You are on page 1of 8

Statistics and Probability Letters 193 (2023) 109654

Contents lists available at ScienceDirect

Statistics and Probability Letters


journal homepage: www.elsevier.com/locate/stapro

Exact simulation for the first hitting time of Brownian motion


and Brownian bridge
Lee Taeho
Korea Advanced Institute of Science and Technology, Republic of Korea1

article info a b s t r a c t

Article history: We propose an exact simulation scheme for the first hitting time of Brownian motion.
Received 16 July 2020 Our method is applicable to expanding boundaries, which are linear or piecewise linear.
Received in revised form 20 July 2022 We first develop an algorithm for symmetric linear boundaries, and then extend it
Accepted 5 August 2022
to other boundaries using a localization technique. The extended algorithm iteratively
Available online 6 September 2022
invokes a symmetric-case algorithm, and we prove that the expected number of
Keywords: iterations is uniformly bounded. We apply our scheme to simulate the first hitting time
Brownian motion of a Brownian bridge.
Brownian bridge © 2022 Published by Elsevier B.V.
First hitting time
Exact simulation

1. Introduction

Simulations of stochastic processes have been well-studied from both theoretical and practical viewpoints (Asmussen
and Glynn, 2007; Pereyra et al., 2015). Practically, the Euler scheme is the most common approach, but its time
discretization is accompanied by inherent biases (Glasserman, 2013). The alternative approach, which generates unbiased
samples, is called the exact simulation scheme. Beskos et al. (2005) proposed an exact simulation scheme for a solution
of SDE, and Chen and Huang (2013) extended this to a wider class of SDEs. Casella and Roberts (2008) developed an exact
simulation for a killed diffusion. Despite these extensive works, even for Brownian motion, few have provided an exact
simulation scheme for the first hitting time to non-constant boundaries. In this paper, we develop one to simulate the
first passage time of Brownian motion and a Brownian bridge to a class of expanding boundaries, including asymmetric
and piecewise linear boundaries.
The distributions of the first hitting time of Brownian motion to certain boundaries have been well-studied in several
works. For the case of a one-sided linear boundary, we have a well-known analytic formula (Karatzas and Shreve, 1998).
For the cases of two-sided linear boundaries, Abundo (2002) included the most general results. For the cases of piecewise
linear boundaries, the formulae usually appeared as convolutions (Wang and Pötzelberger, 2007; Jin and Wang, 2017;
Dong and Cui, 2019). There is no closed-form solution applicable to a general nonlinear boundary, but there are studies
that apply particular boundaries (Salminen, 1988; Novikov et al., 1999). However, owing to the complexity of their analytic
formulae, exact sampling from these distributions cannot be directly followed.
For a standard Brownian motion, B = {Bt }t ≥0 , Burq and Jones (2008) established an exact simulation for the first hitting
time, τ := inft ≥0 { t : |Bt | = 1 }, and Lee (2017) extended this to an exact simulation scheme for τa,b := inft ≥0 { t : |Bt | =
a + bt }. Our work extends the scope of these works. However, our work differs from previous works in two ways. First,
instead of only sampling the first hitting time, τ , we sample the tuple (τ , Bτ ) simultaneously. Next, we consider hitting

E-mail address: highseeker@kaist.ac.kr.


1 Currently with Block Crafters Capital Ltd.

https://doi.org/10.1016/j.spl.2022.109654
0167-7152/© 2022 Published by Elsevier B.V.
T. Lee Statistics and Probability Letters 193 (2023) 109654

times under a finite time horizon, where the simultaneous sampling becomes more difficult, as well as under an infinite
time horizon. These differences are required for our localization technique, which is a key tool to extend the scope of
our algorithm considerably. Moreover, they make our algorithm useful for certain applications, including asset pricing in
mathematical finance, where it is more realistic to assume a finite time horizon and sometimes necessary to know its
state Bτ , at the hitting time.
The remainder of this paper is organized as follows. In Section 2, we develop exact simulation scheme for two-sided
symmetric linear boundaries with a finite time horizon. In Section 3, we construct an iterative algorithm that is applicable
to asymmetric linear or other boundaries. In Section 4, we show that our algorithm can simulate the first hitting time of
a Brownian bridge. In Section 5, we present numerical results. Finally, we conclude the paper in Section 6.

