Formalisation of the Theory of integration applied to stochastic processes. More specifically, we are able to calculate where:

  • is a Semimartingale.
  • is adapted to the Filtration generated by . More specifically, is measurable with respect to . In full generality must be predictable (for example left-continuous and adapted) and satisfy an integrability condition such as .

260 page Intro by Gregory F. Lawler

Ito vs Stratonovich

The Ito integral and the Stratonovich Integral are both limits of Riemann sums over partitions with mesh going to zero, taken in probability. With :

For a deterministic Riemann-Stieltjes integral the evaluation point is irrelevant. Here it matters because the Wiener Process has non-zero quadratic variation, . For semimartingales the two are related by

where is the quadratic covariation. Example with :

For SDEs the covariation term becomes a drift correction, is equivalent to ; see Converting between Ito vs Stratonovich Calculus.

ItoStratonovich
Evaluation pointLeft endpoint, non-anticipatingAverage of endpoints, looks ahead
Martingale property is a (local) martingale with mean zeroNot a martingale in general
IsometryIto Isometry No direct analogue
Chain ruleIto Formula, extra term Ordinary chain rule
IntegrandsAny predictable, square integrable must itself be a semimartingale
Coordinate changesNot coordinate invariantCoordinate invariant, natural for SDEs on manifolds
Physical modellingNatural for decisions at discrete times (finance, filtering)Wong-Zakai: limit of ODEs driven by smooth, coloured noise as the correlation time goes to zero
Simple numerical schemeEuler-Maruyama converges to the Ito solutionHeun (stochastic midpoint) converges to the Stratonovich solution

Neither is more correct: they describe the same processes with different parameterisations of the drift. Example: has solution read as an Ito SDE, and read as a Stratonovich SDE. The first dies out almost surely when , the second when . The difference lies in the meaning of : in the Ito reading is the growth rate of the mean, ; in the Stratonovich reading it is the growth rate of the typical path, . The choice should follow the modelling: Ito when the noise is genuinely white and the integrand cannot see the future, Stratonovich when the white noise idealises a physical noise with small but non-zero correlation time. (Overview: Ito vs Stratonovich, OATML blog.)