MATH 5050: Practice In what follows we will need a multidimensional variant of Ito’s formula. But first, let us recall Taylor’s expansion in multi dimensions: Let f (y1 , . . . , yd ) be a function of d variables that has two continuous derivatives in each coordinate. Let (δ1 , . . . , δd ) be d small increments. Then d X ∂f f (y1 + δ1 , . . . , yd + δd ) = f (y1 , . . . , yd ) + (y1 , . . . , yd ) × δi ∂yi i=1 + 1 2 d X d X i=1 j=1 ∂2f × δi δj + errors smaller than any δi δj . ∂yi ∂yj Now let B1 (t), . . . , Bd (t) be d independent standard Brownian motions. Let f (y1 , . . . , yd ) be a function of d variables that has two continuous derivatives in each coordinate. Let Z(t) = f (B1 (t), . . . , Bd (t)). Then applying the above Taylor expansion, replacing every (dBi (t))2 by dt and ignoring anything that is smaller than dt (e.g. (dt)2 or (dt)(dBi (t))), we arrive at d X 1 ∂f dZ(t) = (B1 (t), . . . , Bd (t)) dBi (t) + ∆f (B1 (t), . . . , Bd (t)) dt . ∂yi 2 i=1 P 2 (Recall that ∆f (y1 , . . . , yd ) = di=1 ∂∂yf2 (the sum of all second derivatives). i Make sure you understand the above derivation before you keep going (like where the 2f ∆f came from, and why things like ∂y∂1 ∂y are not there). 2 Problem 1 Let d ≥ 2 and let B1 (t), . . . , Bd (t) be d independent standard Brownian motions. Let x1 , . . . , xd be numbers not all 0 at once. Abbreviate B(t) = (B1 (t), . . . , Bd (t)) (the d-dimensional standard Brownian motion) and x = (x1 , . . . , xd ) (the d-dimensional vector). Let M (t) = kB(t) − xk2−d = ((B1 (t) − x1 )2 + · · · + (Bd (t) − xd )2 )1−d/2 . (a) Prove that M (t) is a martingale relative to the filtration Ft given by the Brownian motions B1 (s), . . . , Bd (s), s ≤ t. Hint: Apply the above Ito formula to the function f (y1 , . . . , yd ) = ky − xk2−d = ((y1 − x2 )2 + · · · + (yd − xd )2 )1−d/2 . (b) What is E[kB(t) − xk2−d ]? 1 2 The next problem is a bit long, but the steps are not too difficult. So be patient. Problem 2 Let W (t) be standard Brownian motion. Consider a process D(t) that is the solution of the stochastic differential equation d−1 dD(t) = (1) dt + dW (t). 2D(t) Note that D(t) is a one dimensional process. The number d is just a parameter here. We will only consider the case of integer d ≥ 2, although the process is well defined for any real number d (even negative!). Such a process is called a Bessel process. Note that it has diffusion coefficient 1, but it also has a drift that is proportional to the value of the process. I.e. the closer is D(t) to 0 the more it is pushed away from 0. Furthermore, the larger d is the larger the drift. We want to prove that if d ≥ 3 then the drift is large enough to make D(t) transient, while if d = 2, then the drift is not large enough and D(t) is neighborhood recurrent. I.e. we want to show that: when d ≥ 3 the process has a positive chance of going above any specified number R and never going back below it, no matter how large R is chosen, while when d = 2, the process will get eventually smaller than any number ε > 0, no matter how small ε is chosen to be. To achieve the above, we will work on recognizing what D(t) really is. Let d ≥ 2 (the same d as above) and let B1 (t), . . . , Bd (t) be d independent standard Brownian motions. Define q D(t) = B12 (t) + . . . + Bd2 (t). This D(t) is the distance of the d-dimensional Brownian motion B(t) = (B1 (t), . . . , Bd (t)) to the origin (0, . . . , 0). For now, it is not clear that this D(t) is the same as the Bessel process defined in (1). But this is what we will prove in parts (b)-(f). (a) Prove that if it is true that the Bessel process in (1) is indeed simply the distance of a standard d-dimensional Brownian motion to the origin 0, then the transience/recurrence properties we are after do hold. Now we will prove that the distance of a standard d-dimensional Brownian motion is in fact a Bessel process with parameter d, as in (1). (b) For any given t > 0, explain why P {D(t) = 0} = 0. (c) Using the multidimensional Ito formula prove that dD(t) = d−1 B1 (t) Bd (t) dt + dB1 (t) + · · · + dBd (t) . 2D(t) D(t) D(t) The above equation can be written as dD(t) = d−1 dt + dW (t) 2D(t) 3 where Z W (t) = 0 t B1 (s) dB1 (s) + · · · + D(s) Z 0 t Bd (s) dBd (s) . D(s) To show that D(t) satisfies (1) we need to prove that W (t) is in fact a Brownian motion. Being a stochastic integral, W (t) is continuous and a martingale with respect to the filtration Ft of the Brownian motions B1 (s), . . . , Bd (s), s ≤ t. (d) If s < t, what is E[W (t) − W (s)]? (e) If s < t, what is E[(W (t) − W (s))2 ]? Hint: you can use the fact that stochastic integrals with respect to independent Brownian motions are themselves independent. Hence, 2 i h Z t B (s) 2 i h Z t B (s) 1 1 2 dB1 (s) + ··· + E dBd (s) . E[(W (t) − W (s)) ] = E 0 D(s) 0 D(s) (The averages of the cross terms cancel out because they split by independence into products of averages and each of the averages is that of a martingale that is 0 at time 0. Now write this down carefully and in details!) Then, recall that stochastic integrals have the following Rt Rt isometry property: E[( 0 Y (s) dB(s))2 ] = 0 E[Y (s)2 ] ds. Apply this property to each of the above averages of squares, and then combine into one integral of one average of a quantity. See what that becomes! (f) [Within the scope of the course, but beyond the level of the exam.] Prove that W (t) − W (s) is Normal with mean 0 and variance t − s and is independent of the Brownian motions B1 (r), . . . , Bd (r) for r ≤ s and hence is independent of W (r), r ≤ s. Hint: Fix a real number λ. Use Ito’s formula to show that eλW (t)−λ 2 This shows that E[eλ(W (t)−W (s)) | Fs ] = eλ (t−s)/2 . Conclude. 2 t/2 is a martingale. As a conclusion, we see that the distance D(t) of a standard d-dimensional Brownian motion from the origin is in fact a diffusion process that satisfies the stochastic differential equation (1). This fact helped us deduce transience/recurrence properties of the solution of (1), as you have done in (a). However, this connection is also useful when studying the distance of Brownian motion to the origin: describing it as just the distance to the origin makes it a complicated quantity to analyze, while describing it as a diffusion process (i.e. as a solution to (1)) shows that it has many nice properties that diffusion processes have.