2. Two-sided symmetric linear boundaries with a finite time horizon

In this section, we develop an exact simulation scheme for the tuple (τ , Bτ ) where τ is the first hitting time of standard
Brownian motion, {Bt }t ≥0 , to a boundary composed of two symmetric lines and a finite time horizon. Before we start, we
briefly discuss previous works that dealt with the cases having an infinite time horizon.
Let τa,b = inft ≥0 { t : |Bt | = a + bt } be the first hitting time of Brownian motion to symmetric linear boundaries, and
let fτa,b be the density function of τa,b . For the two-sided constant boundaries with (a, b) = (1, 0), Burq and Jones (2008)
showed that the density, fτ1,0 , is dominated by a scaled-up density function of gamma distribution, using this to sample
τ1,0 via the acceptance–rejection method. Lee (2017) showed that for any nonnegative coefficients, a, b, the same can be
done with the gamma distribution whose coefficients depend on b (see Theorem 1.1, Lee and Lee (2020)). Thus for any
given nonnegative coefficients, a, b, τa,b can be simulated using the acceptance–rejection method with the corresponding
gamma distribution (see Algorithm 2 and 3 in Lee and Lee (2020) for details).
Next, we consider a boundary consisting of two symmetric lines, ±(a + bt), with nonnegative coefficients a, b, and
a finite time horizon, T > 0. Let τ be the first hitting time of a standard Brownian motion to the boundary. Note that
there exists an exact simulation scheme for τa,b , thus for τ = τa,b ∧ T . Nevertheless, the exact sampling of Bτ needs
additional consideration. In the event of [τ = τa,b ], because the boundary is symmetric, we have P[Bτ = (a + bτ )] =
P[Bτ = −(a + bτ )] = 1/2. Thus, it remains to sample Bτ on [τa,b > T ], which is the tricky part of our extension from an
infinite time horizon to a finite one. For this sampling, we apply the boundary crossing probability of a Brownian bridge
from (Abundo, 2002):
( ⏐ ) ∞
(−1)n · e(−2an((b+a/T )n−z /T )) .

∪ [|Bt | ≥ a + bT ] ⏐ BT = z = 1 − (1)

P
t ∈[0,T ]
n=−∞

Let h and h̃ be two conditional densities of BT defined by

h(z ; a, b, T ) := P[BT ∈ dz | τa,b > T ] and


h̃(z ; a, b, T ) := P[BT ∈ dz | |BT | < a + bT ].

Our goal is to sample from the density h(z ; a, b, T ) and the following theorem characterizes the likelihood ratio between
the two densities h and h̃.

Theorem 2.1. Let a, b, T > 0. For any |z | < a + bT , we have


h(z ; a, b, T )
∝ H̃(z , a, b, T ) ≤ 1 (2)
h̃(z ; a, b, T )

where H̃(z , a, b, T ) is defined by



(−1)n · e−2an((b+a/T )n−z /T ) .

H̃(z , a, b, T ) = (3)
n=−∞

Proof. From Eq. (1), we have


( ⏐ )
P[τa,b > T , BT ∈ dz ] = P ∩ [|Bt | < a + bT ] ⏐ BT = z · P[BT ∈ dz ] = H̃(z , a, b, T ) · φT (z) (4)

t ∈[0,T ]

√ 1 e−x /2T .
2
where φT (z) := P[BT ∈ dz ] = Dividing both sides by the normalizing constant Ca,b,T = P[τa,b > T ] yields
2π T

h(z ; a, b, T ) = H̃(z , a, b, T ) · φT (z)/Ca,b,T (5)

Similarly, we have for h̃(z ; a, b, T ),

h̃(z ; a, b, T ) = φT (z)/C̃a,b,T (6)


