Solving the Ornstein–Uhlenbeck SDE

We are given

$$ dr_t = \lambda(\theta-r_t)\,dt+\sigma\,dW_t, \qquad r_0 \text{ given}. $$

This is a linear stochastic differential equation. When $\lambda>0$, it is the classic Ornstein–Uhlenbeck process (or Vasicek model in interest-rate modeling).

1Rearrange the SDE

Expand the drift:

$$ dr_t = \lambda\theta\,dt - \lambda r_t\,dt + \sigma\,dW_t. $$

Move the $r_t$ term to the left:

$$ dr_t+\lambda r_t\,dt = \lambda\theta\,dt+\sigma\,dW_t. $$

This now looks exactly like a first-order linear ODE.

2Use an integrating factor

For the deterministic equation $y'(t)+\lambda y(t)=\cdots$, the integrating factor is

$$ e^{\lambda t}. $$

Multiply the whole SDE by $e^{\lambda t}$:

$$ e^{\lambda t}dr_t + \lambda e^{\lambda t}r_t\,dt = \lambda\theta e^{\lambda t}\,dt + \sigma e^{\lambda t}\,dW_t. $$

3Recognize a product derivative

Using Itô's product rule,

$$ d\left(e^{\lambda t}r_t\right) = e^{\lambda t}dr_t + \lambda e^{\lambda t}r_t\,dt. $$

There is no quadratic variation correction here because $e^{\lambda t}$ is deterministic.

Therefore,

$$ d\left(e^{\lambda t}r_t\right) = \lambda\theta e^{\lambda t}\,dt + \sigma e^{\lambda t}\,dW_t. $$

4Integrate from $0$ to $t$

$$ e^{\lambda t}r_t-r_0 = \lambda\theta \int_0^t e^{\lambda s}\,ds + \sigma \int_0^t e^{\lambda s}\,dW_s. $$

For $\lambda\neq0$,

$$ \lambda \int_0^t e^{\lambda s}\,ds = e^{\lambda t}-1. $$

Hence

$$ e^{\lambda t}r_t = r_0 + \theta(e^{\lambda t}-1) + \sigma\int_0^t e^{\lambda s}\,dW_s. $$

Final solution

Multiply by $e^{-\lambda t}$:

$$ \boxed{ r_t = \theta + (r_0-\theta)e^{-\lambda t} + \sigma \int_0^t e^{-\lambda(t-s)} \,dW_s } $$

The solution has three pieces:

Long-run level $$ \theta $$ The value toward which the process is pulled when $\lambda>0$.
Initial condition $$ (r_0-\theta)e^{-\lambda t} $$ The effect of $r_0$ decays exponentially.
Random noise $$ \sigma \int_0^t e^{-\lambda(t-s)}\,dW_s $$ Past Brownian shocks decay over time.

Distribution of $r_t$

The stochastic integral has a deterministic integrand, so it is Gaussian. Therefore $r_t$ is normally distributed.

Mean

Since an Itô integral has mean zero,

$$ \boxed{ \mathbb E[r_t] = \theta+(r_0-\theta)e^{-\lambda t} }. $$

Variance

Using Itô isometry,

$$ \begin{aligned} \operatorname{Var}(r_t) &= \sigma^2 \int_0^t e^{-2\lambda(t-s)} \,ds \\[4pt] &= \boxed{ \frac{\sigma^2}{2\lambda} \left(1-e^{-2\lambda t}\right) } \qquad (\lambda\neq0). \end{aligned} $$

Thus

$$ \boxed{ r_t \sim \mathcal N \left( \theta+(r_0-\theta)e^{-\lambda t}, \frac{\sigma^2}{2\lambda} (1-e^{-2\lambda t}) \right) } $$

Intuition when $\lambda>0$

$$ dr_t=\lambda(\theta-r_t)dt+\sigma dW_t. $$

If $r_t<\theta$, then

$$ \theta-r_t>0 \quad\Rightarrow\quad \text{positive drift}. $$

If $r_t>\theta$, then

$$ \theta-r_t<0 \quad\Rightarrow\quad \text{negative drift}. $$

The process is pulled back toward $\theta$.

Long-run distribution

When $\lambda>0$,

$$ e^{-\lambda t}\to0 \qquad\text{as }t\to\infty. $$

Therefore,

$$ \mathbb E[r_t]\to\theta $$

and

$$ \operatorname{Var}(r_t) \to \frac{\sigma^2}{2\lambda}. $$

Hence the stationary distribution is

$$ \boxed{ r_\infty \sim \mathcal N \left( \theta, \frac{\sigma^2}{2\lambda} \right) }. $$

Interview way to recognize the solution

The important pattern is

$$ dX_t+aX_t\,dt = \text{something}. $$

Your immediate thought should be:

$$ \boxed{\text{Integrating factor }e^{at}} $$

So for this problem:

$$ dr_t+\lambda r_tdt = \lambda\theta dt+\sigma dW_t $$
$$ \text{multiply by }e^{\lambda t} $$
$$ d(e^{\lambda t}r_t) = \lambda\theta e^{\lambda t}dt + \sigma e^{\lambda t}dW_t $$
$$ \boxed{ r_t = \theta+(r_0-\theta)e^{-\lambda t} + \sigma\int_0^t e^{-\lambda(t-s)}dW_s }. $$