Part 1 - An Introduction to Stochastic Processes

Introduction

Most SDEs take the form,

where $x(t)$ is a stochastic process and $\beta(t)$ is Brownian motion ¹.

The SDE in Equation (1) is shorthand for the (Itô, teaser!) integral equation,

Suppose $f(x, t) = 2, L(x, t) = 0$, $\Rightarrow \frac{dx}{dt} = 2$.

We also assume that $x(0) \sim \mathcal{N}(0, 2)$. See Figure 1.

Suppose $f(x, t) = 0, L(x, t) = 2$, $\Rightarrow dx(t) = 2 \ d \beta(t)$.

We also assume that $x(0) \sim \mathcal{N}(0, 1)$. See Figure 2.

To sample $x(t)$, we can draw $x(0) \sim p(x(0))$ and then repeat,

We can use Equation (2) to sample from,

which corresponds to $f(x, t) = 0.5x$ and $L(x, t) = 1$. See Figure 3.

Motivation and Dynamic Models

SDEs are commonly used to model object dynamics (cars, airplanes, etc).

Often, time-discretized versions are used for filtering, smoothing, and prediction.

Constant Velocity Models (Scalar Example)

Suppose $\mathbf{x}(t) = \begin{bmatrix} x(t) & v(t) \end{bmatrix}^T$ represents the position and velocity of an object.

According to the constant velocity model,

The position follows the velocity, whereas the velocity is a Brownian motion.

The Ornstein-Uhlenbeck process ²,

is used in several fields, e.g., in physics to model particles moving with friction.

The Black-Scholes equation ³ is widely used in finacne to model the value of “financial instruments”.

A basic version of this equation is,

where $s(t)$ is the stock price.

The stochastic Lotka-Volterra ⁴ model describes the predator-prey dynamics,

Here, $x(t)$ and $y(t)$ are the prey and the predator populations, respectively, and $\beta_1(t)$ and $\beta_2(t)$ are independent Brownian motions.

Diffusion can be used as a (deep) generative model as well (and this is what we are mostly interested in).

The basic idea is,

• Recursively remove structure from data (make it noisier!).
• Train network to remove noise and obtain structure again.
• Generate (new) data by feeding noise to the trained network.

The forward process is generally an SDE,

We are going to get a solid understanding of theoretical results used to train and perform inference in these models.

Image generation is arguably the most well-known application. A few other examples are,

• Besides bridging noise and data distributions we can bridge, e.g., low- and high-resolution images.
• Molecular and drug design, e.g., protein structures.
• Sequential data, e.g., human pose sequences and financial data.
• Motion planning and reinforcement learning with diffusion policies.

The Engineering Approach: White Noise

Recall that we study SDEs on the form,

In engineering literature, the notation $d \beta(t)$ is uncommon in integrals.

Imagine for a second that if $\beta(t)$ were differentiable (it isn’t!), we could introduce $w(t) = \frac{d \beta(t)}{dt}$,

where $w(t)$ is white noise.

In engineering literature, this is the standard form to express SDEs.

However, what is white noise? We know that $\beta(t)$ has independent increments.

It is common to assume that $w(t)$ is a Gaussian process with,

However, if we naively use these assumptions to derive, e.g., $\mathbb{E}[X(t)]$ and $\mathrm{Cov}(x(t))$, we get the wrong expression for $\mathrm{Cov}(x(t))$.

“The white Gaussian process is mathematical fiction”. - AH Jazwinski.

We’ll take the (more) mathematical approach and express SDEs using $d \beta(t)$ and learn about Itô calculus ⁵ to enable us to derive the correct expressions.

Basic Results on Random Variables

Formally, we can introduce a probability space with,

• A probability (or sample) space $\Omega$, whose elements $\omega \in \Omega$ are called sample points,
• A $\sigma$-algebra $\mathcal{F}$ ⁶, and,
• A probability measure ⁷ $Pr(F)$ for $F \in \mathcal{F}$.

A random variable is then a function $x(\omega)$.

We’ll occasionally use the notation $x(\omega)$, but generally characterize random variables using,

• $Pr[x]$, a probability mass function, for discrete random variables,
• $p(x)$, a probability density function, for continuous random variables.