2
T. Lee Statistics and Probability Letters 193 (2023) 109654

where C̃a,b,T = P[|BT | ≤ a + bT ]. By taking the ratio of Eqs. (5) and (6), we obtain Eq. (2) as

h(z ; a, b, T ) C̃a,b,T
= · H̃(z , a, b, T ) ∝ H̃(z , a, b, T ) ≤ 1.
h̃(z ; a, b, T ) Ca,b,T

where the last inequality is satisfied because H̃(z , a, b, T ) is the probability, P[τa,b > T | BT = z ]. □

Because the conditioning event [|BT | < a + bT ] only depends on BT , we can sample BT ∼ h̃ by sampling BT from
N(0, T ), repeatedly until |BT | ≤ a + bT . Through Theorem 2.1, using the acceptance–rejection method with the proposal
density h̃, we can sample BT ∼ h and complete the simulation of the tuple (τ , Bτ ). See Algorithm 1 for further details.

Algorithm 1 Exact simulation of (τ , Bτ ) where τ is the first hitting time of a Brownian motion to a boundary consisting
of two symmetric lines, ±(a + bt), and a finite time horizon, T .
1: Generate S ∼ fτa,b . ▷ Use Algorithms 2 and 3 in Lee and Lee (2020)
2: if S < T then ▷ Case [τ < T ]
3: B = ±(a + bS) where the sign is chosen with probability 1/2 each.
4: else ▷ Case [τ = T ]
5: S=T
6: Accept = FALSE
7: repeat
8: Generate Z ∼ N(0, T ) and U ∼ U [0, 1]
9: if |Z | < a + bT & U < H̃(Z , a, b, T ) then
10: B=Z
11: Accept =TRUE
12: end if
13: until Accept
14: end if
15: return (S , B)

However, to assert that Algorithm 1 is an exact simulation scheme, an exact evaluation of the inequality, U <
H̃(Z , a, b, T ), is required. Because H̃(Z , a, b, T ) is an alternative series,
∑n we cank determine the inequality, U < H̃(Z , a, b, T ),
−2ak((b+a/T )k−z /T )
exactly in a finite time from its partial sum, H̃n (Z , a, b, T ) := k=−n (−1) · e , thanks to the following
criteria learned from (Chen and Huang, 2013):

H̃2m−1 (Z , a, b, T ) ∧ H̃2m (Z , a, b, T ) > U for some m ≥ 1 ⇒ H̃(Z , a, b, T ) > U


H̃2m−1 (Z , a, b, T ) ∨ H̃2m (Z , a, b, T ) < U for some m ≥ 1 ⇒ H̃(Z , a, b, T ) < U .

Note that the above criteria must end in a finite step unless H̃(Z , a, b, T ) = U, which is a negligible event.

3. Extension to asymmetric boundaries and other boundaries

Next, we extend the scope of our scheme to boundaries that are not necessarily symmetric. For any region bounded
by two-sided boundaries that are asymmetric and linear, one can define the largest symmetric subset of the region. We
call the boundary of this subset the ‘‘sub-symmetric boundary’’. Two types of sub-symmetric boundaries exist. If the sub-
symmetric boundary and the original boundary share the finite time horizon, we call the case ‘‘Type I’’. Otherwise, we
call it ‘‘Type II’’. We illustrate both types in Fig. 1 and use the dotted line to represent the segments of the sub-symmetric
boundary that are not included in the original boundary.
Intuitively, our idea is to repeatedly generate the first hitting time to the sub-symmetric boundary until the tuple
(τ , Bτ ) reaches the original boundary. Because the sub-symmetric boundary is symmetric and linear, we use Algorithm
1 to generate the tuple (τ1 , Bτ1 ), where τ1 is the first hitting time to the sub-symmetric boundary. The algorithm ends if
(τ1 , Bτ1 ) is in the original boundary. Otherwise, we move to (τ1 , Bτ1 ) as a new starting point and generate (τ2 , Bτ2 ) with the
sub-symmetric boundary redefined there and repeat. Theoretically, we decompose the original boundary into a random
sequence of sub-symmetric boundaries based on the strong Markov property of Brownian motion. We call this technique
as ‘‘boundary localization technique’’. Our technique is inspired by Chen and Huang (2013), which used a localization
technique to extend the scope of the method developed by Beskos et al. (2005). See Algorithm 2 for details.
Because Algorithm 2 is an iterative algorithm, we estimate the cost of iterations to assert its efficiency. Theorem 3.1
shows that the expected number of iterations in Algorithm 2 is uniformly bounded.

