From random walks to a continuous-time limit
The symmetric random walk of the gambler's-ruin notes moves by at each integer time. If the step size is shrunk to and time is sped up by the factor , the rescaled walk has, at each fixed , variance close to and, by the central limit theorem, an approximately normal law. The limiting continuous-time process is Brownian motion. This unit gives the definition, the first consequences, and the reflection principle, which transfers the symmetry of the walk to the continuous setting and yields explicit laws for the running maximum and for hitting times. It builds on the unit on martingales and bounded stopping, whose optional stopping argument is used again here, and on Gaussian laws and independence.
Definition and first properties
Definition 5.1.1
(Standard Brownian motion)
A standard Brownian motion is a stochastic process with:
- ;
- independent increments: for , the variables are independent;
- Gaussian stationary increments: for ;
- continuous sample paths: is continuous with probability one.
That a process with these properties exists is a theorem of Wiener; we use it without proof. Let be the information generated by .
Proposition 5.1.2
(Covariance, scaling, symmetry and martingales)
For :
- , and every finite family is jointly Gaussian with mean zero.
- (Scaling) For , the process is again a standard Brownian motion.
- (Symmetry) is a standard Brownian motion.
- , , and for each real are martingales with respect to .
Proof
(Increments, rescaling and conditional expectations)
(1) For write . The two summands are independent, so . Any vector is a linear image of the independent Gaussian increments, hence jointly Gaussian.
(2) The scaled process has , continuous paths, independent increments, and .
(3) Negation preserves all four properties, because is symmetric.
(4) For , . Next , and the independence of from gives , so is a martingale. For the exponential, independence gives , the Gaussian moment generating function, so .
Part (1) has a converse, which gives a second working definition of the same process.
Concept lensStructural
Brownian motion as a Gaussian process
A process with continuous paths whose finite families are jointly Gaussian with mean zero and is a standard Brownian motion. Indeed, gives . For the covariance of the increments is , and . Jointly Gaussian variables that are uncorrelated are independent, so the increments are independent with law . This description replaces a check of independence by a computation of one covariance function; it is how the scaling example below can be verified in one line, since .
The reflection argument needs one further property, which we use without proof.
Theorem 5.1.3
(Strong Markov property)
If is a stopping time, then conditionally on the process is a standard Brownian motion independent of .
The sample paths, though continuous, are nowhere differentiable; we do not need this.
Exit from an interval
Let and .
Proposition 5.1.4
(Exit probabilities and mean exit time)
The exit time of is finite almost surely, and
Proof
(Optional stopping at a truncated exit time)
Fix . Optional stopping at the bounded stopping time , the continuous-time analogue of the bounded theorem (used here without proof), gives and . Since , we get , so by monotone convergence and almost surely. Dominated convergence now gives and . With :
These are the same formulas as for the symmetric random walk in the gambler's-ruin problem, with and replaced by .
The reflection principle
For let be the first hitting time of level , and let . By continuity, if and only if .
Theorem 5.1.5
(Reflection principle)
For and ,
where is the standard normal distribution function. More generally, for and ,
Proof
(Reflecting the path after the hitting time)
The time is a stopping time, possibly infinite. On the event define the reflected path
By the strong Markov property, is a Brownian motion independent of , and by symmetry its negative has the same law; is the path obtained by negating the post- increments. Therefore has the same law as , and is also the first time reaches . On we have if and only if . Hence
the second equality because and have the same law and the same hitting time. Since , the event is contained in by continuity (the path must reach on the way to ), so the right side equals . This proves the second formula. For the first, take : . Since and ,
Finally , so .
Proof X-ray
Where the strong Markov property and continuity enter
The strong Markov property is used once, and for a specific purpose: because the increments after form a Brownian motion independent of everything before , negating them leaves the joint law of the whole path unchanged. Independence matters here; negating a piece that depended on the past could change the law. If were replaced by a time that looks into the future, such as the time at which the maximum over is attained, the increments after it are nonpositive up to time and form no fresh Brownian motion, so the argument has no starting point.
The key move is a bijection between paths: among paths that reach by time , ending at or below corresponds to ending at or above . Continuity is used twice: the reflected path first reaches at the same time , and a path ending at or above must already have crossed . The argument closes with the case , where the null event lets the two halves and be added to give the factor .
The reflection principle gives the whole law of the hitting time .
Corollary 5.1.6
(Hitting times and the running maximum)
(a) for every , but . (b) has the density for . (c) has the same law as .
Proof
(Reading off the tail of the hitting time)
(a) as . Since as and , . Since and diverges, .
(b) Differentiate in .
(c) By the reflection principle, for , and both variables are nonnegative.
Part (a) is the continuous counterpart of the unbounded stopping example in the martingale notes, and it shows exactly which hypothesis of optional stopping cannot be dropped.
Counterexample mode
Optional stopping at an almost surely finite time
Claim: for a martingale and an almost surely finite stopping time , . Take and . Part (a) gives , and continuity gives on that event, so . The claim fails because is unbounded and the stopped process , although at most , is not bounded below; moreover . The repaired statement is the one used for the exit problem: stop at a bounded time such as , and pass to the limit only when the stopped process is bounded, as was there.
Worked examples
Values used: , .
Worked example 5.1.7
(The running maximum)
Find . By the reflection principle with , :
Worked example 5.1.8
(A scaling check)
The same number arises from scaling. Take : is a standard Brownian motion, and it reaches at time exactly when reaches at time . So for equals , where is the hitting time of for , and has the law of . Hence , in agreement with the running-maximum example.
Worked example 5.1.9
(Maximum and endpoint together)
For , find . The reflection principle with , gives . As a consistency check, , so . This is larger than , as it must be, since a path with may end anywhere above .
Worked example 5.1.10
(Exit from an interval)
For , , the probability that reaches before is and the expected exit time is . As with fixed, the probability of reaching first tends to while the mean exit time tends to infinity, consistent with from the corollary and symmetry, because the exit time never exceeds .
Worked example 5.1.11
(A Gaussian vector)
with independent summands, so it is . The covariance formula gives the same answer: .
Common mistakes and subtle points
Common mistake
(Reflecting with only the simple Markov property)
The reflection principle needs the strong Markov property at a hitting time, which is a random time. The ordinary Markov property at a fixed time says nothing about the increments after , so it would not justify reflecting there.
Common mistake
(Dropping the factor 2)
Writing counts only the paths that end above . Every path that touches and ends below is matched by a reflected path that ends above , so the two classes have equal probability. For , the wrong formula gives instead of .
Common mistake
(Scaling time like space)
Scaling multiplies positions by and times by . The claim that has the law of would give , not the correct ; the correct relation is .
Two further points. The formula is for the first hitting of a level by a process starting at ; for a level use symmetry. And is a martingale, but is not; the paths of have unbounded variation on every interval, so integrals against are not defined path by path.
Exercises
Quick check
(Compute in terms of and numerically.)
Apply the reflection principle with .
Solution · Answer
By the reflection principle with , : .
Quick check
(Find the distribution of .)
Use the covariance .
Solution · Answer
The sum is a mean-zero Gaussian with variance . So .
Quick check
(Compute and use it to check the reflection principle at .)
Split according to whether .
Solution
By the reflection principle, . Check: , which equals .
Quick check
(For and , find the probability that hits before and the mean exit time of .)
Use the exit-time proposition.
Solution · Answer
The probability of reaching before is , and .
Quick check
(Show that using the density in part (b) of the corollary.)
Bound the exponential factor from below for large .
Solution
By (b), . For the exponential is at least , so the integral is at least , which diverges. Hence .