Some (very) basic concepts from probability theory,

• Expected values: $\mathbb{E}[f(x)]$ = $\int f(x) p(x) \ dx$,
• Variance: $\mathrm{Var}(x) = \mathbb{E}[(x - \mathbb{E}[x])^2]$,
• Covariance: $\mathrm{Cov}(\mathbf{x}) = \mathbb{E}[(\mathbf{x} - \mathbb{E}[\mathbf{x}])(\mathbf{x} - \mathbb{E}[\mathbf{x}])^T]$,
• Jointly distributed random variables $p(x, y)$, and,
  • Conditional distributions $p(x | y) = \frac{p(x, y)}{p(y)}$,
  • Law of total probability (marginalization) $p(x) = \int p(x, y) \ dy$,
  • Independent variables $p(x, y) = p(x) p(y)$.
  • Law of iterated/total expectations $\mathbb{E}[x] = \mathbb{E}[\mathbb{E}[x | y]]$.

Gaussian random variables are key to SDEs. A Gaussian (normal) random variable $\mathbf{x} \sim \mathcal{N}(\mathbf{\mu}, \mathbf{P})$ has density,

The most important property of Gaussian random variables is that any linear combination of Gaussian random variables is also Gaussian.

If $\mathbf{x_1} \sim \mathcal{N}(\mathbf{\mu_1}, \mathbf{P_1})$ and $\mathbf{x_2} \sim \mathcal{N}(\mathbf{\mu_2}, \mathbf{P_2})$, and $\mathbf{A_1}$, $\mathbf{A_2}$, and $\mathbf{b}$ are constants, then,

Introduction to Stochastic Processes

SDEs describe a class of stochastic processes (SDEs cannot describe all stochastic processes).

Processes described by SDEs have continuous states $x$ and parameters $t$.

For a given time, $t$, we have stochastic variables, $x(t)$. Note that we do not separate the notation for the stochastic/random variable and its realization.

Often, we also use a shorthand for its density, $p(x, t) = p(x(t))$ and allow our notation for the argument to specify the density that we evaluate.

Convergence, Integrals and Mean Square Calculus

• Convergence is an essential concept in real analysis.
• Continuity: A function $f(x)$ is continuous at $a$ if $\lim_{x \to a} f(x) = f(a)$.
• Derivative: If it exists, the derivative of $f$ at $x$ is,

• Riemann integrals:

Let $P(t)$ be a partition of $[a, b]$,

and let its norm be $|P| = \underset{i}{\max}(t_{i + 1} - t_i)$.

• If the limit exists, the Riemann integral is,

for any $t_i^{\star} \in [t_i, t_{i + 1}]$.

How is convergence defined?

A deterministic sequence $x_n \in \mathbb{R}$ converges to $a$ if $\forall \epsilon > 0, \exists N$,

That is, for sufficiently large $n$, $x_n$ is roughly $a$.

How should one define convergence for a sequence of random variables?

A few (decently) intuitive ways are,

The sequence of random variables $x_n$ is said to converge to $x$,

• with probability 1 (almost surely) if,

• in mean square if,

• in probability if for all $\epsilon > 0$,

• in distribution if for all $A \subset \mathbb{R}^n$,

The “weakest” of these is convergence in distribution, and the “strongest” is convergence almost surely.

Continuous and Differentiable Random Variables

Consider $x(t) = a \ sin(t + \phi)$ where $a \sim \text{unif}[1, 2]$ and $\phi \sim \text{unif}[0, 2\pi]$.

The random function is differentiable. Note that the derivative is itself random.

See Figure 7 for a visualization of the random function and its derivative at $x = 1$.

For $x(t) = a \ sin(t + \phi)$, consider the integral,

Note, the integral is a random variable.

Summary

Most SDEs take the form,

where $x(t)$ is a stochastic process and $\beta(t)$ is a Brownian motion.

SDEs describe a class of stochastic processes useful in many areas (economics, physics, etc). Where we specifically (will) focus on (deep) generative models.

We can use mean square convergence to define convergence, derivates, and Riemann integrals for stochastic processes.

Equation (3) is shorthand for,

Note, we can use mean square Riemann integrals to define $\int_{t_0}^{t} f(x(s), s) \ ds$ but not $\int_{t_0}^{t} L(x(s), s) \ d\beta(s)$.

We need Itô calc for that ;).