Theorem 3.1. Let M be the number of sub-symmetric boundaries in Algorithm 2. Then, we have

E[M ] ≤ 8.
3
T. Lee Statistics and Probability Letters 193 (2023) 109654

Fig. 1. (a) If the reflection of one linear boundary is dominated by the other linear boundary on [0, T ], the sub-symmetric boundary and the original
boundary have the same finite time horizon T . We call the case Type I. (b) Else, the sub-symmetric boundary has a new boundary on time which
is prior to the original boundary on time. We call the case Type II.

Algorithm 2 Exact simulation of the first hitting time of Brownian motion to two-sided asymmetric linear boundaries
1: Set (T , B) = (0, 0) and k = 0
2: Touched = FALSE
3: repeat
4: Generate (Tk , Bk ) the first hitting time to the sub-symmetric boundary. ▷ Algorithm 1
5: T = T + Tk , B = B + Bk .
6: if (T , B) is in the original boundary then
7: Touched = TRUE
8: else
9: k=k+1
10: Set (T , B) the origin and reset the boundary there.
11: end if
12: until Touched
13: return (T , B)

Proof. In each iteration of Algorithm 2, the algorithm is in one of the following three states.

State I: The sub-symmetric boundary is of Type I;


State II: The sub-symmetric boundary is of Type II;
State III: It reaches the original boundary, and thus the algorithm terminates.

We write [i → j] to represent the event where the state changes from state i to state j. Let M be the number of
state changes from the initial state to State III, which is equal to the number of sub-symmetric boundaries required in
Algorithm 2. If the algorithm is on State I, it hits the original boundary with a probability above 1/2. Thus, intuitively, the
algorithm ends in a finite number of steps unless the algorithm is on State II repeatedly. Our claim is that P[M ≤ 2] ≥ 1/4,
regardless of its initial state. We prove this claim. If Algorithm 2 starts from State I, the algorithm terminates if our sample
hits the linear boundary of the original boundary or the finite time horizon, T (see Fig. 2(a)). Therefore, we have

P[M ≤ 2] ≥ P[M = 1] = P[I → III ] ≥ 1/2.

Else, Algorithm 2 starts from State II, event [II → III ] occurs if and only if the sample hits the original boundary.
Otherwise, if the sample hits half of the vertical line segment opposite the steeper linear boundary, the next state is of
Type I since the boundary with a gentle slope is reflected. These observations are shown in Fig. 2(b). Because the sum of
these two segments amounts to half of the sub-symmetric boundary, we have P[II → (I or III)] ≥ 1/2. Therefore, when
Algorithm 2 starts from State II, we have

P[M ≤ 2] ≥ P[II → III ] + P[II → I → III ]


≥ P[II → III ] + P[II → I ] · P[I → III ]
4
T. Lee Statistics and Probability Letters 193 (2023) 109654

Fig. 2. With a given initial state, the next state is determined by the segment hit.

Fig. 3. Boundary localization technique can be applied to the cases of a one-sided linear boundary and piecewise linear boundaries. (a) For the
case of a one-sided linear boundary, the sub-symmetric boundary is composed of the original boundary and its reflection. As in Algorithm 2, we
iteratively use Algorithm 1 until it reaches the original boundary. (b) For the case of piecewise linear boundaries, we use Algorithm 2 to generate
the first hitting time to the boundary restricted on [0, T1 ]. The algorithm terminates if the sample hits the upper or lower boundaries. Else, the
algorithm moves forward to the next interval and repeats.

