional Brownian motion, starting from 0.
[DZ98, Lemma 5.2.1] gives the following estimate: for any c, T > 0
!
1 c2
 
c
P sup |Wt | ≥ c = P (τ ≤ T ) ≤ 4d · exp −
0≤t≤T 2d T

while our estimate (3) gives for any c, T > 0

c2
P (τ c ≤ T ) ≤ 1 if ≤d
T
and  !2 
1 d c2 2
 if c > d.
P (τ c ≤ T ) ≤ exp − 1− c2
8 T T
T

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 9 / 24
Times τ c , x0 ∈ [0, c), - representation
In the paper [Ken80], using eigenvalus of the infinitesimal operator of the
Bessel process, Kent representd the law of τ c as a convolution of
’elementary mixtures of exponential distributions’. Namely,
law
X
τc = c
τν,n , (4)
n≥1

where ν = δ/2 − 1, τν,n c , n = 1, 2, . . ., are independent and τ c has the


ν,n
following law:
!
 x02 2
x02 jν,n
 
c c

P τν,n = 0 = 2 , P τν,n > t = 1 − 2 exp − 2 t , t > 0.
c c 2c
(5)
jν,n , n = 1, 2, . . ., denote consecutive zeros of Bessel functions of the 1st
kind Jν :
+∞
 y ν X (−1)m  y 2m
Jν (y ) = .
2 m!Γ (m + 1 + ν) 2
m=0

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 10 / 24
Moments of elementary mixtures of exponential
distributions

Using results of Efim Gluskin and Stanislaw Kwapień on moments of linear


combinations of symmetric random variables with logarythmic concave
tails, [GK95], and the Bernstein inequality, we found ’easy’ formulas for
central and ordinary moments of elementary mixtures of exponential
distributions.

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 11 / 24
Central moments of elementary mixtures of exponential
distributions
Lemma (1)
Let κn be independent variables with the 0 − 1 law

P (κn = 0) = 1 − P (κn = 1) = α ∈ [0, 1),

θn be independent (also from κn ) with exponential law P Exp(1),


n = 1, 2, . . . and let a1 ≥ a2 ≥ . . . ≥ 0 be such that n≥1 an2 < +∞ then
for p ≥ 2

X
(an κn θn − Ean κn θn )
n≥1
p
 1/p

X s X
1/p
∼ p (1 − α)  αn−1 anp  + p (1 − α) an2 .
n≥1 n≥1
Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 12 / 24
Ordinary moments of elementary mixtures of exponential
distributions
Lemma (2)
Let κn be independent variables with the 0 − 1 law

P (κn = 0) = 1 − P (κn = 1) = α ∈ [0, 1),

θn be independent (also from κn ) with exponential law P Exp(1),


n = 1, 2, . . . and let a1 ≥ a2 ≥ . . . ≥ 0 be such that n≥1 an < +∞ then
for p ≥ 1

X
an κn θn
n≥1
p
 1/p
X X
∼ p (1 − α)1/p  αn−1 anp  + (1 − α) an .
n≥1 n≥1
Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 13 / 24
Central moments of τ c
Using representation (4), (5), Lemma 1 and fact that jν,n ∼ ν + n when
n = 2, 3, . . ., denoting α = z0 /c 2 = x02 /c 2 we obtain
if αν+2 ≥ 1/2 then

τ c − Eτ c
c2 p
( ) √ √
1/p 1 1 p 1−α
∼ p (1 − α) 2
+ + √ ;
jν,1 (ν + 2)2−1/p (ν + 1) ν + 2

if αν+2 < 1/2 but α > 1/2 then

τ c − Eτ c
c2 p
( ) √ √
1/p 1 1 p 1−α
∼ p (1 − α) 2
+ + √ ;
jν,1 (ν + 2)2 (log2 (1/α))1/p (ν + 1) ν + 2

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 14 / 24
Central and ordinary moments of τ c

if α ≤ 1/2 then

τ c − Eτ c
c2 p
√ √ √
1/p 1 p 1−α p p
∼ p (1 − α) 2
+ √ ∼ 2 + √ .
jν,1 (ν + 1) ν + 2 jν,1 (ν + 1) ν + 2

Formulas for ordinary moments are very similar, with the only difference
that the term √
p

(ν + 1) ν + 2
is replaced by
1−α
.
ν+1

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 15 / 24
Estimates of the right tails obtained via ordinary moments
of τ c

Using ordinary moments and the Chebyshev inequality one can naturally
obtain estimates of the right tails of τ c , but they will be optimal (up to
universal multiplicative constants) only for α = x02 /c 2 ≈ 0, say α ≤ 1/2
(hipercontraction).
However, conditioning with resp. to the variable
N := min {n ≥ 1 : κn = 1} one can obtain optimal (up to universal
multiplicative constants) estimates of the right tails of τ c .

