White Noise Theory

An introduction of white noise theory

Introduction

Recently, I have picked up some background of white noise theory after ready this paper . White noise theory developed by Hida generalizes the white noise in functional views.

Starting from a classical results for a Browian montion, \(B_t\), the following holds

\[dB_t = (dt)^2\]

This result leads to a new branch of calculus, stochastic calculus or Ito calculus. One of prominent lemmas is Ito’s lemma which is considered as a chain rule to compute the derivative of a function, \(f(t, B_t)\). A simple way to intepret Ito’s lemma is to use Talor approximation up to second order and then obtain

\[df(t, B_t)=\left(\frac{\partial f}{\partial t} +\frac{1}{2}\frac{\partial^2 f}{\partial x^2}\right)dt + \frac{\partial f}{\partial x}dB_t\]

Now, what if the stochastic process here is not a Brownian motion? How do we do stochastic calculus for this case?

Generalize with characteristic function

Let us start with a simple example of characteristic function of a univarite Normal random variable, \(X \sim \mathcal{N}(0, 1)\). The characteristic function is under the form

\[\varphi(t) = \mathbb{E}[\exp(itX)] = \exp\left(-\frac{1}{2}t^2\right)\]

Extending to finite dimesion where we have to deal with isotropic multivariate Gaussian distribution, the characteristic function is

\[\varphi(t) = \mathbb{E}[\exp(i\langle t, X \rangle)] = \exp\left(-\frac{1}{2}\lVert t \rVert^2\right)\]

where \(t = [t_1, \dots, t_n]^\top\) and \(\lVert t\rVert = \sqrt{t_1^2 +\dots, t_n^2}\). Here the inner product \(\langle x, y \rangle = x_1y_1 + \dots + x_ny_n\).

Obviously, this case is equivalent to Gaussian white noise with finite sample. However, stochastic process can be defined over a contiuous index set. It is natural to generalize even more.

Definition (White noise probability measure)

\[\int_{\mathcal{S}'(\mathbb{R})} \exp(i \langle \omega, f \rangle) \mu d\omega = \exp \left(-\frac{1}{2}\lVert f \rVert^2_{L^2(\mathbb{R})}\right),\]

then the measure $\mu$ is called the white noise probability measure.

The Lévy process also is defined based on its characteristic function in Lévy-Khintchine formula.

We have the following properties:

Brownian motion case

We can see that Brownian motion can fit under the definition of white noise theory.

\[\chi_{[0, t]}(s)=\begin{cases} 1 & \text{ if } 0 \leq s \leq t\\ - 1 &\text{ if } t \leq s \leq 0 \\ 0 &\text{ otherwise} \end{cases}\]

We can define \(\tilde{B}_t=\langle \omega, \chi_{[0, t]}(\cdot)\rangle\). Furthermore, we can say the \(\tilde{B}_t\) is a Gaussian process with zero mean and covariance

\[\mathbb{E}[\tilde{B}_t, \tilde{B}_s] = \min(t,s).\]

This is exactly the definition of Brownian motion. We can verify this by examining the characteristic function

\[\begin{align*} \mathbb{E}\left[\exp(i \sum_{j=1}^n c_j\tilde{B}_{t_j} )\right] = & \mathbb{E}\left[\exp\left(i \langle \omega, \sum_{j=0}^n c_j\chi_{[0, t_j]} \rangle \right) \right] = \exp \left(-\frac{1}{2}\lVert \sum_j c_j\chi_{[0, t_j]}\rVert^2 \right) \\ =& \exp\left(-\frac{1}{2}\sum_{i,j}c_i c_j\int \chi_{[0, t_i]}(s)\chi_{[0, t_j]}(s)ds\right) \\ =& \exp\left(-\frac{1}{2}\sum_{i,j}c_i c_j \min(t_i, t_j)\right) \end{align*}\]

This is the charateristic of a multivariate Gaussian distribution.

Fractional Brownian Motion (fBm)

The covariance function of fBm is

\[k(s, t) = |t|^{2H} + |s|^{2H} - |t-s|^{2H}\]

To derive the similar form like Brownian motion, we need an operator $M$

\[Mf(y) = |y|^{1/2 - H}\hat{f}(y), \quad y \in \mathbb{R}\] \[Mf(x) = C_H\int_\mathbb{R}\frac{f(x-t) - f(x)}{|t|^{3/2 - H}}dt\]

We define \(M\chi_{[0, t]}(x) = M[0, t](x)\)

The fractional Brownian motion is

\[\tilde{B}^{(H)}(t) = \langle \omega, M[0,t](\cdot)\rangle\]

We also can reobtain the covariance kernel for

\[\begin{aligned} \mathbb{E}[\tilde{B}^{(H)}(s)\tilde{B}^{(H)}(t)] & = \int_{\mathbb{R}} M[0,s](x) M[0, t](x) dx \\ & = \end{aligned}\]

Conclusion

Although the mathematical formulation presented aboved is elegant, having applications in finance multifractional, it is not clear if there is any implication yet. One potential direction is to use in generative model for time series where Brownian motion is the main source of random noise.