1
≥ · (P[II → III ] + P[II → I ])
2
1
= · P[II → (I or III)] ≥ 1/4 .
2
Then we complete the proof of our claim by combining the results with both initial states. Consequently, from the strong
Markov property of Brownian motion, the number of iterations required to reach State III is dominated by 2N, where N
is the geometric random variable with p = 1/4, and we conclude that E[M ] ≤ E[2N ] = 8. □

Finally, we describe how we apply our scheme to one-sided or piecewise linear boundaries, under the assumption that
the boundaries are expanding (i.e., the upper boundary is non-decreasing, and the lower boundary is non-increasing). In
each case, we use the boundary localization technique similarly but slightly differently.
For the case of a one-sided linear boundary, we define the sub-symmetric boundary as the union of the original
boundary and its reflection (see Fig. 3(a)). As in Algorithm 2, we use Algorithm 1 to generate the first hitting time to
the sub-symmetric boundary until it hits the original boundary iteratively. Then, our sample hits the original boundary
with a probability above 1/2 in each iteration, whether it is bounded in time or not. Hence the expected number of
iterative use of Algorithm 1 is bounded by two.
For the case of piecewise linear boundaries that are expanding with time nodes 0 < T1 < T2 · · · , we restrict the
boundaries to the first time interval [0, T1 ] and use Algorithm 2 to generate the first hitting time to the restricted
boundaries with the finite time horizon T1 . The algorithm terminates if our sample hits either the upper or lower boundary.
Otherwise, we move forward to the next time interval [T1 , T2 ] and repeat. In this case, the number of iterative use of
Algorithm 2 is bounded by the number of pieces that make up the piecewise linear boundary.
5
T. Lee Statistics and Probability Letters 193 (2023) 109654

As a result, our algorithm can be applied to a boundary whether it is symmetric or asymmetric, linear or piecewise
linear, bounded in time or not. All we need for the boundary in our scheme is that it is locally linear and expanding.

4. Extension to a Brownian bridge

Let BT ,K be the Brownian bridge pinned at (0, 0) and (T , K ). We recall that a Brownian bridge can be represented as a
time-changed Brownian motion with a linear drift (see Borodin and Salminen, 2012 Ch.4):
( )
T ,K T −t
Bt = BtT /(T −t) + (t /T )K = (T − t)Wt /T (T −t) + (t /T )K , 0≤t≤T (7)
T
where B and W are standard Brownian motions, and the second equality comes from the scaling property of Brownian
motion. Eq. (7) allows us to connect the hitting time of the Brownian bridge to the hitting time of Brownian motion.

Lemma 4.1. Assume that the pinned points (0, 0) and (T , K ) of a Brownian bridge, BT ,K , are located between a lower boundary,
−(A1 + B1 t), and an upper boundary, (A2 + B2 t). Let τ be the first hitting time of the Brownian bridge, BT ,K , to these boundaries.
Then we have
d τ̃ · T 2
τ= (8)
(τ̃ · T + 1)
where τ̃ is the first hitting time of a standard Brownian motion to the two-sided linear boundaries, −(Ã1 + B̃1 t) and (Ã2 + B̃2 t),
whose coefficients are determined by

(Ã1 , Ã2 , B̃1 , B̃2 ) = (A1 /T , A2 /T , A1 + B1 T + K , A2 + B2 T − K ). (9)

Proof. Using Eq. (7), we have


[ ] [
−A1 − B1 t < BTt ,K < A2 + B2 t = −A1 − B1 t < (T − t)Wt /T (T −t) + (t /T )K < A2 + B2 t .
d ]
(10)