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 16 / 24
Estimates of the right tails of τ c

Theorem (1)
Denote α = z0 /c 2 = x02 /c 2 . There exist universal constants
γ1 , γ2 , γ3 , γ4 ∈ (0, +∞), such that for t ≥ γ3 Eτ c
 ! 
j 2  2


ν,1
X (ν + n)
P (τ c ≥ γ1 t) > (1 − α) exp − 2 t + αn−1 exp − t
 2c 2c 2 
n≥2

and for t ≥ γ4 Eτ c
 ! 
j 2  2


ν,1
X (ν + n)
P (τ c ≥ γ2 t) ? (1 − α) exp − 2 t + αn−1 exp − t .
 2c 2c 2 
n≥2

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 17 / 24
Estimates of the right tails of τ c
Corollary
Using notation of Theorem 1, for t > 0 define

(ν + n)2
 
n1 (t) := min n ∈ {2, 3, . . .} : t ≥ 1 ≥ 2,
2c 2
n 2
o
(ν+n1 (t))2
n2 (t) := min n ∈ {2, 3, . . .} : (ν+n)
2c 2 t − 2c 2 t ≥ 1 ≥ n1 (t) + 1

and

F (t; ν, c, z0 )
2 !
(ν+2)2
jν,1
 
− t − t 1
:= (1 − α) e 2c 2 + αe 2c 2 min , n2 (t)
ln(1/α)

then for t ≥ max (γ3 , γ4 ) Eτ c

P (τ c ≥ γ1 t) > F (t; ν, c, z0 ), P (τ c ≥ γ2 t) & F (t; ν, c, z0 ).


Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 18 / 24
Estimates of the left tails of τ c
Estimates of moments are not easy to apply to obtain estimates of left
tails, unfortunately. Using the Laplace transform one can obtain estimates
of the left tails from below. This, together with estimate (3) gives

Fact (1)
Using notation of Theorem 1 for η ∈ (0, 1) and δ = 2(ν + 1) > 0 the
following estimate holds

δ c 2 − x02 η 2
 
c c
P (τ ≤ (1 − η) Eτ ) ≤ exp −
8 c2 1 − η

while for δ = 2(ν + 1) > 1 the following estimate holds


   
c c − x0 1 c − x0 1
P τ ≤ 2c (1 − η) ≥ exp −2(δ − 1) .
δ−1 3 c 1−η

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 19 / 24
Estimates of the left tails of τ c , cont.

Unfortunately, lower estimates in Fact 1 do not work when δ ≤ 1 (and are


rather weak when δ > 1, δ ≈ 1). However, using the following interlacing
property of zeros of the Bessel function:

jν,1 < jν+1,0 < jν,2 < jν+1,1 < jν,3 < jν+1,2 < . . .

and representation (4), (5) we get the following strengthening of Fact 1


for small δs:

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 20 / 24
Estimates of the left tails of τ c , cont.
Fact
Using notation of Theorem 1 for η ∈ (0, 1) and δ = 2(ν + 1) > 0 the
following estimates hold

c 2 − x02
 
c
P τ ≤ (1 − η)
δ
 ( 2
!)
c 2 − x02 η 2 x02 jν,1 c 2 − x02
  
≤ exp −δ ∧ 1 − 1 − 2 exp − (1 − η)
8c 2 1 − η c δ 2c 2

and
 
c c − x0
P τ ≤ 5c (1 − η)
δ
( 2
!)
x02 jν,1
  
1 c − x0 1 c − x0
≥ exp −2δ 1 − 1 − 2 exp − (1 − η) .
3 c 1−η c 2δ c

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 21 / 24
References I
T. Byczkowski and M. Ryznar, Hitting distibution of geometric
Brownian motion, Stud. Math. 173 (2006), no. 1, 19–38.
Z. Ciesielski and S. J. Taylor, First passage times and sojourn times
for brownian motion in space and the exact Hausdorff measure of the
sample path, Transactions of the American Mathematical Society 103
(1962), no. 3, 434–450.
A. Dembo and O. Zeitouni, Large deviations techniques and
applications, Springer, 1998.
E. D. Gluskin and S. Kwapień, Tail and moment estimates for sums of
independent random variables with logarithmically concave tails, Stud.
Math. 114 (1995), no. 3, 303–309 (English).
Y. Hamana and H. Matsumoto, The probability densities of the first
hitting times of bessel process, J. of Math-for-Industry 4 (2012),
91–95.
Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 22 / 24
References II

, The probability distributions of the first hitting times of


Bessel processes, Trans. Amer. Math. Soc. 365 (2013), 5237–5257.
John T. Kent, Eigenvalue expansions for diffusion hitting times, Z.
Wahrscheinlichkeitstheor. Verw. Geb. 52 (1980), 309–319 (English).
P. Lévy, La mesure de Hausdorff de la courbe du mouvement
brownien., Giorn. Ist. Ital. Attuari 16 (1953), 1–37.
G. Serafin, Exit times densities of the Bessel process, Proc. Amer.
Math. Soc. 145(7) (2017), 3165–3178.
J. Malecki T. Byczkowski and M. Ryznar, Hitting times of Bessel
processes, Potential Analysis 38 (2013), 753–786.
P. Graczyk T. Byczkowski and A. Stós, Poisson kernels of half-spaces
in real hyperbolic space, Rev. Mat. Iberoam. 23 (2007), 85–126.

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 23 / 24
Thank you for your attention!

Rafal Lochowski, WSE Hitting times of Bessel processes GPSD Mannheim 2021 24 / 24

You might also like