Unlike ordinary Riemann-Stieltjes integrals, a stochastic integral against a Wiener Process depends on where in each interval the integrand is evaluated.
Statement
Let be a partition with mesh , fix , and evaluate at . Then
in mean square, that is .
Proof
Write , and , so . The increments , for all are independent, with and .
Each of the three sums converges in mean square, because each is a sum of independent terms whose variances add up to :
| Sum | Mean | Variance | Limit |
|---|---|---|---|
Hence .
Choosing
Each defines a different stochastic calculus. See Converting between Ito vs Stratonovich Calculus to move between them.
| Choice | Advantages | Disadvantages |
|---|---|---|
| Ito Calculus, | The integrand does not look into the future, so it suits causal models such as finance and filtering. The integral is a martingale with zero mean, and the Ito Isometry gives its variance directly. | The ordinary chain rule fails: changes of variables need Ito’s lemma and its second-order correction. |
| Stratonovich calculus, | The ordinary chain rule holds, so it is coordinate independent and works on manifolds. It is the limit of equations driven by smooth noise with a short correlation time (Wong-Zakai), which makes it natural for physical models. | The integral is not a martingale, so expectations are harder to compute. The integrand anticipates the future by half a step. |
| (backward, or Hänggi-Klimontovich) | Appears in some models of diffusion with a position-dependent diffusion coefficient, where it gives the right equilibrium distribution. | Rarely used, and has the disadvantages of Stratonovich calculus. |