By substituting s = t /(T (T − t)) and t = sT 2 /(sT + 1), the lemma is proved by the following equivalencies:
[ ]
−A1 − B1 t < (T − t)Wt /T (T −t) + (t /T )K < A2 + B2 t
sT 2 sT 2
( ) ( )]
[ T (Ws + sK )
⇔ −A1 − B1 < < A2 + B2
sT + 1 sT + 1 sT + 1
( ) ( )
[ sT + 1 sT + 1 ]
⇔ −A1 − B1 (sT ) < Ws + sK < A2 + B2 (sT )
T T
[ ]
⇔ −(A1 /T ) − (A1 + B1 T + K )s < Ws < (A2 /T ) + (A2 + B2 T − K )s . □
Lemma 4.1 shows that the hitting time of the Brownian bridge can be represented as the time-changed hitting time
of a standard Brownian motion, while maintaining the linearity of the boundaries. The coefficients A1 , A2 , (A1 + B1 T + K ),
and (A2 + B2 T − K ) are all positive if and only if the pinning points (0, 0) and (T , K ) of the Brownian bridge are located
between its boundaries, −(A1 + B1 t) and (A2 + B2 t). Hence, we have the following theorem.

Theorem 4.1. An exact simulation scheme exists for the first hitting time of the Brownian bridge, BT ,K , to two-sided linear
boundaries such that the pinning points of the Brownian bridge are located between the boundaries.

5. Numerical examples

To demonstrate our algorithms numerically, we generated samples from them and compared the empirical distribution
to its true distribution. For this, we considered the cases in which their analytic forms of the hitting distributions are
known. Specifically, we considered the hitting time distributions of Brownian motion to the following two boundaries:

B1. Two-sided symmetric linear boundaries ±(1 + t) with a finite time horizon, T = 1;
B2. Two-sided asymmetric linear boundaries −(1 + t /2) and (1/2 + t) with a finite time horizon, T = 1.
For their analytic formulae of true distributions, we refer to Abundo (2002) (see proposition 3.5).
The experiment was simulated using MATLAB, and the results are shown in Fig. 4. In the first row, we consider the first
hitting time of Brownian motion to boundary B1, and the samples are generated using Algorithm 1. In the second row, we
consider the first hitting time of Brownian motion to boundary B2, and the samples are generated using Algorithm 2. Each
column shows the experimental results when the number of samples N is 102 , 103 , and 104 , respectively. The distance
between the empirical distribution FN (x) and its true distribution F (x) is measured by the Kolmogorov–Smirnov statistic
D = supx |FN (x) − F (x)| (Smirnov, 1948). However, since the distribution of stopping time has a jump discontinuity at its
6
T. Lee Statistics and Probability Letters 193 (2023) 109654

Fig. 4. In the first row, we consider the first hitting time of Brownian motion to the boundary, B1, comprises two symmetric lines ±(1 + t) and
a finite time horizon, T = 1. In the second row, we consider the first hitting time of Brownian motion to the boundary, B2, which comprises two
asymmetric lines, −(1 + 2t ) and ( 21 + t), and a finite time horizon, T = 1. In each column, we generate 102 , 103 , and 104 samples, respectively, and
the number of samples is indicated by N. In each figure, we compare two hitting time distributions: (1) the empirical distribution of τ from the
samples generated by our method; (2) the true distribution √ of τ from its analytic formula. D is an estimated Kolmogorov–Smirnov statistic (Smirnov,
1948), and all the cases satisfy the criteria [DN < 1.63/ N ] for the Kolmogorov–Smirnov test with a 99% confidence level.

Table 1
When using Algorithm 2, we estimate the average number of iterations, Kav g , and the
maximum number of iterations, Kmax . In the experiment, we consider the boundary B2,
composed of two asymmetric lines, −(1 + 2t ) and ( 12 + t), and a finite time horizon, T = 1.
In each column, we generate 102 , 103 , 104 , 105 , and 106 samples, respectively.
N 102 103 104 105 106
Kav g 1.820 1.767 1.709 1.718 1.716
Kmax 5 6 7 8 9

boundary on time, plotting of the plain empirical distribution can be distorted by the jump. Thus, for clarity, we plot the
conditional distribution, F̃ (t) = P(τ ≤ t | τ < 1) instead. Fig. 4 shows that as N increases, the empirical distribution gets
closer to its true distribution while the Kolmogorov–Smirnov
√ statistic decreases to 0. In addition, we observe that all the
cases in Fig. 4 satisfy the criteria [DN < 1.63/ N ] for the Kolmogorov–Smirnov test with a 99% confidence level.
Separately, with boundary B2, we provide additional experiments by increasing the number of samples up to N = 106
to numerically demonstrate the upper bound of Theorem 3.1. Table 1 lists the numerical results for the average and
maximum numbers of iterations. We observe that the average number of iterations, Kav g , is bounded by the theoretical
upper bound, and the maximum number of iterations, Kmax , has logarithmic growth with respect to N.

