$W_t$ is a brownian motion and $X_t:= \sqrt{t}Z$, where: $Z\sim N(0,1)$.

How to show that for a bounded continuous $f$ process, $$U_t = \int_0^t (f(W_s))ds$$ and $$V_t = \int_0^t (f(X_s))ds$$ have the same expectation but not the same variance in general?

  • $\begingroup$ I am not sure to get the point of your question. Basically, Brownian motion and your $X$ stochastic process are equal in distribution. So all their moments will be equal, as well as as the moments of your two processes $U$ and $V$. They can be ‘different’ in the sense of non-indistinguishability (a.s. equality), just as two identically distributed random variables can be different ‘event-wise’ $\endgroup$ – siou0107 Apr 20 '20 at 15:07
  • 3
    $\begingroup$ The thing though is that the $X_t$ process is very smooth, while $W_t$ is very rough. The quadratic variation will be very different. $\endgroup$ – noob2 Apr 20 '20 at 15:55
  • $\begingroup$ I don’t agree with that. $dX_t = \sqrt{dt}Z$, which has the same distribution and properties (such as quadratic variation) than $dW_t$ $\endgroup$ – siou0107 Apr 20 '20 at 17:15

The means are equal

Suppose $f$ is analytic so that we can give it a Taylor series that works everywhere such that $f(x) = \sum a_n x^n$, and then let us let this be bounded too. To show that the expectations are the same consider taking the expectation of $U_t$ \begin{equation} \mathbb{E}(U_t) = \mathbb{E}\left(\int_0^t f(W_s) \,\mathrm{d}s\right)\\ \end{equation} as $f$ if bounded we can use Fubini's theorem to move the expectation inside the integral, giving \begin{align} \mathbb{E}(U_t) & = \int_0^t \mathbb{E}\left(f(W_s)\right) \,\mathrm{d}s\\ & = \int_0^t \sum_n a_n \mathbb{E}\left(W_s^n\right) \,\mathrm{d}s \\ & = \int_0^t \sum_n a_n s^{n/2} M_n \,\mathrm{d}s \end{align} where $M_n$ denotes the $n$-th moment of a standard Gaussian distribution. Identically we could do this for the expectation of $V_t$ and see that the two are equal, and hence that $U_t$ has the same expectation of $V_t$.

The variances are not equal

For the variance of $U_t$ we need to evaluate the expected value of \begin{align} U_t^2 = \left(\int_0^t f(W_s) \,\mathrm{d}s \right)^2 &= \int_0^t \int_0^t f(W_s) f(W_u) \,\mathrm{d}u \,\mathrm{d}s \\ &= 2 \int_0^t \int_0^s f(W_s) f(W_u) \,\mathrm{d}u \,\mathrm{d}s. \end{align}

Taking the expectation and putting this inside the integral again gives \begin{equation} \mathbb{E}\left(U_t^2\right) = 2 \int_0^t \int_0^s \sum_{n, m, k} a_n a_m \binom{n}{k} \mathbb{E}(W_s^n W_u^m) \,\mathrm{d}u \,\mathrm{d}s. \end{equation} Within the integrand $u \leq s$ and so expand $W_s = W_u + (W_s - W_u)$ where as $W$ is a Weiner process the increment $W_s - W_u$ is independent to the value of $W_u$ and has variance $s - u$. This then simplifies to give \begin{equation} \mathbb{E}\left(U_t^2\right) = 2 \int_0^t \int_0^s \sum_{n, m, k} a_n a_m \binom{n}{k} u^{\frac{n + m}{2}}(s - u)^{\frac{n-k}{2}}M_{n+m}M_{n-k} \,\mathrm{d}u \,\mathrm{d}s. \end{equation}

An identical process can be applied for $V_t$ but instead of expanding $W_s$ we just use $X_t = \sqrt{t}Z$ and obtain \begin{equation} \mathbb{E}\left(V_t^2\right) = 2 \int_0^t \int_0^s \sum_{n, m, k} a_n a_m \binom{n}{k} s^{\frac{n}{2}}u^{\frac{m}{2}}M_{n+m} \,\mathrm{d}u \,\mathrm{d}s. \end{equation}

The second expression is more easily integrated out than the first, but both are doable and I think they evaluate to different quantities.

A numeric demonstration

Take $f \equiv \sin$ and you can simulate this, as the example python code does, the key result is then:

Variable = U,   Mean = 0.00206326180593 +/- 0.00228420660865,   Standard deviation = 0.228420660865 
Variable = V,   Mean = -0.0177668651186 +/- 0.0380135632954,    Standard deviation = 3.80135632954 

We see this confirms that the means are the same but variances differ.

enter image description here

Python code

import numpy as np
from scipy.stats import norm

f = np.sin

M = 10000
T = 2*np.pi
dt = T/M
N = 10000
mean_u, mean_v = 0, 0
var_u, var_v = 0, 0
for _ in range(N):
    w = np.cumsum(np.concatenate([[0], norm.rvs(scale=dt, size=M)]))
    z = norm.rvs()
    t = np.cumsum([0] + [dt]*M)
    x = np.sqrt(t) * z
    u = np.sum(f(w))*dt
    v = np.sum(f(x))*dt
    mean_u += u
    mean_v += v
    var_u += u ** 2
    var_v += v ** 2
mean_u /= N
mean_v /= N
var_u /= N
var_v /= N
var_u -= mean_u ** 2
var_v -= mean_v ** 2
std_u = np.sqrt(var_u)
std_v = np.sqrt(var_v)
std_error_u = np.sqrt(var_u / N)
std_error_v = np.sqrt(var_v / N)

for n, m, s, se in [["U", mean_u, std_u, std_error_u], ["V", mean_v, std_v, std_error_v]]:
    print("Variable = {},\tMean = {} +/- {},\tStandard deviation = {} ".format(n, m, se, s))
  • $\begingroup$ Your $X$ process does not seem much random to me... It is just a random variable which, once known, grows at $\sqrt{t}$, but then it is not a stochastic process. $\endgroup$ – siou0107 Apr 24 '20 at 15:15
  • $\begingroup$ That's exactly the definition of the process $X_t$ in the original post: $X_t:=\sqrt{t}\times Z$. $\endgroup$ – Daneel Olivaw Apr 24 '20 at 15:17
  • $\begingroup$ @siou0107 It's no more or less random than $W_t$. Once a sample $\omega$ is realised both $X_t$ and $W_t$ are known, it's just that one is much smoother than the other. $\endgroup$ – oliversm Apr 24 '20 at 15:48
  • $\begingroup$ @oliversm Thanks so much for your answer... $\endgroup$ – Alicia Apr 27 '20 at 0:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.