6. Conclusion

We proposed an exact simulation scheme for the first hitting time of Brownian motion and a Brownian bridge. We
developed our scheme for symmetric linear boundaries with a finite time horizon, and extended it to more complicated
boundaries via the boundary localization technique. As a result, our scheme is applicable to any boundary that is locally
linear and expanding. We also applied our scheme to generate the first hitting time of a Brownian bridge to linear
boundaries. Although exact simulation schemes often tend to be inefficient due to their complexity, our method is simple
and efficient, as demonstrated by numerical experiments.

Data availability

Data will be made available on request.


7
T. Lee Statistics and Probability Letters 193 (2023) 109654

References

Abundo, M., 2002. Some conditional crossing results of Brownian motion over a piecewise-linear boundary. Statist. Probab. Lett. 58 (2), 131–145.
Asmussen, S., Glynn, P.W., 2007. Stochastic Simulation: Algorithms and Analysis. vol. 57. Springer Science & Business Media.
Beskos, A., Roberts, G.O., et al., 2005. Exact simulation of diffusions. Ann. Appl. Probab. 15 (4), 2422–2444.
Borodin, A.N., Salminen, P., 2012. Handbook of Brownian Motion-Facts and Formulae. Birkhäuser.
Burq, Z.A., Jones, O.D., 2008. Simulation of Brownian motion at first-passage times. Math. Comput. Simulation 77 (1), 64–71.
Casella, B., Roberts, G.O., 2008. Exact Monte Carlo simulation of killed diffusions. Adv. Appl. Probab. 40 (1), 273–291.
Chen, N., Huang, Z., 2013. Localization and exact simulation of Brownian motion-driven stochastic differential equations. Math. Oper. Res. 38 (3),
591–616.
Dong, Q., Cui, L., 2019. First hitting time distributions for Brownian motion and regions with piecewise linear boundaries. Methodol. Comput. Appl.
Probab. 21 (1), 1–23.
Glasserman, P., 2013. Monte Carlo Methods in Financial Engineering. vol. 53. Springer Science & Business Media.
Jin, Z., Wang, L., 2017. First passage time for Brownian motion and piecewise linear boundaries. Methodol. Comput. Appl. Probab. 19 (1), 237–253.
Karatzas, I., Shreve, S.E., 1998. Brownian Motion and Stochastic Calculus. Springer.
Lee, J.M., 2017. Efficient Simulation for Greeks and Bridges Under Diffusion Models (Ph.D. thesis). KAIST.
Lee, J.M., Lee, T., 2020. Exact simulation of the first passage time of Brownian motion to a symmetric linear boundary. arXiv preprint arXiv:2007.05986.
Novikov, A., Frishling, V., Kordzakhia, N., 1999. Approximations of boundary crossing probabilities for a Brownian motion. J. Appl. Probab. 36 (4),
1019–1030.
Pereyra, M., Schniter, P., Chouzenoux, E., Pesquet, J.-C., Tourneret, J.-Y., Hero, A.O., McLaughlin, S., 2015. A survey of stochastic simulation and
optimization methods in signal processing. IEEE J. Sel. Top. Sign. Proces. 10 (2), 224–241.
Salminen, P., 1988. On the first hitting time and the last exit time for a Brownian motion to/from a moving boundary. Adv. Appl. Probab. 20 (2),
411–426.
Smirnov, N., 1948. Table for estimating the goodness of fit of empirical distributions. Ann. Math. Stat. 19 (2), 279–281.
Wang, L., Pötzelberger, K., 2007. Crossing probabilities for diffusion processes with piecewise continuous boundaries. Methodol. Comput. Appl. Probab.
9 (1), 21–40.

You might also like